content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
The Square Kilometre Array\footnote[1]{www.skatelescope.org} (SKA) is a
billion dollar, multinational project which aims to construct the world's
largest radio telescope\cite{Schilizzi08,Lazio09}. A consortium of 17
countries has pledged involvement in this mega-science project which aims
which will collect more
information on the radio sky in the first seven hours of operation than
has been obtained in the first 70 years of radio astronomy. The SKA will
be comprised of thousands of individual radio telescopes, with a total
collecting area of one square kilometre. These telescopes will be arranged in
stations of 20 - 45 antennas. The stations will stretch over thousands of
kilometres in order to provide resolution of the order of milliarcseconds
which will be comparable to the best optical and infrared instruments.
The array will be sited in either Australasia or Southern Africa and bids
to host the telescope are being coordinated by South Africa and Australia
with both countries committed to building `pathfinder' telescopes on the
sites of the proposed SKA core in each country.
Realization of the truly paradigm-shifting nature of the SKA will require
significant challenges in radio engineering, data transport \& storage,
signal processing and power generation to be overcome. Much progress has
been made on many of these fronts but there is much still to do and even
seemingly small communities can play a vital role in the project. In this
paper we describe the current status of the project in New Zealand
with particular emphasis on the research enagement from Aotearoa.
\section{SKA in Australia}
Australian researchers have been involved in the SKA project since its
inception over a decade ago. Radio astronomy is the largest single
discipline in
Astrophysics in Australia accounting for over 30 per cent of the total
astronomical research output in Australia
\cite{decrev06}. Radio astronomy research groups currently exist in seven
Australian universities in addition to the government funded CSIRO division
of the Australia Telescope National Facility and radio astronomy accounted for
nearly 40 per cent of all permanent jobs in Australian Astronomy
in 2005 \cite{decrev06}. With the advances toward the SKA, this fraction is
likely to expand significantly.
To date the Australian Federal Government has committed 125 million AUD for
construction of their pathfinder telescope, the Australian Square Kilometre
Array Pathfinder (ASKAP) and 80 million AUD for a High Performance
Computing (HPC) facility in Perth\footnote[2]{www.ska.gov.au/news/Pages/MinisterCarrnamesPawseyHPCCentre.aspx}. In addition, the Western Australian
Government has committed a further 20 million AUD to establish the
International Centre for Radio Astronomy Research (ICRAR)\footnote[3]
{www.icrar.org} as a joint venture between the state and the University of
Western Australia and Curtin University of Technology.
\section{SKA in New Zealand}
In August 2009 New Zealand and Australia signed an arrangement
to collaborate on the bid to host the SKA \footnote[4]{'Arrangement between
the Australian Government and the New Zealand Government on a Joint Bid
to Secure the Siting of the Square Kilometre Array in Australia and New
Zealand'}. This marked the first formal step of New Zealand's engagement
with the project and will see a joint trans-Tasman effort to work together
to co-host the telescope.
\subsection{NZ SKA Project Office}
The Ministry of Economic Development (MED) is the lead government
agency for New Zealand's involvement in the SKA. A New Zealand SKA
project office has been established as New Zealand's point of contact
for SKA matters and is coordinating New Zealand's Government level
programme for engagment in the SKA project. As part of the coordination effort
the project office is working closely with the Australia-New Zealand SKA
Coordinating Committee (ANZSCC) on which the NZ Government has a
representative.
\subsection{SKA Research \& Development Consortium}
In June 2009, just prior to the announcement of the formal arrangement
with Australia, the New Zealand Square Kilometre Array Research \&
Development Consortium (SKARD) was formed\footnote[5]{www.ska.ac.nz}. The
role of SKARD is to bring
together professional researchers involved in SKA related research in NZ,
to foster collaboration both within NZ and internationally and to liaise
with industry, government and other groups to advance New Zealand's
contribution to the SKA. Members are drawn from all of New Zealand's
major research universities and have interests in antenna design, signal
processing, imaging and inference, high performance computing and
radio astronomy.
\subsection{NZ SKA Industry Consortium}
In addition to SKARD, the New Zealand SKA Industry Consortium
(NZ SKAIC) has formed to achieve positive economic
outcomes for New Zealand from involvement in the SKA project.
The group consists of industry partners representing
IT software, hardware, networks and services, the Ministry of Economic
Development (MED) and New Zealand Trade and Enterprise (NZTE).
\section{Research Engagement in New Zealand}
Although SKARD acts as a coordination point for researchers engaged in
SKA related research, research engagement at the individual level has
been quite active over 2009 even before the formation of the consortium.
Several groups from across the country
have come together to form collaborative projects around radio astronomy and
related engineering. Highlights of this collaborative work include efforts
in radio astronomy, radio engineering and signal processing.
\subsection{Radio Astronomy}
Currently there are only two universities in which radio astronomy
research is conducted in New Zealand.
Consequently, direct involvement in pure radio astronomy research is limited
\footnote[6]{Refereed research publications in radio astronomy for NZ
based researchers in the last decade are confined to only two authors:
Johnston-Hollitt \& Budding.}, however where such engagement has occurred
it has been well placed.
In particular, NZ has good representation on forthcoming major science
and associated technical research associated with the Australian SKA
Pathfinder, pan-NZ capability building in radio astronomy, signal processing
and engineering and an emerging interest in associated High Performance
Computing (HPC).
\subsubsection{ASKAP Surveys}
Australia has committed to build a one per cent demonstrator for the SKA called
the Australian Square Kilometre Array Pathfinder (ASKAP). ASKAP, which is in
construction now and due for completion in 2012, will be a powerful new
telescope in its own right. Located at the core of the Australian \&
New Zealand site for the SKA, the Murchison Radio Observatory in Western
Australia, ASKAP will
comprise 36 15 meter diameter radio dishes spread over a 6 km diameter area.
The configuration is designed to optimize sensitivity on angular scales of
30 arcseconds, whilst still providing good low surface brightness sensitivity
at lower resolutions \cite{Gupta08}. ASKAP will operate over 700 -
1800 MHz with a one degree field of view and has been specifically designed as
a survey instrument for redshifted observations of neutral hydrogen (HI)
\cite{Staveley-Smith08}. ASKAP will act as a survey instrument for
approximately 75 per cent of the first five years of operation
(2013 - 2018) and an open call for survey science proposals in late 2008
resulted in 38
expressions of interest from world-wide teams of researchers. An extensive,
multi-stage process of international peer review to prioritize the
science surveys has just concluded with eight science surveys and two
strategic programs selected\cite{Ball09}.
Several NZ-based researchers are involved in the successful surveys which were
selected for ASKAP. Staff from the Victoria University of Wellington (VUW)
are involved in three ASKAP proposals, the A ranked Evolutionary Map of the
Universe (EMU)\cite{Norris09} and the A- ranked surveys
``POSSUM''and ``VAST''. Staff from the Auckland University of Technology are
part of the strategic project ``The high resolution component of ASKAP:
meeting the long baseline specifications for the SKA''.
Within the commensally observed EMU and POSSUM projects, researchers at
VUW will contribute to science related to clusters of galaxies. In particular,
understanding the role of environment on the production of radio emission
in clusters will be investigated. This will include: the detection and
characterization of low surface brightness diffuse synchrotron
emission associated with dynamical encounters \cite{mjh03},
the use of tailed radio galaxies as barometers of ``cluster weather''
\cite{mjh04} and as probes of high density regions \cite{Mao09a, Mao09b},
and an investigation into the properties of AGN and starforming galaxies in
different cluster environments \cite{Venturi01, Venturi02, mjh08}.
The latter two projects are relatively straightforward in terms of finding
and identifying sources. Detection of diffuse, low surface brightness
sources however will be a challenge and is likely to require new signal
processing techniques. Consequently researchers from VUW are undertaking
a broader program of research in signal processing for astronomical
research to address the challenge (see \cite{Hollitt09}).
The particular interest of NZ members of the VAST team is investigating
coherent emission from active stars and binary systems. Such observations will
probe fundamental parameters of stellar magnetospheres such as magnetic
field intensity and topology and electron energy distribution and provide
vital information on the relation of this emission to the physical
characteristics of the host stars.
\subsubsection{Radio Telescopes in New Zealand}
The have been a range of substantial low-frequency arrays and small
higher frequency ($\nu \leq$ 3 GHz) dishes in NZ over the last
50 years. In particular, there has been a long history of polytechnics
associated with small dishes including a 10m dish at Hamilton built
in 1985 by the Waikato Institute of Technology in association with
Waikato University\cite{Head06} and a 5m dish constructed in 1998 by
the Central Institute of Technology in Upper Hutt \cite{Budding00}.
(See \cite{Head06} and references therein for a full discussion of
the history of small radio dishes in NZ.) The most recent in this
long line of small dishes is the AUT 12m dish at
Warkwork, located just north of Auckland \cite{Gulyaev09}. Although
this dish, which has been constructed but is currently still
undergoing testing,
will primarily undertake geodetic observations \cite{Gulyaev09}, it is
hoped it will contribute to science observations of bright radio
sources through the Australian Long Baseline Array (LBA).
Additionally, there are several low-frequency arrays operating in NZ for the
purposes of undertaking upper atmospheric or meteor detection research,
which are not connected to radio astronomy research.
\begin{figure}
\includegraphics[width=\linewidth]{RFI_Threshold.eps}
\caption{Interference thresholds as a function of frequency for continuum
observations with several different configurations of radio telescopes.
The line for total power describes the spectral power flux density (dB(
W m$^{-2} Hz^{-1}$)) for a single dish radio telescope. Subsequent lines
give the threshold for harmful interference for different configurations
of interferometers going from more compact (VLA(D)) to least compact
(Merlin) configurations. The least stringent requirements are the present
recommendations for Very Long Baseline Interferometry (VLBI). Taken from
the International Telecommunication Union Handbook on Radio Astronomy
\cite{ITUHandbook}. The outer stations of the SKA, such as those
potentially in New Zealand, will be required to conform with VLBI levels
\cite{ITURegs,SKAmemo73}.}
\label{rfi-tab}
\end{figure}
\subsubsection{Low Frequency Transient Network Development}
\label{tread}
Funding has recently been obtained to build a low frequency receiver
network deployed across NZ to investigate properties of the transient radio
sky. The project Transient Radio Emission Array Detector
Prototype (TREAD-P) \cite{Kitaev09} will employ FPGA devices connected across
the Kiwi Advanced Research and Education Network (KAREN) to test collection
and storage of very high data rates and novel techniques for data processing
using Bayesian inference with the hope of characterizing the low frequency
radio sky. The core of the detector network will be the Digital Receiver
Sensor (DRS) which is based around Field Programmable Gate Array (FPGA)
technology, which is becoming increasingly important to radio astronomy.
Depending on the configuration, each DRS will produce up to 10MB of data
per second and the network of 10 devices will produce more than 8.6 Terabytes
daily when running at maximum capacity.
Major international radio telescopes like the Low Frequency Array (LOFAR)
\cite{Huub03} and ASKAP have identified characterization of the transient
radio sky as a high science priority and proposed instruments like the
Long Wavelength Array plan to conduct science specifically on the transient
radio sky (20 - 80 MHz) in the next decade. Instruments like the
DRS will provide vital information on the frequency of both artificial and
naturally occurring radio signals which will be vital to the survey design
of these large low-frequency telescopes.
In order to achieve this goal researchers from the Auckland
University of Technology, University of Otago, Victoria University of
Wellington, University of Auckland and the University of Canterbury and
industry partners have agreed to contribute resources. It is expected
this will provide opportunities for at least three research students at
partner institutions across New Zealand in 2010.
\subsection{HPC Facilities}
New Zealand has capability in High Performance Computing, with dedicated
facilities at the University of Canterbury (BlueFern) and computation
clusters in several universities and government research labs. Many of
these facilities are coordinated via its national grid initiative
BeSTGRID led by the University of Auckland and seven other member
institutions, which operates over a 10 Gb/s national research network the
Kiwi Advanced Research and Education Network (KAREN), providing essential
services for collaboration, computation, and data management.
NZ-based researchers are currently utilizing these tools to
develop capability in relation to the SKA. As an example, the low
frequency transient project (Section
\ref{tread}) involving researchers from five institutions
has been recently funded to exploit this infrastructure to transport, store
and process data at high rates from a network of GPS-synchronized low
frequency antennas across NZ for radio astronomy research.
Additionally, researchers at three of NZs universities (Otago, Auckland and
AUT) have established a collaboration developing new high performance
computing techniques for processing radio astronomy data with a specific
focus on reconfigurable computing. AUT, Otago and VUW maintain HPC clusters
and in addition AUT and Otago have reconfigurable FPGA based systems
(based in physics and engineering respectively) while BeSTGrid (Auckland)
provides an ideal middleware platform to develop techniques for distributed
computations which will be required for SKA research.
While this is a foundation, it should be noted however that host countries
for the SKA will require significant upgrades to e-Infrastructure including
a distributed high-speed fibre optic network between the stations and the
science processing facility, mega-watt power for SKA stations and
specialised hardware for the central digital signal processing facility.
Depending on how the SKA is configured many of these
resources may be located in NZ as part of the international project.
\section{Discussion: Immediate goals}
\subsection{Station Locations}
In partnering with Australia to host the SKA, New Zealand is hoping to
contribute the outer stations of the array. At this stage this is likely
to mean two stations of antennas in NZ, though arguments for more stations
could be made in the future. In order to achieve the science
goals of the SKA even the outer stations of the array will require low
levels of radio frequency interference (RFI)\cite{SKAmemo73}.
Figure \ref{rfi-tab}
shows the currently required levels specified for outer stations of the
SKA in the low to mid frequency bands\cite{ITURegs}. While there have been
a couple of suggested locations put forward for SKA stations none of these
meet the required specifications.
It is probably worth noting here that the vast majority of RFI is
man-made and easiest way to achieve the radio quite levels required for the
SKA is to place the antennas in regions of low population density. Figure
\ref{census} shows the current population density of NZ as determined from
the 2006 national census\footnote[7]
{http://en.wikipedia.org/wiki/File:NewZealandPopulationDensity.png} in
addition to Telecom's current mobile phone
coverage\footnote[8]{www.telecom.co.nz}. Regions of low population density and no mobile coverage exist in
much of the South Island and there are some pockets in the North Island with
similar characteristics.
\begin{figure}
\includegraphics[width=\linewidth]{ENZcon_fig_small.eps}
\caption{LHS: Population density of NZ from the 2006 Census$^7$; RHS current
mobile phone coverage for NZ$^8$. To reduce the levels of
RFI, potential SKA stations should ideally be located in regions of low
population density with little mobile coverage.}
\label{census}
\end{figure}
At present, a comprehensive and independent site selection and survey
process in close collaboration with Australia is currently being
negotiated through the NZ SKA Project Office. As with all telescope site
selection processes this will examine a range of metrics before finalising
potential sites.
\subsection{Research Engagement}
With the signing of the arrangement with Australia and formation of SKARD and
NZSKAIC, New Zealand is now on a firm footing for researchers from the private
and public sector to engage in this project. As the final host nations
for the array will not be decided until the early part of 2012 it is important
for NZ to maximise benefit in the project via early engagement while
simultaneously mitigating the risk that the project will go to Southern
Africa. This means that it is important to identify niche opportunities
which are
independent of the ultimate location of the array.
This can be achieved via expanding existing links between NZ and
Australia in radio astronomy related areas via:\\
i) continued participation of NZ scientists in the science and design
of the SKA,\\
ii) building expertise in NZ through the training of graduate students
in radio astronomical research - a natural complement to NZ's existing
strength in optical astronomy, and\\
iii) undertaking significant and ground breaking research which will be
used for further enhancement of science with next generation radio telescopes.
The latter speaks directly to the identification and exploitation of niche
research opportunities. Such niches have clearly
emerged in NZ around science-driven post-processing via novel signal processing
\cite{Hollitt09}. Work of this nature combines NZ's strengths in radio
astronomy, signal processing, HPC \& middleware. Additionally, capacity
building exercises and pan-NZ collaborative work \cite{Kitaev09} which has
commenced will be
important to realize opportunities that will arise with hosting the telescope.
On the purely astronomical research front, in order to maximize the
science outcomes of both ASKAP and eventually the SKA it is crucial that
knowledge from existing instruments is examined and used to developed
expertise vital for survey design on next generation instruments and the
research teams thus far involved are well placed to make significant
contributions.
\section*{Acknowledgements}
MJ-H acknowledges continuing support towards radio astronomy research by the
Victoria University of Wellington.
|
\section{Introduction}
The dynamics in a liquid become increasingly correlated in space and heterogeneous in time as the system is cooled toward its glass transition. These correlations are most pronounced at time scales on the order of the relaxation time of the system and are typically quantified using a dynamic susceptibility that measures fluctuations in the number of mobile particles~\cite{Glotzer2000b, Richert2002, Andersen2005, Berthier2005, Dalle-Ferrier2007, Berthier2007a, Berthier2007b}. A particle is deemed mobile if it has moved a distance of at least $a$ in a time interval $t$, so the dynamic susceptibility depends on the choice of $a$ and $t$. For $a$ fixed at a fraction of the particle diameter, the dynamic susceptibility exhibits a maximum at times $t$ on the order of the relaxation time. This maximum increases in height as temperature is lowered towards the glass transition and the relaxation time increases~\cite{Matsui1994, Kob1995, Matsui1998, Hiwatari1998, Glotzer2000, Lacevic2003b, Toninelli2005, Szamel2006, Chandler2006, Berthier2007a, Flenner2007, Charbonneau2007}. Thus, the peak in the dynamic susceptibility is viewed as an important signature of the glass transition.
In molecular dynamics simulations of glass-forming liquids, the dynamic susceptibility is typically measured as a function of $t$ with the length scale $a$ fixed. Studies of the length-scale dependence have been more rare~\cite{Lacevic2003b, Chandler2006, Abate2007, Charbonneau2007}.
However, recent experiments on granular systems~\cite{Lechenault2008} suggest that it is instructive to study the dynamic susceptibility as a function of $a$ as well as $t$. This approach has the advantage of avoiding any arbitrariness in the choice of $a$ or $t$.
In this paper, we use the dynamic susceptibility to measure the spatial extent of kinetic heterogeneities as a function of distance $a$ and time $t$ in a model glass-forming liquid. At temperatures where there is a well-defined sub-diffusive plateau in the root mean square displacement, we find that there are
{\emph two} distinct local maxima in the dynamic susceptibility $\chi_{\rm ss}(a,t)$ quantifying the spatial extent of kinetic heterogeneities. These two maxima correspond to crossovers in the dynamics. The larger of the two maxima, which has been observed in previous simulations, occurs at late times at the crossover from sub-diffusive to diffusive motion. This maximum quantifies the strong and well-studied correlations in the dynamics that arise at the scale of the relaxation time. In addition, we find a secondary maximum at much earlier times, corresponding to the crossover from ballistic to sub-diffusive motion. The presence of this second maximum indicates that the dynamics also exhibit significant spatial correlations at time scales corresponding to the trapping of a particle in the cage of its neighbors. The discovery of a second maximum implies that kinetic heterogeneities are not just associated with the onset of slow dynamics near a glass or jamming transition, and that they should be regarded more generally as symptoms of crossovers in the dynamics.
\section{Model}
Our model consists of disks interacting in two dimensions via the purely repulsive pairwise potential
\begin{equation}\label{potential}
V(r_{ij})=
\begin{array}{lr}
\frac{\epsilon}{\alpha}\left(1-\frac{r_{ij}}{\sigma_i+\sigma_j}\right)^2 &\rm{for \ } r_{ij}<\sigma_i+\sigma_j,\\
0 &\rm{for \ } r_{ij}\ge \sigma_i+\sigma_j,
\end{array}
\end{equation}
where $\epsilon$ is the characteristic energy scale and $\sigma_i$ is the radius of the $i$th disk. We use equal mixtures of disks of radii $\sigma$ and $1.4\sigma$ and equal mass $m$. We choose units such that 2$\sigma$, $\epsilon$, m, and $k_B$ set equal to 1. This sets the time unit $2\sigma\sqrt{m/\epsilon}=1$. To solve the dynamics, we perform molecular dynamics simulations at fixed temperature and pressure in square periodic cells containing $N=100$, 400, or 1600 disks. We employ Gaussian constraints\cite{Evans1983a} to fix the instantaneous temperature $T=|\vec{p}_i|^2/2(N-1)$ and hydrostatic pressure $p=(\vec{r}_i\cdot \vec{F}_i/2+NT)/L^2$, where $\vec{F}_i=-\vec\nabla\sum_jV(r_{ij})$, allowing the side length L of the periodic cell to vary. In our runs, we fix $p=10^{-2}$ and measure the dynamics at various temperatures $T=10^{-4}$ to $T=9\times 10^{-4}$ by running 10 to 20 simulations for durations between $\Delta\tau=10^5$ and $\Delta\tau=10^7$ after equilibrating at the temperature for one fifth as long. We arrive at the temperature by quenching at rates $10^{-9}$ or $10^{-8}$ starting from well-equilibrated configurations at $T=10^{-3}$. Results here are shown for a quench rate of $10^{-9}$.
\section{Overlap function and its correlations}
Correlations in the dynamics are measured by using the standard time- and distance-dependent order parameter
\begin{equation}
q_{\rm s}(a, t; i, \tau)\equiv w_a(\vec{r}_i(\tau)-\vec{r}_i(\tau+t)),
\label{orderdef}
\end{equation}
where $w_a(\vec{r})$ is an overlap function~\cite{Glotzer2000}. Here, $w_a(\vec{r})=1$ if $|\vec{r}|<a$ and $0$ otherwise. We denote the average over all particles $i$ by $Q_{\rm s}(a, t; \tau)\equiv(1/N)\sum_{i} q_{\rm s}(a, t; i, \tau)$ and its time average by $\bar{Q}_{\rm{s}}(a, t)\equiv\langle Q_{\rm s}(a, t; \tau)\rangle_\tau.$ At very short times, $\bar{Q}_{\rm{s}}(a, t)=1$ because no particles have moved a distance $a$. At very long times, positions at time $\tau+t$ are uncorrelated with positions at time $\tau$, so $\lim_{t\rightarrow \infty}\bar{Q}_{\rm{s}}(a, t)\simeq\phi a^2/N$, where $\phi$ is the packing fraction. Between these two extremes, $\bar{Q}_{\rm{s}}(a, t)$ decays smoothly from 1 to $\phi a^2/N$.
The spatial extent of fluctuations can be characterized by the variance of the overlap function over different starting times~\cite{Glotzer2000}
\begin{equation}
\chi_{\rm ss}(a, t)\equiv N\left(\langle Q_{\rm s}(a, t; \tau)^2\rangle_\tau-\bar{Q}_{\rm{s}}(a, t)^2\right).
\label{chidef}
\end{equation}
As shown by the counting argument of ref.~\cite{Abate2007}, $\chi_{\rm ss}(a, t)$ is roughly the number of particles that move a distance $a$ over a time $t$ in a correlated manner.
Note that the spatially averaged overlap function $Q_{\rm s}(a, t; \tau)$ is the self part of the order parameter $Q(a, t; \tau)\equiv(1/N)\sum_{ij} w_a(\vec{r}_i(\tau)-\vec{r}_j(\tau+t))$, and $\chi_{\rm ss}(a, t)$ is the self-self part of the dynamic susceptibility $\chi_4(a, t)\equiv N(\langle Q(a, t; \tau)^2\rangle_\tau-(\langle Q(a, t; \tau)\rangle_\tau)^2).$ Glotzer et al showed that the self-self part of the dynamic susceptibility dominates over the distinct-distinct and the self-distinct parts for a model glass-forming liquid~\cite{Glotzer2000, Lacevic2003b}.
\begin{figure}[t]
\onefigure[width=80mm]{figure1.eps}
\caption{Color plots of the value of the dynamic susceptibility $\chi_{\rm ss}(a, t)$ as a function of the lag time, $t$ and the overlap distance, $a$. Results are shown at a fixed pressure, $p=10^{-2}$, and system size, $N=1600$, at three different temperatures: (a) $T=7\times 10^{-4}$, (b) $T=5.5\times 10^{-4}$, and (c) $T=2\times 10^{-4}$. The dynamical glass transition is at $T \approx 5 \times 10^{-4}$. For comparison, we plot the root-mean-square displacement $\Delta r(t)$ (solid curves). The circles mark the locations, $(a^{\star}, t^{\star})$, of observed local maxima of $\chi_{\rm ss}(a, t)$. Filled circles represent cage-escaping local maxima near the end of the sub-diffusive plateau of $\Delta r(t)$. Such maxima are observed for $T\ge5\times10^{-4}$. Open circles represent cage-exploring local maxima near the beginning of the plateau. Such maxima are observed for $T\le5.5\times10^{-4}$, so that there is a range of temperatures over which both local maxima exist.}
\label{maps}
\end{figure}
\section{Results}
The color plots of fig.~\ref{maps} display the spatial and temporal dependence of the dynamic susceptibility $\chi_{\rm ss}(a, t)$. The three plots represent three temperatures: one well above the dynamic glass transition $T_g\approx 5 \times 10^{-4}$, one slightly above it, and one below $T_g$. In each plot, we also display the root-mean-square (rms) displacement, $\Delta r (t)\equiv\sqrt{\langle|\vec{r}_i(0)-\vec{r}_i(t)|^2\rangle}$. Not surprisingly, for a given $t$ we find that $\chi_{\rm ss}(a, t)$ is largest for $a$ near $\Delta r(t)$, the characteristic distance over which some, but not all, particles have moved. This is consistent with granular experiments by Lechenault, et al~\cite{Lechenault2008}. We place circles on the locations $(t^\star, a^\star)$ of the local maxima of $\chi_{\textnormal ss}(a, t)$.
For the highest temperature (fig.~\ref{maps}(a)), well above the glass transition, the rms displacement $\Delta r(t)$ shows a ballistic regime at small $t$, the hint of a sub-diffusive regime at intermediate times, and a diffusive regime at long times. We observe one local maximum of $\chi_{\rm ss}(a, t)$, marked by a solid circle, at the onset of the diffusive regime: the heterogeneities are largest when some particles have begun to diffuse but others remain caged. Note also that the maximum at $(t^*,a^*)$ lies somewhat below the $\Delta r (t)$ curve, indicating that larger clusters of dynamically correlated particles are formed by less mobile particles~\footnote{In all cases, we also find that $\chi_{\rm ss}(a, t)$ increases along $\Delta r(t)$ with decreasing $t$ for the shortest times observed, well within the ballistic regime, reflecting correlations in the velocities. We observe this behavior at all temperatures. This is an artifact of the barostat used; we do not observe this increase when the system is held at fixed area.}.
For the intermediate temperature (fig.~\ref{maps}(b)), the sub-diffusive plateau of $\Delta r (t)$ is well established and spreads out over two decades. Again, we observe a local maximum of $\chi_{\rm ss}(a, t)$ near the onset of the diffusive regime. However, in addition to this ``cage-escaping" maximum we also observe a secondary ``cage-exploring" maximum, marked by an open circle, near the onset of the plateau; the dynamics are also heterogeneous on time scales when particles are becoming constrained by their neighbors. Thus, in the regime where there are two distinct crossovers in the dynamics--one from ballistic to sub-diffusive motion, and one from sub-diffusive to diffusive motion--there are also two maxima in $\chi_{\rm ss}(a,t)$, located at distances $a$ and times $t$ corresponding to the two crossovers.
For the lowest temperature shown (fig.~\ref{maps}(c)), we do not reach a diffusive regime on the time scale of our simulation, which means that this temperature lies below the dynamic glass transition temperature $T_g\approx 5\times 10^{-4}$. We therefore do not observe the cage-escaping maximum associated with the onset of the diffusive regime. However, the location of the cage-exploring maximum remains relatively unchanged, as shown by the open circle. Although the system is out of equilibrium for all $T<T_g$, we observe mild aging effects only for the highest of these temperatures studied, $T=4\times10^{-4}$ and $T=4.5\times10^{-4}$, where the equilibration times are comparable to the quench times, waiting times, and run times of our simulations.
\begin{figure}[t]
\onefigure[width=80mm]{figure2.eps}
\caption{Dynamic susceptibility $\chi_{\rm ss}(a, t)$ vs lag time $t$ for $p=10^{-2}$, $N=1600$, and four different temperatures above but near the glass transition: (a) $T=6.5\times 10^{-4}$, (b) $T=6\times 10^{-4}$, (c) $T=5.8\times 10^{-4}$, and (d) $T=5.5\times 10^{-4}$. At each temperature, we show multiple curves, each corresponding to a different value of the overlap distance $a$. The range of overlap distances represented by the curves is the same as in the color plots of fig.~\ref{maps}. The maximum of $\chi_{\rm ss}$ with respect to $a$ at each value of $t$ is represented by the heavy red curve. Circles mark the lag time $t^\star$ and amplitude $\chi^\star$ of local maxima of $\chi_{\rm ss}(a, t).$ As in fig.~\ref{maps}, filled circles represent cage-breaking local maxima near the end of the sub-diffusive plateau of $\Delta r(t)$, and open circles represent cage-forming local maxima near the beginning of the plateau.}
\label{plots}
\end{figure}
To study the evolution of the maxima with temperature, we plot in Fig.~\ref{plots} $\chi_{\rm ss}(a,t)$ as a function of $t$ for many different fixed values of $a$ covering the same range of $a$ as in Fig.~\ref{maps}. The thick red curve marks the maximum of $\chi_{\rm ss}$ with respect to $a$ at each $t$. Maxima of this curve (marked with circles) therefore represent maxima of $\chi_{\rm ss}$ with respect to both $a$ and $t$. Fig.~\ref{plots} shows that at high temperatures, only one local maximum of $\chi_{\rm ss}$ can be resolved. This is the well-known maximum at times $t$ comparable to the relaxation time; it is represented by a filled circle in fig.~\ref{plots}(a). As $T$ is lowered, however, a shoulder becomes apparent at time scales corresponding to cage exploring, as shown in fig.~\ref{plots}(b). This shoulder evolves into a secondary maximum, denoted by an open circle in fig.~\ref{plots}(c), as $T$ is lowered still further. Note that the primary cage-escaping maximum is quite broad in $t$ and that the secondary cage-exploring maximum is only observed at temperatures at which the cage-escaping maximum has moved to sufficiently long times. Thus, the reason why the observation of two distinct maxima is restricted to temperatures fairly close to the glass transition is because the smaller cage-exploring maximum is obscured by the larger cage-escaping maximum at higher temperatures.
\begin{figure}
\onefigure[width=80mm]{figure3.eps}
\caption{Local maxima of the dynamic susceptibility. We plot (a) the magnitudes of all observed local maxima of $\chi_{\rm ss}(a, t)$, $\chi^\star$, (b) their lag times $t^{\star}$, and (c) their overlap distances $a^\star$ as a function of temperature at fixed pressure $p=10^{-2}$ and system size $N=1600$. As in fig.~\ref{maps}, the filled circles represent cage-escaping maxima, while the open circles represent cage-exploring maxima. The vertical dashed lines denote the range over which we observe both types of maxima. Note that (a) and (c) are presented on a linear scale, while (b) is presented on a log-linear scale. Error bars represent the larger of (1) the standard deviation of the mean calculated from simulations with different initial conditions and (2) the uncertainty associated with the discrete sampling of $a$ and $t$.}
\label{max}
\end{figure}
Fig.~\ref{max} shows the temperature dependence of the two maxima of the dynamic susceptibility. We plot the amplitude of the maxima $\chi^\star$ and their associated locations in time and space, $t^\star$, and $a^\star$, as functions of $T$ at fixed pressure $p=10^{-2}$ and system size $N=1600$. The amplitude $\chi^\star$ and lag time $t^\star$ of the primary, cage-escaping maximum (solid symbols) both increase as $T$ decreases until $t^\star$ passes beyond our observable time window, in accord with many previous observations. The overlap distance $a^\star$ of the cage-escaping maximum is on the order of half a disk diameter for temperatures near the glass transition, a distance about two times larger than the plateau of the rms displacement. This distance presumably reflects the scale of the rearrangements necessary for diffusive motion to occur. At high temperatures above $T=6 \times 10^{-4}$, the plateau in the rms displacement is less well defined and the cage-escaping maximum lies at slightly larger values of $a^\star$.
The temperature dependence of the cage-exploring, secondary maximum is markedly different from that of the primary maximum. Figure~\ref{max} shows that $\chi^\star$ and $t^\star$ remain roughly constant as temperature decreases from $T=5.5\times 10^{-4}$, just above $T_g \approx 5 \times 10{-4}$, to $T=10^{-4}$, which is well below it. This is consistent with the behavior of the rms displacement; the time of the onset of the plateau does not vary much with temperature as long as the temperature is low enough for the plateau to exist. The cage-exploring overlap distance $a^\star$ decreases with decreasing $T$, remaining near the plateau value in the rms displacement.
\begin{figure}
\onefigure[width=80mm]{figure4.eps}
\caption{Dependence of the relaxation time $t^\star$ on the spatial extent of kinetic heterogeneities, $\chi^\star$. We plot the position of the cage-escaping maximum, $t^\star$ vs the magnitude of that maximum, $\chi^\star$ at fixed pressure $p=10^{-2}$. Each point corresponds to a different temperature above $T_g$. Results are shown for three different system sizes $N$, as labelled, on both (a) a log-linear scale and (b) a log-log scale. The straight lines in (a) and (b) are exponential and power law fits, respectively. Error bars represent the larger of (1) the standard deviation of the mean calculated from simulations with different initial conditions and (2) the uncertainty associated with the discrete sampling of $a$ and $t$.}
\label{tstar_chistar}
\end{figure}
The amplitude $\chi^\star$ of the cage-escaping maximum is a measure of the number of disks whose dynamics are correlated at the crossover to diffusion, while $t^\star$ is a measure of the timescale for this longest, or $\alpha$-, relaxation process. In fig.~\ref{tstar_chistar} we plot $\chi^\star$, the number of disks whose dynamics are correlated at the cage-escaping maximum of $\chi_{\rm ss}$, as a function of the time scale corresponding to the cage-escaping maximum, $t^\star$. This time scale is a measure of the $\alpha$-relaxation process. We have plotted $\chi^\star$ vs.~$t^\star$ for three different system sizes, $N=100$, 400, and 1600. Note the substantial system-size dependence in $\chi^\star$~\cite{Karmakar2009}, even for temperatures for which the size of correlated regions $n^\star\approx 4 \chi^\star$ is only on the order of $10$. This indicates that the finite size of the system affects the dynamics even when the relative linear size of the correlated regions is as small as $\sqrt{n^\star/N}\approx 0.1$. This strong effect is likely due to a broad distribution of sizes of correlated regions. We find that while the spatial correlation function of the overlap function has an average correlation length consistent with the value $\sqrt{\pi \langle \sigma^2\rangle n^\star/\phi}$ expected from a compact correlated region, it also has a long-ranged tail, decaying slower than exponentially with distance. This long range tail reflects intermittent periods when the the clusters are extended. We find that $t^\star$ and $a^\star$ also depend on system size. If $a$ is held fixed, $t^\star$ decreases and $\chi^\star$ increases with increasing system size $N$ at all temperatures studied, consistent with Ref.~\cite{Karmakar2009}. However, when we maximize over $a$, we find that at sufficiently high $T$, $t^\star$ increases with $N$ because $a^\star$ increases with $N$ there.
Fig.~\ref{tstar_chistar} shows that for each system size $N$, the dependence of $\chi^\star$ on $t^\star$ is qualitatively similar. Note that the log-linear plot (fig.~\ref{tstar_chistar}(a)) is straighter than the log-log plot (fig.~\ref{tstar_chistar}(b)) for each system size, indicating that the dependence of $t^\star$ on $\chi^\star$ is described better by an exponential than by a power law. Such an exponential dependence of relaxation time on cluster size is expected in scenarios involving cooperative rearrangements~\cite{Toninelli2005}. Previous numerical studies of model glass formers fit $\chi^\star$ and $t^\star$ to power laws in $T-T_c$, where $T_c$ is a critical temperature~\cite{Glotzer2000, Lacevic2003b, Szamel2006}, implying that the relationship between $\chi^\star$ and $t^\star$ is also a power law. However, these studies determined $\chi^\star$ as the maximum of $\chi_4(a, t)$ with respect to $t$ at fixed $a$, instead of searching for a maximum with respect to both $t$ and $a$. Also, the relationship between $\chi^\star$ and $t^\star$ was not shown directly. An analysis of a lower bound of $\chi_4$ for several real glass-forming liquids found an exponential dependence near the glass transition~\cite{Dalle-Ferrier2007}, consistent with our results.
\section{Discussion}
We quantified the size of kinetic heterogeneities for a model glass-forming liquid as a function of temperature in terms of fluctuations of a dynamic order parameter. By measuring the dynamic susceptibility $\chi_{\rm ss}(a, t)$ as a function of both lag time and overlap distance, we identified two distinct local maxima. The primary, ``cage-escaping" maximum, which has been found in previous numerical and experimental studies, occurs at the end of the plateau in the root-mean-square displacement, near the onset of diffusive transport. We also discovered a secondary, ``cage-exploring'' maximum at the beginning of the plateau in the mean-squared displacement, which corresponds to the ballistic to sub-diffusive crossover. Thus, the dynamics are most heterogeneous at displacements and times corresponding to crossovers in the dynamics.
The secondary, cage-exploring maximum at the crossover from ballistic to sub-diffusive motion emerges from the shoulder of the primary cage-escaping maximum when the temperature is low enough that there is a well-defined plateau in the mean-squared displacement. We find that the plateau must extend over at least two orders of magnitude in order for the cage-exploring maximum to appear. We observe both maxima over the range of temperatures that are low enough that the mean-squared displacement exhibits a broad plateau but high enough that the eventual crossover to the diffusive regime is still observable. Note that although both maxima emerge, grow, and shift continuously with temperature, there is an apparent discontinuous drop in the largest observed maximum at the dynamical glass transition. This occurs when the primary maximum passes out of observation range, leaving only the secondary, cage-exploring maximum (see fig.~\ref{max}(b)).
While the crossover from sub-diffusive to diffusive motion measures the approach to the glass transition, the crossover from ballistic to sub-diffusive motion is a general and relatively innocuous feature of dense fluids and solids. Our finding that both crossovers exhibit maxima in the dynamic susceptibility suggests that spatially heterogeneous dynamics are a general feature of dynamical crossovers caused by interactions. Indeed, a crystalline system of monodisperse disks exhibits a cage-exploring maximum at the end of the ballistic regime that is similar to the one shown here. We suspect that more exotic dynamical crossovers, like those between two diffusive regimes with different diffusion coefficients, should also exhibit kinetic heterogeneities. The existence of kinetic heterogeneities is therefore a general feature of dynamical crossovers and should not necessarily be taken as a sign of an impending glass transition.
\acknowledgments
We thank Adam Abate, Olivier Dauchot, Douglas Durian, and Smarajit Karmakar for useful discussions. This work was supported by DE-FG02-05ER46199 and in part by the MRSEC program under NSF-DMR05-20020.
\bibliographystyle{eplbib.bst}
|
\section{INTRODUCTION}
Aim of this contribution is to illustrate a fully
covariant model for describing the electromagnetic
(em) decay of neutral
Vector Mesons (VM), in both light and heavy sectors (see \cite{phipsi} for a
preliminary presentation of our approach). For the present stage,
we have considered only the ground states, to be described by using
Bethe-Salpeter
(BS)
amplitudes with a simple analytic form, namely exhibiting only single poles. This
assumption allows us
to easily perform the analytic integration needed for
calculating the constituent quark (CQ) transverse-momentum distribution
inside the
VMs
(see, e.g. Refs. \cite{melo02} and \cite{pion09} for the pion case)
and then
em decay widths. The free parameters present in the Ansatz for each VM are fixed
by comparing
the so-called {\em valence transverse-momentum distribution}, $n(k_\perp)$,
with the same quantity
evaluated within a Light-Front
Hamiltonian Dynamics (LFHD) approach exploiting the eigenfunctions
of two phenomenological mass operators, presented in Refs. \cite{qcd_tob} and
\cite{GI}, and able to reproduce the VM spectra. It should be pointed out
that the free parameters represent an effective way for including
some non perturbative features in our analytical Ansatz.
A possible form of
the BS amplitude for an interacting $q\bar q$ system with $J=1$, can be
written
as follows
\begin{eqnarray}&&
\Psi_\lambda(k,P)=S(k,m_1)~{\cal V}_{\lambda}(P)~S(k-P,m_2)\times \nonumber \\&&
\Lambda_{VM}(k,P)
\label{bsa}
\end{eqnarray}
where $S(k,m)=
S(k_{on},m) + {\gamma^+ / (2 k^+)}$ is the Dirac propagator
of a constituent quark with mass $m$, $k^\mu_{on}\equiv\{k^-_{on},k^+,{\bf k}_\perp\}$
with
$k^-_{on}=(m^2+|{\bf k}_\perp|^2)/k^+$,
$P^\mu$ the four-momentum of a VM
with mass $P^2=M_{VM}^2$, $\Lambda_{VM}(k,P)$ the
momentum dependence of the BS amplitude, and ${\cal V}_{\lambda}(P)=\epsilon_\lambda(P)
\cdot V(P)$ with $\epsilon^\mu_\lambda(P)$ the polarization
four-vector ($\lambda$ is the helicity) and $V^\mu(k,k-P)$
the Dirac structure given by the following
familiar expression, transverse to $P^\mu$, (see, e.g., \cite{ji92})
\begin{eqnarray}&&
V^\mu(P) ={M_{VM} \over M_{VM}+m_1+m_2}~\left [\gamma^\mu -
{P^\mu ~\slash \! \! \! \! P \over M_{VM}^2} + \right. \nonumber \\&& \left. + ~ \imath ~ {1 \over M_{VM}}~
\sigma^{\mu\nu} P_\nu\right ]
\end{eqnarray}
Then, the scalar product ${\cal V}_{\lambda}(P)$ reduces to
\begin{eqnarray}&&
{\cal V}_{\lambda}(P)={\slash \! \! \! \epsilon_\lambda\left [M_{VM} -
{\slash \! \! \! \! P }\right ]\over M_{VM}+m_1+m_2}
\label{dstruc}
\end{eqnarray}
It should be pointed out that one could have two different $\Lambda_{VM}(k,P)$'s, one
for each Dirac
structure in ${\cal V}_{\lambda}(P)$ (i.e.
$\slash \! \! \! \epsilon_\lambda$ and $\slash \! \! \! \epsilon_\lambda {\slash \! \! \! \! P }$),
but at the present stage we assume a
simpler form, with the same $\Lambda_{VM}(k,P)$ for the two structures, that
leads to the expected Melosh Rotation
factor for a
$^3S_1$ system, in the limit of non interacting systems, i.e. $M_{VM}\to M_{free}$
or $P^\mu \to P^\mu_{free}$
\cite{jaus91}. Therefore the following comparison with the
phenomenological LFHD outcomes can be made more strict.
In the
preliminary calculations presented in this contribution,
the momentum dependence of the BS amplitude is given by the following
Ansatz, with only single poles, viz
\begin{eqnarray}&& \Lambda_{VM}(k,P)= {\cal N}~\nonumber \\&&
\left[k^2-m_{1}^2+(P-k)^2-m_{2}^2\right ]~\Pi_{i=1,3} \times \nonumber \\&&
{1\over \left[k^2-m_{R_i}^2 +\imath\epsilon\right ]
\left [(P-k)^2-m_{R_i}^2 +\imath\epsilon\right ]}
\label{lambda}
\end{eqnarray}
where $m_{R_i}$, $i=1,2,3$ are free parameters (determined as
described below), and ${\cal N}$ the normalization factor, that is obtained
by imposing the normalization
for the BS amplitude in
Impulse Approximation, i.e. by adopting constituent quark free propagators.
The form chosen for $\Lambda_{VM}(k,P)$ allows one: i) to implement
the correct symmetry under the exchange of the quark momenta
(for equal mass constituents), as follows from the charge conjugation of the
neutral mesons; ii) to regularize the integrals needed in our approach for the evaluation of valence wave
functions, decay constants, transverse-momentum distributions, etc.; and iii)
to avoid
any free propagation in the valence wave function, given the presence of the
numerator in
$\Lambda_{VM}(k,P)$ (see also \cite{phipsi,Gross}).
To determine $m_{R_i}$ in Eq. (\ref{lambda}), we have introduced the
transverse-momentum
distribution of a CQ inside the VM, adopting the same definition
already applied by de Melo
et al. \cite{melo02} for a covariant description of
the pion. In a frame where ${\bf P}_\perp={\bf 0}$, one first
defines the
valence component of the VM state in the standard way
(see, e.g., \cite{HK}), viz.
\begin{eqnarray}&&
\Psi^{val}_{VM}(\xi,{\bf k}_\perp; m_{R_i})=\int dk^-~S(k_{on},m_1)~
\times\nonumber \\&&
{\cal V}_{\lambda}(P)~
\Lambda_{VM}(k,P)S\left((k-P)_{on},m_2\right )
\label{val} \end{eqnarray}
with $\xi=k^+/P^+$.
The normalization factor in $\Lambda_{VM}(k,P)$ (cf Eq. (\ref{lambda})) is
evaluated
in Impulse
Approximation by exploiting the total-momentum sum rule, that reads for the plus
component as follows
\begin{eqnarray}&&
P^+ =~{N_c\over 2}
\int\frac{d^4k}{(2\pi)^4}~{\bar\Lambda_{VM}(k,P)\over
\left[(k-P )^2- m^2_2+\imath\epsilon\right]}~
~\times \nonumber \\&&
{ { \bf T}^+_{\lambda,\lambda}(P,k) \over \left[(k-P)^2- m^2_2+\imath\epsilon\right]}
{ \Lambda_{VM}(k,P)\over (k^2 - m^2_1+\imath \epsilon) }
\end{eqnarray}
with $N_c=3$ and
\begin{eqnarray}&&
{ \bf T}^+_{\lambda,\lambda}(P,k)=Tr\left \{{\cal V}_{\lambda}(P)
~\times \right .\nonumber \\&& \left.({\slash \! \! \! k}-\slash \! \! \!{P} +m_2)
\gamma^{+} ~(\slash \! \! \!{k}-\slash \! \! \!{P}+m_2) ~\times \right .\nonumber \\&& \left.
{\cal V}^\dagger_{\lambda}(P)~(\slash \! \! \!{k}+m_1)\right \}=
\nonumber \\&&= -2 (P^+-k^+)\left[\phantom{{(k^- -k^-_{on}) \over 2}}\hspace{-1.5cm}
{\cal T}^+(P,k,on)+\right. \nonumber \\&& \left.
{(k^- -k^-_{on}) \over 2}{ \cal T}^+(P,k,ist)\right]
\label{trace} \end{eqnarray}
In Eq. (\ref{trace}),
$
{\cal T}^+(P,k,on)$
and
${ \cal T}^+(P,k,ist)$ are the on-shell and the instantaneous
contributions, respectively, according to the well-known decomposition of the Dirac
propagator (see below Eq. (\ref{bsa})).
Their explicit expressions will be given elsewhere \cite{fpps10}.
Notice that, after performing the trace, the dependence upon the
helicity disappears, and therefore its presence in ${ \bf
T}^+_{\lambda,\lambda}(P,k)$ becomes dummy.
\begin{table*}[tbh]
\caption{The mass eigenvalues for the ground states of the neutral VMs investigated in this contribution
compared with the
corresponding experimental values. The mass
operator corresponds to the one in Ref. \cite{qcd_tob} (see \cite{fpps10} for
details). The quark masses adopted for the actual evaluation are also listed.}
\label{tab1}
\vspace{.3cm}
\hspace{3.8cm}\begin{tabular}{cccc}\hline
{VM} & $m_q (MeV)$ &$M^{th}_{VM}$ (MeV) & $M^{exp}_{VM}$ (MeV)
\\
\hline
$\rho$ &328 &777.0&775.500 $\pm$ 0.4
\\
$\phi$ &449& 1020.7 &1019.455 $\pm$ 0.020
\\
$J/\psi$& 1559& 3082.9 & 3096.916 $\pm$ 0.011
\\
$\Upsilon$ &4891& 9537.2 &9460.300 $\pm$ 0.268
\\ \hline \\
\end{tabular}
\end{table*}
\begin{table*}[tbh]
\caption{The same as in Table 1, but for the mass
operator of Ref. \cite{GI}.}
\label{tab2}
\vspace{.3cm}
\hspace{.5cm}\begin{tabular}{cccc}\hline
{VM} & $m_q (MeV)$ &$M^{th}_{VM}$ (MeV) & $M^{exp}_{VM}$ (MeV)
\\
\hline
$\rho$ &220 &777&775.500 $\pm$ 0.4
\\
$\phi$ &419& 1016 &1019.455 $\pm$ 0.020
\\
$J/\psi$& 1628& 3091 & 3096.916 $\pm$ 0.011
\\
$\Upsilon$ & 4977& 9460 &9460.300 $\pm$ 0.268
\\ \hline
\end{tabular}
\end{table*}
Finally, let us introduce the valence transverse-momentum distribution, $n_{val}(k_\perp)$,
given by
\begin{eqnarray}&&
n_{val}(k_\perp)={N_c\over (2\pi)^3\left [P^+\right]^2 P^{VM}_{q\bar q}}
\int_0^{2\pi}d\theta_{k_\perp} ~\times \nonumber \\&& \int_0^1 d\xi{ {\cal T}^+(P,k,on)
\over
\xi ~(1-\xi)}~|\Phi^{val}_{VM}(\xi,{\bf k_{\perp}};
m_{R_i})|^2
\label{nk} \end{eqnarray}
where $P^{VM}_{q\bar q}$ is the probability of the valence component,
given by
\begin{eqnarray}&&
P^{VM}_{q\bar q}=
{N_c\over (2\pi)^3\left [P^+\right]^2}
\int_0^1 {d\xi \over
\xi ~(1-\xi)}~\times \nonumber \\&&\int d{\bf k}_\perp ~{\cal T}^+(P,k,on)~
|\Phi^{val}_{VM}(\xi,{\bf k_{\perp}};
m_{R_i})|^2
\label{nvm} \end{eqnarray}
and
\begin{eqnarray}&&
\Phi^{val}_{VM}(\xi,{\bf k_{\perp}};
m_{R_i})=~k^+~(P-k)^+~\int dk^-~\times \nonumber \\&&
{\Lambda_{VM}(k,P) \over (k^2-m^2_1)~\left[(k-P)^2-m^2_2\right]}
\end{eqnarray}
In a LFHD
approach (see, e.g. \cite{melo02} and references quoted therein), one has
\begin{eqnarray}&&
n^{HD}_{VM}(k_{\perp})=
\int_0^{2\pi} d\theta_{ k_\perp}\int_0^1 {d\xi ~M^2_{free}\over
\xi (1-\xi)} ~ \times\nonumber \\&&
|\psi_{VM}(\xi,{\bf k}_{\perp})|^2
\end{eqnarray}
where $\psi_{VM}(\xi,{\bf k}_{\perp})$ is the eigenfunction of a relativistic
mass
operator and $M^2_{free}=(m^2+|{\bf k}_{\perp}|^2)/\xi (1-\xi)$. In order to
implement our procedure for fixing the free parameters in $\Lambda_{VM}$,
we have first
evaluated $n^{HD}_{VM}(k_{\perp})$
by adopting the eigenfunctions obtained by using the model mass operators of both
Refs. \cite{qcd_tob} and \cite{GI}, and then we have minimized the
difference $|n_{val}( k_{\perp})-n^{HD}_{VM}(k_{\perp})|$.
It should be
pointed out that the VM
spectra are satisfactorily reproduced for each VM investigated
in this contribution. As an example of the reliability of the adopted
LFHD models, the masses of the considered neutral VMs
for the ground states are compared with
the corresponding experimental
values in Tables \ref{tab1} and \ref{tab2} for the mass operators
of Refs. \cite{qcd_tob} and \cite{GI}, respectively.
Indeed, also the masses of the
first three/four (depending upon the considered VM) excited states are
well reproduced by the model mass operators.
Once the parameters have been determined, one can calculate both static (e.g., em decay rates)
and dynamical (e.g. parton distributions) quantities. For completing
this section, devoted to the general presentation of the formalism, we give the
expression of
the parton distribution,
i.e.
\begin{eqnarray}&&
q(\xi)= \int dk_\perp ~ k_\perp~f_1(\xi,k_\perp) \end{eqnarray}
in terms of the chiral-even transverse-momentum distribution,
$f_1(\xi,k_\perp)$,
(see, e.g. \cite{fpps10} and \cite{pion09} for the pion case), that yields the
distribution of a constituent
inside the meson with the longitudinal and transverse components
of the LF-momentum. It reads
\begin{eqnarray}&& f_1(\xi,k_\perp)=~{N_c\over 4 P^+}~\int^{2\pi}_0 d\theta_{k_\perp}~
\int\frac{dk^-}{(2\pi)^4}~\times\nonumber \\&&
~{\bar\Lambda_{VM}(k,P )\over \left[(k-P )^2- m^2_2+\imath\epsilon\right]}
{{ \bf T}^+_{\lambda,\lambda}(P,k)\over \left[(k-P)^2- m^2_2+\imath\epsilon\right]}
\times \nonumber \\&&{\Lambda_{VM}(k,P) \over (k^2 - m^2_1+\imath \epsilon) }
\end{eqnarray}
The explicit expression of $f_1(\xi,k_\perp)$ will be presented elsewhere
\cite{fpps10}.
\section{THE MANDELSTAM FORMULA FOR THE EM DECAY CONSTANT}
In order to evaluate the em decay constant, $f_{VM}$, we adopt
a Mandelstam-like formula \cite{mandel} (see also \cite{pionprd} and
\cite{pion09}
for the pion case).
The starting point is the {\em macroscopic} definition of $f_{VM}$, through
the transition matrix element of the em current for a given neutral VM, viz
\begin{eqnarray}&&
\langle 0|J^{\mu}(0)| P,\lambda \rangle =
\imath \sqrt{2}f_{VM} \epsilon_{\lambda}^{\mu}
\label{eq:fv}
\end{eqnarray}
Let us remind that the decay constant $f_{VM}$ is related to the em decay width
as follows
\begin{eqnarray}&&
\Gamma^{VM}_{e^+ e^-}=\frac{8\pi\alpha^2}{3}\frac{|f_{VM}|^2}{M_{VM}^3}
\label{eq:gamma}
\end{eqnarray}
In our model, the transition matrix element in Eq. (\ref{eq:fv}) can be
approximated {\it microscopically}
{\it \`a la} Mandelstam through
\begin{eqnarray}&&
\langle 0|J^{\mu}(0)| P,\lambda \rangle=
{\cal Q}_{VM}~{{N_c}{\cal N} \over (2\pi)^4}\int d^4 k~\times \nonumber \\&&
{\Lambda_{VM}(k,P,m_1,m_2)\over
(k^2-m^2_1+ \imath \epsilon)~[(P-k)^2-m^2_2
+ \imath \epsilon]} \times \nonumber \\&&
{\rm Tr} [ {\cal V}_\lambda(P) ~(\slash \! \! \! k -\slash \! \! \! P +m_2)\gamma^\mu
(\slash \! \! \! k+m_1)]
\label{eq:mandel}
\end{eqnarray}
where $$ {\cal Q}_\rho= {(Q_u-Q_d) \over \sqrt{2}} \quad \quad
{\cal Q}_\phi= Q_s$$ $$
{\cal Q}_{J/\Psi}= Q_c \quad \quad
{\cal Q}_{\Upsilon}= Q_b
$$
with $Q_i$ the quark charge. The lengthy algebra for evaluating the analytic
integrations in Eq. (\ref{eq:mandel}) will be reported
elsewhere \cite{fpps10}.
\section{PRELIMINARY RESULTS}
In Figs. 1 and 2, the valence transverse-momentum distributions (cf Eq. (\ref{nk}))
for
$\rho$, $\phi$, $J/\psi$ and $\Upsilon$ are shown, for different sets of the free
parameters.
In order to get some dynamical input, in Fig. 1
the free parameters of our Ansatz (see Eq. (\ref{bsa})) have been
determined through the model of Ref. \cite{qcd_tob}, while in Fig. 2
through the model of Ref. \cite{GI}. In the figures the distributions are
multiplied by a factor $k_\perp$, to have an immediate intuition of the
transverse-momentum region relevant in the evaluation of the valence
probabilities. It is worth noting that the long tail of the heavy VM is an expected feature.
In Figs. 3 and 4, the parton distributions are presented. As it is shown, the general
behavior at the
end-points and the accumulation of the distribution for $x=1/2$ as
the CQ mass increases are recovered
by using our Ansatz.
Finally, in Tables 3 and 4, the preliminary values for the em decay widths,
for the two
different choices of the free parameters,
are compared
with the corresponding experimental data. It should be pointed out that while
the model of Ref. \cite{qcd_tob} yields reasonable results for the light sector
(let us remind that the confining interaction for the squared mass operator of
\cite{qcd_tob} is parabolic), in the heavy sector the model of Ref. \cite{GI}
seems to behave better (for this model a linear confining interaction is considered
in the mass operator). Such a comparison suggests that an improved description
of the confining interaction for the squared mass operator
of Ref. \cite{qcd_tob} could lead the theoretical predictions of the em
decays closer to the experimental ones.
\begin{table*}[tbh]
\caption{Preliminary VM em decay widths, calculated using our Ansatz,
(\ref{bsa}), and the
free parameters determined through the comparison with the eigenfunctions of the
model \cite{qcd_tob} for the mass operator.}
\label{tab3}
\vspace{.3cm}
\hspace{3.5cm}\begin{tabular}{ccc} \hline
{VM} & {$\Gamma_{e^+ e^-}^{th}$
(keV)}& {$\Gamma_{e^+ e^-}^{exp}$ (keV)} \cite{PDG08}\\
\hline
$\rho$ & 8.579 & 7.04 $\pm$ 0.06
\\
$\phi$ & 1.952 & 1.27 $\pm$ 0.04
\\
$J/\psi$ & 2.526 & 5.55 $\pm$ 0.14
\\
$\Upsilon$ & 0.187 &1.340$\pm$ 0.018
\\ \hline
\end{tabular}
\end{table*}
\begin{table*}[tbh]
\caption{The same as in table 1, but using the eigenfunctions of the
model \cite{GI} to fix the free parameters.}
\label{tab4}
\vspace{.3cm}
\hspace{3.5cm}\begin{tabular}{ccc} \hline
{VM} & {$\Gamma_{e^+ e^-}^{th}$
(keV)}& {$\Gamma_{e^+ e^-}^{exp}$ (keV)} \cite{PDG08}\\
\hline
$\rho$ & 24.791 & 7.04 $\pm$ 0.06
\\
$\phi$ & 4.342 & 1.27 $\pm$ 0.04
\\
$J/\psi$ & 7.102 & 5.55 $\pm$ 0.14
\\
$\Upsilon$ & 0.493 &1.340$\pm$ 0.018
\\ \hline
\end{tabular}
\end{table*}
\begin{figure}
\vspace{.3cm}
{\includegraphics[width=7.5cm]{LC09n_kperp.eps}}
\caption{Valence transverse-momentum distributions for a constituent inside
$\rho$, $\phi$, $J/\Psi$ and $\Upsilon$
vs the quark transverse momentum. The normalization is $
\int k_\perp~d{ k}_\perp ~n_{val}( k_{\perp})=1
$. Solid line: $\rho$; dashed line: $\phi$; dotted line: $J/\Psi$;
double-dot-dashed line: $\Upsilon$. The
free parameters in (\ref{bsa}), have been determined
through the comparison with the eigenfunctions of the
model \cite{qcd_tob}.}
\end{figure}
\begin{figure}
\vspace{-1.8cm}
{\includegraphics[width=7.5cm]{LC09n_kperp_GI.eps}}
\caption{The same as in Fig. 1, but using the eigenfunctions of the
model \cite{GI} to fix the free parameters in (\ref{bsa}).}
\end{figure}
\begin{figure}
{\includegraphics[width=7.5cm]{LC09Qx.eps}}
\caption{Valence distributions for a constituent inside
$\rho$, $\phi$, $J/\Psi$ and $\Upsilon$
vs $x$, with normalization $\int d\xi~q(\xi)=1$. The same line convention as in
Fig. 1 is adopted. The
free parameters in (\ref{bsa}), have been determined through
the comparison with the eigenfunctions of the
model \cite{qcd_tob}.}
\end{figure}
\begin{figure}
{\includegraphics[width=7.5cm]{LC09Qx_GI.eps}}
\caption{The same as in Fig. 3, but using the eigenfunctions of the
model \cite{GI} to fix the free parameters in (\ref{bsa}).}
\end{figure}
\section{CONCLUSIONS}
The preliminary results for
some static and dynamical quantities for neutral VM, obtained
within our covariant description, both in the light and
the heavy sectors, have been shortly presented. After
calculating
the valence wave functions, Eq. (\ref{val}),
and the transverse-momentum
distributions, Eq. (\ref{nk}), we have fixed the three parameters in our Ansatz for the
Bethe-Salpeter amplitude through a comparison with the
transverse-momentum
distribution
obtained within a Light-Front Hamiltonian Dynamics approach. The
eigenvectors of two different mass operators, corresponding to Ref.
\cite{qcd_tob} and Ref. \cite{GI}, have been considered in order to perform the comparison.
Then, we have evaluated the parton distributions and the em
decay constants (cf
Table \ref{tab3} and Table \ref{tab4}).
The work in progress will substantially improve the present calculations,
in two respect: both introducing a more refined Ansatz for the BS amplitude
and considering new phenomenological mass operators.
This work was partially supported by the Brazilian agencies CNPq and
FAPESP and by Ministero della Ricerca Scientifica e Tecnologica.
S.P. acknowledges the "Fondazione Della Riccia" for supporting
her research activity and the hospitality of ITA.
|
\section{\label{}}
\section{Introduction}
The diffuse Galactic emission (DGE) arises from
interactions of cosmic-rays (CRs) with interstellar gas and radiation
field in the Galaxy. Due to the smooth nature of the
interstellar radiation field and the CR flux after propagation, the fine structure of the DGE is
determined by the structure of the interstellar gas. Getting the
distribution of the interstellar gas correct is therefore crucial when
modeling the DGE.
It is generally assumed that Galactic CRs are accelerated in interstellar
shocks and then propagate throughout the Galaxy \citep[see e.g.][for a
recent review.]{Strong2007}. In
this paper, CR propagation and corresponding diffuse emission is
calculated using the GALPROP code \citep[see][and references
within.]{Strong2004}. We use the so-called
conventional GALPROP model \citep{Strong2004}, where the CR injection spectra and the
diffusion parameters are chosen such that the CR flux agrees
with the locally observed one after propagation. The gas distribution is
given as Galacto-centric annuli and the diffuse emission is calculated
for those same annuli. The distribution of H I is
determined from the 21-cm LAB line survey \citep{Kalberla2005} while
distribution of molecular hydrogen, H$_2$,
is found using the CO ($J=1\rightarrow0$) survey of \cite{Dame2001}
assuming $N_{\rm{H}_2} = X_{CO}(R) W_{CO}$.
While converting observations of the 21-cm H I line to column density is
theoretically possible, it is not practically feasible. To correctly
account for the optical depth of the emitting H I gas, one must know its
spin temperature, $T_S$ \citep[see e.g.][]{Kulkarni1988}.
Under the assumption of a constant $T_S$ along the line of sight, the
column density of H I can be calculated from the observed brightness
temperature $T$ using
\begin{equation}
N_{H I}(v,T_S) = -\log\left(1-\frac{T}{T_S-T_{bg}}\right) T_S C,
\label{eq:OpticalDepthCorrection}
\end{equation}
where $T_{bg}$ is the background continuum temperature and $C =
1.83\times10^{18}$ cm$^{-2}$
K (km/s)$^{-1}$. The assumption of a constant $T_S$ along the line of
sight is known to be wrong for many directions in the Galaxy \citep[see
e.g.][]{Dickey2009}. The $T_S$ values derived in this paper are
therefore only a global average and should not be taken at face value.
Figure~\ref{fig:TsRatio} shows how changing $T_S$
affects $N_{H I}$ in a non-linear way, mainly affecting areas with $T$
close to $T_S$ in the Galactic plane. This figure was created under the
assumption of a fixed $T_S$ for the whole Galaxy that is known to be
wrong but has been used for DGE analysis from the days of
COS-B \citep{Strong1985}. Note that for
equation~(\ref{eq:OpticalDepthCorrection}) to be valid the condition
$T_S > T+T_{bg}$ must hold. When generating the gas annuli, this
condition is forced by clipping the value of $T$.
While the assumption of a constant spin temperature $T_S = 125\,\rm{K}$ for the whole
Galaxy may have been sufficient for older instrument, it is no longer
acceptable for a new generation experiment like Fermi-LAT
\citep{Atwood2009}. This has
been partially explored for the outer Galaxy in \citet{2ndQuadrant}. In this
paper we will show a better assumption for $T_S$ can be easily found and
also show that direct observations of $T_S$ using absorption measurement
of bright radio sources are needed for accurate DGE modeling.
\begin{figure}
\begin{center}
\includegraphics[width=75mm]{rbands_hi12_Ts125-Ts200_Ratio_Healpix.png}
\end{center}
\caption{The ratio $N_{H I}(125\,
\rm{K})/N_{H I}(200\,\rm{K})$ in
Galactic coordinates. The figure clearly shows the non-linearity of the
correction that can be as high as a factor of 2 in this case.}
\label{fig:TsRatio}
\end{figure}
\section{Method}
We assume the source distribution of CR nuclei and electrons are the
same. CR propagation is handled by GALPROP and we use the conventional
model so that after the propagation the CR spectra agree with local
observations. The GALPROP diffuse emission is output in Galacto-centric
annuli, split up into different components corresponding to different
processes (bremsstrahlung, $\pi^0$-decay, and inverse Compton scattering). To
allow for radial variations in CR intensity we perform a full sky maximum
likelihood fit,
preserving the spectral shape of each component. We allow for one
global normalization factor for the electron to proton ratio.
Additionally, we also allow for radial variation in the $X_{CO}$ factor.
This accounts for uncertainties in the CR source distribution and
$X_{CO}$ factor.
The maximum likelihood fits were performed on the whole sky using the
GaRDiAN package \citep{Abdo2009} after preparing the Fermi-LAT data with
the science tools. We use the same dataset as \cite{Ackermann2009} that
has special cuts to reduce CR background contamination compared to the
standard event selection \citep{Atwood2009}. In addition to the
DGE model, we also include all sources from the 1 year Fermi-LAT source
list \citep{Abdo2010} and an isotropic component to account
for EGB emission and particle contamination. This fit is
performed for different assumptions of $T_S$ and a likelihood
ratio test is used to compare the quality of the fits.
\section{Results}
The simplest assumption is that of a constant $T_S$ for the whole Galaxy
and it deserves some attention
for historical reasons. It will also serve as a baseline model for
comparison with other assumptions.
To get an approximation for the best model, we scan $T_S$ from 110 K to
150 K in 5 K steps. Our results show that $T_S=130\,\rm{K}$ gives the
maximum likelihood for this setup.
One of the problems with the constant global $T_S$ approximation, apart
from the fact that observations of the interstellar gas have shown it
to be wrong, is that the
maximum observed brightness temperature in the LAB survey is $\sim$150 K
which is greater than our best fit global $T_S$. This is solved by
clipping the observations when generating the gas annuli, which is not
an optimal solution. A different possibility is to use
the assumption
\begin{equation}
T_S = \max(T_{S,\rm{min}}, T_{\rm{max}}+\Delta T_S).
\label{eq:LinearTS}
\end{equation}
Here, $T_{\rm{max}}$ is the maximum observed brightness temperature for each line
of sight. This ensures $T_S$ is always greater than $T$.
Scanning the values of
$T_{S,\rm{min}}$ and $\Delta T_S$ with a step size of 10 K and 5 K,
respectively, gives us a maximum likelihood for
$T_{S,\rm{min}} = 110\,\rm{K}$ and $\Delta T_S=
10\,\rm{K}$. While this assumption still does not account for the
complexity of the interstellar medium, the log likelihood ratio between the best fit linear
relation model and the best fit constant $T_S$ model is of the order of
1000, a significant change.
The most accurate $T_S$ estimates come from observations of H
I in absorption against bright radio sources. We gathered over 500
lines of sight with observed $T_S$ from the literature
\citep{Strasser2004, Heiles2003, Dickey2009}. This
covers about 0.2\% of the pixels in the LAB survey, allowing for accurate
column density estimates only in those pixels. After taking our best
fit linear relation model and correcting the pixels with known $T_S$ the
fit was redone for the whole sky. Note that we did not change the values of
$T_{S,\text{min}}$ and $\Delta T_S$. The log likelihood ratio of --105
tells us
that this model is worse than the best fit linear relation. This is not
unexpected, since the gamma rays are generated from CR interactions with
the gas and if the gas distribution is wrong, we won't get the correct CR
distribution from the fit. To limit the
uncertainty involved with the linear relation assumption, we did another
fit, limiting ourselves to the region $-10^\circ < |b| < 10^\circ$,
$15^\circ < l < 165^\circ$ that
covers the observations made in the Canadian Galactic plane survey
(CGPS) where the density of $T_S$ observations is the highest and is large
enough to get a good fit to the LAT data. The fit in this region
results in a log likelihood ratio of 28 indicating a statistically
significant improvement in the fit. This is despite the observed $T_S$
lines of sight only covering 25\% of the fitted region and the values of
$T_{S,\rm{min}}$ and $\Delta T_S$ not being adjusted after correcting for known $T_S$ values.
\section{Discussion}
Our small exercise here has shown that for accurate DGE modeling we
need to know more about the distribution of gas in the Galaxy,
especially the H I distribution. The standard constant $T_S$ assumption
is not sufficient for current instruments and small adjustments cause
large differences in the quality of the resulting model. We also show
that direct observations of $T_S$ help in creating a better model
of the DGE. Unfortunately, direct observations of $T_S$ are difficult
since they require high resolution telescopes and bright radio continuum
sources. Some assumptions will therefore have to be made for the
regions in between bright radio sources.
It must be stated here that all of the above results are model
dependent. The Fermi-LAT data can only provide us with the intensity of
gamma-rays from a particular direction of the sky. Uncertainties in our
modeling of contribution other than those directly related to the H I
distribution will affect the value obtained for $T_S$. We are currently
studying the systematic effects this will have on our results. We also
note that even for the best fit models, the residuals show signs of
structure, strongly indicating our models are less than
perfect.
|
\section{Introduction}
Although QCD is well established as the theory of strong interactions,
the hadronic realm is beyond the reach of perturbative calculations.
In that regime, lattice methods are a useful tool to
calculate QCD observables, but results on light meson
resonances are few and usually
obtained at very large quark masses compared with their
physical values.
Very recently \cite{Hanhart:2008mx},
an alternative technique, based on
Chiral Perturbation Theory (ChPT) and dispersion relations,
has been applied to calculate the
dependence of the $f_0(600)$ (or ``sigma'') and
$\rho(770)$ resonances on the averaged u and d quark mass, $\hat{m}$.
In this talk we report our progress extending this study to include the strange quark
within a unitarized SU(3) ChPT formalism \cite{Nebreda}. Our aim is threefold:
to confirm previous results within a more general formalism, to analyze the dependence on the
$\hat m$
of the $K^*(892)$ and
$\kappa(800)$ strange resonances and then,
to study the dependence of all the
$f_0(600)$, $\kappa(800)$, $\rho(770)$ and
$K^*(892)$ parameters in terms of the strange quark mass $m_s$.
For unitarization we use the well-known one-loop elastic
Inverse Amplitude Method (IAM) that provides a remarkable description of
$\pi\pi$ and $K\pi$ data up to $\sim 1$ GeV, while generating the
poles associated to the $f_0(600)$, $\rho(770)$, $K^*(892)$
and $\kappa(800)$ resonances \cite{Dobado:1996ps}.
This is achieved using Low Energy
Constants (LECs)
compatible with those of standard ChPT.
Previous descriptions come from fitting
only to experiment (see details in \cite{GomezNicola:2001as}),
and therefore are mostly sensitive to the LECs
that govern the $s$ dependence of partial waves.
In order to get better determinations of the LECs that
carry an explicit meson mass
dependence, now we have fitted also to lattice results on $M_\pi$, $M_K$,
$f_\pi$, $f_K$ and scattering lenghts \cite{lattice}.
In Fig.~\ref{fig:fits} we present two new fits: "Fit I" describes best
the data, but we show also "Fit II" to give an idea of the typical size
of the systematic uncertainties. It also makes clear that we
don't need any fine tuning of the LECs
to describe the experimental and lattice results together.
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=.45,angle=-90]{experimentalesNuevosFits.ps}\\
\includegraphics[scale=.45,angle=-90]{latticeNuevosFits.ps}
\end{tabular}
\caption{
Two upper rows: results of our IAM fits versus experimental data on $\pi\pi$ and $\pi K$
scattering. Two lower rows: result of the unitarized fits to lattice calculations of $f_\pi$, $f_K$,
$m_\pi/f_\pi$ and the $\pi^+\pi^+$, $K^+ K^+$, $K^+\pi^+$
scattering lengths.
The continuous and dashed lines correspond, respectively, to Fits I and II. For comparison
we show the results of the IAM
if we used the ChPT LECs obtained from $K_{l4}$ \cite{Amoros:2001cp} (dotted line) and the results of
standard non-unitarized ChPT with the sets of LECs given in \cite{Pelaez:2004xp} (dot-dashed line). The lattice data come from \cite{lattice} and the references for the experimental data are given in \cite{GomezNicola:2001as}.
}
\label{fig:fits}
\end{figure}
\vspace*{-.4cm}
\section{Dependence on the mass of the light quarks}
The $\hat m$ is a parameter of ChPT closely related to
the pion mass that we can vary to study the corresponding
change in the resonances.
It is known \cite{Dobado:1996ps} that the IAM works for Goldstone bosons masses
at least as high as 500 MeV. Since we want pions
always lighter than kaons and etas in order to apply the elastic
approximation, we will show results up to
$M_\pi< 440\,\rm MeV$ but not beyond, since then
$M_K\simeq 600\,\rm MeV$.
In terms of quark masses, this means $\hat m/\hat m_{\rm phys}\leq 9$.
\vspace*{-.3cm}
\paragraph{Light vector mesons}
The $\rho(770)$ and $K^*(892)$ vector
resonances are well established $q \bar q$ states belonging to
an SU(3) octet. In the first two rows of Fig.~\ref{fig:mhatdependence}, we show their dependence
on the non-strange quark masses. For each resonance,
its mass and width are defined
from the position of the associated pole in the second Riemann sheet,
through the usual Breit-Wigner identification $\sqrt{s_{pole}}\equiv M-i\,\Gamma/2$, and
its coupling to two mesons is given by the residue of the amplitude at the pole position.
The results obtained for the $\rho(700)$ resonance are very consistent
with the previous SU(2) results in \cite{Hanhart:2008mx} (dotted line in Fig.~\ref{fig:mhatdependence}), and the estimations
for the two first coefficients of the $M_\rho$ chiral expansion \cite{bruns}.. Besides,
the similarity of the behavior of $\rho(770)$ and $K^*(892)$ is evident.
Their masses increase smoothly
as the quark mass increases, but much slower than the pion mass.
This means that
the thresholds grow faster than the masses of the resonances, and
as a consequence there is a strong phase space suppression.
In fact, the decrease of their widths agrees remarkably well with
the expected reduction coming only from phase space suppression without a dynamical effect
through the vector coupling to two mesons (thin lines).
Accordingly, we see that $g_{\rho\pi\pi}$ and
$g_{K^*\pi K}$ are remarkably constant,
which is an assumption made in lattice studies
of the $\rho(770)$ width \cite{Aoki:2007rd}.
\vspace*{-.3cm}
\paragraph{Light scalar mesons}
The $f_0(600)$, or sigma, and the $\kappa(800)$ scalar mesons
are still somewhat controversial. Their huge width makes their experimental identification complicated and there are no present lattice calculations
with realistic quark masses.
It is therefore more interesting to obtain predictions on their
quark mass dependence.
In the third and fourth rows of Fig.~\ref{fig:mhatdependence}
we show their dependece on the non-strange quark mass.
As before, we find that for the $f_0(600)$ the results are
in good agreement with the existing SU(2) calculation of
\cite{Hanhart:2008mx}.
The most prominent feature of the scalars behavior
is the appearance of two branches for the mass.
The reason is that for physical values of the quark mass,
the poles associated to resonances appear as
conjugated poles in the second Riemann sheet. As the quark mass increases these poles
move closer to the real axis
until they join in a single pole below threshold and then
split again, now remaining in the real axis.
Despite the evident qualitative similarities, the quantitative behavior of the $f_0(600)$
and the kappa is rather different.
In particular, the growth of the $f_0(600)$ mass before the ``splitting point'' is
much faster than that of the $\kappa(800)$.
Regarding their width decrease, we show that for the scalars it
cannot be attributed to the phase space reduction
due to the increase of pion and kaon masses.
Related to this, we see that their coupling constants to two mesons show a strong quark mass dependence. Moreover, they
increase dramatically near the ``apparent splitting point''.
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=.53]{rhokstodo.ps}\\
\includegraphics[scale=.53]{sigmakappatodo.ps}
\caption{sigmakappa}
\end{tabular}
\label{fig:mhatdependence}
\caption{Dependence of the $\rho(770)$, $K^*(892)$, $f_0(600)$
and $\kappa(800)$ mass, width and coupling to two mesons with respect to
the non-strange quark mass $\hat m$ (horizontal upper scale), or
the pion mass (horizontal lower scale).
Note that we give all quantities normalized to their
physical values.
The thick continuous and dashed lines correspond to
Fit I and Fit II, respectively.
For the $\rho(770)$ and the $f_0(600)$ these results are very
compatible with those in \cite{Hanhart:2008mx}
using SU(2) ChPT (dotted line).
The continuous (dashed) thin line shows the $M_\pi$ dependence of
the widths from the change of phase space only, assuming a constant
coupling of the resonances to two mesons, $\rho(770)$ and $f_0(600)$
to $\pi\pi$ and
$K^*(892)$ and $\kappa(800)$ to $\pi K$, calculated from the
dependence of masses and momenta given by Fit I (II).
}
\end{figure}
\vspace*{-.2cm}
\section{Dependence on the mass of the strange quark}
We will only vary the strange quark mass
in the limited range $0.8<m_s/m_{s\,\rm phys}<1.4$ to ensure that the kaon does not become
too heavy to spoil the ChPT convergence nor too light to require a coupled channel
formalism
\vspace*{-.3cm}
\paragraph{Light vector mesons}
In the two upper rows of Fig.~\ref{fig:msdependence}
we show the kaon (or strange quark) mass dependence
of the $\rho$ and $K^*(892)$.
As it could be expected, the properties of the $\rho(770)$ being non-strange
are almost independent of the strange quark mass within the range of study.
Obviously, the $K^* (892)$ shows a strong $m_s$ dependence.
As the kaon mass is made heavier, the $K^*(892)$ mass grows much
faster than it did when increasing the light quark mass.
On the other hand, the $K^*(892)$ mass grows much slower than the kaon mass,
so that phase space shrinks and the resonance width
decreases almost exactly as it would be expected from phase space suppression only.
Accordingly, its coupling to $K\pi$ is almost constant.
\vspace*{-.3cm}
\paragraph{Light scalar mesons}
In the lower two rows of Fig.~\ref{fig:msdependence}, we show the variation of the sigma
and $\kappa(800)$ properties.
Again, the effect on the resonance in the $\pi\pi$ chanel, the sigma, is very small.
On the contrary, the mass and width of the $\kappa(800)$
change by as much as 16\% within the range of study. Finally we see again
that the width decrease deviates significantly from what expected from only phase space
suppression, in agreement with $g_{\kappa\pi K}$ depending quite strongly on the strange
quark mass.
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=.53]{rhoksmstodo.ps}\\
\includegraphics[scale=.53]{sigmakappamstodo.ps}
\caption{sigmakappa}
\end{tabular}
\label{fig:msdependence}
\caption{Dependence of the $\rho(770)$, $K^*(892)$, $f_0(600)$
and $\kappa(800)$ mass, width and coupling to two mesons with respect to
the strange quark mass $\hat m$ (horizontal upper scale), or
the kaon mass (horizontal lower scale).
Note that we give all quantities normalized to their
physical values.
The thick continuous and dashed lines correspond to
Fit I and Fit II, respectively.
The continuous (dashed) thin line shows the $M_K$ dependence of
the widths from the change of phase space only, assuming a constant
coupling of the resonances to two mesons, $\rho(770)$ and $f_0(600)$
to $\pi\pi$ and
$K^*(892)$ and $\kappa(800)$ to $\pi K$, calculated from the
dependence of masses and momenta given by Fit I (II).
}
\end{figure}
\vspace*{-.3cm}
\section*{Acknowledgments}
\vspace*{-.1cm}
Work partially supported by Spanish Ministerio de
Educaci\'on y Ciencia contracts: FPA2007-29115-E,
FPA2008-00592 and FIS2006-03438,
U.Complutense/Banco Santander grant PR34/07-15875-BSCH and
UCM-BSCH GR58/08 910309 as well as the European Community-Research Infrastructure
Integrating Activity
``Study of Strongly Interacting Matter''
(HadronPhysics2, Grant Agreement
n. 227431)
under the Seventh Framework Programme of the EU.
\vspace*{-.2cm}
|
\section{Introduction}
Short hard $\gamma$- ray bursts (SHBs) are now suspected to be caused
by the merging of two compact objects, such as neutron stars or
black holes, which would release large amounts of energy over short
time intervals (e.g. Goodman, 1986). Collapse of a single object has also
been proposed to give rise to a similar situation (Berezinsky 1987).
Eichler et al (1989) suggested that GRBs could be observed in
coincidence with GW signals when two neutron stars merge.
The huge
isotropic equivalent energy requirements implied by the
BATSE observations of GRB isotropy and submaximal $V/V_{max}$,
suggested that GRBs might be highly collimated (e.g. Levinson and
Eichler, 1993) and this would make them a bad bet to corroborate GW
signals from such mergers, as gravitational radiation is unlikely to
be strongly collimated. This might be fatal (e.g. Guetta and
Stella, 2009) to the original proposal
that LIGO signals would be coincident with GRBs. It is now accepted that GRBs are indeed highly collimated.
Alternative ways to confirm LIGO signals from
coalescing neutron stars [and, according some suggestions (Van
van Putten 1999a,b), unstable collapsing disks] are therefore all the
more desirable.
The horizons of first generation LIGO and Virgo for NS-NS,
and (BH-NS) mergers are $\sim 20$ and $43$ Mpc,
respectively, while advanced LIGO/Virgo should detect them out to a
distance of $\sim 300$ and $650$ Mpc (for a review see Cutler \&
Thorne 2002). GW signals from NS-NS mergers are expected at a rate
of one in 10-150 years with Virgo and LIGO and one every 1-15 days
with Advanced LIGO/Virgo class interferometers (Berezinsky et al.
2002, 2007). The BH-NS and BH-BH merger rates in the Galaxy are
highly uncertain. Berezinsky et al. (2007) estimate 1\% and 0.1\% of
the NS-NS merger rate, respectively, implying that BH-NS and BH-BH
mergers contribute marginally to the GW event rate, despite the
greater distance out to which they can be detected.
Ultahigh energy (UHE) neutrinos may come from nearby supernova even if an associated
GRB is shaded from our view or entirely smothered by the envelope of the host star (Eichler and
Levinson,
1999). A fluence $F$ of $10^{-4}$erg/cm$^2$ in muon neutrinos at
$10^{12} \le E_{\nu} \le 10^{14}$ eV yields roughly a single
neutrino detection in a gigaton detector such as ICECUBE, the
exact
expectation value depending somewhat upon the energy. An UHE
neutrino signal from a nearby supernova or supernova/GRB could
therefore be detectable at a distance of $D\sim 10^2
E_{iso,\nu,50}^{1/2}$ Mpc. Note that $E_{iso}$ for $\gamma$-rays can
be as high as $10^{54}$ erg [$E_{iso,\gamma,50} = 10^4$]. We face
the following interesting possibility: If the UHE neutrinos from GRB
are beamed into a wider beam than the $\gamma$-rays, then even if
the neutrino efficiency is high, the value of $E_{iso,\nu}$ may be
too low to be seen from any given burst unless it is close.
More importantly, most UHE neutrino events from GRB sources
would not coincide with observed GRBs, as the latter would be most
likely beamed away from us. For corroborating UHE neutrino signals,
as is the case for GW corroboration, we therefore seek
electromagnetic signals that have broader angular reach than the
primary $\gamma$-rays, even at the expense of $E_{iso}$.
We note that several wide
angle manifestations of nearby GRBs have been proposed. Eichler and
Levinson (1999) have suggested both high energy neutrino signals and scattered
photons [i.e. scattered off material moving at Lorentz factor less
than the intrinsic opening angle of the primary emission, see also
Eichler and Manis (2007)], each of which could corroborate LIGO
events, at large viewing offsets. Levinson et al (2002) have
suggested orphan afterglows, though there might be some problem
establishing uniqueness via coincidence because of their long
timescales.
As it
happens, evidence for a high degree of collimation is more
convincing for the long GRBs, while SHBs are the ones now believed to be
associated with mergers. SHBs frequently show a lower $E_{iso}$, a somewhat broader
$V/V_{max}$ distribution, and less evidence of a narrow opening
angle from jet breaks. This could be understood, for example, if the giant envelope in the case of long bursts provides better collimation than when it is absent.
The presence or absence of the envelope may be responsible for other differences between short and long GRBs. For example, it may be that there is intrinsic spread in the timescales of the central engines accretion timescale, and that only long bursts are sustained enough to break through a massive envelope, whereas mergers, perhaps for different reasons, also produce a spread in timescales while allowing all of them to be observed, though this would explain neither spectral differences nor differences in spectral lags and sub-pulse time scales between short and long GRBs. Neither would it by itself explain why the short duration, hard emission of short GRB lie off to one side of the Amati relation while long bursts, X-ray flashes and the X-ray tails of short GRB obey it.
Eichler and Manis (2007, 2008) and Eichler,
Guetta and Manis (EGM, 2009) noted that the unusually hard spectrum
displayed by SHBs, their unusually soft X-ray tail (as compared to the emission of long GRBs), and
their short duration were consistent with a smaller Lorentz factor at the time the short, hard emission is last scattered,
and a larger viewing angle. The larger viewing angle is, {\it a priori},
statistically expected if the line of sight is not obscured by an
extended stellar envelope which is known to exist in the case of
long GRBs, and which would be less likely in the case of NS-NS
mergers. Less collimation and larger allowed viewing offset angle make a coincidence with a
GW signal more likely. While larger viewing
angle and/or less collimation means smaller $E_{iso}$ and therefore
less $V_{max}$, that is not
a problem for LIGO collaboration, where the sources
would in any case be very close.
\footnote{The suggestion of Eichler and Manis (2007, 2008) and EGM (2008) that viewing angle affects the perceived durations both of the SHBs hard emission and X-ray tail is compatible with an additional intrinsic spread in durations of central engine activity for mergers (van Putten 1999b).
The long duration of GRB060614 can be attributed to the prolonged activity of a rotating black hole
(Van Putten, 2008). The hypothesis of EGM can also accommodates events such as GRB 060614, which was of long duration while resembling SHBs in other respects. Also, an X-ray tail that lasts $10^2$ s in observer time can result from a SHB whose intrinsic duration is only 1 s. In this paper, however, we are concerned only with the angular spread of the X-ray tail, not the intrinsic duration of the central engine activity that causes it, and consider the possibility that the X-ray tails of SHBs may have broader angular spreads and encompass more observers than the short, hard emission.}
Admittedly, the typical viewing angle for SHBs, though perhaps larger than for long GRBs, is uncertain and could
be small compared to unity.
There exists by now some evidence that SHBs are beamed, like long GRBs, into a
modest solid angle. Fox et al. (2005)
interpreted the steepening of the optical afterglow light curve of
GRB 050709 and GRB 050724 in terms of a jet break that translates into
a beaming factor $f_b^{-1}\sim 50$ (with $f_b$ the fraction of the
$4\pi$ solid angle within which the GRB is emitted). Soderberg et al
(2006) found a beaming factor of $\sim 130$ for GRB 051221A.
Therefore with the present data the beaming angles of SHB seem to lie
in the range of $\sim 0.1-0.2$ radian.
The discovery that X-ray flashes (XRFs) are a class of long GRBs was made by the Wide
Field Camera (WFC) on BeppoSax (Heise et al. 2001). The XRFs are GRBs characterized by no or faint signal in the
$\gamma$-ray energy range, are isotropically distributed in the sky
and have an average duration of $\sim 100$ sec like long GRBs. There
is strong evidence that classical GRBs and XRFs are closely related
phenomena, and understanding what makes them differ could yield
important insights into their origin. The redshift distribution of
XRFs is very similar to normal GRBs and therefore the high redshift
hypothesis, which might otherwise justify the softness of the burst, cannot
account for all XRFs. D'Alessio et al. (2006) have concluded that the
off-axis hypothesis seems to be the best hypothesis for now.
Many SHBs show a bright X-ray tail (XRT) that follows the short
prompt $\gamma$-ray emission and lasts for $\sim 100$ s (e.g. Nakar
2007). This X-ray component is evident in SHBs 050709 and 050724,
where the X-ray energy is comparable to or even larger than the energy
in the prompt $\gamma$-rays. It seems that XRTs, though not detected in all SHBs, are rather common among them. Extrapolation of the late afterglow back to
early times suggests that these tails cannot be interpreted as the
onset of the afterglow emission (Nakar 2007). These X-ray tails have
spectra and durations that are similar to those of the know XRFs,
and maybe both can be attributed to an off-axis line of sight. In this case, they could encompass more observers than the hard emission of the SHB, and could thus be more opportune for corroborating non-electromagnetic manifestations of mergers and/or core collapses. EGM (2009) made rough estimates of order 0.1 to 0.2 radian offset
from the periphery of the primary fireball, but with large uncertainties.
In this paper we consider the possibility that a wider opening angle of X-ray tails, relative to the hard SHB emission, enhances their likelihood of corroborating non-electromagnetic signals from merger and collapse events.
In section 2 we report all the properties of the X-ray tails. In section 3
we present a method to determine the XRFs rate. In section 4 we
compare this rate with the XRT rate and discuss our results.
\section{X-ray tail properties}
We have considered all
the short bursts detected by Swift from its launch (November 2004)
until August 2009, this
constitutes a sample of $\sim$ 40 bursts. In Table 1 we report the
observed data.
We report the properties of SHB prompt and X-ray tail emission as
detected by the Swift X-ray telescope and HETE-2. The X-ray flux is
estimated at 60-100 sec after the burst and is given in the 0.3-10
keV energy range. In the last column we also report what the X-ray
flux would be if the SHBs were at a distance of LIGO (advanced
version) detectability (300 Mpc if SHBs come from NS-NS mergers)
\begin{table}[t]
\caption{Properties of SHBs prompt and afterglow emission as
detected by Swift X-ray telescope and HETE-2 (indicated with the *),
the + indicates that
they could be detected by the WFC. The X-ray flux is estimated at
60-100 sec after the burst and is given in the 0.3-10 keV energy
range. In the last column we report what the X-ray flux would be if
the SHBs were at a distance of LIGO (advanced version) detectability
(300 Mpc if SHBs come from NS-NS mergers). }
\begin{tabular}{lllllll}
\hline
GRB & $z$ & $S_{\gamma}$ & $E_{\gamma, iso}$ & $F_x $& $ E_{x,iso}$
& $F_x $ (@ 300 Mpc) \\
& & $10^{-7}$ erg/cm$^2$ & $10^{49}$ erg & $10^{-11}$ erg/cm$^{2}/s$& $10^{49}$ erg
& $10^{-11}$ erg/cm$^{2}/s$\\
\hline
050709*$^+$&0.16 & 3$\pm$ 0.38 & 1.4 & 800 & 3.9 & 3.9 $\times 10^3$ \\
050724$^+$ & $0.258$ & 6.3 $\pm$ 1 & 7.2 & 1200 & 14.6 &1.4 $\times 10^4$\\
051210$^+ $& & 1.9$\pm$ 0.3 & & 90 & & \\
051221$^+$ & 0.546 & 22.2 $\pm$ 0.8 & 84 & 20 & 0.9 & 923\\
060313 $^+$& & 32.1 $\pm$ 1.4 & & 30& &\\
071227$^+$ & 0.383 & 2.2 $\pm$ 0.3 & 4.0 & 46& 1.1 & 1.1 $\times 10^3$\\
050509B & 0.225& 0.23$\pm$ 0.09& 0.2 & 0.06 & 5.6$\times 10^{-4}$ &0. 55 \\
050813 & 0.7& 1.24 $\pm$ 0.46 & 5.2 & 0.6& 0.042 & 42\\
050906 & & 0.84 $\pm$ 0.46& & $<0.007$ & &\\
050925 & & 0.92 $\pm$ 0.18 & & $ <0.003$ & & \\
060502B& 0.287 & 1 $\pm$ 0.13 & 1.15 & 0.1 & 0.001 & 1.4\\
060801 & & 0.8 $\pm 0.1$& & 0.1 & &\\
061201 & 0.11 & 3.3 $\pm$ 0.3 & 0.7 & 10 & 0.02& 23\\
061217 & 0.827 & 0.46 $\pm$ 0.08 & 2.4 & 0.1 & 0.005 &9.1\\
070429B & 0.904 & 0.63 $\pm 0.1$ &3.5 & 0.11& 0.006 & 10.4\\
070724 & 0.45 & 0.3 $\pm 0.2$ & 0.6 & 0.05 & 0.0012 &1.7\\
070729 & 0.904 & 1.0 $\pm$ 0.2& 5.6 & 0.024 & 0.001& 2.5\\
070809 & & 1.0 & & 0.179 & & \\
071112B & & 0.48 & &$<0.02$& &\\
080426 & & 3.7 $\pm$ 0.3 & & 0.91 & & \\
080702A & & 0.36 $\pm$ 0.1 & & 1.0 & & \\
080905A & &1.4 $\pm$ 0.2 & & 31 & & \\
081226A& & 0.99 $\pm$ 0.18 & & 0.047 & & \\
090305A & & 0.75 $\pm$ 0.13 & & 0.55 & & \\
090426 &2.6 & 1.8 $\pm$ 0.3 &71 & 1.2 & 0.48 & 470 \\
090621B & & 0.7 $\pm$ 0.1 & & 0.045 & & \\
\hline
\label{t:fit}
\end{tabular}
\end{table}
In order to give an estimate of the GW events expected from XRFs, XRTs
and SHBs, their cosmic event rates per unit volume and their beaming
factors must be known.
Measuring the relative detection rates and
distribution of distances of each of the categories of events
reduces the number of parameters. The universal central engine
hypothesis discussed in EGM (2009), in its
simplest and most naive form, together with the offset viewing
hypothesis for XRFs posit that SHBs and long GRBs have the same
energetics and that they present XRTs to offset observers the same
way long GRBs display XRFs to such observers.
\section{The rate per unit volume of XRFs in the Local Universe}
\label{rate}
A burst is classified as an XRF when the
Softness Ratio (SR) between the fluences
in the 2-30 KeV band to the 30-400 KeV band is
greater than unity (Lamb \& Graziani 2003).
In this section we give a method to estimate the rate of XRFs (by
which we mean rate per unit volume) following Pelangeon et al. 2008.
This method is valid both for the WFC that works in the 2-28 keV
energy range and for WXM on HETE-2 that works in a similar energy
range. Moreover, the sensitivity of the WFC $\sim 4\times 10^{-9}
{\rm erg cm}^{-2}{\rm s}^{-1}$ (De Pasquale et al. 2006) is
comparable to that of WXM $\sim 9\times 10^{-9} {\rm erg
cm}^{-2}{\rm s}^{-1} $ (Ricker et al. 2002).
In order to estimate the rate of XRFs we need to know the redshift of the sources, therefore we select only the XRFs that have determined redshift. These are only the XRFs detected by the WXM as no XRF of
the WFC has a known redshift. Our sample contain 6 long bursts and we report the relevant information about these bursts in table 2.
\begin{table}[t]
\caption{Properties of the X-ray flashes detected by WXM on HETE-2. S is the fluence in 2-25 kev}
\begin{tabular}{llll}
\hline
XRF& $T_{90}$ & $S$ & $z$ \\
& sec & $10^{-7}$ erg/cm$^2$ & \\
\hline
011130 & 39.5 & 5.9 & 0.5 \\
020317 & 7.14 & 2.2 & 2.11 \\
020903 & 10 & 0.8 & 0.25 \\
030429 & 12.95 & 4.7 & 2.68 \\
030528 & 62.8 & 62 & 0.782 \\
040701 & 11.67& 5.44 & 0.214 \\
\hline
\label{t:fit}
\end{tabular}
\end{table}
The steps to compute the rate are:
\begin{itemize}
\item
Determine the maximum redshift $z_{max}$ at which the source, an XRF, could have been
detected by the instruments, by first comparing the
measured flux with the threshold flux of the instrument for an XRF with known distance (z):
$F_x/F_T=(D_{\rm max}(z_{\rm max})/D(z))^2$
\item assume that the GRB rate follows the star formation rate.
For this we have adopted the model SFR$_2$ of Porciani \& Madau (2001)
that reproduces a fast evolution between $z=0$ and $z=1$
and remains constant beyond $z\ge2$
\begin{equation}
R_\mathrm{SFR_2}(z)\propto\,0.15 h_{65} \frac{exp(3.4z)}{exp(3.4z)+22}
\end{equation}
\item derive for each burst the number of GRBs per year within its visibility volume
\begin{equation}
\label{NV}
N_{\rm Vmax}\propto\,\int_0^{z_{\rm max}} dz\,\frac{dV(z)}{dz}\,\frac{R_{SFR_2}(z)}{1+z}
\end{equation}
In this equation $dV(z)/dz$ is the comoving element volume, described by
\begin{equation}
\frac{dV(z)}{dz}=\frac{c}{H_0}\,\frac{4\pi\,dl^2(z)}{(1+z)^2\,[\Omega_{M}(1+z)^3+\Omega_{K}(1+z)^2+\Omega_{\Lambda}]^{1/2}}
\end{equation}
where $H_0$ is the Hubble constant, $\Omega_{K}$ is the curvature contribution
to the present density parameter ($\Omega_{K}=$~1$-\Omega_{M}-\Omega_{\Lambda}$),
$\Omega_{M}$ is the matter density and $\Omega_{\Lambda}$ is the vacuum density.
Throughout this paper we have assumed a flat $\mathrm{\Lambda}$CDM universe where
($H_0$, $\Omega_{M}$, $\Omega_{\Lambda}$)$=$(65$h_{65}$~km~s$^{-1}$Mpc$^{-1}$, 0.3, 0.7).\\
This procedure allows us to give each burst a weight ($W_b$)
inversely proportional to $N_{Vmax}$. The rationale of weighting
each burst by $1/N_{Vmax}$ is the following: the visibility volume
is different for each XRF of our sample. Moreover, each XRF observed
is randomly taken from all the bursters present in its visibility
volume. In this way, rare bright XRFs, having a large visibility
volume, will have low weights, while faint local XRFs will have
higher weights. This procedure also takes into account the fact that
the XRF rate evolves with redshift, leading to the fact that XRFs are
about ten times more frequent at $z\sim1$ than at present.
\end{itemize}
This study also allows us to derive the rate of XRFs detected
and localized by HETE-2 ($R_0^\mathrm{H2}$).
For that purpose, we consider that each XRF in our sample contributes
to the local rate in proportion to:
\begin{equation}
h_b=\frac{N_{Vloc}(z=0.1)}{N_{Vmax}(z=z_{b,max})}
\end{equation}
with the $N_V(z)$ computed according to Eq.~\ref{NV}, and we obtain
the number per unit volume of HETE-2 GRBs during the mission,
\begin{equation}
\tau=\frac{1}{V_{loc}}\sum_{b=1}^{n_{burst}}h_b
\end{equation}
the value of which we calculated to be
~ 7.7 Gpc$^{-3}$, similar to what found by Pelangeon et al. 2008.\\
\\
In order to normalize this in terms of rate per year,
we took into account the effective monitoring time of the WXM, obtained from:
\begin{equation}
T_{m} = \frac{T_\mathrm{mission} \times T_\mathrm{\epsilon}}{4\pi} \times S_\mathrm{cov}
\end{equation}
where $T_m$ is the effective monitoring time of the WXM, $T_{mission}=$~69~months
is the duration of the mission, $T_{\epsilon}=$~37\% is the mean observation
efficiency during $T_{mission}$, and $S_{cov}=2\pi(1-cos~45\degr)=1.84$~sr
the sky-coverage of the WXM\footnote{Recall that throughout this study,
we only consider the XRF localized by the WXM.}.\\
The effective monitoring time of the WXM is hence $T_{m}=$~0.31~yr.
Using this, the rate of detectable XRFs in the Local Universe per
Gpc$^3$ and per year can be found to be
$\sim$25~Gpc$^{-3}$yr$^{-1}$. This result is a lower limit because
HETE-2 missed the bursts with intrinsic peak-energy $E_p$ lower than 1
or 2~keV as well as bursts
occuring at very high redshifts (Pelangeon et al. 2008).\\
Note that the main contribution to the rate comes from XRF 020903,
which has a maximum detectability redshift $z_{max}\sim 0.3$,
implying a small visibility volume and therefore a large
weight.
\subsection{Swift XRFs}
Because the Swift BAT instrument , which provide the trigger conditions, has an energy band
(15-150 kev) narrower than WXM+Fregate on HETE2, we have to find another way to define an XRF.
Sakamoto et al. (2008) define an XRF to be a burst with the ratio of the fluence in the 25-50 KeV band to that in the 50-100 KeV band greater than 1.32 and we use this definition to construct a sample
of XRF with redshift.
The relevant properties are given in table 3
\begin{table}[t]
\caption{Properties of the X-ray flashes detected by Swift, S is the fluence in 15-150 kev and F is the 1-sec Peak Photon Flux}
\begin{tabular}{lllll}
\hline
XRF& $T_{90}$ & S& $z$ & F\\
& sec & $10^{-7}$ erg/cm$^2$ & &ph/cm$^2$/s\\
\hline
050315 & 95.6 & 32.2 & 1.949 & 1.93 \\
050319 & 152 & 13.10 & 3.24 & 1.52 \\
050406 & 6.4 & 0.79 & 2.44 & 0.36 \\
050416 & 3.0 & 3.7 & 0.6535 & 4.88 \\
050824 & 26.6 & 2.7 & 0.83 & 0.5 \\
051016 & 4.0 & 1.7 & 0.9364 & 1.3 \\
060108 &14.3 & 3.69 & 2.08 & 0.77 \\
060218 & 2100 & 15.7 & 0.033& 0.25 \\
060512 & 9.7 & 2.3 & 0.44 & 0.88 \\
\hline
\label{t:fit}
\end{tabular}
\end{table}
In order to estimate the rate of XRFs from
the Swift data we can repeat the analysis described above taking the
threshold of Swift in 15-150 kev of $\sim $ 0.25 ph/cm$^2$/s (Band
2006). We then consider 4.5 years of activity with a sky coverage of
0.17 sr and find a rate of R$\sim$ 130 Gpc$^{-3}$ yr$^{-1}$. Note again
that the main contribution to the rate comes from XRF 060218, which
has a maximum redshift $z_{max}\sim 0.03$ implying a small
visibility volume and therefore a large weight. This rate
determined by XRF 060218 is similar to what found in Guetta \& Della
Valle (2007). The soft $\gamma$-rays of XRF 060218, which show a
spectral evolution similar to many other GRBs and subpulses therein,
may be photons scattered off relatively slow ambient material
(Mandal and Eichler, in preparation).
Comparing this rate and the rate found with the XRFs of the WXM
with the rate found by Guetta et al. (2005) and Guetta \& Piran (2007)
for long GRBs, $\sim$ 0.1-0.4 ~Gpc$^{-3}$yr$^{-1}$,
we infer that the population of $\gamma$-ray bursts
is dominated by the X-ray flashes at $z<0.1$. This is understandable, as XRFs
are soft but also faint in the observer frame, according to the
hardness-intensity relation derived by Barraud et al.
(2003). Therefore, if the rate of detected bursts is in
fact higher for classical GRBs than for XRFs, we can guess that this
is because the classical GRBs can be seen out to greater distances.
That XRFs have a much higher rate per unit volume than classical GRBs,
within the framework of our hypothesis that they are the same
phenomenon viewed from different angles, suggests that the
opening angle of XRF is significantly wider than for the GRB.
Significantly, the supernova-associated GRB and XRF event rate is
much larger not only than the classical GRB rate (Guetta \& Della
Valle 2007) but also larger than the estimated rate of NS merging.
The main contribution to the supernova-associated GRB event rate comes from GRB 980425 at z=0.0085. This burst was detected by the
{\it Beppo}SAX WFC (Pian et al. 1999) and by BATSE (Kippen 1998).
The peak flux in the 50-300 kev band was $F_{50-300}=4.48$ ph cm$^{-2}$s$^{-1}$.
Given the threshold of BATSE, $\sim 0.25 $ ph/cm$^2$/s we find that D$_{\rm max}$ =160 Mpc.
The {\it Beppo}SAX sky coverage was about 0.08 and the operation
time $\sim$ 4 years. Therefore the rate of 980425-like events is $R\sim 182$ Gpc$^{-3}$yr$^{-1}$
which is $\sim $ 10 times higher than the XRFs rate and more than 100 times higher than
the high luminous "classical" GRB rate.
This high rate suggests that if classical GRB emit GW, e.g. from an
unstable protocollapsar tori (van Putten 1999a,b) then combined signals
from wide angle electromagnetic emission and GW might be the most
common sort of electromagnetic plus non-electromagnetic multi-detections of mergers/collapses.
\section{Rate of XRTs and discussion}
The rate of XRFs is an upper limit to the rate of XRTs. For the lower limit
we can take the one of SHBs derived by Guetta and Stella (2009).
In this paper they find evidence in
favor of a bimodal origin of SHB progenitors where a fraction of SHBs
comes from the merging of primordial neutron star-neutron star (black hole)
and a fraction comes from the merging of dynamically formed
binaries in galaxy clusters.
In particular they find that the redshift distribution of SHBs
is best fitted when the incidences of primordial and dynamical
mergers among the SHB population are 40\% and 60\% respectively.
In this case the rate of SHB is
$R_0\sim 2.9$~Gpc$^{-3}$yr$^{-1}$.
For a fiducial value of $f_b^{-1}\sim 100$, we derive a beaming-corrected rate of
$\rho_0 =f_b^{-1} R_0\sim 290(f_b^{-1}/100)$.
Therefore the rate of XRTs is $2.9<R<130$ Gpc$^{-3} $yr$^{-1}$. This rate is left
uncorrected for the beaming as we don't know the beaming angle of this X-ray emission
which can be the same or larger than the beaming angle of the $\gamma$-ray emission.
Another way to estimate the rate of the XRTs is to consider the X-ray tails
detected by the Swift X-ray telescope that could be detected by the WFC
if they were at a distance of 300 Mpc. These are four
GRBs (GRB 050709, GRB 050724, GRB 051221, GRB 071227).
Considering the threshold of the detector that triggered them (Fregate for 050709
and BAT for the other three bursts) and
using the same procedure described above for the XRFs rate, we find a
rate of $\sim 7$ Gpc$^{-3}$yr$^{-1}$.
Our suggestion that some XRTs of SHBs are XRFs,
combined with
the hypothesis that they correspond to offset viewing of
a long burst in some
other direction (Eichler, Guetta and Manis, 2009) predicts that a
large enough sample of XRFs, even if unbiased by any $\gamma$-ray
trigger, should have a subset that correlates with SHBs. A careful
analysis, however, shows that BATSE should have detected less than
one SHB coincident with any XRF recorded by other contemporaneous
detectors. A larger sample of XRFs detected while a SHB detector is
operating would give tighter constraints.
In summary, we have considered the possibility that SHBs have
larger opening angles than long GRBs, and that the XRT associated
with the SHB has a wider angular extent than the harder emission.
Very rough estimates of the opening angle of XRTs, based on their
hypothesized similarity to XRFs, is an opening angle of 0.1 to 0.2
radians [EGM 2009], which may be somewhat larger than estimates of
the opening angles for the hard emission, which are typically 0.03
to 0.1 radians (Bloom et al. 2003).
While this does not constitute proof that the XRTs have
larger opening angles than the hard $\gamma$-ray emission from the SHBs,
the fact that extended soft emission is a reliable indicator for
SHBs (Donaghy al. 2006) suggests that the solid angle in which the
soft photons are detectable by HETE II is at least as large as that
from which the hard $\gamma$-beam is detectable. { On the other
hand, our estimate for the rate per unit volume of XRTs that would
be visible from the typical distance of an advanced LIGO source,
$\sim 300 Mpc$, is less than the estimated rate of neutron star
mergers, which leaves open the possibility that even the XRTs are
beamed and could not corroborate most LIGO signals. Further
information on the relative detectabilities of XRTs and the
corresponding short hard $\gamma$-ray emission could be obtained by a wide
field X-ray camera and $\gamma$-ray detectors working together.
In any case, the rate of WFC-detectable XRTs per sphere of radius
300$R_{300}$ Mpc is at least about 0.8$R_{300}^3$ per year, meaning
that a $2\pi$ detector would detect one per 2.5 years for
$R_{300}=1$. This is a non-negligible if modest fraction of the
total expected rate
of LIGO
signals from mergers, about 3 per year with advanced LIGO.
Including the other two WFC-detectable XRTs, though their redshift is unknown,
would raise the estimate
to about 1 $\times R_{300}^3$ per year. This suggests that some fraction of LIGO
signals, if not most or all, could be corroborated by wide field
X-ray cameras.
Coincidentally, this rate of 1 per several years is about the rate
of supernova-associated GRB within 300 Mpc, as estimated from the
prototypes GRBs 980425 and 060218, which could be looked for in
coincidence with UHE neutrinos. We also find the event rate per unit
volume for supernova-associated GRBs and XRFs to be about $10^2$
higher than for cosmologically distant GRBs. If gravitational waves
and/or neutrinos are emitted by such events, then nearby
SN-associated GRBs, corroborated by wide angle EM emission such as
XRFs or scattered $\gamma$-rays, may be the most frequent collapse
events observed simultaneously in both electromagnetic and
non-electromagnetic modes.
We acknowledge support
from the U.S.-Israel Binational Science Foundation, the Israel
Academy of Science, and the Robert and Joan Arnow Chair of
Theoretical Astrophysics, and in part from the National Science Foundation under grant number NSF PHY05-51164.
|
\section{Introduction}
\subsection{}
\label{subsec:Sn}
The symmetric group $S_n$ acts on $V=\mathbb C^n$ and then on the
symmetric algebra $S^*V$ naturally. It is well known that the
algebra of $S_n$-invariants on $S^*V$ is a polynomial algebra in
$n$ generators of degree $1,2, \ldots, n$. More generally,
consider the graded multiplicity of a given Specht module $S^\lambda$
for a partition $\lambda=(\lambda_1, \lambda_2, \ldots)$ of $n$ in the graded
algebra $S^*V$, which has a generating function $P_\lambda (t):
=\sum_{j\geq 0} m_\lambda (S^jV) t^j$. Kirillov \cite{Ki} has obtained
the following elegant formula for $P_\lambda (t)$
(also compare Steinberg~\cite{S}):
\begin{align*}
P_\lambda (t) =\frac{t^{n(\lambda)}} {\prod_{(i,j)\in \lambda}(1-t^{h_{ij}})},
\end{align*}
where $h_{ij}$ is the hook length associated to a cell $(i,j)$ in
the Young diagram of $\lambda$, and
\begin{eqnarray} \label{n lambda}
n(\lambda)=\sum_{i\geq 1}(i-1)\lambda_i.
\end{eqnarray}
The generating function for the bi-graded $S_n$-invariants in
$S^*V \otimes \wedge^* V$ was computed in Solomon \cite{So}, see
\eqref{Solomon}. More generally, Kirillov and Pak \cite{KP}
obtained the bi-graded multiplicity of the Specht module $S^\lambda$
for any $\lambda$ in $S^*V \otimes \wedge^* V$, see \eqref{eq:KPak}.
\subsection{}
According to Schur \cite{Sch}, the symmetric group $S_n$ affords a
double cover $\widetilde{S}_n$:
$$
1 \longrightarrow \mathbb Z_2 \longrightarrow \widetilde{S}_n
\longrightarrow S_n \longrightarrow 1.
$$
Let us write $\mathbb Z_2 =\{1,z\}$. The spin (or projective)
representation theory of $S_n$, or equivalently, the
representation theory of the spin group algebra $\mathbb C S_n^- =\mathbb C
\widetilde{S}_n/\langle z+1\rangle$, has been systematically
developed by Schur (see J\'ozefiak \cite{Jo1} for an excellent
modern exposition via a systematic use of superalgebras). Rich
algebraic combinatorics of Schur $Q$-functions and shifted
tableaux has been developed by Sagan \cite{Sa1} and Stembridge
\cite{St} (also see Nazarov \cite{Naz}) in relation to the
irreducible spin representations and characters of $S_n$.
\subsection{}
The goal of this paper is to formulate and prove the spin analogue
of the graded multiplicity formulas in \ref{subsec:Sn}.
The results of this paper, though looking classical, have not
appeared in literature to our knowledge; however, it is expected
that such simple results, once formulated, can be also derived by
other approaches. It strongly suggests that the spin invariant
theory of Weyl groups, or of finite groups in general, in the
sense of this paper is a very interesting research direction to
pursue. It is also natural to ask for the spin double counterpart,
a spin analogue of Kostka polynomials, an interpretation of the
graded multiplicity for the spin coinvariant algebra as generic
degrees for (quantum) Hecke-Clifford algebras, etc. We hope to
return to these topics at another occasion.
\subsection{}
It is known \cite{Jo2, Se2, St2, Ya} (see Kleshchev
\cite[Chap.~13]{Kle}) that the representation theory of spin
symmetric group (super)algebra $\mathbb{C}S_n^-$ is
super-equivalent to its counterpart for Hecke-Clifford
(super)algebra $\mathcal{H}_n := \mathcal{C}_n \rtimes \mathbb C S_n$; See
Section~\ref{subsec:super equiv} for notations and precise
formulations. (All the algebras and modules in this paper are
understood to admit a $\mathbb Z_2$-graded structure; however we will
avoid using the terminology of supermodules.) Let $D^\lambda_-$ denote
the simple $\mathbb{C}S_n^-$-module and $D^\lambda$ denote the simple
$\mathcal{H}_n$-module, associated to a strict partition $\lambda$ of $n$. The
Clifford algebra $\mathcal{C}_n$ is naturally a simple module over the
algebra $\mathcal{H}_n$ (which is identified with $D^{(n)}$), and it is the
counterpart of the basic spin $\mathbb{C}S_n^-$-module
$\mathcal{B}_n:=D^{(n)}_-$.
In Proposition~\ref{prop:mult.equiv} we show that,
for an arbitrary $S_n$-module $M$, the multiplicity problem for a simple
$\mathbb{C}S_n^-$-module $D^\lambda_-$ in $\mathcal B_n\otimes M$ is
essentially equivalent to the multiplicity problem for a simple
$\mathcal{H}_n$-module $D^\lambda$ in $\mathcal{C}_n\otimes M$.
Therefore, in this paper, we shall mainly work with the algebra
$\mathcal{H}_n$, keeping in mind that the results can be transferred to
the setting for $\C S_n^-$.
\subsection{}
Our first main result provides the graded multiplicity of the
simple $\mathcal{H}_n$-module $D^\lambda$ in $\mathcal{C}_n
\otimes S^* V$ for $V=\mathbb C^n$. For a partition $\lambda$ of $n$ with
length $\ell(\lambda)$, we set
\begin{eqnarray*}
\delta(\lambda)= \left \{
\begin{array}{ll}
0,
& \text{ if }\ell(\lambda) \text{ is even}, \\
1
, & \text{ if }\ell(\lambda) \text{ is odd}.
\end{array}
\right.
\end{eqnarray*}
If $\lambda$ is moreover a strict partition, we denote by $\lambda^*$ the
shifted diagram of $\lambda$, by $c_{ij}$ the content, and by
$h^*_{ij}$ the shifted hook length of the cell $(i,j) \in \lambda^*$
(see Section~\ref{sec:special} for precise definitions).
\begin{thma} \label{th:S(V)}
Let $\lambda$ be a strict partition of $n$. The graded multiplicity of
$D^{\lambda}$ in the $\mathcal{H}_n$-module $\mathcal{C}_n\otimes
S^* V$ is
\begin{align} \label{hook}
2^{-\frac{\ell(\lambda)+\delta(\lambda)}{2}} \frac{t^{n(\lambda)}\prod_{(i,j)\in
\lambda^*}(1+t^{c_{ij}})} {\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.
\end{align}
\end{thma}
The lowest degree term in \eqref{hook} is
$2^{\frac{\ell(\lambda)-\delta(\lambda)}{2}} t^{n(\lambda)}$, thanks to the
contribution $2^{\ell(\lambda)}$ from the product over the main diagonal
of $\lambda^*$ in the numerator. Theorem~A can be reformulated in
terms of the graded multiplicity of a coinvariant algebra which is
isomorphic to a graded regular module of $\mathcal{H}_n$, see
Theorem~\ref{Cliff.tensor coin.}. In the spirit of a classical
theorem of Borel which identifies the coinvariant algebra of a
Weyl group with the cohomology ring of the corresponding flag
variety, the coinvariant algebra of $\mathcal{H}_n$ should be regarded as the
cohomology ring (which has yet to be developed) of the flag
variety for the queer Lie supergroup.
To prove Theorem~A, we first obtain an expression of the graded
multiplicity in terms of the principal specialization of the Schur
$Q$-function, $Q_\lambda (t^\bullet) :=Q_\lambda (1,t,t^2,\ldots)$, and
then apply the following formula.
\begin{thmb}\label{th:t-Schur Q}
For a strict partition $\lambda$ of $n$, we have
\begin{align}
Q_\lambda (t^\bullet) =\frac{t^{n(\lambda)}\prod_{(i,j)\in
\lambda^*}(1+t^{c_{ij}})} {\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.
\label{eqn:t-Schur Q}
\end{align}
\end{thmb}
It is well known (cf. \cite{Sa1, St}) that a Schur $Q$-function
can be written as a sum over the so-called marked shifted tableaux
of a given shape.
We establish in Theorem~\ref{thm:bijection} a bijection between marked shifted tableaux and
certain new combinatorial objects which we call {\em colored
shifted tableaux}. Theorem~B is an easy consequence of such a bijection.
We show in Proposition~\ref{equiv} that the
formula \eqref{eqn:t-Schur Q} is equivalent to another formula of
Rosengren \cite[Proposition 3.1]{R}, who derived it from a Schur
function identity of Kawanaka~\cite{Ka}.
\subsection{}
Another result of this paper is a formula for the bi-graded
multiplicity of the simple $\mathcal{H}_n$-module $D^\lambda$ in
$\mathcal{C}_n\otimes S^* V \otimes \wedge^* V$:
$$
\sum_{p=0}^\infty \sum_{q=0}^n t^p s^q m_\lambda(\mathcal{C}_n \otimes S^pV \otimes \wedge^qV).
$$
We shall adopt the following
short-hand notation for a specialization of Schur $Q$-function in
$2$-variables $s$ and $t$:
$$
Q_\lambda (t^\bullet; st^\bullet) :=Q_\lambda (1,t,t^2,\ldots;
s,st,st^2,\ldots).
$$
\begin{thmc}\label{th:Koszul}
Let $\lambda$ be a strict partition of $n$. The bi-graded multiplicity
of $D^{\lambda}$ in the $\mathcal{H}_n$-module
$\mathcal{C}_n\otimes S^{*}V\otimes \wedge^{*}V$ is
\begin{equation} \label{eq:bi mult}
2^{-\frac{l(\lambda)+\delta(\lambda)}{2}}
Q_{\lambda}(t^{\bullet};st^{\bullet}).
\end{equation}
\end{thmc}
Setting $s=0$, we recover
Theorem~A from Theorem C. On the other hand, setting $t=0$, we
obtain a graded multiplicity formula of $D^\lambda$ in $\mathcal{C}_n \otimes
\wedge^*V$, see Corollary~\ref{mult wedge}.
We may also consider a Koszul $\mathbb Z$-grading
which counts the standard generators of $S^*V$ as
degree $2$ and the standard generators of $\wedge^*V$
as degree $1$. It follows by Theorem~C that, for the Koszul
grading, the graded multiplicity in $\mathcal{C}_n\otimes
S^*V\otimes \wedge^*V$ is given by the
same formula \eqref{hook} above. It will be nice to
obtain a closed formula for $Q_\lambda
(t^\bullet; st^\bullet)$.
\subsection{}
The paper is organized as follows. In Section~\ref{sec:special},
we provide a bijective proof of Theorem~B. The graded multiplicities in $\mathcal{C}_n\otimes
S^*V$ are studied and Theorem~A is proved in
Section~\ref{sec:mult}. In Section~\ref{sec:bigraded}, we study
the bi-graded multiplicities in $\mathcal{C}_n\otimes
S^*V\otimes \wedge^*V$, and prove
Theorem~C.
{\bf Acknowledgments.} J.W. is supported by a semester dissertation fellowship
from the Department of Mathematics, University of Virginia.
W.W. is partially supported by NSF grant DMS-0800280.
\section{Principal specialization of Schur $Q$-functions}\label{sec:special}
In this section, we shall provide a bijective proof for Theorem~B,
after first recalling some basics on strict partitions and Schur
$Q$-functions (\cite{Sa1, St, Mac}).
\subsection{Strict partitions and shifted diagrams}
Let $n\in \mathbb Z_+$. We denote a composition
$\lambda=(\lambda_1,\lambda_2,\ldots)$ of $n$ by $\lambda\models
n$, and denote a partition $\lambda$ of $n$ by $\lambda \vdash n$. A
partition $\lambda$ will be identified with its Young diagram, that
is, $\lambda=\{(i,j)\in\mathbb{Z}^2~|~1\leq i\leq \ell(\lambda), 1\leq
j\leq \lambda_i\}$. To each cell $(i,j)\in \lambda$, we associate its content
$c_{ij}=j-i$ and
hook length $h_{ij}=\lambda_i+\lambda_j'-i-j+1$, where
$\lambda'=(\lambda_1',\lambda_2',\ldots)$ is the conjugate
partition of $\lambda$.
Suppose that the main diagonal of the Young diagram $\lambda$ contains
$r$ cells. Let $\alpha_i=\lambda_i-i$ be the number of cells in
the $i$th row of $\lambda$ strictly to the right of $(i,i)$, and
let $\beta_i=\lambda_i'-i$ be the number of cells in the $i$th
column of $\lambda$ strictly below $(i,i)$, for $1\leq i\leq r$.
We have $\alpha_1>\alpha_2>\cdots>\alpha_r\geq0$ and
$\beta_1>\beta_2>\cdots>\beta_r\geq0$. Then the Frobenius notation
for a partition is
$\lambda=(\alpha_1,\ldots,\alpha_r|\beta_1,\ldots,\beta_r)$. For
example, if $\lambda=(5,4,3,1)$, then $\alpha =(4,2,0),
\beta=(3,1,0)$ and hence $\lambda=(4,2,0|3,1,0)$ in Frobenius
notation.
Suppose that $\lambda$ is a strict partition of $n$, denoted by
$\lambda\vdash_s n$. Let $\lambda^*$ be the associated shifted
Young diagram, that is,
$$
\lambda^*=\{(i,j)~|~1\leq i\leq \ell(\lambda), i\leq j\leq\lambda_i+i-1
\}
$$
which is obtained from the ordinary Young diagram by shifting the
$k$th row to the right by $k-1$ squares, for each $k$.
Given $\lambda\vdash_s n$ with $\ell(\lambda)=\ell$, define its double
partition $\widetilde{\lambda}$ to be
$\widetilde{\lambda}=(\lambda_1,\ldots,\lambda_{\ell} |
\lambda_1-1,\lambda_2-1,\ldots,\lambda_{\ell}-1)$ in Frobenius
notation. Clearly, the shifted Young diagram $\lambda^*$ coincides
with the part of $\widetilde{\lambda}$ that lies above the main
diagonal. For each cell $(i,j)\in \lambda^*$, denote by $h^*_{ij}$
the associated hook length in the Young diagram
$\widetilde{\lambda}$, and set the content $c_{ij}=j-i$.
For example, let $\lambda= (4, 2, 1)$. The corresponding shifted
diagram and double diagram are
$$
\lambda^*=\young(\,\,\,\,,:\,\,,::\,)
\qquad \qquad
\widetilde{\lambda}=\young(\,\,\,\,\,,\,\,\,\,,\,\,\,\,,\,)
$$
The hook lengths and contents for each cell in $\lambda$
are respectively as follows:
$$
\young(6541,:32,::1)
\qquad \qquad \young(0123,:01,::0)
$$
\subsection{Schur $Q$-functions}
Let $\lambda$ be a strict partition with $\ell(\lambda)=\ell$. Suppose
$m\geq \ell$. The associated Schur $Q$-function
$Q_{\lambda}(z_1,z_2,\ldots,z_m)$ is defined by
\begin{align}
Q_{\lambda}(z_1,z_2,\ldots,z_m)=2^{\ell}\sum_{w\in
S_m/S_{m-\ell}}w\Big(z_1^{\lambda_1}\cdots z_{\ell}^{\lambda_{\ell}}
\prod_{1\leq i\leq \ell}\prod_{i<j\leq
m}\frac{z_i+z_j}{z_i-z_j}\Big),\label{SchurQ1}
\end{align}
where the symmetric group $S_m$ acts by permuting the variables
$z_1,\ldots,z_m$ and $S_{m-\ell}$ is the subgroup acting on
$z_{\ell+1},\ldots,z_{m}$. The definition of $Q_{\lambda}(z_1,z_2,\ldots,z_m)$
stabilizes as $m$ goes to infinity, and we write
$Q_{\lambda}(z) =Q_{\lambda}(z_1,z_2,\ldots)$, the symmetric functions
in infinitely many variables $z=(z_1,z_2,\ldots)$.
For $y=(y_1, y_2,\ldots)$, the following identity holds (see
\cite[III, \S 8]{Mac}):
\begin{align}
\prod_{i,j}\frac{1+y_iz_j}{1-y_iz_j} =\sum_{\lambda\text{:
strict}}2^{-\ell(\lambda)}Q_{\lambda}(y)Q_{\lambda}(z).
\label{Cauchy identity}
\end{align}
It will be convenient to introduce another family of symmetric
functions $q_{\nu}(z)$ for any composition
$\nu=(\nu_1,\nu_2,\ldots)$ as follows:
\begin{align}
q_0(z)&=1\notag,\\
q_r(z)&=Q_{(r)}(z), \quad \text{ for }r\geq 1,\notag\\
q_{\nu}(z)&=q_{\nu_1}(z)q_{\nu_2}(z)\cdots.\notag
\end{align}
The generating function for $q_r(z)$ is
\begin{align}
\sum_{r\geq 0}q_r(z)u^r=\prod_i\frac{1+z_iu}{1-z_iu}.\label{gen.fun.qr}
\end{align}
We will write $q_r =q_r(z)$, etc.,
whenever there is no need to specify the variables.
Let $\Gamma_\mathbb C$ be the $\mathbb C$-algebra generated by $q_r, r\geq 1$, that is,
\begin{align}
\Gamma_\mathbb C =\mathbb C [q_1,q_2,\ldots].\label{Gamma}
\end{align}
Then $Q_{\lambda}$ for strict partitions $\lambda$ form a basis of
$\Gamma_\mathbb C$.
\subsection{Marked shifted tableaux and Schur $Q$-functions}
Denote by $\mathbf{P}'$ the ordered alphabet
$\{1'<1<2'<2<3'<3\cdots\}$. The symbols $1',2',3',\ldots$ are said
to be marked, and we shall denote by $|a|$ the unmarked version of
any $a\in\mathbf{P}'$; that is, $|k'| =|k| =k$ for each $k \in
\mathbb N$. For a strict partition $\lambda$, a {\it marked shifted tableau}
$T$ of shape $\lambda$, or a {\it marked shifted $\lambda$-tableau}
$T$, is an assignment
$T:\lambda^*\rightarrow\mathbf{P}'$ satisfying:
\begin{itemize}
\item[(M1)] The letters are weakly increasing along each row and
column.
\item[(M2)] The letters $\{1,2,3,\ldots\}$ are strictly increasing
along each column.
\item[(M3)] The letters $\{1',2',3',\ldots\}$ are strictly
increasing along each row. \label{M3}
\end{itemize}
For a marked shifted tableau $T$ of shape $\lambda$, let $\alpha_k$
be the number of cells $(i,j)\in \lambda^*$ such that $|T(i,j)|=k$ for
$k\geq 1$. The sequence $(\alpha_1,\alpha_2,\alpha_3,\ldots)$ is
called the {\em weight} of $T$. The Schur $Q$-function associated
to $\lambda$ can be interpreted as (see \cite{Sa1, St, Mac})
\begin{align}
Q_{\lambda}(x)=\sum_{T}x^{T},\label{SchurQ}
\end{align}
where the summation is over all marked shifted $\lambda$-tableaux, and
$x^T=x_1^{\alpha_1}x_2^{\alpha_2}x_3^{\alpha_3}\cdots$ if $T$ has
weight $(\alpha_1,\alpha_2,\alpha_3,\ldots)$.
Denote by $|T|=\sum_{k\geq1}k\alpha_k$ if the weight of $T$ is $(\alpha_1,\alpha_2,\ldots)$.
\begin{example} \label{markedTableau0}
Suppose $\lambda=(5,4,2)$.
The following is an example of a marked shifted tableau of shape
$\lambda$ and its weight is $(2,5,4)$:
\begin{align}
T=\begin{tabular}{cccccc}
\cline{1-5}
\multicolumn{1}{|c|}{$1'$}&\multicolumn{1}{|c|}{$1$}
&\multicolumn{1}{|c|}{$2'$}&\multicolumn{1}{|c|}{$2$}&\multicolumn{1}{|c|}{$2$}
\\
\cline{1-5}
&\multicolumn{1}{|c|}{$2'$}&\multicolumn{1}{|c|}{$2$}&
\multicolumn{1}{|c|}{$3'$}&\multicolumn{1}{|c|}{$3$}\\
\cline{2-5}
&&\multicolumn{1}{|c|}{$3'$}&\multicolumn{1}{|c|}{$3$}\\
\cline{3-4}
\end{tabular}\notag
\end{align}
\end{example}
A {\em shifted reverse plane tableau} $S$
of shape $\lambda$ is a labeling of cells in the shifted diagram
$\lambda^*$ with nonnegative integers so that the rows and columns
are weakly increasing. Denote by $|S|$ the summation of the
entries in $S$. It is known (cf. \cite[Theorem 6.2.1]{Sa2}) that
\begin{align}
\sum_St^{|S|}=\prod_{(i,j)\in
\lambda}\frac{1}{1-t^{h^*_{ij}}},\label{Sagan}
\end{align}
summed over all shifted reverse plane
tableaux of shape $\lambda$.
\subsection{A bijection theorem}
Let $\lambda$ be a strict partition. A {\it colored shifted
tableau} $C$ is an assignment $C:\lambda^*\rightarrow\mathbf{P}'$ such
that the associated assignment
$\overline{C}:\lambda^*\rightarrow\mathbb{Z}_+$ defined by
\begin{align}
\overline{C}(i,j)&=\left\{
\begin{array}{ll}
|C(i,j)|-j, &\text{ if } C(i,j) \text{ is marked}, \\
|C(i,j)|-i, &\text{ if } C(i,j) \text{ is unmarked}
\end{array}
\right.\notag
\end{align}
is a shifted reverse plane tableau of shape $\lambda$. The weight of a colored
shifted tableau is defined in the same way as for marked shifted
tableaux. Denote by $|C|=\sum_{k\geq1}k\alpha_k$ if the weight of $C$ is $(\alpha_1,\alpha_2,\ldots)$.
\begin{thm}\label{thm:bijection}
Suppose that $\lambda$ is a strict partition of $n$ and
$\alpha=(\alpha_1,\alpha_2,\ldots)$ is a composition of $n$. Then
there exists a bijection between the set of marked shifted
$\lambda$-tableaux of weight $\alpha$ and the set of
colored shifted $\lambda$-tableaux of weight $\alpha$.
\end{thm}
\begin{proof}
Suppose that $T$ is a marked shifted tableau of shape $\lambda$ and
weight $\alpha=(\alpha_1,\alpha_2,\ldots)$.
Set $m=\max\{|T(i,j)|~|~(i,j)\in \lambda^*\}$. For each $1 \le k \le m$,
$\lambda^{k,*} =\{(i,j)\in\lambda^*~|~|T(i,j)|\leq k\}$ is a shifted diagram of a strict partition $\lambda^k$,
and $\lambda^{1} \subseteq \lambda^{2}\subseteq\cdots\subseteq \lambda^{m}$.
We shall construct by induction on $k$ a chain of colored shifted tableaux $T^k$ of shape
$\lambda^{k}$ and weight $\alpha^k=(\alpha_1,\ldots,\alpha_k)$, for $1\leq k\leq m$.
Set
$T^1: \lambda^{1,*}\rightarrow \mathbf{P}'$ to be the restriction of $T$ to $\lambda^{1,*}$.
Since $T$ is a marked shifted tableau, $\lambda^{1}$ is a one-row partition
and hence $T^1$ is already a colored shifted tableau of weight $\alpha^1=(\alpha_1)$.
Suppose that $T^{k-1}$ is a colored shifted tableau of shape
$\lambda^{k-1}$ and weight $\alpha^{k-1}=(\alpha_1,\ldots,\alpha_{k-1})$.
In order to construct $T^k$ from $T^{k-1}$,
we start with an intermediate tableau $T_k$ defined by
\begin{align*}
T_k: \lambda^{k,*} &\longrightarrow \mathbf{P}'\notag\\
(i,j)&\mapsto\left\{\begin{array}{ll}
T^{k-1}(i,j), & \text{ if }(i,j)\in \lambda^{(k-1),*} \\
T(i,j),& \text{ if }(i,j)\in \lambda^{k,*} / \lambda^{(k-1),*}.
\end{array}\right.
\end{align*}
There is at most one cell labeled by
$k'$ in each row of $T_{k}$ since $T$ satisfies~(M3).
Suppose that the cells labeled by $k'$ in
$T_{k}$ are $(i_1,j_1),(i_2,j_2),\ldots,(i_p,j_p)$ with
$i_1<i_2<\cdots <i_p$. Start with the topmost cell $(i_1,j_1)$ and label
its left and upper neighboring cells $(i_1,j_1-1)$ and
$(i_1-1,j_1)$, if they exists, as
\begin{center}
\begin{tabular}{cc}
\cline{2-2}&\multicolumn{1}{|c|}{$c$}\\
\cline{1-2}
\multicolumn{1}{|c|}{$b$}&\multicolumn{1}{|c|}{$k'$}\\
\cline{1-2}
\end{tabular}
\end{center}
(In case when either the left or the upper neighboring cell is missing, the
exchange procedure below is simplified in an obvious manner).
Set
\begin{align*}
\bar{b}=\left\{
\begin{array}{ll}
|b|-(j_1-1), &\text{ if } b \text{ is marked}, \\
|b|-i_1, &\text{ otherwise;}
\end{array}
\right.
\end{align*}
\begin{align*}
\bar{c}=\left\{
\begin{array}{ll}
|c|-j_1, &\text{ if }c \text{ is marked}, \\
|c|-(i_1-1), &\text{ otherwise.}
\end{array}
\right.
\end{align*}
If $k-j_1<\bar{b}$ or
$k-j_1<\bar{c}$, exchange $k'$ and $b$ or $c$ in $T_k$ as follows:
\begin{align}
\text{Case I}\; (\bar{b}\geq\bar{c}):\qquad & \begin{tabular}{ccc}
\cline{2-3}&\multicolumn{2}{|c|}{$c$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$b$}&\multicolumn{2}{|c|}{$k'$}\\
\cline{1-3}
\end{tabular}
\longrightarrow
\begin{tabular}{ccc}
\cline{3-3}&&\multicolumn{1}{|c|}{$c$}\\
\cline{1-3}
\multicolumn{2}{|c|}{$k'$}&\multicolumn{1}{|c|}{$b$}\\
\cline{1-3}
\end{tabular} \notag\\
\text{ Case II }\;
(\bar{c}>\bar{b}):\qquad &\begin{tabular}{ccc}
\cline{2-3}&\multicolumn{2}{|c|}{$c$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$b$}&\multicolumn{2}{|c|}{$k'$}\\
\cline{1-3}
\end{tabular}\longrightarrow\begin{tabular}{ccc}
\cline{2-3}&\multicolumn{2}{|c|}{$k'$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$b$}&\multicolumn{2}{|c|}{$c$}\\
\cline{1-3}
\end{tabular} \notag
\end{align}
Note that $b$ is unmarked in Case~I and
$c$ is unmarked in Case~II.
Hence the resulting diagram
\begin{align*}
\text{I: } \quad \begin{tabular}{ccc}
\cline{2-3}&\multicolumn{2}{|c|}{$c$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$k'$}&\multicolumn{2}{|c|}{$b$}\\
\cline{1-3}
\end{tabular}
\qquad \text{ or } \qquad
\text{II: } \quad \begin{tabular}{ccc}
\cline{2-3}&\multicolumn{2}{|c|}{$k'$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$b$}&\multicolumn{2}{|c|}{$c$}\\
\cline{1-3}
\end{tabular}
\end{align*}
satisfies the requirement for colored shifted tableaux.
Keep repeating the above procedure for the new cells occupied with this $k'$, until it stops.
Then move on to apply
the same procedure above to the cells $(i_2,j_2), \ldots, (i_p,j_p)$
one by one, and denote by $T^k$ the resulting tableau in the end.
We claim that $T^k$ is a colored shifted tableau. By induction
hypothesis, $T^{k-1}$ is a colored shifted tableau. Clearly, the
exchange procedure above by definition ensures that the
requirement being a colored shifted tableau is already fulfilled
for the cells in $T^k$ other than those occupied by $k$. So it
remains to check the conditions on each cell $(i,j)\in \lambda^{k,*}$
with $T^k(i,j)=k$. Assume that the cell $(i,j-1)$ in $T^k$, if
it exists, is labeled by $d\in\mathbf{P}'$. Note that $j-1\geq i$
and $|d|\leq k$. If $d$ is unmarked, then $|d|-i\leq k-i$. If $d$
is marked, then $|d|-(j-1)\leq|d|-i\leq k-i$. Similarly, assume
that the cell $(i-1,j)$ in $T^k$, if it exists, is labeled by
$e\in\mathbf{P}'$. For unmarked $e$, we have $|e|<k$ or
equivalently $|e|-(i-1)\leq k-i$, since there is at most one
unmarked $k$ in each column of $T^k$. For marked $e$, we have
$|e|\leq k$ and hence $|e|-j \leq k-i$, since $j\geq i$. This
proves the claim.
Hence, we have constructed a colored shifted tableau $T^m$
of the same shape and weight as for $T$ which we started with.
We claim the exchange procedure above from $T$ to $T^m$ is reversible.
It suffices to show that the above procedure from $T^{k-1}$
to $T^k$ is invertible for each $k$. Denote by $T^{k,0}$ the resulting tableau
after removing cells labeled by unmarked $k$ from $T^k$.
There exists at most one cell labeled by marked $k'$ in each row of $T^{k,0}$,
and suppose that these cells are
$(i_1,j_1),(i_2,j_2),\ldots,(i_p,j_p)$ in $T^{k,0}$ with $i_1>i_2>\ldots>i_p$.
Start with the lowest cell $(i_1,j_1)$ and suppose that
its right and lower neighboring cells $(i_1,j_1+1)$ and $(i_1+1,j_1)$, if they exist, in $T^{k,0}$
are labeled by $b,c\in\mathbf{P}'$ as follows:
\begin{align}
\begin{tabular}{ccc}
\cline{1-3}\multicolumn{2}{|c|}{$k'$}&\multicolumn{1}{|c|}{$b$}\\
\cline{1-3}
\multicolumn{2}{|c|}{$c$}&\\
\cline{1-2}
\end{tabular}\notag
\end{align}
Set
\begin{align}
\tilde{b}=\left\{
\begin{array}{ll}
|b|-(j_1+1), &\text{ if } b \text{ is marked}, \\
|b|-i_1, &\text{ otherwise},
\end{array}\notag
\right.
\end{align}
\begin{align}
\tilde{c}=\left\{
\begin{array}{ll}
|c|-j_1, &\text{ if }c \text{ is marked}, \\
|c|-(i_1+1), &\text{ otherwise}.
\end{array}\notag
\right.
\end{align}
If $k'>b$ or $k'>c$, exchange $k'$ and $b$ or $c$ in $T^{k,0}$ as follows:
\begin{align}
\text{Case I} \; (\tilde{b}\leq \tilde{c}):\qquad &\begin{tabular}{ccc}
\cline{1-3}\multicolumn{2}{|c|}{$k'$}&\multicolumn{1}{|c|}{$b$}\\
\cline{1-3}
\multicolumn{2}{|c|}{$c$}&\\
\cline{1-2}
\end{tabular}
\longrightarrow
\begin{tabular}{ccc}
\cline{1-3}\multicolumn{1}{|c|}{$b$}&\multicolumn{2}{|c|}{$k'$}\\
\cline{1-3}
\multicolumn{1}{|c|}{$c$}&&\\
\cline{1-1}
\end{tabular}\notag\\
\text{Case II} \; (\tilde{b}>\tilde{c}):\qquad &\begin{tabular}{ccc}
\cline{1-3}\multicolumn{2}{|c|}{$k'$}&\multicolumn{1}{|c|}{$b$}\\
\cline{1-3}
\multicolumn{2}{|c|}{$c$}&\\
\cline{1-2}
\end{tabular}
\longrightarrow
\begin{tabular}{ccc}
\cline{1-3}\multicolumn{2}{|c|}{$c$}&\multicolumn{1}{|c|}{$b$}\\
\cline{1-3}
\multicolumn{2}{|c|}{$k'$}&\\
\cline{1-2}
\end{tabular}\notag
\end{align}
Keep repeating the above procedure to the new cell occupied by this $k'$,
until it stops. Then move on to the cells $(i_2,j_2), \ldots, (i_p,j_p)$
one by one and apply
the same procedure. Denote by $T^{k,1}$ the resulting tableau in the end.
Removing the cells labeled by $k'$ from $T^{k,1}$, we recover the tableau $T^{k-1}$.
\end{proof}
\begin{example} Suppose $\lambda=(5,4,2)$ and
$T$ is the marked shifted tableau given by Example~\ref{markedTableau0}.
Then the colored shifted tableau corresponding to $T$ is
$T^3$, where
\begin{align*}
T^1&=\begin{tabular}{cc}
\cline{1-2}
\multicolumn{1}{|c|}{$1'$}&\multicolumn{1}{|c|}{$1$}
\\
\cline{1-2}
\end{tabular}
\qquad\quad T^2=\begin{tabular}{ccccc}
\cline{1-5}
\multicolumn{1}{|c|}{$1'$}&\multicolumn{1}{|c|}{$2'$}
&\multicolumn{1}{|c|}{1}&\multicolumn{1}{|c|}{2}&\multicolumn{1}{|c|}{$2$}\\
\cline{1-5}
&\multicolumn{1}{|c|}{$2'$}&\multicolumn{1}{|c|}{$2$}\\
\cline{2-3}
\end{tabular}\notag\\ \\
& T^3 =\begin{tabular}{ccccc}
\cline{1-5}
\multicolumn{1}{|c|}{$1'$}&\multicolumn{1}{|c|}{$2'$}
&\multicolumn{1}{|c|}{$3'$}&\multicolumn{1}{|c|}{$1$}&\multicolumn{1}{|c|}{2}\\
\cline{1-5}
&\multicolumn{1}{|c|}{$2'$}&\multicolumn{1}{|c|}{$2$}
&\multicolumn{1}{|c|}{$2$}&\multicolumn{1}{|c|}{$3$}\\
\cline{2-5}
&&\multicolumn{1}{|c|}{$3'$}&\multicolumn{1}{|c|}{$3$}\\
\cline{3-4}
\end{tabular}\notag
\end{align*}
\end{example}
\subsection{Proof of Theorem~B}
\begin{proof}
It follows by Theorem~\ref{thm:bijection} that
\begin{align} \label{T=C}
\sum_T t^{|T|} =\sum_{C}t^{|C|},
\end{align}
where the first summation is over all marked shifted $\lambda$-tableaux $T$
and the second summation is over all colored shifted $\lambda$-tableaux $C$.
The left hand side of \eqref{T=C} is equal to $Q_{\lambda}(t,t^2,t^3, \ldots)$ by (\ref{SchurQ}).
It follows from the definition of colored shifted tableaux and then \eqref{Sagan} that
\begin{align*}
\sum_{C}t^{|C|}
&=\Big(\prod_{(i,j)\in
\lambda^*}(t^i+t^j)\Big)\sum_{S}t^{|S|} \\
&=\frac{\prod_{(i,j)\in \lambda^*}(t^i+t^{j})}
{\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})},
\end{align*}
where the summation on $S$ is taken over all shifted reverse plane tableaux of shape $\lambda$.
Putting everything together, we obtain that
\begin{align}
Q_{\lambda}(t^{\bullet}) &=\frac{1}{t^n}Q_{\lambda}(t,t^2,t^3,\ldots)
=\frac{1}{t^n}\sum_T t^{|T|} \\
&=
\frac{1}{t^n}\frac{\prod_{(i,j)\in \lambda^*}(t^i+t^{j})}
{\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}
=\frac{t^{n(\lambda)}\prod_{(i,j)\in \lambda^*}(1+t^{c_{ij}})}
{\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.\notag
\end{align}
This completes the proof of Theorem~B.
\end{proof}
\begin{rem}
It follows from the proof above that Theorem B can be restated as
\begin{align*}
Q_\lambda (t^\bullet)
=\frac{\prod_{(i,j)\in\lambda^*}(t^{i-1} +t^{j-1})} {\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.
\end{align*}
\end{rem}
\subsection{Another formula for $Q_{\lambda}(t^{\bullet})$}
For $k\in\mathbb N$, we set
$$
(a;t)_k=(1-a)(1-at)\cdots(1-at^{k-1}).
$$
Rosengren \cite[Proposition~3.1]{R} has obtained the following
formula for $Q_{\lambda}(t^{\bullet})$, starting from a Schur
function identity of Kawanaka:
\begin{align}
Q_{\lambda}(t^{\bullet})
=\prod_{1\leq i\leq \ell(\lambda)}\frac{(-1;t)_{\lambda_i}}{(t;t)_{\lambda_i}}
\prod_{1\leq i<j\leq \ell(\lambda)}
\frac{t^{\lambda_j}-t^{\lambda_i}}{1-t^{\lambda_i+\lambda_j}}.
\label{eqn(R):t-Schur Q}
\end{align}
\begin{prop}\label{equiv}
The formula~(\ref{eqn:t-Schur Q}) is equivalent to~Rosengren's
formula (\ref{eqn(R):t-Schur Q}).
\end{prop}
\begin{proof}
Set $\ell =\ell(\lambda)$. It is known (cf. \cite[III, \S 8, Ex.~12]{Mac}) that in
the $i$th row of $\lambda^*$, the hook lengths $h^*_{ij}$ for $i\leq
j\leq\lambda_i+i-1$ are
$1,2,\ldots,\lambda_i,\lambda_i+\lambda_{i+1},\lambda_i+\lambda_{i+2},\ldots,\lambda_{i}+\lambda_{\ell}$
with exception
$\lambda_i-\lambda_{i+1},\lambda_i-\lambda_{i+2},\ldots,\lambda_{i}-\lambda_{\ell}$. Hence we
have
\begin{align}
\prod_{(i,j)\in\lambda^*}\frac{1}{1-t^{h^*_{ij}}}
=\frac{1}{\prod_{1\leq i\leq \ell}(t;t)_{\lambda_i}}\prod_{1\leq i<j\leq
\ell} \frac{1-t^{\lambda_i-\lambda_j}}{1-t^{\lambda_i+\lambda_j}}.
\label{shifted hook}
\end{align}
The equivalence between~(\ref{eqn:t-Schur Q})
and~(\ref{eqn(R):t-Schur Q}) can now be deduced by applying
\eqref{shifted hook} and noting that the contents $c_{ij}$ for
$i\le j \le \lambda_i +i-1$ are $0,1,\ldots, \lambda_i-1$.
\end{proof}
\begin{rem}
By~(\ref{SchurQ}), $Q_{\lambda}(1^m)$ is
equal to the number of marked shifted Young tableaux of shape
$\lambda$ with fillings by letters $\le m$. On the other hand, it
follows from~\cite[Theorem~4]{Se1} that
$2^{\frac{\delta(\lambda)-\ell(\lambda)}{2}}Q_{\lambda}(1^m)$ gives the dimension
of the irreducible representation of the queer Lie superalgebra
$\mathfrak{q}(m)$ of highest weight $\lambda$.
\end{rem}
\section{The graded multiplicity in $\mathcal{C}_n\otimes
S^{*}V$} \label{sec:mult}
The goal of this section is to establish Theorem~A. In addition, a
tensor identity in Lemma~\ref{lem:equiv.tensor} allows us to
translate a multiplicity problem for $\mathcal{H}_n$ to $\C S_n^-$, and vice versa
(see Proposition~\ref{prop:mult.equiv}).
\subsection{Some basics about superalgebras}
We shall recall some basic notions of superalgebras, referring the
reader to~\cite[Chapter 12]{Kle}. Let us denote by
$\bar{v}\in\mathbb{Z}_2$ the parity of a homogeneous vector $v$ of a
vector superspace. A superalgebra $\mathcal{A}$ is a
$\mathbb{Z}_2$-graded associative algebra.
An $\mathcal{A}$-module always means a $\mathbb{Z}_2$-graded
left $\mathcal{A}$-module in this paper. A homomorphism $f:V\rightarrow W$ of
$\mathcal{A}$-modules $V$ and $W$ means a linear map such that $
f(av)=(-1)^{\bar{f}\bar{a}}af(v).$ Note that this and other such
expressions only make sense for homogeneous $a, f$ and the meaning
for arbitrary elements is attained by extending linearly from
the homogeneous case. Let $V$ be a finite dimensional
$\mathcal{A}$-module. Let $\Pi
V$ be the same underlying vector space but with the opposite
$\mathbb{Z}_2$-grading. The new action of $a\in\mathcal{A}$ on $v\in\Pi
V$ is defined in terms of the old action by $a\cdot
v:=(-1)^{\bar{a}}av$.
Denote by $\mathcal{A}\text{-smod}$ the category of finite dimensional $\mathcal{A}$-modules.
Given two superalgebras $\mathcal{A}$ and $\mathcal{B}$,
the tensor product $\mathcal{A}\otimes\mathcal{B}$
is naturally a superalgebra.
Suppose that $V$ is an $\mathcal{A}$-module and $W$ is a
$\mathcal{B}$-module. Then the tensor space $V\otimes W$ affords an $\mathcal{A}\otimes \mathcal{B}$-module,
denoted by $V\boxtimes W$, via
$$
(a\otimes b)(v\otimes w)=(-1)^{\bar{b}\bar{v}}av\otimes bw,\quad a\in \mathcal{A},
b\in \mathcal{B}, v\in V, w\in W.
$$
\iffalse
Given two superalgebras $\mathcal{A}$ and $\mathcal{B}$, we view
the tensor product of superspaces $\mathcal{A}\otimes\mathcal{B}$
as a superalgebra with multiplication defined by
$$
(a\otimes b)(a'\otimes b')=(-1)^{\bar{b}\bar{a'}}(aa')\otimes (bb')
\qquad (a,a'\in\mathcal{A}, b,b'\in\mathcal{B}).
$$
Suppose $V$ is an $\mathcal{A}$-module and $W$ is a
$\mathcal{B}$-module. Then $V\otimes W$ affords $A\otimes B$-module
denoted by $V\boxtimes W$ via
$$
(a\otimes b)(v\otimes w)=(-1)^{\bar{b}\bar{v}}av\otimes bw,\quad a\in A,
b\in B, v\in V, w\in W.
$$
If $V$ is an irreducible $\mathcal{A}$-module and $W$ is an
irreducible $\mathcal{B}$-module, $V\boxtimes W$ may not be
irreducible. Indeed, we have the following standard lemma (cf.
\cite[Lemma 12.2.13]{Kle}).
\begin{lem}\label{lem:tensor smod}
Let $V$ be an irreducible $\mathcal{A}$-module and $W$ be an
irreducible $\mathcal{B}$-module.
\begin{enumerate}
\item If both $V$ and $W$ are of type $\texttt{M}$, then
$V\boxtimes W$ is an irreducible
$\mathcal{A}\otimes\mathcal{B}$-module of type $\texttt{M}$.
\item If one of $V$ or $W$ is of type $\texttt{M}$ and the other
is of type $\texttt{Q}$, then $V\boxtimes W$ is an irreducible
$\mathcal{A}\otimes\mathcal{B}$-module of type $\texttt{Q}$.
\item If both $V$ and $W$ are of type $\texttt{Q}$, then
$V\boxtimes W\cong X\oplus \Pi X$ for a type $\texttt{M}$
irreducible $\mathcal{A}\otimes\mathcal{B}$-module $X$.
\end{enumerate}
Moreover, all irreducible $\mathcal{A}\otimes\mathcal{B}$-modules
arise as constituents of $V\boxtimes W$ for some choice of
irreducibles $V,W$.
\end{lem}
If $V$ is an irreducible $\mathcal{A}$-module and $W$ is an
irreducible $\mathcal{B}$-module, denote by $V\circledast W$ an
irreducible component of $V\boxtimes W$. Thus,
$$
V\boxtimes W=\left\{
\begin{array}{ll}
V\circledast W\oplus \Pi (V\circledast W), & \text{ if both } V \text{ and } W
\text{ are of type }\texttt{Q}, \\
V\circledast W, &\text{ otherwise }.
\end{array}
\right.
$$
\fi
\subsection{Spin symmetric group algebras $\mathbb{C}S_n^-$
and Hecke-Clifford algebras $\mathcal{H}_n$}\label{subsec:super equiv}
Recall that the spin symmetric group algebra $\mathbb C S_n^-$ is the algebra generated by
$t_1,t_2,\ldots,t_{n-1}$ subject to the relations:
\begin{align}
t_i^2&=1,\quad 1\leq i\leq n-1\notag\\
t_it_{i+1}t_i&=t_{i+1}t_it_{i+1},\quad 1\leq i\leq n-2\notag\\
t_it_j&=-t_jt_i,\quad 1\leq i,j\leq n-1,|i-j|\geq 1.\notag
\end{align}
$\mathbb{C}S_n^-$ is a superalgebra with each $t_i$ being odd, for $1\leq i\leq n-1$.
Denote by $\mathcal{C}_n$
the Clifford superalgebra generated by the odd elements $c_1,\ldots,c_n$,
subject to the relations $c_i^2=1,c_ic_j=-c_jc_i$ for $1\leq i\neq
j\leq n$. Observe that $\mathcal{C}_n$ is a simple superalgebra
and there is a unique (up to isomorphism) irreducible
$\mathcal{C}_n$-module $U_n$.
Define the Hecke-Clifford
algebra $\mathcal{H}_n=\mathcal{C}_n\rtimes\mathbb C S_n$ to be the
superalgebra generated by odd elements $c_1,\ldots,c_n$ and even
elements $s_1,\ldots,s_{n-1}$, subject to the relations:
\begin{align*}
s_i^2&=1, s_is_j=s_js_i,\quad 1\leq i,j\leq n-1, |i-j|>1,\\
s_is_{i+1}s_i&=s_{i+1}s_is_{i+1},\quad 1\leq i\leq n-2,\\
c_i^2&=1,c_ic_j=-c_jc_i,\quad 1\leq i\neq
j\leq n,\\
s_ic_i&=c_{i+1}s_i, s_ic_j=c_js_i,\quad 1\leq i,j\leq n-1, j\neq i,i+1.
\end{align*}
There is a superalgebra isomorphism (cf.
\cite{Se1, Ya}):
\begin{align}
\label{map:isorm.HC}
\begin{split}
\mathbb C S_n^-\otimes\mathcal{C}_n&\longrightarrow\mathcal{H}_n \\
c_i&\mapsto c_i, \quad 1\leq i\leq n, \\
t_j&\mapsto \frac{1}{\sqrt{-2}}s_j(c_j-c_{j+1}),\quad 1\leq j\leq n-1.
\end{split}
\end{align}
The two exact functors
\begin{eqnarray*}
\mathfrak{F}_n :=-\boxtimes U_n:
& \mathbb C
S_n^-\text{-smod} \rightarrow\mathcal{H}_n\text{-smod},\notag\\
\mathfrak{G}_n :={\rm Hom}_{\mathcal{C}_n}(U_n,-):
& \mathcal{H}_n
\text{-smod} \rightarrow\mathbb C S_n^-\text{-smod} \notag
\end{eqnarray*}
define Morita super-equivalence between the superalgebras $\mathcal{H}_n$ and $\C S_n^-$
(cf. Kleshchev \cite[Proposition~13.2.2]{Kle} for precise details).
\iffalse
\begin{lem}\cite[Proposition~13.2.2]{Kle}\label{lem:two functors} The following holds.
\begin{enumerate}
\item Suppose that $n$ is even. Then the two functors
$\mathfrak{F}_n$ and $\mathfrak{G}_n$ are equivalences of
categories with
\begin{align*}
\mathfrak{F}_n\circ\mathfrak{G}_n\cong{\rm id},\quad
\mathfrak{G}_n\circ\mathfrak{F}_n\cong{\rm id}.
\end{align*}
\item Suppose that $n$ is odd. Then
\begin{align*}
\mathfrak{F}_n\circ\mathfrak{G}_n\cong{\rm id}\oplus\Pi,\quad
\mathfrak{G}_n\circ\mathfrak{F}_n\cong{\rm id}\oplus\Pi.
\end{align*}
Moreover, $\mathfrak{F}_n$ induces a bijection between the
isoclasses of irreducible $\C S_n^-$-modules of type \texttt{M} and the
isoclasses of irreducible $\mathcal{H}_n$-modules of type \texttt{Q}.
Meanwhile $\mathfrak{G}_n$ induces a bijection between the
isoclasses of irreducible $\mathcal{H}_n$-modules of type \texttt{M} and
the isoclasses of irreducible $\C S_n^-$-modules of type \texttt{Q}.
\end{enumerate}
\end{lem}
\fi
It is known \cite{Jo2, Se1, St2} (cf. \cite{Kle}) that for each
strict partition $\lambda$ of $n$, there exists an irreducible
$\mathcal{H}_n$-module $D^{\lambda}$ and $\{D^{\lambda}~|~\lambda\vdash_s n\}$
forms a complete set of non-isomorphic irreducible
$\mathcal{H}_n$-modules.
We have a complete set of
non-isomorphic irreducible $\C S_n^-$-modules
$\{D^\lambda_-~|~\lambda\vdash_sn\}$,
and by \cite[Proposition~13.2.2]{Kle},
\begin{align}
\mathfrak{G}_n(D^{\lambda}) = \left\{\begin{array}{ll}
D^\lambda_-,
& \text{ if }n\text{ or } \ell(\lambda)\text{ is even},\\
D^\lambda_-\bigoplus\Pi D^\lambda_-,
& \text{ otherwise}.
\end{array}
\right. \label{corresp:spin HC}
\end{align}
Denote the trivial representation by ${\bf 1}$
and the sign representation of $S_n$ by ${\rm sgn}$. Note that
$\mathcal{C}_n\cong{\rm ind}^{\mathcal{H}_n}_{\mathbb{C}S_n}{\bf
1}$ is the irreducible $\mathcal{H}_n$-module $D^{(n)}$
\cite[Lemma~22.2.4]{Kle}. It follows from
~(\ref{corresp:spin HC})~
that the irreducible
$\mathbb{C}S_n^-$-module $\mathcal{B}_n:=D^{(n)}_-$ satisfies that
\begin{align}
{\rm Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n)\cong\left\{
\begin{array}{ll}
\mathcal{B}_n,\quad &\text{ if } n \text{ is even},\\
\mathcal{B}_n\bigoplus\Pi\mathcal{B}_n, \quad &\text{ if } n \text{ is odd}.
\end{array}
\right.\label{basic spin}
\end{align}
It can be shown that the $\mathbb{C}S_n^-$-module
$\mathcal{B}_n$ coincides with the basic spin representation $L_n$
defined in~\cite[2C]{Jo1}.
\iffalse
\begin{rem}
By~\cite[Lemma~22.2.4]{Kle}, $\mathcal{C}_n$ is the irreducible $\mathcal{H}_n$-module
corresponding to the partition $(n)$, that is, $\mathcal{C}_n\cong D^{(n)}$.
Meanwhile we have $\mathcal{B}_n\cong S^{(n)}_-$ as
$\C S_n^-$-modules. Actually, the $\mathbb{C}S_n^-$-module
$\mathcal{B}_n$ coincides with the basic spin representation $L_n$
defined in~\cite[2C]{Jo1}. Identify $\mathcal{C}_n\otimes\C S_n^-$ with $\mathcal{H}_n$
using the isomorphism $\mathcal{H}_n\cong\mathcal{C}_n\otimes\C S_n^-$. By
Lemma~\ref{lem:tensor smod}, $\mathcal{C}_n\cong U_n\circledast L$ for some
irreducible $\C S_n^-$-module $L$ and moreover ${\rm
res}^{\mathcal{H}_n}_{\C S_n^-}\mathcal{C}_n$ is isomorphic to a direct sum of copies of
$L$. On the other hand, by computing characters
{\bf one can check that} ${\rm
res}^{\mathcal{H}_n}_{\C S_n^-}\mathcal{C}_n\cong L_n^{\oplus\frac{n}{2}}$ if $n$ is even
and ${\rm res}^{\mathcal{H}_n}_{\C S_n^-}\mathcal{C}_n\cong L_n^{\oplus\frac{n+1}{2}}$ if
$n$ is odd. ({\bf Details about the calculation on characters will be added.})
This implies that $L\cong L_n$ and $\mathcal{C}_n\cong
U_n\circledast L=U_n\boxtimes L_n$ . Then we have as
$\C S_n^-$-modules:
\begin{align*}
{\rm Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n)
\cong{\rm Hom}_{\mathcal{C}_n}(U_n, U_n\boxtimes L_n)\\
\cong\left\{
\begin{array}{ll}
L_n,\quad &\text{ if } n \text{ is even},\\
L_n\oplus\Pi L_n, \quad &\text{ if } n \text{ is odd}.
\end{array}
\right.
\end{align*}
Therefore $\mathcal{B}_n\cong L_n$.
\end{rem}
\fi
\subsection{The multiplicity problem of $\mathcal{H}_n$ vs $\C S_n^-$}
Given a $\mathbb C S_n$-module $M$ and a $\mathbb C S_n^-$-module $E$,
the tensor product $E\otimes M$ affords a $\mathbb C S_n^-$-module as follows:
\begin{align}
t_j(u\otimes x)=(t_ju)\otimes (s_jx),
\quad 1\leq j\leq n-1, u\in E, x\in M.\label{eqn:spin.tensor sym}
\end{align}
Meanwhile, the tensor product $F\otimes M$ of a $\mathcal{H}_n$-module $F$
and a $\mathbb C S_n$-module $M$ naturally affords a $\mathcal{H}_n$-module
with
\begin{align}
c_i(u\otimes x)=(c_iu)\otimes x,\qquad s_j(u\otimes x)=(s_ju)\otimes (s_jx),
\label{eqn:HC.tensor sym}
\end{align}
for $1\leq i\leq n, 1\leq j\leq n-1, u\in F, x\in M.$
\begin{lem} [A tensor identity]
\label{lem:equiv.tensor}
Suppose that $M$ is a $\mathbb C S_n$-module.
Then we have an isomorphism of $\mathbb C S_n^-$-modules:
$\mathfrak G_n (\mathcal{C}_n) \otimes M \cong \mathfrak G_n (\mathcal{C}_n \otimes M)$; that is,
\begin{align}
{\rm Hom}_{\mathcal{C}_n}(U_n,\mathcal{C}_n)\otimes M\cong{\rm
Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n\otimes M) .\notag
\end{align}
\end{lem}
\begin{proof}
Observe that by~(\ref{map:isorm.HC}) the action of $\C S_n^-$ on
${\rm Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n\otimes M)$ is given
by
\begin{align}
(t_j*f)(u)=(\frac{1}{\sqrt{-2}}s_j(c_j-c_{j+1}))(f(u)), \quad
f\in{\rm Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n\otimes M), u\in
U_n\label{eqn:action1}
\end{align}
while by~(\ref{eqn:spin.tensor sym}) the $\C S_n^-$-module structure
of ${\rm Hom}_{\mathcal{C}_n}(U_n,\mathcal{C}_n)\otimes M$ is given by
\begin{align}
t_j*(f\otimes x)=(t_j*f)\otimes(s_j x),\quad f\in{\rm
Hom}_{\mathcal{C}_n}(U_n,\mathcal{C}_n), x\in
M.\label{eqn:action2}
\end{align}
Define a map
\begin{align}
\phi:{\rm Hom}_{\mathcal{C}_n}(U_n,\mathcal{C}_n)\otimes M&\longrightarrow
{\rm Hom}_{\mathcal{C}_n}(U_n, \mathcal{C}_n\otimes M),\notag\\
f\otimes x&\mapsto (u\mapsto f(u)\otimes x).\notag
\end{align}
Clearly $\phi$ is injective and thus an isomorphism of vector spaces by a dimension
counting argument. It remains to show that $\phi$ is a
$\C S_n^-$-module homomorphism. Indeed, for $u\in U_n$, $f\in{\rm
Hom}_{\mathcal{C}_n}(U_n,\mathcal{C}_n)$ and $ x\in M$, we have
\begin{align*}
\phi(t_j*(f\otimes x))(u)
&=\phi(t_j*f\otimes s_jx)(u) \quad \text{ by}~~(\ref{eqn:action2})\notag\\
&=(t_j*f)(u)\otimes s_jx\notag\\
&=((\frac{1}{\sqrt{-2}}s_j(c_j-c_{j+1}))f(u))\otimes s_jx\notag\\
&=(\frac{1}{\sqrt{-2}}s_j(c_j-c_{j+1}))(f(u)\otimes x)
\quad \text{ by~}(\ref{eqn:HC.tensor sym})\notag\\
&=(t_j*\phi(f\otimes x))(u) \quad \text{ by}~(\ref{eqn:action1}).
\end{align*}
\end{proof}
\begin{prop}\label{prop:mult.equiv}
Suppose that $M$ is a $\mathbb C S_n$-module. Let $m_{\lambda}$ and
$m_{\lambda}^-$ be the multiplicities of $D^{\lambda}$ and
$D^\lambda_-$ in the $\mathcal{H}_n$-module $\mathcal{C}_n\otimes M$ and
$\C S_n^-$-module $\mathcal{B}_n\otimes M$, respectively. Then,
\begin{align*}
m_{\lambda}^- =\left\{\begin{array}{ll}
m_{\lambda}, & \text{ if }n \text{ is even},\\
m_{\lambda}, & \text{ if }n \text{ is odd} \text{ and } \ell(\lambda) \text{ is odd},\\
\frac12 m_{\lambda}, & \text{ if }n \text{ is odd} \text{ and } \ell(\lambda)
\text{ is even}.
\end{array}
\right.
\end{align*}
\end{prop}
\begin{proof}
It follows by definition that
\begin{align}
\mathcal{C}_n\otimes M\cong\bigoplus_{\lambda\vdash_s n}m_{\lambda} D^{\lambda},
\qquad
\mathcal{B}_n\otimes M\cong\bigoplus_{\lambda\vdash_s n}m_{\lambda}^-
D^\lambda_-\label{eqn:Bl M}.
\end{align}
By~(\ref{basic spin}) and Lemma~\ref{lem:equiv.tensor}, we have
\begin{align*}
\mathfrak{G}_n(\mathcal{C}_n\otimes M)\cong\left\{
\begin{array}{ll}
\mathcal{B}_n\otimes M,\quad &\text{ if } n \text{ is even},\\
2\mathcal{B}_n\otimes M,\quad &\text{ if } n \text{ is odd}.
\end{array}
\right.
\end{align*}
This together with (\ref{eqn:Bl M}) implies
that
\begin{align*}
\bigoplus_{\lambda\vdash_s n}m_{\lambda} \mathfrak{G}_n(D^{\lambda})\cong
\left\{\begin{array}{ll}
\bigoplus_{\lambda\vdash_s n}m_{\lambda}^- D^\lambda_-, & \text{ if }n \text{ is even}, \\
\bigoplus_{\lambda\vdash_s n}2m_{\lambda}^- D^\lambda_-, & \text{ if }n
\text{ is odd}.
\end{array}
\right.
\end{align*}
The proposition now follows by comparing the multiplicities of
$D^\lambda_-$ on both sides and using~(\ref{corresp:spin HC}).
\end{proof}
\subsection{Proof of Theorem~A}
The symmetric group $S_n$ acts naturally on the ($\mathbb Z_+$-graded)
symmetric algebra on $V =\mathbb C^n$:
$$
S^*V =\bigoplus_{j \ge 0} S^jV.
$$
As $S_n$-modules, we will identify $ S^*V$ with the algebra
of polynomials in $n$ variables over $\mathbb C$. Note that
$\mathcal{C}_n\otimes S^* V={\rm ind}^{\mathcal{H}_n}_{\mathbb C
S_n} S^* V$ is naturally a $\mathbb Z_+$-graded
$\mathcal{H}_n$-module, with the grading inherited from the one on
$ S^* V$.
\begin{lem} \label{lem:induced}
We have the following isomorphism of
$\mathcal{H}_n$-modules for $j \ge 0$:
\begin{align}
\mathcal{C}_n\otimes S^jV\cong \bigoplus_{\nu\models
n,n(\nu)=j} {\rm ind}^{\mathcal{H}_n}_{\mathbb C S_{\nu}} {\bf 1}. \notag
\end{align}
\end{lem}
\begin{proof}
Identify $S^*V \equiv \mathbb C[x_1,\ldots, x_n]$.
The definition \eqref{n lambda} of $n(\nu)$ makes sense for any composition $\nu$.
The representatives of
the $S_n$-orbits
on the set of all monomials of degree $j$ in $\mathbb C[x_1,\ldots, x_n]$
can be chosen to be
$$
(x_1\ldots x_{\nu_1})^0 (x_{\nu_1+1} \ldots x_{\nu_1+\nu_2})^1
(x_{\nu_1+\nu_2+1} \ldots x_{\nu_1+\nu_2+\nu_3})^2 \ldots
$$
where $\nu =(\nu_1, \nu_2, \ldots)$ runs over all compositions of $n$ such that $n(\nu)=j$.
Then, as $\mathbb{C}S_n$-modules,
$$
S^jV\cong\bigoplus_{\nu\models n,n(\nu)=j} {\rm ind}^{\mathbb C
S_n}_{\mathbb C S_{\nu}} {\bf 1}.
$$
Hence, as $\mathcal{H}_n$-modules, we have
\begin{align*}
\mathcal{C}_n\otimes S^jV
\cong \bigoplus_{\nu\models n,n(\nu)=j}
{\rm ind}^{\mathcal{H}_n}_{\mathbb C S_n} {\rm ind}^{\mathbb C S_n}_{\mathbb C S_{\nu}} {\bf 1}
\cong \bigoplus_{\nu\models
n,n(\nu)=j} {\rm ind}^{\mathcal{H}_n}_{\mathbb C S_{\nu}} {\bf 1}.
\end{align*}
\end{proof}
Below, we shall denote by $[u^n] f(u)$ the coefficient of $u^n$
of a formal power series $f(u)$ in a variable $u$. We are
ready to prove Theorem~A in Introduction.
\begin{proof}[Proof of Theorem~A]
Denote by $K(\mathcal{H}_n\text{-smod})$ the complexified Grothendieck group of the
category $\mathcal{H}_n\text{-smod}$. Recall the definition of the
algebra $\Gamma_\mathbb C$ from~(\ref{Gamma}).
There exists
an isomorphism called the characteristic map~\cite{Se1, St2, Jo2}
\begin{align*}
{\rm ch}: \bigoplus_{n\geq
0}K(\mathcal{H}_n\text{-smod})&\longrightarrow\Gamma_\mathbb C,
\end{align*}
which sends $D^{\lambda}$ to
$2^{-\frac{\ell(\lambda)-\delta(\lambda)}{2}}Q_{\lambda}$ for all strict
partitions $\lambda$. It is known that
\begin{align}
{\rm ch} \big ( {\rm ind}^{\mathcal{H}_n}_{\mathbb{C}S_{\nu}}{\bf
1}\big ) = q_{\nu}, \quad \forall \nu \models n. \label{char.
map}
\end{align}
By~Lemma~\ref{lem:induced}, (\ref{char. map}) and (\ref{gen.fun.qr}), we have
\begin{align}
\displaystyle \sum_{j}t^j\text{ch}(\mathcal{C}_n\otimes S^jV)
&=\sum_jt^j\sum_{\nu\models n, n(\nu)=j}
q_{\nu}(z)\notag\\
&=\sum_{\nu\models n}
\prod _{r\geq 0}q_{\nu_{r+1}}(z)(t^{r})^{\nu_{r+1}}\notag \\
&=[u^n] \prod_{r\geq 0}\sum_{s\geq 0}q_s(z)(t^{r}u)^s\notag\\
&=[u^n] \prod_{i\geq 1,r\geq 0}\frac{1+z_it^ru}{1-z_it^ru}.
\notag
\end{align}
Recall the notation $Q_{\lambda}(t^{\bullet})$ from the introduction.
It follows from~(\ref{Cauchy identity}) that
\begin{align}
\displaystyle \prod_{i\geq 1,j\geq 0}\frac{1+z_it^ju}{1-z_it^ju}
=\sum_{\lambda:\text{ strict}}2^{-\ell(\lambda)}u^{|\lambda|}Q_{\lambda}(t^{\bullet})Q_{\lambda}(z).\notag
\end{align}
Hence
\begin{align}
\displaystyle \sum_{j}t^j\text{ch}(\mathcal{C}_n\otimes S^jV)=
\sum_{\lambda\vdash_s
n}2^{-\ell(\lambda)}Q_{\lambda}(t^{\bullet})Q_{\lambda}(z).\notag
\end{align}
Since the characteristic map ch is an isomorphism, we have an isomorphism of $\mathcal{H}_n$-modules:
\begin{align*}
\mathcal{C}_n\otimes S^* V\cong\bigoplus_{\lambda\vdash_s n}
2^{-\frac{\delta(\lambda)+\ell(\lambda)}{2}}Q_{\lambda}(t^{\bullet})
D^{\lambda}.
\end{align*}
This together with Theorem~B implies Theorem~A.
\end{proof}
The following corollary follows directly from Theorem~A
and Proposition~\ref{prop:mult.equiv}.
\begin{cor}
The graded multiplicity of $D^\lambda_-$ in the $\C S_n^-$-module
$\mathcal{B}_n\otimes S^* V$ is given by \eqref{hook} unless $n$ is
odd and $\ell(\lambda)$ is even; in this case, the graded multiplicity
is
$$
2^{-\frac{\ell(\lambda)}{2}-1}
\frac{t^{n(\lambda)}\prod_{(i,j)\in \lambda^*}(1+t^{c_{ij}})}
{\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.
$$
\end{cor}
\subsection{A graded regular $\mathcal{H}_n$-module}
It is well known that the algebra $ (S^* V)^{S_n}$ of
$S_n$-invariants in $S^* V$ is a free polynomial algebra
whose Hilbert series $P(t)$ is given by
\begin{align}
\displaystyle P(t)=\frac{1}{(1-t)(1-t^2)\cdots(1-t^n)}.\label{Poincare series}
\end{align}
Define the ring of coinvariants $(S^* V)_{S_n}$ to be the
quotient of $ S^* V$ by the ideal generated by the homogeneous invariant
polynomials of positive degrees.
It is well known that $S^* V$ is a free module over the algebra $(S^* V)^{S_n}$,
and so we have an isomorphism of graded $\mathbb C S_n$-modules
\begin{align}
S^* V&\cong (S^* V)_{S_n}
\otimes_{\mathbb C} (S^* V)^{S_n}.
\label{coin.tensor inv.}
\end{align}
\begin{thm}\label{Cliff.tensor coin.}
The graded multiplicity of $D^{\lambda}$ in $\mathcal{C}_n\otimes (S^* V)_{S_n}$ is
\begin{align*}
\displaystyle 2^{-\frac{\ell(\lambda)+\delta(\lambda)}{2}}
\frac{t^{n(\lambda)}(1-t)(1-t^2)\cdots(1-t^n)\prod_{(i,j)\in
\lambda^*}(1+t^{c_{ij}})}{\prod_{(i,j)\in \lambda^*}(1-t^{h^*_{ij}})}.
\end{align*}
\end{thm}
\begin{proof}
Follows directly from Theorem~A, (\ref{Poincare series}), and
\eqref{coin.tensor inv.}.
\end{proof}
\begin{rem}
Recall the isomorphism of $\mathcal{H}_n$-modules $D^{(n)} \cong \mathcal{C}_n$ (cf. \cite{Kle}).
It follows that the graded multiplicity
of $\mathcal{C}_n$ in $\mathcal{C}_n\otimes (S^* V)_{S_n}$ is
$
(1+t)(1+t^2)\cdots(1+t^{n-1}),
$
and the graded multiplicity
of $\mathcal{C}_n$ in $\mathcal{C}_n\otimes S^*
V$ is
\begin{align} \label{basic mult}
\frac{(1+t)(1+t^2)\cdots(1+t^{n-1})}{(1-t)(1-t^2)\cdots(1-t^n)}.
\end{align}
\end{rem}
\begin{rem}
The number $g^{\lambda}$ of standard shifted Young tableaux of shape
$\lambda$ is known to be (cf. \cite{Sa2, Mac})
\begin{align}
g^{\lambda}=\frac{n!}{\prod_{(i,j)\in \lambda^*}h^*_{ij}}.\notag
\end{align}
By the isomorphism of $\mathbb C S_n$-modules
$(S^* V)_{S_n} \cong\mathbb C S_n$, the $\mathcal{H}_n$-module
$\mathcal{C}_n\otimes (S^* V)_{S_n}$ is isomorphic to the regular representation of $\mathcal{H}_n$. It is
known (cf. \cite{Kle}) that the multiplicity of $D^{\lambda}$ in the regular
representation of $\mathcal{H}_n$ is given by
$\frac{1}{2^{\delta(\lambda)}} \ {\rm dim} D^{\lambda}$. By
specializing $t=1$ in Theorem~\ref{Cliff.tensor coin.},
we recover the dimension formula
$
{\rm dim} D^{\lambda}=2^{n-\frac{\ell(\lambda)-\delta(\lambda)}{2}} g^\lambda.
$
\end{rem}
\section{The graded multiplicity in $\mathcal{C}_n\otimes S^* V\otimes\wedge^*
V$} \label{sec:bigraded}
\subsection{The $S_n$-module $ S^* V\otimes\wedge^* V$}
The $S_n$-action on $V=\mathbb C^n$ induces a natural $S_n$-action on the exterior algebra
$$
\wedge^{*}V
=\bigoplus_{q=0}^n \wedge^{q}V.
$$
This gives rise to a $\mathbb Z_+ \times \mathbb Z_+$ bi-graded $\mathbb C S_n$-module structure on
\begin{align*}
S^*V\otimes\wedge^*V=\bigoplus_{p\geq 0, 0\leq
q\leq n}S^pV\otimes\wedge^qV.
\end{align*}
According to Kirillov and Pak \cite{KP}, the bi-graded multiplicity of the Specht module
$S^{\lambda}$ for $\lambda\vdash n$ in $S^*V\otimes
\wedge^*V$ is given by
\begin{align} \label{eq:KPak}
\sum_{p=0}^\infty \sum_{q=0}^n t^p s^q m_\lambda (S^pV \otimes \wedge^qV)
= \frac{\prod_{(i,j)\in\lambda}(t^{i-1}+st^{j-1})}{\prod_{(i,j)\in\lambda}(1-t^{h_{ij}})},
\end{align}
which can be rewritten as
\begin{align*}
\frac{t^{n(\lambda)}\prod_{(i,j)\in\lambda}(1+st^{c_{ij}})}{\prod_{(i,j)\in\lambda}(1-t^{h_{ij}})}.
\end{align*}
In particular, this recovers Solomon's formula \cite{So}
for the generating function for the bi-graded $S_n$-invariants
in $S^*V\otimes
\wedge^*V$:
\begin{align} \label{Solomon}
\frac{(1+s)(1+st)\cdots(1+st^{n-1})}{(1-t)(1-t^2)\cdots(1-t^n)}.
\end{align}
The formal similarity between the graded multiplicities \eqref{hook} and \eqref{eq:KPak}
in very different settings is rather striking.
Also compare the similarity between \eqref{basic mult} and \eqref{Solomon}.
\subsection{Proof of Theorem~C}
\label{subsec:Cliff.Sym.Exter}
\begin{lem}\label{lem:cliff.sgn}
The following holds
as $\mathcal{H}_n$-modules:
\begin{align*}
{\rm ind}^{\mathcal{H}_n}_{\mathbb{C}S_n}{\rm sgn}\cong {\rm
ind}^{\mathcal{H}_n}_{\mathbb{C}S_n}{\bf 1}.
\end{align*}
\end{lem}
\begin{proof}
Define a $\mathbb C$-linear map
\begin{align*}
f: {\rm ind}^{\mathcal{H}_n}_{\mathbb{C}S_n}{\rm sgn}=\mathcal{C}_n\otimes {\rm sgn}&\rightarrow
{\rm ind}^{\mathcal{H}_n}_{\mathbb{C}S_n}{\bf 1}\notag\\
c\otimes 1&\mapsto c\cdot(c_1c_2\cdots c_n \otimes 1)
\end{align*}
It is straightforward to show that $f$ is actually a $\mathcal{H}_n$-module isomorphism.
\end{proof}
\begin{lem} \label{Cliff.Epq}
For $p\geq 0, 0\leq q\leq n$, as $\mathcal{H}_n$-modules, we have
\begin{align*}
\mathcal{C}_n \otimes S^{p}V\otimes\wedge^qV\cong
\bigoplus_{\alpha,\beta}{\rm ind}^{\mathcal{H}_n}_{\mathbb C
(S_{\alpha}\times S_{\beta})}\mathbf{1},
\end{align*}
summed over all
$\alpha\models n-q, \beta\models q$
with $n(\alpha)+n(\beta)=p$.
\end{lem}
\begin{proof}
Arguing similarly as in the proof of Theorem~A,
we have an isomorphism of $\mathbb{C}S_n$-modules:
\begin{align*}
S^pV\otimes\wedge^qV
\cong\bigoplus_{\alpha,\beta}{\rm ind}^{\mathbb C S_n}_{\mathbb C
(S_{\alpha}\times S_{\beta})} (\mathbf{1}\otimes{\rm sgn}),
\end{align*}
summed over all
$\alpha\models n-q, \beta\models q$
with $n(\alpha)+n(\beta)=p$.
Now the lemma follows by applying $\text{ind}_{\mathbb C S_n}^{\mathcal{H}_n}$ to the above isomorphism and
using Lemma~\ref{lem:cliff.sgn}.
\end{proof}
We are ready to prove Theorem C from the Introduction.
\begin{proof}[Proof of Theorem~C]
It follows by~(\ref{char. map}) and Lemma~\ref{Cliff.Epq} that
\begin{align*}
{\rm ch}(\mathcal{C}_n \otimes S^{p}V\otimes\wedge^qV)
=\sum_{\alpha,\beta}q_{\alpha}(z)q_{\beta}(z),
\end{align*}
summed over all
$\alpha=(\alpha_1, \alpha_2,\ldots) \models n-q,
\beta =(\beta_1,\beta_2,\ldots) \models q$
with $n(\alpha)+n(\beta)=p$.
Hence,
\begin{align*}
\sum_{p\geq 0,0\leq q\leq n} & t^ps^q{\rm ch}(\mathcal{C}_n\otimes
S^{p}V\otimes\wedge^qV) \label{eqn:Cliff.sym exterior} \\
&=\sum_{p\geq 0,0\leq q\leq n}t^ps^q
\sum_{\alpha\models n-q,\beta\models q, n(\alpha)+n(\beta)=p}q_{\alpha}(z)q_{\beta}(z)\notag\\
&=[u^n]\sum_{\alpha_1, \alpha_2,\ldots,\beta_1,\beta_2,\ldots}
\prod _{r\geq 0}q_{\alpha_{r+1}}(z)(t^{r}u)^{\alpha_{r+1}}
\prod _{k\geq 0}q_{\beta_{k+1}}(z)(t^{k}su)^{\beta_{k+1}}\notag \\
&=[u^n] \prod_{i\geq 1,r\geq 0} \frac{1+z_it^ru}{1-z_it^ru}
\frac{1+z_it^rsu}{1-z_it^rsu} \notag\\
&=\sum_{\lambda\vdash_s n}2^{-\ell(\lambda)}
Q_{\lambda}(t^{\bullet};st^{\bullet})Q_{\lambda}(z) \nonumber
\end{align*}
where we have used the short-hand notation $Q_{\lambda}(t^{\bullet};st^{\bullet})$
from Introduction and \eqref{Cauchy identity} in the last equation.
Theorem~C follows.
\end{proof}
\begin{rem}
It is an interesting open problem to find an explicit formula for
$Q_{\lambda}(t^{\bullet};st^{\bullet})$. Consider a
Koszul $\mathbb Z_+$-grading which counts the standard generators of
$S^*V$ as degree $2$ and the standard generators of
$\wedge^*V$ as degree $1$. This corresponds precise
to setting $t=s^2$, and hence,
$Q_{\lambda}(t^{\bullet};st^{\bullet})=Q_{\lambda}(s^{\bullet})$.
Therefore, for the Koszul grading, the graded multiplicity of
$D^\lambda$ in $\mathcal{C}_n\otimes S^*V\otimes
\wedge^*V$ is given by the same formula \eqref{eq:bi mult}.
\end{rem}
\subsection{Some consequences of Theorem~C}
Recall the notation of $Q_{\lambda}(t^{\bullet};st^{\bullet})$ from Introduction.
By Proposition~\ref{prop:mult.equiv},
we have the following corollary as a counterpart of Theorem~C for $\C S_n^-$.
\begin{cor}
The bi-graded multiplicity of $D^\lambda_-$ in the $\C S_n^-$-module
$\mathcal{B}_n\otimes S^*V\otimes\wedge^*V$
is given by \eqref{eq:bi mult} unless $n$ is odd and $\ell(\lambda)$ is even;
in this case, the bi-graded multiplicity is
$$
2^{-\frac{\ell(\lambda)}{2}-1}Q_{\lambda}(t^{\bullet};st^{\bullet}).
$$
\end{cor}
\begin{cor} \label{mult wedge}
The graded multiplicity of $D^{\lambda}$ in the $\mathcal{H}_n$-module
${\rm ind}^{\mathcal{H}_n}_{\mathbb C
S_n}(\wedge^*V)$ is given by $ \displaystyle
2^{-\frac{\ell(\lambda)+\delta(\lambda)}{2}} Q_{\lambda}(1,s). $
Moreover,
\begin{align}
\displaystyle Q_{\lambda}(1,s)= \left \{
\begin{array}{ll}
\displaystyle \frac{2^{\ell(\lambda)}(1+s)(s^l-s^k)}{1-s},
& \text{ if }\lambda=(k,l) \text{ with }k>l\geq 0,\\
0, \text{ otherwise.}
\end{array}\notag
\right.
\end{align}
\end{cor}
\begin{proof}
The first statement is obtained by setting $t=0$ in Theorem~C.
By~(\ref{SchurQ1}), we see that
\begin{align}
\displaystyle Q_{\lambda}(z_1,z_2)= \left \{
\begin{array}{ll}
\displaystyle \frac{2^{\ell(\lambda)}(z_1+z_2)(z_1^kz_2^l-z_1^lz_2^k)}{z_1-z_2},
& \text{ if }\lambda=(k,l) \text{ with }k>l\geq 0,\\
0, \text{ otherwise.}
\end{array}\notag
\right.
\end{align}
The Corollary follows by setting $z_1=1$ and $z_2=s$.
\end{proof}
Setting $s=0$ in Theorem~C,
we recover Theorem~A.
|
\section{Introduction}
Understanding the response of a solid to applied magnetic or electric fields is
of both fundamental and applied interest. Two standard examples are that
metals can be distinguished from insulators by their screening of an applied
electric field, and superconductors from metals by their exclusion of
magnetic field (the Meissner effect). Magnetoelectric response in insulators
has been studied for many years and is currently undergoing a renaissance
driven by the availability of new materials. The linear response of this type
is the magnetoelectric polarizability: in ``multiferroic'' materials that break
parity and time-reversal symmetries, an applied electric field creates a
magnetic dipole moment and a magnetic field creates an electric dipole moment,
and several applications have been
proposed for such responses.
Such responses are observed in a variety of
materials and from a variety of mechanisms.\cite{spaldin-s05,fiebig-jpdp05}
From a theoretical point of view, the most
intriguing part of the polarizability is that due to the
orbital motion of electrons, because the orbital motion couples
to the vector potential rather than the more tangible
magnetic field.
The orbital magnetoelectric polarizability has also been studied recently in
non-magnetic materials known as ``topological insulators.''
These insulators have Bloch wavefunctions with unusual
topological properties, that lead to
a magnetoelectric response described by an ${\bf E} \cdot {\bf B}$ term in
their effective electromagnetic Lagrangians,~\cite{qilong}
with a quantized coefficient. Qi, Hughes, and Zhang~\cite{qilong} (QHZ)
gave a formula for the coefficient of this term. For the specific
case of topological band insulators, their result reproduces earlier
formulas for the relevant topological
invariant,~\cite{fu&kane&mele-2007,moore&balents-2006,rroy3d} but it
is more generally valid: it describes a contribution to the
magnetoelectric polarizability non just in topological insulators but
in any band insulator. Their formula has a periodicity or ambiguity
by $e^2/h$ that is related to the possibility of surface quantum Hall
layers on a three-dimensional sample and generalizes the ambiguity of
ordinary polarization.
The same ${\bf E} \cdot {\bf B}$ coupling, known as ``axion
electrodynamics'' and originally studied in the
1980s,~\cite{wilczekaxion} was obtained in a previous paper by three
of the present authors~\cite{essinomp} using a semiclassical
approach~\cite{xiao} to compute $dP/dB$, the polarization response to
an applied magnetic field. However, in a general material, that
semiclassical approach leads to an explicit formula for only part of
the orbital magnetoelectric polarizability, the part found by
QHZ.~\cite{qilong} The remainder, which is generically nonvanishing
in materials that break inversion and time-reversal symmetries,
is expressed only implicitly in terms of the modification of
the Bloch wavefunctions by the magnetic field.
In this paper,
we develop a more microscopic approach that enables us to compute all terms in
the orbital response explicitly in terms of the unperturbed wavefunctions,
thereby opening the door to realistic calculations using modern
band-structure methods (e.g., in the context of density-functional
theory). Moreover, beyond
its importance for computation, this expression clarifies the physical
origins of the orbital magnetoelectric polarizability and resolves some issues
that arose in previous efforts to describe the ``toroidal moment'' in periodic
systems.
In the remainder of this introduction,
we review some macroscopic features of the magnetoelectric response, while
subsequent sections will be devoted mainly to a detailed treatment of
microscopic features.
The magnetoelectric tensor can be decomposed into trace and traceless
parts as
\begin{equation} \label{eq:alphadef}
\frac{\partial P^i}{\partial B^j} = \frac{\partial M_j}{\partial E_i} = \alpha^i_j
= \tilde{\alpha}^i_j + \alpha_\theta \delta^i_j,
\end{equation}
where $\tilde{\alpha}$ is traceless and
\begin{equation}
\alpha_\theta = \frac{\theta}{2\pi} \frac{e^2}{h}
\end{equation}
is the trace part expressed in terms of the dimensionless parameter
$\theta$; $\alpha$ has the physical dimension of conductance. The
trace is the most difficult term to determine, as its physical
effects are elusive.
It should be noted that equality between $\partial P^i/\partial B^j$ and
$\partial M_j/\partial E_i$ only holds in the absence of dissipation and
dispersion, which describes the low frequency, low temperature
responses of an insulator.~\cite{expttheta,hornreichshtrikman}
The placement of the indices in Eq.~(\ref{eq:alphadef})
is not essential for the arguments and calculations
in this paper, and the reader can choose to treat $\alpha$ as a Cartesian tensor
$\alpha_{ij}$ if desired.~\cite{foot1} As a Cartesian tensor,
the traceless part decomposes further into symmetric
and antisymmetric parts
\begin{equation}
\tilde{\alpha}^S_{ij} =
\frac{1}{2} \left( \tilde{\alpha}_{ij} + \tilde{\alpha}_{ji} \right), \quad
\tilde{\alpha}^A_{ij} =
\frac{1}{2} \left( \tilde{\alpha}_{ij} - \tilde{\alpha}_{ji} \right)
=-\epsilon_{ijk} T_k,
\end{equation}
where $T_i = -\epsilon_{ijk} \tilde{\alpha}_{jk}/2$ is the toroidal response.
(Unless otherwise stated, in our work repeated indices are implicitly summed.)
The terminology reflects that this part of the orbital magnetoelectric response is related to the ``toroidal moment'', which is an order parameter that has recently been studied intensively; in a Landau effective free energy, the toroidal moment and the toroidal part of the magnetoelectric response are directly
related.~\cite{edererspaldin,aligia}
The primary goal of this paper is to compute the contribution to $\alpha$ that
arises solely from the motion of electrons due to their couplings to the
electromagnetic potentials $\rho\phi$ and $-\bf{j}\cdot\bf{A}$. We call this
contribution the orbital magnetoelectric polarizability, or OMP for short.
Other effects, such as those mediated by
lattice distortions or the Zeeman coupling to the electron's spin, are calculable
with known methods.\cite{wojdel} We shall only treat the
polarization response to an applied magnetic field here; concurrent work by
Malashevich, Souza, Coh, and one of us obtains an equivalent formula by
developing methods to compute the orbital magnetization induced by an
electrical field.\cite{malashevichOMP}
The magnetoelectric tensor's physical consequences arise through the bound
current and charge,\cite{qilong,expttheta} given by
$\rho_b=-\mathrm{div} \,\bf{P}$ and
$\mathbf{J}_b=\partial_t \bf{P} + \mathrm{curl} \,{\bf M}$.
Besides having a ground state value,
each moment responds
(instantaneously and locally, as appropriate for the low-frequency
response of an insulator)
to applied electric and magnetic fields,
\textit{e.g.}, $P^i=P_0^i+\chi_E^{ij}E_j+\alpha^i_jB^j$; we will concentrate
on the magnetoelectric response. (Unless otherwise stated, in this article
repeated indices are implicitly summed.) It is useful
to allow $\alpha^i_j$, a material property, to vary in space and time by allowing
the electronic Hamiltonian to vary; this leads
to a formula that covers the effects of boundaries and time-dependent
shearing of the crystal, for example. Then the relevant terms are
\begin{alignat}{2} \label{eq:boundcharge}
J^i_b
&= (\tilde{\alpha}^l_j \epsilon^{ikj} - \tilde{\alpha}^i_j \epsilon^{jkl} )
\partial_k E_l &&+ (\partial_t \alpha^i_j) B^j + \epsilon^{ijk} (\partial_j \alpha^l_k) E_l
\notag \\
\rho_b &= - \tilde{\alpha}^i_j\partial_i B^j &&- (\partial_i \alpha^i_j)B^j,
\end{alignat}
We have used two of
Maxwell's equations to simplify the first term in each line.
The most important point to notice here is that $\alpha_\theta$ does not appear
except in derivatives, so that any uniform and static contribution to $\theta$ has
no effect on electrodynamics. Hence
in a uniform, static crystal, the components of $\tilde{\alpha}$ can be
computed or measured from the current or charge response to spatially varying
fields, given by the first term in each line. On the other hand, if we wish
similarly to obtain $\alpha_\theta$ from charge
or current responses to applied fields, we need to
consider a crystal that varies either spatially or temporally, so that ${\bf E}$
or ${\bf B}$ will couple to $\partial_i \alpha_\theta$ or $\partial_t \alpha_\theta$, as in
the second terms of Eqs.~\eqref{eq:boundcharge}. These considerations, which
we will elaborate later, motivate our theoretical approach to
the OMP in this paper.
We will proceed as follows.
In Section II, we present the results of our calculation of the OMP
in the independent-electron approximation.
This section includes a review of known results, followed by
a discussion of the new contributions we compute and when those contributions
can be expected to vanish (so that the OMP reduces to the form found in
the literature previously). We follow these discussions with a detailed
presentation of
the calculation in Section III. This calculation involves a novel method for
dealing with a uniform magnetic field in a crystal.
An alternative derivation is presented in the Appendix.
\section{General features of orbital magnetoelectric response}
In this section we discuss properties of the OMP and its
explicit expression in the independent electron approximation. There
is a natural decomposition into two parts, which is, however, not
equivalent to the standard symmetry decomposition given in
Eq.~\eqref{eq:alphadef} of the Introduction.
The first part is the scalar
``Chern-Simons'' term $\alpha_{CS}$ obtained by QHZ~\cite{qilong}
that contributes only to the trace part $\alpha_\theta$. It is
formally similar to the Berry-phase expression for
polarization~\cite{ksv} in that it depends only on the wavefunctions,
not their energies, which explains the terminology ``magneto-electric polarization'' introduced by QHZ for $\alpha_{CS}$.~\cite{qilong}
The second part of the response is not simply scalar. It
has a different mathematical form that is not built from the Berry
connection, looking like a more typical response function in that it
involves cross-gap contributions and is not a ``moment'' determined
by the unperturbed wavefunctions. We label this term $\alpha_G$
because of its connection with cross-gap contributions.
This term does not seem to have been obtained previously although its physical
origin is not complicated.
\subsection{The OMP expression and the origin of the cross-gap term $\alpha_G$}
We first give the microscopic expression of the new term in the OMP
and discuss its interpretation.
The later parts of this section explain why the new term vanishes in
most of the models that have been introduced in the literature to study
the OMP, and discuss to what extent the two terms in the OMP expression
are physically separate.
The OMP expression that we
discuss here will be derived later in Section III as follows:
we compute the bulk current in
the presence of a small, uniform magnetic field as the crystal Hamiltonian is
varied adiabatically. The result is a total time derivative which can be
integrated to obtain the magnetically induced bulk polarization.
The most obvious property of the new term $\alpha_G$ in the response is that,
unlike the Chern-Simons piece, it has off-diagonal components;
for instance, $\partial P^x/\partial B^y\neq0$.
To motivate the expression for $\alpha_G$
intuitively, we note that it is very similar to what one would
expect based on simple response theory: An electric
dipole moment, $e{\bf r}$, is induced when a magnetic field is
applied. This field couples linearly to the magnetic dipole moment
$(e/4)({\bf r}\times{\bf v}-{\bf v}\times{\bf r})$
(this form takes care of the operator ordering when we go to operators on Bloch states).
The expression we actually get for the OMP is
expressed in terms of the periodic part of the Bloch wave functions
$u_{l{\bf k}}$ and the energies $E_{l{\bf k}}$ describing the electronic structure of a
crystal:
\begin{widetext}
\begin{subequations}
\begin{align}
\alpha^i_j &= (\alpha_G)^i_j + \alpha_{CS} \delta^i_j \\
(\alpha_G)^i_j &=
\sum_{\begin{subarray}{c}n\,\mathrm{occ}\\m\,\mathrm{unocc}\end{subarray}}
\int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} \mathrm{Re} \left\{
\frac{ \me{u_{n{\bf k}}}{e \!\not\!r^i_{{\bf k}}}{u_{m{\bf k}}}
\me{u_{m{\bf k}}}{
e(\mathbf{v}_{{\bf k}}\times\!\not\!{\bf r}_{{\bf k}})_j-e(\!\not\!{\bf r}_{{\bf k}}\times\mathbf{v}_{{\bf k}})_j
-2i\partial H'_{{\bf k}}/\partial B^j
}{u_{n{\bf k}}}
}{E_{n{\bf k}}-E_{m{\bf k}}}
\right\} \label{eq:alphaI} \\
\alpha_{CS} &= -\frac{e^2}{2\hslash} \, \epsilon_{abc} \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} \, \mathrm{tr}
\left[ \mathcal{A}^a \partial^b \mathcal{A}^c
- \frac{2i}{3} \mathcal{A}^a \mathcal{A}^b \mathcal{A}^c \right].
\label{eq:alphaCS}
\end{align} \label{eq:alphatot}
\end{subequations}
\end{widetext}
Here the Berry connection
$\mathcal{A}^a_{nn'}({\bf k}) = i \ip{u_{n{\bf k}}}{\partial_{k_a}u_{n'{\bf k}}}$ is a matrix on the
space of occupied wave functions $u_{n{\bf k}}$, and the derivative with an upper index
$\partial^a = \partial_{k_a}$ is a $k$-derivative, as opposed to the spatial derivative
$\partial_i$ in Eq.~\eqref{eq:boundcharge}. The velocity operator is related
to the Bloch Hamiltonian, $\hslash v^i({\bf k})=\partial^i H_{{\bf k}}$, while the
operator $\!\not\!r^i_{{\bf k}}$ is defined as the derivative $\partial^i\mathcal{P}_{{\bf k}}$
of the projection $\mathcal{P}$ onto the occupied bands at ${\bf k}$.
This operator is closely related to the position operator;
its ``cross-gap'' matrix elements between occupied and unoccupied bands
are $\me{u_{m{\bf k}}}{\negthickspace\not\negmedspace r^i_{{\bf k}}}{u_{n{\bf k}}} =
\me{u_{n{\bf k}}}{\negthickspace\not\negmedspace r^i_{{\bf k}}}{u_{m{\bf k}}}^* = -i\me{u_{m{\bf k}}}{r^i}{u_{n{\bf k}}}$,
while its ``interior'' matrix elements between two occupied bands or two
unoccupied bands
vanish. Finally, the operator $H'$ is introduced for generality, as
discussed in Section \ref{sec:mag}; it vanishes for the continuum
Schr\"{o}dinger
Hamiltonian and for tight-binding Hamiltonians whose hoppings are all
rectilinear, and so will be ignored for most of the analysis that follows.
Neglecting this subtlety, the form of $\alpha_G$ is nearly what would be
expected for the response in electric dipole moment to a field coupling
linearly to the magnetic dipole moment. In the derivation
presented in Section III, the term $\alpha_G$ appears in abbreviated
form at Eq.~\eqref{eq:alphaI2}, and $\alpha_{CS}$ follows immediately
from Eq.~\eqref{eq:JCSfinal}.
The main difference between the explicit form of $\alpha_G$ and the na\"ive
expectation from the dipole moment argument above is that $\alpha_G$ excludes
terms of the form $\me{n}{{\bf r}}{m}\me{m}{{\bf v}}{n'}\me{n'}{{\bf r}}{n}$,
for example, that include interior matrix elements of ${\bf r}$. In some
sense, this
omission is compensated for by the extra factor of 2
relative to the na\"ive expectation and by a
remainder term, namely,
$\alpha_{CS}$, the Chern-Simons part.
The Chern-Simons term $\alpha_{CS}$ alone has appeared
previously.\cite{qilong,essinomp} The next subsection reviews the
properties of $\alpha_{CS}$ and gives a geometrical picture for its discrete
ambiguity, which is not present in the $\alpha_{G}$ term. We then explain how
the existence of the previously unreported $\alpha_G$ can be reconciled with
previous
studies on model Hamiltonians that found only $\alpha_{CS}$, and then show that
the two terms are more intimately related than they first appear.
\subsection{The Chern-Simons form, axion electrodynamics, and topological
insulators} \label{sec:CS}
The Chern-Simons response $\alpha_{CS}$ has been discussed at some length in
the literature.\cite{qilong,essinomp} It does not emerge as clearly as
$\alpha_G$ from the intuitive argument above about dipole moment in a field;
rather, in Ref.~\onlinecite{essinomp}, it was derived by treating
the vector potential as a background inhomogeneity and utilizing a general
formalism for computing the polarization in such a background.~\cite{xiao}
The most important feature of the microscopic expression for the
\emph{isotropic} OMP is that
it suffers from a discrete ambiguity. The dimensionless
parameter $\theta$ quantifying the isotropic susceptibility contains
the term
\begin{equation}
\theta_{CS} = -\frac{1}{4\pi} \, \epsilon_{abc} \int \! d^3k \, \mathrm{tr}
\left[ \mathcal{A}^a \partial^b \mathcal{A}^c
- \frac{2i}{3} \mathcal{A}^a \mathcal{A}^b \mathcal{A}^c \right],
\end{equation}
which
is only defined up to integer multiples of $2\pi$. This is tied to a ``gauge''
invariance: ground state properties of a band insulator should only be
determined by the ground state density matrix $\rho^g_{{\bf k}}$, which is invariant
under unitary transformations $U_{nn'}(\bf{k})$ that mix the occupied bands.
Now, the Berry connection ${\cal A}$ is not invariant under such a
transformation, but there is no inconsistency because,
in the expression for $\theta_{CS}$, all
the terms produced by the gauge transformation
cancel except for a multiple of $2\pi$.
An analogous phenomenon, slightly easier to understand,
is found in the case of electric polarization\cite{ksv}
\begin{equation}
P^i = e\int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3}\, {\cal A}^i \;,
\end{equation}
which has invariance only up to a discrete ``quantum,'' or ambiguity, which
counts the number of times $U(\bf{k})$ winds around the Brillouin zone
(e.g., if $U_{11}=e^{ik_x a}$ and $U_{ii}=1$, $i\neq 1$, then $P^x$ changes
by one quantum). The
Chern-Simons response $\alpha_{CS}$ behaves similarly, although the
``winding'' that leads to the ambiguity is more complicated (in particular,
it is non-Abelian).
These ambiguities can be understood from general arguments, without relying
on the explicit formulae.
In the case of
the polarization, the quantum of uncertainty of $P^x$,
$e/S_x$, depends on the lattice structure, with $S_x$
the area of a surface unit cell normal to $x$. The ambiguity results
because the bulk polarization
does not completely determine
the surface charge:
isolated surface bands can be filled or emptied, changing the number of
surface electrons per cell by an integer.
For the magnetoelectric response, the quantum of magnetoelectric polarizability
is connected with the fact that $\theta$ gives a surface Hall
conductance, as can be seen from the term $\bf{J}_b = (\bm{\nabla}
\alpha_\theta )\times {\bf E}$ in Eq.~\eqref{eq:boundcharge}. Therefore, the
ambiguity in $\alpha_\theta$ is just $e^2/h$, the ``quantum of
Hall conductance,'' because it is
possible to add a quantum Hall layer to the surface. (This
remains a theoretical possibility even if no intrinsic quantum Hall
materials have yet been found.)
Now let us show that this ambiguity afflicts only the trace of
the susceptibility. This can be seen directly by measuring the bound
charge and currents. For example,
all the components of $\tilde{\alpha}$ can be deduced from a measurement
of $\rho_b$ in the presence of a nonuniform magnetic field [see
Eq.~(\ref{eq:boundcharge})], but $\alpha_\theta$ itself does not determine
any bulk properties.
More concretely, one can
derive the ambiguities in the magnetoelectric response from the ambiguities
in the surface polarization.
In a periodic system, which for simplicity we take to have a cubic unit cell,
the smallest magnetic field that can be applied without destroying
the periodicity of the Schr\"odinger equation corresponds to
one flux quantum per unit cell,
or $B = h / (e S)$, where $S$ is again a
transverse cell area. The ambiguity in the polarization of the system in this
magnetic field corresponds to an ambiguity in $dP/dB$ of
\begin{equation}
{\Delta P \over B} = {e/S \over h/(eS)} = {e^2 \over h}.
\end{equation}
Hence on purely geometrical grounds there is
a natural quantum $e^2/h$ of the {\it diagonal} magnetoelectric
polarizability.~\cite{essinomp}
In order to see that this uncertainty remains the same when
a \emph{small} magnetic field is applied (after all, $\alpha$ is defined
as a \emph{linear} response), we will have to construct
large supercells in a direction perpendicular to the applied $B$
(Fig.~\ref{polargeomfig}).
\begin{figure}
\includegraphics[width=0.8\columnwidth]{polargeomfig}
\caption{A supercell admits a small magnetic flux, and the quantum of
polarization transverse to the long direction is correspondingly small, but
the quantum for polarization along the long direction is much larger.
\label{polargeomfig}}
\end{figure}
While a supercell of $N$ fundamental cells has a
less precisely defined polarization
(the quantum decreases by a factor
$N$, so the \emph{uncertainty} increases), the minimum field that can be applied also decreases by this factor, so
that the uncertainty in the polarizability $dP^i/dB^i$ (no sum) remains
constant. On
the other hand, if we consider the off-diagonal response, we can consider a
supercell with its long axis parallel to the applied $B$.
In this case, the polarization quantum remains constant as the supercell grows
large and the minimum applied flux becomes small; the quantum in $dP^i/dB^j$
(for $i \not = j$) then becomes large, which means that the uncertainty
vanishes. For this geometry, a small $B$ acts like a continuous parameter,
and the change in polarization induced by $B$ can be continuously tracked, even
if the absolute polarization remains ambiguous.
Thus, with or without interactions, there is a fundamental
difference between the isotropic response and the other components
of the response.
For the trace-free components, we indeed do not find a quantum of
uncertainty in the polarizability formula. In particular, if
the toroidal response is defined by
$T_i = -\epsilon_{ijk} \tilde{\alpha}_{jk}/2$, then we
believe that a
``quantum of toroidal moment''~\cite{aligia} can only exist
when there is a spin direction with conserved ``up'' and ``down''
densities. (This toroidal moment is typically defined as
$\bm{t} = (1/2)\int {\bf r}\times\bm{\mu}d{\bf r}$, with $\bm{\mu}$ the
magnetization density,\cite{edererspaldin} or more generally in terms of a
tensor $\mathcal{T}^{ij}$ such that
$\partial_i\mathcal{T}^{ij}=-2\mu^j$.\cite{aligia})
It then reduces to the polarization difference between up
and down electrons.
A particular class of materials for which the
ambiguity in $\alpha_{\theta}$ is extremely important
is the strong topological
insulators,\cite{fu&kane&mele-2007,moore&balents-2006,rroy3d} in which
$\theta = \pi$ (Ref.~\onlinecite{qilong}).
These are time-reversal ($T$) symmetric band
insulators. At first blush, $T$ invariance should rule out magnetoelectric
phenomena at linear order, since ${\bf M}$ and ${\bf B}$ are $T$-odd. However, the
ambiguity by $2\pi$ in $\theta$ provides a loophole, since $-\pi$ is equivalent
to $\pi$.
Here we regard the ambiguity/periodicity of $\theta$ as a consequence of its microscopic origin (alternately, its coupling to electrons); because $\theta$ can be modified by $2 \pi n$ by addition of surface integer quantum Hall layers, only $\theta$ modulo $2 \pi$ is a meaningful bulk quantity for systems with charge-$e$ excitations. This is consistent with the gauge-dependence of the integral for $\alpha_{CS}$. An alternate approach is to derive an ambiguity in $\theta$ by assuming that the $U(1)$ fields are derived from a non-Abelian gauge field.\cite{wilczekaxion}
The view here that periodicity of $\theta$ results from the microscopic coupling to electrons is similar to the conventional understanding of the polarization quantum.
\subsection{Conditions causing $\alpha_G$ to vanish}
It is worthwhile to understand in more detail the conditions
under which the response $\alpha_G$ is allowed. It is forbidden in systems with inversion ($P$) or time-reversal ($T$) symmetry, which can be seen explicitly from the presence
of three $k$-derivatives acting on gauge-invariant matrices
in the formula written in terms of $\mathcal{P}_{{\bf k}}$ and
$H_{{\bf k}}$.~\cite{foot2}
However, this alone is not sufficient to explain why $\alpha_G$ did not appear
in the $T$-breaking models
previously introduced to study the OMP.\cite{essinomp,qilong,dynamicalaxion}
This is explained by the fact that
the interband contribution $\alpha_G$ [Eq.~(\ref{eq:alphaI})] will also vanish
if dispersions satisfy the following ``degeneracy" and ``reflection"
conditions:
\begin{itemize}
\item At a given $\bf k$, all the occupied valence bands have the
same energy $E^{\rm v}_{\bf{k}}$.
\item Similarly, all the unoccupied conduction bands have the
same energy $E^{\rm c}_{\bf{k}}$.
\item $E^{\rm v}_{\bf{k}}+E^{\rm c}_{\bf{k}}$ is independent of $\bf k$
(and can be taken to be zero).
\end{itemize}
This can be seen immediately in an expanded form of the integrand of
$\alpha_G$, [see Eqs.~\eqref{eq:unoccb} and \eqref{eq:unoccc}]
\begin{multline}
-\sum_{\begin{subarray}{c}n,n'\,\mathrm{occ}\\m\,\mathrm{unocc}\end{subarray}}
(E_n-E_{n'})\frac{\ip{\partial^b n}{n'}\ip{\partial^a n'}{m} \ip{m}{\partial^i n} }{E_n-E_m} \\
+\sum_{\begin{subarray}{c}n\,\mathrm{occ}\\m,m'\,\mathrm{unocc}\end{subarray}}
(E_m-E_{m'})\frac{\ip{\partial^b n}{m'}\ip{m'}{\partial^a m} \ip{m}{\partial^i n} }{E_n-E_m} \\
-\sum_{\begin{subarray}{c}n\,\mathrm{occ}\\m\,\mathrm{unocc}\end{subarray}}
\partial^b(E_n+E_m)\frac{\ip{\partial^a n}{m} \ip{m}{\partial^i n} }{E_n-E_m},
\end{multline}
where $\ket{n} = \ket{u_{n{\bf k}}}$ and $E_n = E_{n{\bf k}}$, etc.
Such a structure is automatic when only two orbitals (with
both spin states) are taken into account and
the system has particle-hole and $PT$ symmetries.
$PT$ symmetry guarantees
that the bands remain spin-degenerate even if spin is not a good quantum
number. To see this, recall that $T$ acts on wave functions as $i\sigma^y K$
and maps $\bf{k} \rightarrow -\bf{k}$. Here, $K$ is complex conjugation and
$\sigma^y$ takes the form of the usual Pauli matrix in the $z$ basis of spin.
Then $P$ maps $\bf{k} \rightarrow -\bf{k}$ again, so that $PT$ effectively acts
as ``$T$ at each ${\bf k}$.''\cite{asss}
Then particle-hole symmetry implies that the dispersion is
reflection-symmetric, $E^{\rm v}_{\bf{k}}=-E^{\rm c}_{\bf{k}}$.
\begin{figure}
\includegraphics[width=0.8\columnwidth]{phsymspec2}
\caption{Schematic band structure that leads to vanishing $\alpha_G$. The
bands below the chemical potential are degenerate with energy
$E^{\rm v}_{\bf{k}}$, while
the bands above the chemical potential have energy
$E^{\rm c}_{\bf{k}} = \mathrm{const}. - E^{\rm v}_{\bf{k}}$.
\label{fig:symspec} }
\end{figure}
Most model Hamiltonians discussed in the literature that access the
topological insulator
phase,\cite{fu&kane&mele-2007,qilong,essinomp,hosur,dynamicalaxion}
as well as the Dirac Hamiltonian (in the context of which the axion electrodynamics
was first discussed\cite{wilczekaxion}),
can be defined in terms of a Clifford algebra,~\cite{foot3}
and this ensures that the dispersions are degenerate and
reflection symmetric.
The only exception of which we are aware is the model of Guo and Franz
on the pyrochlore lattice, which has four orbitals per unit cell.\cite{guo}
The topological insulator
phase itself will not have a contribution from $\alpha_G$, since it is
$T$-invariant, and so the Guo and Franz model will not show such a response;
however, the addition of any $T$-breaking perturbation to their model
should produce off-diagonal magnetoelectric responses.
Finally, there is a simple mathematical condition that will cause $\alpha_G$
to vanish. Namely, $\alpha_G$ decreases as the gap becomes large without
changing the wave functions,
and in the limit of infinite bulk gap the only magnetoelectric response comes
from the Chern-Simons part, which is not sensitive to the energies and depends
only on the electron wave functions.
\subsection{Is the Chern-Simons contribution physically distinct?}
Apart from the ambiguity in $\alpha_{CS}$ that is not present in
$\alpha_G$, there seems to be no real physical
distinction between the two terms of the linear magnetoelectric response.
We discuss two aspects that relate to this observation below.
\emph{Localized vs. itinerant contributions}
The ambiguity in $\theta_{CS}$ can be interpreted as a manifestation of
the fact that bulk quantities cannot determine the surface quantum Hall
conductance, since a two-dimensional quantum Hall layer could appear on a
surface independent of bulk properties. This suggests, perhaps, that the
Chern-Simons term appears only in bulk systems with extended wave functions,
and is a consequence of the itinerant electrons,
while $\alpha_G$ is a localized molecule-like contribution.
However, this turns out not to be the case.
Consider a periodic array of isolated molecules, which is an extreme limit of
the class of crystalline insulators. Such a system has flat bands, with
energies equal to the energies of the molecular states, since the
electrons cannot propagate. It is certainly possible to construct a molecular
system where all the unoccupied states have
one energy and all the occupied states have another, by tuning the potentials.
In this case $\alpha_G$
will vanish. However, such a molecule can still display a magnetoelectric
response; it will therefore have to be given by $\alpha_{CS}$ (and so
restricted to diagonal responses). For example, consider the ``molecule''
of Fig.~(\ref{fig:molecule}) with the shape of a regular tetrahedron.
If the two low-energy levels are occupied, the magnetoelectric response is
\begin{equation*}
\frac{\partial P^i}{\partial B^j} = \pm \delta^i_j \frac{1}{\sqrt{6}} \frac{e^2}{\hslash},
\end{equation*}
where $P^i$ here is the electric dipole moment divided by the volume of the
tetrahedron; the sign of the polarizability reverses when the complex phases
are reversed. This shows
that the Chern-Simons term does not require delocalized orbitals.
\begin{figure}
\includegraphics[width=0.6\columnwidth]{molecule2}
\caption{A tetrahedral tight-binding molecule for spinless electrons, with one
orbital per site and
complex hoppings. The hopping integrals are all equal, except that those
around one face have a phase of $i$ relative to the other three. There are
then two pairs of degenerate levels.
\label{fig:molecule}}
\end{figure}
\emph{Additivity}
Another argument against distinguishing
between the Chern-Simons part and the rest of the susceptibility
is based on band additivity.
When interactions are not taken into account, each occupied band can
be regarded as an independent physical system
(at least if
there are no band crossings). Applying a magnetic
field causes each band $n$ to become polarized by a certain amount $\bf{P}_n$,
and so the net polarization should be $\bf{P}=\sum_n \bf{P}_n$. The
Pauli exclusion principle does not lead to any ``interactions"
between pairs of bands, because the polarization (like any single-body
operator) can be written as the sum of the mean polarization in
each of the orthonormal occupied states.
Now the Chern-Simons form does not look particularly
additive in this sense, and is not by itself.
Because it is the trace of a
matrix product in the occupied subspace, it necessarily involves matrix
elements between different occupied states, while an additive formula would
not. Nevertheless, $\alpha_G$ and $\alpha_{CS}$ are together additive, as can
be seen most simply in Eq.~\eqref{eq:compactalpha}, where the two terms combine
into a single sum over occupied bands.
In terms of $\alpha_G$ and $\alpha_{CS}$ separately, one finds that when the
values of $\alpha_G$, assuming just band
1 or 2 is occupied, are added together, some terms occur that
are not present in the expression for $\alpha_G(1+2)$ (where
both bands are occupied), and vice-versa.
Using Eqs.~\eqref{eq:unoccb} and \eqref{eq:unoccc}
we see, in fact, that
$\alpha_G$ is a sum of contributions which depend on three bands, as
$\alpha_G=\sum_{n,m,m'} C(n;m,m') + \sum_{n,n',m} D(n,n';m)$. Terms such
as $C(1;2,m')$ are not
present in the expression for $\alpha_G(1+2)$. (Likewise
$D(1,2;m)$ appears in $\alpha_G(1+2)$ but not in
$\alpha_G(1)$ and $\alpha_G(2)$.) Adding up the
discrepancies,
one finds that the energy-denominators all cancel,
and the non-diagonal terms from the Chern-Simons form appear!
Seemingly paradoxical is the fact that for band structures satisfying
the degeneracy and reflection conditions of the last subsection,
the magnetoelectric susceptibility
is given by the Chern-Simons term alone, which does not seem to be additive.
However, the additivity property applies only to bands that do not cross.
It does not make any sense to ask whether the susceptibility is the sum over
the susceptibilities for the systems in which just one of the degenerate bands
is occupied, since those systems are not gapped.
\section{The OMP as currents in response to Chemical Changes}
Now we will tackle the problem of deriving the formula for
the OMP $\alpha$ discussed in the
last section. There are two impediments we need to overcome, a physical
one, and a more technical one (which we will overcome starting
from an insight of
Levinson).\cite{levinson}
In order to determine $\alpha$, we would like to carry out a thought
experiment in which a crystal is exposed to appropriate electromagnetic
fields. For specificity, we will apply a uniform magnetic field.
To make the calculation of the response
clean, we wish to deal with an infinite crystal.
Then the polarization does not simply reduce to the first moment of the
charge density,\cite{ksv}
so we will instead have to calculate the current or
charge distribution induced by the fields, and then use
Eq.~(\ref{eq:boundcharge}) to deduce $\alpha$. If both the crystal
and the electromagnetic fields are independent of space and time,
there is no macroscopic charge
or current density. We will assume \emph{spatial} uniformity,
so that there are two choices for how to proceed. Either the magnetic
field can be varied in time or the crystal parameters, and thus
$\alpha$, can be varied. In either case,
we measure the current that flows through the bulk and try
to determine $\alpha$.
As ever, the diagonal response $\alpha_\theta$ is the
most difficult to capture: while either time-dependent
experiment can be used to determine $\tilde{\alpha}$, only
the latter approach sheds light on the value of
$\alpha_\theta$.
To see why $\alpha_\theta$ can be determined only in this
way (given that we want to work with a spatially homogeneous geometry),
let us discuss how currents flow through the
crystal. The necessity of varying the crystal
in time can be deduced
from Maxwell's equations (see below) but we will give a more intuitive
discussion here. Suppose that $\tilde{\alpha}=0$.
Then in an applied magnetic field there is a polarization
$\bf{P}=\alpha_\theta {\bf B}$; thus the crystal
gets charged at the surface. As the magnetic field is turned on,
this surface charge has to build up (charge density
$\hat{\bf{n}}\cdot\bf{P}$).
This
occurs entirely due to flows of charge {\it along the surface}.
Suppose, for example, that the sample is a cylinder (radius $R$) with the
magnetic field along its $z$ axis, as illustrated in
Fig.~\ref{fig:currents}(a).
\begin{figure}
\includegraphics[width=0.85\columnwidth]{currents}
\caption{As outlined in the text, (a) turning on a magnetic field produces
a macroscopic polarization through the flow of surface currents, while (b)
varying the crystal Hamiltonian in the presence of a fixed magnetic field
produces a polarization through the flow of current through the bulk.
\label{fig:currents}}
\end{figure}
Then an
electric field ${\bf E}_{\mathrm{ind}}=-\dot{B}R\hat{\bm{\phi}}/2$ is induced at
the surface according to Faraday's law. Besides being the magnetoelectric
response, $\theta$ also represents the Hall coefficient for surface currents.
Therefore, a current of $J_s=\alpha_\theta\dot{B}R/2$ flows to the
top of the cylinder, adding up to a surface charge of
$2\pi R\int \!J_s(t)dt=\alpha_\theta B_f\pi R^2$ and producing the entire
polarization
$\alpha_\theta B_f$. No current flows through the bulk!
In fact, the Hall conductance on the circular face
produces a radial current as well, so that the charge distributes over the
surface rather than just accumulating in a ring.
Note that the surface current
grows with the radius of the cylinder. This sounds like a nonlocal
response, but it can be understood as follows:
the electric field is determined by the non-local
Faraday law, but the \emph{crystal's} response to the electric field (namely,
the surface current) is local.
The current distribution can be understood directly from
Maxwell's equations: there are two contributions to the bulk
current, $\partial_t \mathbf{P}$ and $\mathrm{curl}\ \mathbf{M}$.
The polarization is $\alpha_\theta \mathbf{B}$ while the
magnetization is indirectly produced by the induced
electric field, $\alpha_\theta \mathbf{E}_{\mathrm{ind}}$. The two contributions
thus cancel by Faraday's law in the bulk:
$\mathbf{J}_{b}^{\mathrm{bulk}}
=\alpha_\theta(\partial_t \mathbf{B}+\mathrm{curl}\ \mathbf{E})=0$.
There \emph{is} a surface
current because
$\alpha_\theta$ is discontinuous there.
On the other hand, if $\theta$ changes in time, while
the magnetic field is time-independent (as in Fig.~\ref{fig:currents}b),
the polarization at the
ends of the cylinder builds up entirely by means of flows
of charge \emph{through the bulk}. Surface
flows cannot be large enough to explain the net polarization in
this situation. Since there is no induced electric field, the
surface current is just proportional to the lateral surface area and
is negligible compared to the bulk
current. Therefore the bulk current is equal to
$(\partial_t\alpha_\theta) B$ and can be integrated
to give $\alpha_\theta B$.
For the other component of the OMP, $\tilde{\alpha}$,
either thought experiment can be used.
The simplest approach, however, is still the crystal-variation method,
since the surface currents
are negligible in that case,~\cite{foot4} and in
any case this method allows us to find all the components
of $\alpha$ simultaneously.
\emph{Difficulties with the operator ${\bf r}$ and uniform magnetic fields}
There are two technical difficulties in the theory.
First, the operator ${\bf r}$ has unbounded
matrix elements and thus the matrix elements of
the magnetic dipole
moment $(e/4)(\mathbf{v}\times{\bf r}-{\bf r}\times\mathbf{v})$ are not
well-defined. This rules out the straightforward use of perturbation
theory to calculate
the electric dipole moment of an infinite crystal
in a uniform magnetic field. Second, if we consider
a crystal in a uniform magnetic field, Bloch's
theorem does not hold. Although
the magnetic field is uniform, the vector potential that appears in
the Hamiltonian depends on ${\bf r}$.
We avoid the problems of ${\bf r}$ as follows.
The key idea is to work with the ground state density
matrix, rather than wave functions. The individual eigenstates change
drastically when a magnetic field, no matter how small,
is applied (consider the difference
between a plane wave and a localized Landau level). However, the
\emph{density matrix} of an insulator
summed over the occupied bands only changes by a small
amount when $\mathbf{B}$ is applied; over short distances the magnetic
field cannot have a strong effect (even in the example of Landau levels),
and the density matrix has only short-range correlations because
it describes an insulating state.
More technically, we show (subsection \ref{sec:mag})
that the broken translation invariance
of any single-body operator $\mathcal{O}$ (such as the density
matrix) can be dealt with
by factoring out an Aharonov-Bohm-like phase from
its matrix $\mathcal{O}_{{\bf r}_1{\bf r}_2}$. This solves the problem of the
nonuniform gauge field and
leads to
expressions that depend only on differences between $\mathbf{r}$'s.
In addition, since the exponentially decaying
ground-state density matrix appears multiplying every expression,
the factors of $\mathbf{r}_1-\mathbf{r}_2$
are suppressed.
The calculation then proceeds as follows. First, using the
symmetries of the electron Hamiltonian in a uniform magnetic field, we find
how the density matrix changes in a weak magnetic field.
Next we compute the current response to an adiabatic variation of the
crystal Hamiltonian. Finally, we show that this current can be expressed as
a total time derivative, and therefore can be integrated to give the
polarization; at linear order in ${\bf B}$ we can read off the coefficients,
the magnetoelectric tensor $\alpha$.
\subsection{Single-body operators for a uniform magnetic field} \label{sec:mag}
Recall the form of the Schr\"{o}dinger Hamiltonian for a single electron in
a crystal and under the influence of a magnetic field,
\begin{equation}
\label{eq:eAoverc}
H_S(\bf{p},{\bf r}) = \frac{1}{2m} \left[ \bf{p} - e \bf{A}({\bf r}) \right]^2 + V({\bf r}),
\end{equation}
where $V({\bf r}+\bf{R}) = V({\bf r})$ for lattice vectors $\bf{R}$. The necessity
of using the
vector potential $\bf{A}$ seems at first to spoil
the lattice translation symmetry one would expect in a uniform
magnetic field. However, as
noted by Brown\cite{brown} and Zak,\cite{zak} a more subtle form of translation
symmetry remains. In particular, choosing the gauge
\begin{equation}
\bf{A} = \frac{1}{2} {\bf B} \times {\bf r},
\end{equation}
the Hamiltonian has ``magnetic translation symmetry'':
\begin{equation}
H_S(\bf{p},{\bf r}+\bf{R})
= e^{ie{\bf B}\cdot(\bf{R}\times{\bf r})/2\hslash} H_S(\bf{p},{\bf r})
e^{-ie{\bf B}\cdot(\bf{R}\times{\bf r})/2\hslash}.
\end{equation}
This condition defines magnetic translation symmetry for general single-body
operators.
Any operator $\mathcal{O}$ possessing this symmetry can be written
in the position basis as
\begin{subequations}
\begin{equation}
\mathcal{O}_{{\bf r}_1{\bf r}_2} = \bar{\mathcal{O}}_{{\bf r}_1{\bf r}_2}
e^{-ie {\bf B}\cdot({\bf r}_1\times{\bf r}_2)/2\hslash },
\end{equation}
where $\bar{\mathcal{O}}$ has lattice translation invariance,
\begin{equation}
\bar{\mathcal{O}}_{{\bf r}_1+\bf{R},{\bf r}_2+\bf{R}} = \bar{\mathcal{O}}_{{\bf r}_1{\bf r}_2}.
\end{equation}
\end{subequations}
Note that the phase is just $(ie/\hslash)\int d\bm{\ell}\cdot\bf{A} $ calculated
along the straight line from ${\bf r}_2$ to ${\bf r}_1$, which agrees with the intuition
that comes from writing the second-quantized form of the operator,
\begin{equation} \label{eq:straightline}
\mathcal{O} = \int\! d^3r_1 d^3r_2 \, \bar{\mathcal{O}}_{{\bf r}_1{\bf r}_2}
c^\dag_{{\bf r}_1}e^{-ie {\bf B}\cdot({\bf r}_1\times{\bf r}_2)/2\hslash }c_{{\bf r}_2}.
\end{equation}
This argument shows how to couple general Hamiltonians to uniform fields:
$H=\exp[(ie/\hbar)\int_{\mathbf{r}_2}^{\mathbf{r}_1}
d\bm{\ell}\cdot\mathbf{A}]
[H_{0\mathbf{r}_1\mathbf{r}_2}+H'_{\mathbf{r}_1\mathbf{r}_2}(B)]$.
The vector potential appears explicitly only in $\mathbf{A}$,
while $H'(B)$ gives the rest of the dependence on the magnetic field.
The Schr\"{o}dinger Hamiltonian (\ref{eq:eAoverc})
is obtained if we take
\begin{equation}
\bar{H}_{0,{\bf r}_1{\bf r}_2} = \left[ -\frac{\hslash^2}{2m}\nabla^2_{{\bf r}_2}
+ V({\bf r}_2) \right] \delta^{(3)}({\bf r}_2-{\bf r}_1).
\end{equation}
and set $H'=0$. Our results also apply to tight-binding models.
We introduce
$H'$ to capture the possibility that in a tight-binding model the hoppings will
not be rectilinear, and hence that the phases in Eq.~\eqref{eq:straightline}
do not capture the full field dependence of the Hamiltonian.
\subsection{The ground state density operator}
We find it convenient to work with the one-body density matrix $\rho^g$, or
equivalently the projector onto the occupied states, whenever possible,
because it is a
basis-independent object. Also, in an insulator, $\rho^g_{{\bf r}_1{\bf r}_2}$ is
exponentially suppressed in the distance $|{\bf r}_2-{\bf r}_1|$, which tempers the
divergences that arise from the unboundedness of ${\bf r}$.\cite{kohnloc}
In any case, if the ground state is translationally symmetric,
the structure described
above will apply to $\rho_g$ and we can be sure that \emph{the density
matrix has translational symmetry apart from a phase}:
\begin{equation}
\rho^g_{{\bf r}_1{\bf r}_2} = \bar{\rho}^g_{{\bf r}_1{\bf r}_2}
e^{-ie {\bf B}\cdot({\bf r}_1\times{\bf r}_2)/2\hslash },\label{eq:ahronov}
\end{equation}
where $\bar{\rho}^g$ possesses the translation symmetry of the crystal lattice
and hence should connect smoothly to the field-free density matrix.
Hence we will
write
\begin{equation}
\bar{\rho}^g = \rho_0 + \rho',
\end{equation}
where $\rho_0$ is the density operator of the crystal in the absence
of the
magnetic field.
\emph{Density matrix perturbation theory:}
Now we have to calculate $\rho'$, using a kind of
perturbation theory that focuses on density matrices rather
than wave functions, since the wave-functions suffer from the
problems discussed above. This perturbation theory
starts from two characteristic properties of the
density matrix: it commutes with $H$, and for fermions at zero
temperature it is a projection operator.
The latter means that
all states are either occupied or unoccupied, so the eigenvalues of
the density operator are 0 and 1, which is formalized as
\begin{equation}
\rho^g\rho^g = \rho^g
\end{equation}
(idempotency).\cite{diracdensity} Expressed in the position basis,
\begin{align}
\rho^g_{{\bf r}_1{\bf r}_3} &= \int d{\bf r}_2 \rho^g_{{\bf r}_1{\bf r}_2}\rho^g_{{\bf r}_2{\bf r}_3} \notag \\
\bar{\rho}^g_{{\bf r}_1{\bf r}_3}
&= \int d{\bf r}_2 \bar{\rho}^g_{{\bf r}_1{\bf r}_2}\bar{\rho}^g_{{\bf r}_2{\bf r}_3}
e^{-(ie/2\hslash){\bf B}\cdot({\bf r}_1\times{\bf r}_2+{\bf r}_2\times{\bf r}_3+{\bf r}_3\times{\bf r}_1)}.
\end{align}
The exponent is just $-i \phi_{123}/\phi_0$, proportional to
the magnetic flux through triangle
123, and the exponential can be expanded for small $B$. At first order this
gives
\begin{multline} \label{eq:rhoprime}
\rho'_{{\bf r}_1{\bf r}_3} = \int d{\bf r}_2 [\rho'_{{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3}
+ \rho_{0{\bf r}_1{\bf r}_2} \rho'_{{\bf r}_2{\bf r}_3} \\
- \rho_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \left(i\phi_{123}/\phi_0\right) ].
\end{multline}
\emph{The problem of the unbounded }${\bf r}$'s \emph{is resolved} in
this equation because the area $A_{123}$ of the triangle is finite and
independent of the origin, and also suppressed by the factor of $\rho$.
\emph{Calculation of} $\rho'$:
In the last term of Eq.~\eqref{eq:rhoprime}, we can rewrite
$2A_{123}={\bf r}_1\times{\bf r}_2+{\bf r}_2\times{\bf r}_3+{\bf r}_3\times{\bf r}_1 = ({\bf r}_2-{\bf r}_1)\times({\bf r}_3-{\bf r}_2)$
and then use $({\bf r}_2-{\bf r}_1)\rho_{0{\bf r}_1{\bf r}_2} = [\rho_0,{\bf r}]_{{\bf r}_1{\bf r}_2}$, etc., to
obtain
\begin{equation} \label{eq:exclusion}
(1-\rho_0)\rho'(1-\rho_0) - \rho_0\rho'\rho_0
= -i\frac{e}{2\hslash}{\bf B}\cdot([\rho_0,{\bf r}]\times[\rho_0,{\bf r}]).
\end{equation}
If we define
\begin{equation}
\bar{H} = H_0 + H',
\end{equation}
then analogous manipulations (including
$({\bf r}_1-{\bf r}_2)H_{{\bf r}_1{\bf r}_2} = i\hslash \bf{v}_{{\bf r}_1{\bf r}_2}$) on the equation
$[H,\rho^g] = 0$ give
\begin{equation}
[\rho',H_0] = \frac{e}{2}\mathbf{B}\cdot
([\rho_0,{\bf r}]\times\mathbf{v}-\mathbf{v}\times[\rho_0,{\bf r}]) - [\rho_0,H'].
\label{eq:Lorentz}
\end{equation}
Eqs.~(\ref{eq:exclusion}) and (\ref{eq:Lorentz}) have an
intuitive meaning. The former equation determines the ``interior'' matrix
elements of $\rho'$, those between two occupied or two unoccupied
states of the zero-field Hamiltonian.
An perturbation with the full crystal symmetry does not change the interior
matrix elements of the density matrix because
of the exclusion principle.\cite{mcweeny} In our case, however, multiplying
$\rho_0$ by the phase $e^{(ie/2\hslash){\bf B}\cdot{\bf r}_1\times{\bf r}_2}$ gives a density
matrix
with the correct magnetic translation symmetry, but also changes the momentum
of the states and so results in a small probability for states to be douby
occupied. Therefore $\rho'$ must
correct for this ``violation of the exclusion
principle.'' On the other hand Eq.~(\ref{eq:Lorentz}) determines
the ``cross-gap'' matrix elements of $\rho'$ (those between unoccupied
and occupied states). These matrix elements capture the
expected ``transitions across the gap'' induced by the field.
The rest of this section is devoted to calculating all these matrix elements.
The results are given in Eqs.~\eqref{eq:Xanswer} and \eqref{eq:Lanswer}; the
derivations could be skipped on a first reading.
\emph{Calculation of\ }$\rho'$.
Precisely speaking, Eq.~(\ref{eq:exclusion})
gives the matrix elements of $\bar{\rho}^g$
between pairs of occupied ($n$ and $n'$) or unoccupied ($m$ and $m'$) states:
\begin{align} \label{eq:Xanswer}
\me{\psi_{n{\bf k}}}{\bar{\rho}^g}{\psi_{n'{\bf k}}} &= \delta_{nn'}
-\frac{e}{4\hslash} B^j \epsilon_{jab} {\cal F}^{ab}_{nn'} ({\bf k}) \notag\\
\me{\psi_{m{\bf k}}}{(1-\bar{\rho}^g)}{\psi_{m'{\bf k}}} &= \delta_{mm'}
-\frac{e}{4\hslash} B^j \epsilon_{jab} \check{{\cal F}}^{ab}_{mm'} ({\bf k}),
\end{align}
where ${\cal F}$ is the non-Abelian Berry curvature associated with the occupied
bands,
\begin{align}
{\cal F}^{ab}_{nn'}
&= i \me{u_{n{\bf k}}}{\partial^a\mathcal{P}_{\bf k}\partial^b\mathcal{P}_{\bf k}-\partial^b\mathcal{P}_{\bf k}\partial^a\mathcal{P}_{\bf k}}{u_{n'{\bf k}}} \notag \\
&= \partial^a {\cal A}^b_{nn'} - \partial^b {\cal A}^a_{nn'} -i [{\cal A}^a,{\cal A}^b]_{nn'},
\end{align}
and $\check{{\cal F}}$ is the corresponding quantity for the unoccupied bands.
To derive these relations, we use
\begin{equation}
\rho^g = \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} e^{i{\bf k}\cdot{\bf r}} \mathcal{P}_{{\bf k}} e^{-i{\bf k}\cdot{\bf r}}
\end{equation}
where $\mathcal{P} = \sum_{n\,\mathrm{occ}} \ket{u_{n{\bf k}}}\bra{u_{n{\bf k}}}$ is the
projector onto filled bands at ${\bf k}$. This gives
\begin{align} \label{eq:notr}
i[\rho^g,{\bf r}]
&= \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} e^{i{\bf k}\cdot{\bf r}} (\bm{\nabla}_{{\bf k}}\mathcal{P}_{{\bf k}} ) e^{-i{\bf k}\cdot{\bf r}}
\notag \\ &= \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} e^{i{\bf k}\cdot{\bf r}} \not{\bf r}_{\bf k} e^{-i{\bf k}\cdot{\bf r}}
\end{align}
after discarding a total derivative. The notation
$\not\!\!{\bf r} = \bm{\nabla}_{{\bf k}}\mathcal{P}$ was introduced in Eq.~\eqref{eq:alphatot}.
By contrast, Eq.~\eqref{eq:Lorentz}
describes to what extent
$\bar{\rho}^g$ fails to commute with $H_0$, the crystal Hamiltonian, and
gives the matrix elements of $\rho'$
between occupied and unoccupied states. In this sense it is analogous to the
more usual results for density-matrix perturbation theory.\cite{mcweeny}
In the basis of unperturbed energy eigenstates,
\begin{multline} \label{eq:Lanswer}
\me{\psi_{n{\bf k}}}{\rho'}{\psi_{m{\bf k}}} = i\frac{e}{2\hslash} B^j \epsilon_{jab}
\frac{ \me{u_{n{\bf k}}}{\{\partial^a \mathcal{P}_{{\bf k}},\partial^b H_{{\bf k}} \}}{u_{m{\bf k}}} }{E_{n{\bf k}}-E_{m{\bf k}}}
\\
+ \frac{ \me{u_{n{\bf k}}}{H'_{{\bf k}}}{u_{m{\bf k}}} }{E_{n{\bf k}}-E_{m{\bf k}}}.
\end{multline}
Recall that $\hslash v^b = \partial^b H_{{\bf k}}$ and that $H'$ is introduced
only to capture unusual situations such as
tight-binding models with non-straight hoppings, and vanishes for the
continuum problem.
Eqs.~\eqref{eq:Xanswer} and \eqref{eq:Lanswer} are the key technical results
of this formalism, good to linear order in the magnetic field.
\subsection{Adiabatic current}
Now we need to calculate the current as the Hamiltonian is changed
slowly as a function of time, as in the ordinary theory of
polarization.\cite{resta,ksv}
We have to be careful, however, since the current vanishes in the
zero-order adiabatic ground state described by density matrix
$\rho^g(t)$. It is
necessary to go to first order in adiabatic perturbation theory, which
takes account of the fact that the true dynamical density matrix $\rho(t)$
has an extra contribution proportional to $dH/dt=\dot{H}$.
However, once the current has been expressed in terms of
$\dot{\rho}$, which is proportional to $\dot{H}$, the distinction
is no longer important and the adiabatic approximation can be made.
\emph{Preparing for the adiabatic approximation:}
We can write the current as
\begin{equation}
\mathbf{J}(t) = \frac{e}{\Omega} \mathrm{Tr} \rho(t)\mathbf{v}
= \frac{e}{\Omega} \frac{i}{\hslash}\mathrm{Tr} \rho[H,{\bf r}]
\label{eq:comm1}
\end{equation}
where $\Omega$ is the crystal volume.
Here $H$ is the full Hamiltonian
including the magnetic field. By unitarity of time
evolution, it remains a projector if the initial state describes filled
bands only. The operator ${\bf r}$ appears here, but in a commutator. Since
$\me{{\bf r}_1}{[\mathcal{O},{\bf r}]}{{\bf r}_2} = ({\bf r}_2-{\bf r}_1)\mathcal{O}_{{\bf r}_1{\bf r}_2}$, such
expressions do not suffer from the difficulties of an ``unprotected'' ${\bf r}$,
namely its unboundedness. We
can only use cyclicity of the trace to the extent that this property can be
preserved. In particular, the expression $\mathrm{Tr}\, {\bf r}[\rho,H]$,
which seems formally equivalent to Eq.~(\ref{eq:comm1}), poses problems,
but
\begin{equation}
\mathbf{J}(t)= \frac{e}{\Omega} \frac{i}{\hslash}
\mathrm{Tr} [\rho,[\rho,\mathbf{r}]][\rho,H]
\end{equation}
does not. This expression can be derived from Eq.~(\ref{eq:comm1}) using
again the idempotency of $\rho$ ($\rho\rho=\rho$).
Using the equation of motion for the density matrix,
\begin{equation}
i\hbar\dot{\rho}(t)=[H(t),\rho(t)], \label{eq:schrodinger}
\end{equation}
and making the approximation $\rho\approx \rho^g$ on the right-hand
side at this stage,
the current becomes
\begin{equation} \label{eq:threepigs}
\mathbf{J} = \frac{e}{\Omega} \int\! d{\bf r}_1d{\bf r}_2d{\bf r}_3({\bf r}_1-2{\bf r}_2+{\bf r}_3)
\rho^g_{{\bf r}_1{\bf r}_2}\rho^g_{{\bf r}_2{\bf r}_3}\dot{\rho}^g_{{\bf r}_3{\bf r}_1}.
\end{equation}
\emph{Magnetic field dependence of the current}:
The considerations given in the last subsection make the integrand
\begin{equation}
\rho^g_{{\bf r}_1{\bf r}_2}\rho^g_{{\bf r}_2{\bf r}_3}\dot{\rho}^g_{{\bf r}_3{\bf r}_1}
= \bar{\rho}^g_{{\bf r}_1{\bf r}_2}\bar{\rho}^g_{{\bf r}_2{\bf r}_3}\dot{\bar{\rho}}^g_{{\bf r}_3{\bf r}_1}
e^{-i\phi_{123}/\phi_0},
\end{equation}
where, again,
$\phi_{123} = {\bf B}\cdot({\bf r}_1\times{\bf r}_2+{\bf r}_2\times{\bf r}_3+{\bf r}_3\times{\bf r}_1)/2$ is
the magnetic flux through the triangle with vertices ${\bf r}_1{\bf r}_2{\bf r}_3$ and does
not suffer from the pathologies of ${\bf r}$ itself, which allows us to expand
$e^{-i\phi_{123}/\phi_0} = 1 - i\phi_{123}/\phi_0$ to lowest order in $B$
(recall again that the matrix elements of $\rho$ are exponentially suppressed
with distances).
Recalling our division $\bar{\rho}^g = \rho_0 + \rho'$ where $\rho'$ is of
first order in the magnetic field, Eq.~\eqref{eq:threepigs} becomes
\begin{multline} \label{eq:fullcurrent}
\mathbf{J} = \frac{e}{\Omega} \int d{\bf r}_1d{\bf r}_2d{\bf r}_3 ({\bf r}_1-2{\bf r}_2+{\bf r}_3)
\bigg[ \rho_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \dot{\rho}'_{{\bf r}_3{\bf r}_1} \\
+ \rho'_{{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \dot{\rho}_{0{\bf r}_3{\bf r}_1}
+ \rho_{0{\bf r}_1{\bf r}_2} \rho'_{{\bf r}_2{\bf r}_3} \dot{\rho}_{0{\bf r}_3{\bf r}_1} \\
-i\frac{\phi_{123}}{\phi_0}
\rho_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \dot{\rho}_{0{\bf r}_3{\bf r}_1} \bigg]
\end{multline}
at first order.
The rest of the calculation involves substituting the expressions
for the magnetic-field dependence of $\rho^g$ obtained earlier,
and integrating the result to obtain $\alpha$.
The energy-dependent part of $\alpha$, namely $\alpha_G$, will come from
the mixing of the occupied and unoccupied bands, Eq.~(\ref{eq:Lanswer}).
The Chern-Simons
term will come from the ``exclusion-principle--correcting'' terms,
Eq.~\eqref{eq:Xanswer}, as well as the $\phi_{123}$ term in the previous equation.
\emph{Integrating the results}:
\begin{subequations}
The four terms in the current can be collected and rearranged into
the form
\begin{equation}
\bf{J} = \bf{J}_G + \bf{J}_{CS1} + \bf{J}_{CS2}
\end{equation}
and integrated with respect to time as follows.
The first term in Eq.~\eqref{eq:fullcurrent} can be rewritten with
$\rho_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \dot{\rho}'_{{\bf r}_3{\bf r}_1}
= \partial_t (\rho_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \rho'_{{\bf r}_3{\bf r}_1})
- \dot{\rho}_{0{\bf r}_1{\bf r}_2} \rho_{0{\bf r}_2{\bf r}_3} \rho'_{{\bf r}_3{\bf r}_1}
- \rho_{0{\bf r}_1{\bf r}_2} \dot{\rho}_{0{\bf r}_2{\bf r}_3} \rho'_{{\bf r}_3{\bf r}_1}$
and combined with the next two terms to give
\begin{align}
\bf{J}_G &= \frac{e}{\Omega} \partial_t \mathrm{Tr} [\rho_0,{\bf r}][\rho',\rho_0]
\label{eq:totder} \\
\bf{J}_{CS1} &= - 3 \frac{e}{\Omega} \mathrm{Tr} \rho'[\dot{\rho}_0,[\rho_0,{\bf r}]],
\label{eq:cscurra}
\end{align}
while the final term in Eq.~\eqref{eq:fullcurrent} (\textit{i.e.}, the
term involving $\phi_{123}$) becomes
\begin{equation}
\bf{J}_{CS2} = -i \frac{e^2}{2\hslash\Omega} B^j \epsilon_{jab} \mathrm{Tr} [\rho_0,{\bf r}]
[\rho_0,r^a] [r^b,\dot{\rho}_0] + \mathrm{c.c.} \label{eq:cscurrb}
\end{equation}
\end{subequations}
upon rewriting
\begin{multline}
({\bf r}_1-2{\bf r}_2+{\bf r}_3) ({\bf r}_1\times{\bf r}_2+{\bf r}_2\times{\bf r}_3+{\bf r}_3\times{\bf r}_1) \\
= ({\bf r}_1-{\bf r}_2) [({\bf r}_1-{\bf r}_3)\times({\bf r}_2-{\bf r}_3)] \\
+ ({\bf r}_3-{\bf r}_2) [({\bf r}_1-{\bf r}_2)\times({\bf r}_1-{\bf r}_3)]. \notag
\end{multline}
The total derivative term $\bf{J}_G$ [Eq.~\eqref{eq:totder}] can be written
\begin{equation}
J^i_G = \partial_t (\alpha_G)^i_jB^j
\end{equation}
with $\alpha_{G}$ as given in Eq.~\eqref{eq:alphaI},
\begin{multline} \label{eq:alphaI2}
(\alpha_G)^i_j = \frac{e^2}{\hslash}\mathrm{Re} \negthickspace
\sum_{\begin{subarray}{c}n\,\mathrm{occ}\\m\,\mathrm{unocc}\end{subarray}}
\negthickspace
\frac{ \me{n}{\partial^i\mathcal{P}}{m}
\me{m}{
\epsilon_{jab} \{ \partial^a H, \partial^b\mathcal{P}\}
}{n}
}{E_n-E_m} \\
+ 2e\,\mathrm{Im} \negthickspace
\sum_{\begin{subarray}{c}n\,\mathrm{occ}\\m\,\mathrm{unocc}\end{subarray}}
\negthickspace
\frac{ \me{n}{ \partial^i\mathcal{P}}{m}
\me{m}{ \partial H'/\partial B^j }{n}
}{E_n-E_m},
\end{multline}
where the BZ integral and the dependence on ${\bf k}$ have been
suppressed, and $\ket{n} = \ket{u_n}$, etc.
This result
follows immediately upon taking the trace in the basis of energy
eigenstates. Matrix elements of $[\rho_0,r^i]$ appear as
$\not\! r^i_{{\bf k}} = \partial^i\mathcal{P}_k$, from Eq.~\eqref{eq:notr}, and the cross-gap
matrix elements of $\rho'$ in are given in Eq.~\eqref{eq:Lanswer}.
Note that since $\bf{J}_G$ is a total time derivative, $\alpha_G$ is
uniquely defined for a given Hamiltonian (this assumes the existence
of a reference Hamiltonian with $\alpha_G = 0$, that is, the existence
of a topologically trivial, time-reversal-invariant band insulator).
In $\bf{J}_{CS2}$ [Eq.~\eqref{eq:cscurrb}],
we can replace $r \rightarrow [[r,\rho_0],\rho_0]$
in the third commutator.
This has the same cross-gap matrix elements as $r$; the interior
matrix elements do not contribute to the
trace because the other three factors,
$\dot{\rho}_0$ and
two components of $[\rho_0,r^i]$,
have only cross-gap matrix elements.
Then
\begin{equation} \notag
\bf{J}_{CS2} = i \frac{e^2}{2\hslash\Omega} B^j \epsilon_{jab} \mathrm{Tr} [\rho_0,{\bf r}]
[\rho_0,r^a] [[[\rho_0,r^b],\rho_0],\dot{\rho}_0] + \mathrm{c.c.}
\end{equation}
or
\begin{equation} \notag
\bf{J}_{CS2} = -\frac{e^2}{2\hslash} B^j \epsilon_{jab} \mathrm{Tr} \bm{\nabla}_k \mathcal{P}_{{\bf k}} \partial^a \mathcal{P}_{{\bf k}}
[ [\partial^b \mathcal{P}_{{\bf k}},\mathcal{P}_{{\bf k}}],\dot{\mathcal{P}}_{{\bf k}}] +\mathrm{c.c.}, \\
\end{equation}
where an integral over ${\bf k}$ is suppressed for brevity and the trace is taken
in the Hilbert space at ${\bf k}$. Dropping the subscripts
${\bf k}$ everywhere, this can be expanded and rearranged to give
\begin{multline} \tag{\ref{eq:cscurrb}'}
\bf{J}_{CS2} =
\frac{e^2}{2\hslash} B^j \epsilon_{jab} \mathrm{Tr} \mathcal{P} \{
[\bm{\nabla}\mathcal{P},\dot{\mathcal{P}}]\partial^a\mathcal{P}\partial^b\mathcal{P} \\
+2[\dot{\mathcal{P}},\partial^b\mathcal{P}][\bm{\nabla}\mathcal{P},\partial^a\mathcal{P}] \\
+3(\dot{\mathcal{P}}\partial^b\mathcal{P}\partial^a\mathcal{P}\bm{\nabla}\mathcal{P}
+\partial^b\mathcal{P}\dot{\mathcal{P}}\bm{\nabla}\mathcal{P}\partial^a\mathcal{P}) \}.
\end{multline}
In manipulating these strings of projection operators and their derivatives,
it is very useful to realize that derivatives of projectors only have
cross-gap matrix elements: $\mathcal{P} \partial^a\mathcal{P} \mathcal{P} = \mathcal{Q} \partial^a\mathcal{P} \mathcal{Q} = 0$, where
$\mathcal{Q}=1-\mathcal{P}$ is the projector onto unoccupied bands. This means, for example,
that $\mathcal{P} (\bm{\nabla}\mathcal{P})(\dot{\mathcal{P}}) = \mathcal{P}(\bm{\nabla}\mathcal{P})\mathcal{Q}(\dot{\mathcal{P}})\mathcal{P}$.
To $\bf{J}_{CS2}$
we must add $\bf{J}_{CS1}$ [Eq.~\eqref{eq:cscurra}],
\begin{multline} \tag{\ref{eq:cscurra}'}
\bf{J}_{CS1} =
\frac{3e^2}{2\hslash} B^j \epsilon_{jab} \mathrm{Tr}
(\mathcal{P}-\mathcal{Q})[\dot{\mathcal{P}},\bm{\nabla}\mathcal{P}] \partial^a\mathcal{P}\partial^b\mathcal{P} \\
= \frac{3e^2}{2\hslash} B^j \epsilon_{jab} \mathrm{Tr} \mathcal{P} \{
[\dot{\mathcal{P}},\bm{\nabla}\mathcal{P}] \partial^a\mathcal{P}\partial^b\mathcal{P} \\
-(\dot{\mathcal{P}}\partial^b\mathcal{P}\partial^a\mathcal{P}\bm{\nabla}\mathcal{P}+\partial^b\mathcal{P}\dot{\mathcal{P}}\bm{\nabla}\mathcal{P}\partial^a\mathcal{P}) \},
\end{multline}
to get
\begin{align}
\bf{J}_{CS} &= \bf{J}_{CS1} + \bf{J}_{CS2} \notag \\
&= \frac{e^2}{\hslash} B^j \epsilon_{jab} \mathrm{Tr} \mathcal{P} \{
[\dot{\mathcal{P}},\bm{\nabla}\mathcal{P}]\partial^a\mathcal{P}\partial^b\mathcal{P} \notag \\
&\qquad+[\dot{\mathcal{P}},\partial^b\mathcal{P}][\bm{\nabla}\mathcal{P},\partial^a\mathcal{P}] \}.
\end{align}
By checking the different components
explicitly one can see that this is
\begin{multline}
\bf{J}_{CS} =
{\bf B} \frac{e^2}{\hslash} \mathrm{Tr} \mathcal{P} \{ [\dot{\mathcal{P}},\partial^x\mathcal{P}][\partial^y\mathcal{P},\partial^z\mathcal{P}] \\
+ [\dot{\mathcal{P}},\partial^y\mathcal{P}][\partial^z\mathcal{P},\partial^x\mathcal{P}] + [\dot{\mathcal{P}},\partial^z\mathcal{P}][\partial^x\mathcal{P},\partial^y\mathcal{P}] \},
\end{multline}
so we get the ``topological current''
\begin{equation}
\bf{J}_{CS} = -{\bf B} \frac{e^2}{\hslash} \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} \mathrm{tr}
( {\cal F}^{tx}{\cal F}^{yz} + {\cal F}^{ty}{\cal F}^{zx} + {\cal F}^{tz}{\cal F}^{xy} ),
\end{equation}
where the lower-case trace ($\mathrm{tr}$) is only over the occupied bands, and the
Brillouin-zone integral has been restored.
It remains only to show that $\bf{J}_{CS}$ is a total time derivative
that integrates to $\alpha_{CS}{\bf B}$.
Allowing the indices to run over
$t, x, y, z$, in that order (so that $\epsilon_{t x y z} = +1$),
\begin{align}
&\bf{J}_{CS} =-{\bf B} \frac{e^2}{8\hslash} \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3}
\epsilon_{abcd} \mathrm{tr} {\cal F}^{ab}{\cal F}^{cd} \notag \\
&= -{\bf B} \frac{e^2}{2\hslash} \epsilon_{abcd} \int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} \partial^a \mathrm{tr}
\left( \mathcal{A}^b \partial^c \mathcal{A}^d
- i\frac{2}{3} \mathcal{A}^b \mathcal{A}^c \mathcal{A}^d \right).
\end{align}
The derivatives with respect to $k_x,k_y,k_z$ will vanish when integrated over
the Brillouin zone assuming that ${\cal A}$ is defined smoothly and periodically over
the zone, leaving just
\begin{equation} \label{eq:JCSfinal}
\mathbf{J}_{CS} \! = \! -{\bf B} \frac{e^2}{2\hslash} \partial_t \!
\int_{\rm BZ}\!\frac{d^3k}{(2\pi)^3} \epsilon_{abc} \mathrm{tr} \!
\left( \mathcal{A}^a \partial^b \mathcal{A}^c
- i\frac{2}{3} \mathcal{A}^a \mathcal{A}^b \mathcal{A}^c \right),
\end{equation}
where the indices now only run over $xyz$, as originally.
This obviously gives $\alpha_{CS}$ as in Eq.~\eqref{eq:alphaCS}, completing
the proof.
It must be reiterated that this integral is not always entirely trivial. In
particular, if the adiabatic evolution brings the crystal back to its initial
Hamiltonian in a nontrivial way, the Brillouin zone integral need not return
to its initial value because ${\cal A}$ is not uniquely defined. In other words,
$\int dt \mathbf{J}_{CS}$ can be
multivalued as a function of the Hamiltonian deformation parameters. However,
the change can only be such that $\theta$ changes by an integer multiple of
$2\pi$, as discussed in subsection \ref{sec:CS}.
\section{Summary}
The theoretical calculation of the magnetoelectric polarizability in insulators
presents a difficulty similar to that known well from the theory of
polarization; both quantities suffer an inherent ambiguity in the bulk.
The magnetoelectric polarizability
adds another level of difficulty because the vector potential is unbounded and
breaks lattice translation symmetry. However, we have developed a formalism
that allows
us to deal directly with a uniform magnetic field. In the appendix, we further
show that a
long-wavelength regularization of the vector potential together with a suitable
generalization of the polarization (to deal with the broken crystal symmetry)
provides a (relatively) simple, though less rigorous, way to compute the
response function.
The final expression for the
OMP rederives known results for particular model systems and topological
insulators and completes the
picture with additional terms that have a relatively straightforward and
intuitive interpretation. We hope that these results and the method of their
derivation will be valuable for future work on magnetoelectric effects and
topological electronic phases.
The authors gratefully acknowledge useful discussions with
S.~Coh, A.~Malashevich and I.~Souza.
The work was supported by the Western Institute of Nanoelectronics
(AME), DARPA OLE (AMT), NSF DMR-0804413 (JEM), and NSF DMR-0549198 (DV).
|
\section{Introduction}
\begin{figure*}
\vspace{40mm}
\begin{picture}(50,50)
\put(-210,-30){\special{psfile=f1a.eps hoffset=0 voffset=0 hscale=43 vscale=43 angle=0}}
\put(50,-63){\special{psfile=f1b.eps hoffset=0 voffset=0 hscale=35 vscale=35 angle=0}}
\end{picture}
\caption{Left: 1420 MHz continuum image of CTB 109 and the near HII region Sh 152. Right: CO emission at $-$51 km s$^{-1}$ with contours (8, 15, 21, 29, 100 K) of the continuum emission. The left panel shows boxes (both green and red) used for extraction of HI spectra (see text) and centered at [$l=108.80^\circ$,$b=-0.96^\circ$] for box 1, [$108.76^\circ$,$-0.98^\circ$] for box 2, [$109.43^\circ$,$-1.36^\circ$] for box 3, [$108.76^\circ$,$-0.96^\circ$], [$108.84^\circ$,$-1.02^\circ$] for box 4, [$108.77^\circ$,$-0.95^\circ$] for Sh 152 and [$108.93^\circ$,$-1.22^\circ$] for CTB 109, respectively}
\end{figure*}
The Supernova Remnant (SNR) CTB 109 hosts an anomalous X-ray
pulsar (AXP) 1E 2259+586 with a superstrong magnetic field, i.e. a magnetar. 1E 2259+586 emitted a Soft Gamma-ray Repeater-like outburst in 2002 \citep{Kaset03}, so CTB 109 is an especially interesting target for further study. CTB 109 is believed to strongly interact with the large molecular cloud ($l$=108.77$^\circ$, $b$=-0.95$^\circ$) adjacent to its west side, based on its semi-circular shape in both X-rays and radio \citep{Tatet90, Coeet89}. CTB 109 has an extended X-ray bright region in the interior, which is thermal in origin \citep*{Rhoet97} and has no morphological correlation with AXP 1E 2259+586 \citep{Patet01}. The lobe likely originates from an interaction of the SNR shock with a cloud at $-$55 km s$^{-1}$ \citep*{Saset04, Saset06}.
Distance measurement to a SNR/AXP system is extremely important but usually quite difficult. Many efforts have previously been aimed at distance determination of SNR/pulsar systems \citep{Camet06, Gotet09, Leaet08, McCet05, Rivet10, Suet09, Tiaet06, Tiaet08a, Weiet08}. \citet[K02 hereafter]{Kotet02} analyzed HI absorption and emission spectra of CTB 109 and near compact radio sources, and estimated a distance of 3.0 kpc to CTB 109.
On the other hand, \citet*[DvK06 hereafter]{Duret06} used
the red clump stars method to measure distances to AXPs.
This method allows deriving the function of reddening versus distance in any given line of sight, based on field stars over a relatively small area. DvK06 excluded a distance of 3 kpc for AXP 1E 2259+586 based on its high extinction, instead preferring a much larger value of 6 kpc.
The conflict between these two distances
motivates us to re-analyze distance evidence for CTB 109 using 1420 MHz continuum, 21 cm HI-line and $^{12}$CO data.
We provide a revised distance to CTB 109 by employing improved HI spectral analysis methods, and by analyzing possible HI Narrow Self-Absorption (HINSA) and HI Self-Absorption (HISA) from the adjacent molecular cloud complex. Based on similar considerations, the methods have recently been used to measure kinematic distances to Galactic molecular clouds \citep{Romet09, Krcet08}.
We compare our new results to the previous conflicting distance estimates of K02 and DvK06.
\section{Results and Analysis}
We show the 1420 MHz continuum image (left) and the CO image at $v=-$51 km s$^{-1}$ (right) in Figure 1, both centered on CTB 109.
As suggested by previous studies, there is a good correlation between CTB 109 and CO features in the velocity range $-$50 to $-$57 km s$^{-1}$ (see Fig. 4 of K02), but no respective convincing HI features in the velocity range associated with CTB 109.
\begin{figure}
\vspace{105mm}
\begin{picture}(80,80)
\put(-10,180){\special{psfile=f2a.eps hoffset=0 voffset=0 hscale=45 vscale=40 angle=0}}
\put(-10,-15){\special{psfile=f2b.eps hoffset=0 voffset=0 hscale=45 vscale=40 angle=0}}
\end{picture}
\caption{HI and $^{12}$CO spectra of CTB 109 and Sh 152}
\end{figure}
\subsection{HI absorption}
We construct HI emission and absorption spectra for CTB 109 and for Sh 152.
Since CTB 109 is extended rather than a point source, the standard formula for HI absorption spectra does not apply. The continuum emission from CTB 109 extends into our background region because the background region is chosen to be near the continuum peak. This minimizes the difference in the HI distribution along the two lines of sight (source and background, see \citet*{Tiaet08b}.
The resulting difference in brightness temperature is given by: $\Delta T$ = $T_{off}(\it{v})$ - $T_{on}(\it{v})$ = ($T^{c}_{s}$-$T^{c}_{bg}$)(1-$e^{-\tau_{c}(\it{v})}$).
Here $T_{on}(\it{v})$ and $T_{off}(\it{v})$ are the average HI brightness temperatures of spectra from a selected area on a strong continuum emission region of the source and an adjacent background region which contains weaker continuum emission. $T^{c}_{s}$ and $T^{c}_{bg}$ are the average continuum brightness temperatures for the same regions respectively. $\tau_{c}(\it{v})$ is the HI optical depth from the continuum source to the observer.
The regions used for source and background spectra are shown by the solid-line and dashed-line boxes (red; the background area excludes the source area) in the left panel of Fig. 1. Fig. 2 gives the HI and $^{12}$CO spectra.
The spectrum from the south filament of CTB 109 (top in Fig. 2) shows a highest HI absorption velocity of $-$56 km s$^{-1}$. The HI absorption spectrum from Sh 152 (botton) is the same shape as that given by K02, and shows a highest absorption velocity of $-$65 km s$^{-1}$.
Sh 152 sits at the peak of a large high brightness-temperature CO molecular cloud at velocities of $-$48 to$ -$56 km s$^{-1}$ (see Fig. 4 of K02, and bottom in Fig. 2). There is clear HI absorption against Sh 152 in the same velocity range. This means cold HI gas surrounding the CO cloud
is responsible for the HI absorption and is located in front of Sh 152.
The CO spectrum of CTB 109 shows narrow CO emission at $-$48 to $-$54 km s$^{-1}$ in the direction of the southern filament. The fact that the highest radial velocities of HI absorption features towards both CTB 109 ($-$56 km s$^{-1}$) and Sh 152 ($-$65 km s$^{-1}$) are larger than the recombination line velocity ($-$50 km s$^{-1}$, \citet{Loc89} of Sh 152 confirms the existence of a velocity reversal within the Perseus arm. This reversal explains why the HI gas at higher radial velocity is in front of Sh 152 which is at lower radial velocity.
Based on Fig. 2, the HI column densities integrating over all velocity in front of CTB 109 and Sh 152 are N$^{total}_{HI}$ $\sim$ 4 and 6 $\times$10$^{19}$$T_{s}$ cm$^{-2}$ respectively (using N$_{HI}$= 1.82$\times$10$^{18}\tau\Delta{v}T_{s}$, \citet*{Dicet90}). For CTB 109, the column density of HI from $-$48 to $-$54 km s$^{-1}$ is $\sim$3$\times$10$^{19}$$T_{s}$ cm$^{-2}$ (3$\times$10$^{21}$ cm$^{-2}$ if $T_{s}$$\sim$ 100K).
\subsection{HI self-absorption and narrow self-absorption}
HI self-absorption (HISA) is generally produced by temperature fluctuations in the cold neutral medium (CNM). HISA features in the near Perseus arm have been widely revealed in the CGPS \citep{Gibet05} due to the above-mentioned velocity reversal caused by spiral density waves \citep{Rob72}. A particular kind of HI self-absorption, referred to as HI narrow self-absorption (HINSA, \citet*{Liet03}) is produced by cold HI mixed with molecular clouds and cooled by its molecular environment. HINSA is shown to share the spatial and kinematic structure of cold molecular clouds traced by CO emission lines \citep{Krcet08}. HINSA can only be detected in specific molecular cloud regions which have narrower line width (1 to 2 km s$^{-1}$) than background HI emission from CNM components. HINSA bears special relevance to this work because the originating location of the cold HI can be constrained by studying CO clouds at the same velocity given their co-location.
The CO image and spectrum in the direction of Sh 152 show a large molecular cloud complex in the velocity range of $-$48 to $-$56 km s$^{-1}$.
Part of the CO cloud interacting with CTB 109 has been believed to be a major factor causing CTB 109 to have an asymmetric semi-circular shape structure, i.e. a relatively bright eastern part, well-defined in radio and X-ray images, and no emission in the west. Interaction of the SNR shock with the part of this molecular cloud complex at $-55$ km s$^{-1}$ has been observed \citep{Saset06}. HINSA could be produced if these CO clouds are in the front of warm HI background.
\begin{figure}
\vspace{140mm}
\begin{picture}(80,80)
\put(0,350){\special{psfile=f3a.eps hoffset=0 voffset=0 hscale=40 vscale=26 angle=0}}
\put(0,230){\special{psfile=f3b.eps hoffset=0 voffset=0 hscale=40 vscale=26 angle=0}}
\put(0,110){\special{psfile=f3c.eps hoffset=0 voffset=0 hscale=40 vscale=26 angle=0}}
\put(0,-10){\special{psfile=f3d.eps hoffset=0 voffset=0 hscale=40 vscale=26 angle=0}}
\end{picture}
\caption{HI and $^{12}$CO emission spectra from boxes 1 - 4}
\end{figure}
We searched the HI channel maps and HI emission spectra over the large region covered by the CO cloud complex, and found no HISA whose HI spectra show dips of $\ge$ 10 K with the FWHM $\Delta{v}$ of the absorption line 5 km s$^{-1}$. This is consistent with the non-detection of HISA features in the region in the CGPS HISA census \citep{Gibet05}. As an example, Fig. 3 shows the HI emission spectra and respective CO spectra extracted from boxes 1 to 4 shown in Fig. 1 (green). The top two rows of Fig. 3 clearly show no HISA features from the CO cloud at $-$51$\pm$5 km s$^{-1}$ against the warm HI background (T$_{HI}$$\sim$ 100K in the range of $-$40 to $-$60 km s$^{-1}$).
In contrast, a HISA feature from box 3 nearby CTB 109 is seen (the third row).
We found one area of the CO cloud complex showing narrow CO lines and still bright enough CO emission (at $l$=108.85$^\circ$ and $b$=-1.02$^\circ$ that is labelled "4" in Fig. 1) to satisfy the conditions needed to produce HINSA. The CO emission spectrum from box 4 shown in the bottom of Fig. 3 has a narrow line width of 2 km s$^{-1}$. The respective HI emission spectrum shows no HINSA features.
\section{Discussion and Conclusion}
\subsection{The velocity/distance relation within the Perseus arm}
21 cm HI observations can provide distance estimates
to Galactic objects using the radial velocity-distance relation based on the flat circular rotation curves model. However, this rotation model may lead to serious overestimation of an objects' distance, especially at the Perseus arm where the velocity field is strongly influenced by the spiral arm shock \citep{Rob72}.
The velocity reversal inside the Perseus arm is caused by a spiral shock, while in other parts of the outer Galaxy the radial velocity decreases monotonically with distance.
This velocity reversal creates a distance ambiguity for objects with radial velocities of $-$50 km s$^{-1}$ to $-$60 km s$^{-1}$ in the line-of-sight to CTB 109 \citep[TLF07 hereafter]{Tiaet07}.
We use the two-arm HI model of \citet*{Foset06} (FM06 hereafter), which is an updated version of the \citet{Rob72} model. FM06 fit an empirical model of Galactic structure and density-wave motions to observations.
A comparison between photometric and the HI model distances to 22 objects (19 HII regions and 3 SNRs), shows individual point errors of 0.3 to 0.8 kpc (Fig. 8 of FM06). The 20\% error quoted in FM06 is dominated by systematic errors. The velocity of -55km s$^{-1}$ for CTB 109 gives 3.2 kpc in FM06 model if it is at the near side of the velocity reversal (Fig. 6 of TLF07). Since our new result is that CTB 109 is beyond the velocity reversal, the FM06 model gives 4.0 kpc distance. The difference in distance is 0.8 kpc. The error in this value is dominated by the systematic error of 0.16 kpc.
Because of the velocity reversal in the Perseus arm, HISA and HINSA should be produced if the cold CO clouds associated with CTB 109 are in front of the warm HI background, i.e. at distance of 3.2 kpc.
Absence of any HISA and HINSA features indicate that the cold CO cloud is at the far distance of 4 kpc in order to be behind the warm HI at the same velocity. We analyze this quantitatively below.
\subsection{Is the amount of cold HI gas in the molecular cloud enough to produce significant HISA or HINSA against warm HI background?}
First we consider the conditions for HISA.
The intensity of HISA is given by $\Delta T_{HISA}$ = ($T_{B}(v)-T_{s}$)(1-$e^{-\tau}$).
With a normal background HI temperature of about 100k, to produce a measurable HISA signal
at about 10 K level \citep*{Andet09} requires an optical depth of HISA above 0.1.
Such optical depth is produced by a cold HI column of 1$\times$10$^{20}$ cm$^{-2}$ according to N$_{HI}$=1.82$\times$10$^{19}$$T_{s}$$\tau$$\Delta{v}$,
assuming the cold gas has a velocity dispersion of 5 km s$^{-1}$. The continuum brightness temperature of the southern bright spot of CTB109 is 30K, so the absorption feature at $-$51 km s$^{-1}$ should be produced by cold HI clouds at $-$51 km s$^{-1}$ (with $T_{s}$$\le$ 30K). From the absorption spectra, the column density of the HI absorbing clouds is 3$\times$10$^{19}$$T_{s}$ cm$^{-2}$, i.e. 3$\times$10$^{20}$ cm$^{-2}$ when $T_{s}$=10 K. This is more than enough to produce measurable HISA if the cloud is in front of the warm HI emission (100 K).
Next we calculate the optical depth of an HI absorption line which is required to produce a HINSA feature with peak intensity of 3 $\sigma$. The rms of the HI emission data is about 1.7 K per channel calculated from the HI spectrum of G108.77-0.95 (Fig. 2). \citet{Tayet03} give that the rms noise level in brightness temperature for an empty channel ranges from 2.1 K to 3.2 K throughout the survey.
We take the value of 2.1 K here. Based on a three component radiative transfer calculation and the assumption that the HI emission is optically thin, we have (Equation 8 of \citet*{Liet03},
\begin{equation}
T_{ab}= [pT_{HI} + T_c - T_x](1-e^{-\tau}) \>\> ,
\end{equation}
where the depth of absorption $T_{ab}$ is determined by the fraction of HI in the background $p$, the emission temperature $T_{HI}$, the continuum background $T_c$, and the cold HI excitation temperature $T_x$. Assuming a uniform Galactic HI disk with a 17 kpc radius, it is estimated that $p=0.815$ toward the direction of CTB 109. Using the reasonable assumption of $T_c = 3.5 K$ (including CMB and interstellar radiation field see \citet*{Liet03}), $T_{HI}=100 K$ and $T_x = 10 K$, an optically depth of $\tau = 0.084$ is required to produce an absorption line of 6.3 K peak depth (3 $\sigma$). For a FWHM line-width of 2.0 km s$^{-1}$ of HINSA, this corresponds to an upper limit of the column density of cold HI $N_{HI}$ = 3.2$\times10^{18}$ cm$^{-2}$.
Theoretically, the abundance of atomic HI inside molecular clouds is in the range of 0.05 to 0.27\% \citep*{Golet05}. Observationally, using an improved technique to measure the ratio, \citet{Krcet08} obtained a precise value for two clouds, i.e. 0.09\% for L134 and 0.53\% for L1757. Taking the theoretical mean HI/H$_{2}$ ratio of 0.16\%, the non-detection of HINSA toward the large molecular cloud interacting with CTB 109 puts a 3$\sigma$ upper limit on the cloud $H_2$ column density at 2.0$\times 10^{21}$cm$^{-2}$.
The $^{12}$CO spectrum from box 4 (Fig. 3) gives W$_{^{12}CO}$ = 12 $K km s^{-1}$ for the cloud. The $^{12}$CO to H$_{2}$ conversion factor $X$ is in the range of 2.2 to 2.8$\times$ 10$^{20} [cm^{-2}/(K km s^{-1}$)] based on three calibration techniques \citep*{Solet91}. This gives an H$_{2}$ column density of 3 $\times$10$^{21}$ cm$^{-2}$ for this cloud ($X$ = 2.5$\times$10$^{20}$ cm$^{-2}/(K km s^{-1}$)). This is enough to produce measurable HINSA at 4.5$\sigma$ level if the cloud is in front of the warm HI at the same velocity.
Fig. 3 (third row) shows a good example that a HISA feature at $-$50 km s$^{-1}$ is clearly detected nearby CTB 109. The CO cloud at $-$50 km s$^{-1}$ has low brightness temperature of 0.3 K and shows no physical relation with the remnant. The W$_{^{12}CO}$ from the cloud is about 1 K km/s so it has an H$_{2}$ column density of 2.5 $\times$10$^{20} cm^{-2}$. This is about 30 times lower than that from the cloud complex associated with Sh 152. This indicates that the cold HI cloud at $-$50 km s$^{-1}$ is likely responsible for the absorption feature.
\subsection{Distance of 4 kpc to CTB 109/AXP 1E 2259+586}
The non-detection of both HISA and HINSA leads to the conclusion that the cloud at $-$55 km s$^{-1}$ is behind the warm HI and thus CTB 109, interacting with the cloud, is behind the velocity reversal in the Perseus arm.
As noted above, the HI model distance of 4 kpc has a uncertainty of 20\% (i.e. 0.8 kpc). \citet*{Braet93} have studied the observed velocity field of the outer galaxy and found residuals from circular motions. They found that these velocity residuals are consistent with streams motions due to spiral arms, supporting the Roberts (and thus FM06) models.
K02 estimated a distance for CTB 109 of 3$\pm$0.5 kpc based on two arguments. First, the HI absorption at $-$45 km s$^{-1}$ in the extragalactic source is not seen in the spectra of CTB 109 or Sh 152, so they argued CTB 109 and Sh 152 must be at the nearer distance in the spiral shock region. However, the extragalactic source is not near to CTB 109 (outside of the area shown in Fig.1) and thus spatial variations in the HI distribution can result in lack of significant HI at $-$45 km s$^{-1}$ in front of CTB 109 and invalidate that conclusion.
Secondly they considered the spectroscopic distances and radial velocities of 11 HII regions which in the Perseus arm nearby on the sky to CTB 109. These HII regions have an average distances of 3 kpc. However the two HII regions, Sh 152 ([108.77,-0.95]) and Sh 149 ([108.34, -1.12]), closest to CTB 109 on the sky have spectroscopic distances of 3.6$\pm$1.1 kpc and 5.4$\pm$1.7 kpc, respectively. This does not favour the small distance to CTB 109 of 3 kpc. These two HII regions have respective radial velocities of $-$50.4$\pm$0.5 km s$^{-1}$ (or $-$49.1$\pm$0.9 km s$^{-1}$ from \citet{Loc89} and $-$53.1$\pm$1.3 km s$^{-1}$ from \citet*{Braet93} which are similar to the radial velocity ($-$55 km s$^{-1}$) of the molecular cloud complex with which CTB 109 is interacting. This also support the larger distance for CTB 109 of 4.0 kpc.
CTB 109 is host to AXP 1E 2259+586. DvK06 excluded small distance of 3 kpc for 1E 2259+586 due to absence of red clump stars sufficiently highly reddened to be consistent with 1E 2259+589's column density. There are two jumps (at $A_{V}$ 3 and 8) in their Fig. 7 which shows a redding--distance relation along the line-of-sight to AXP 1E 2259+586. They favour a large distance of 6 kpc for the AXP based on the large extinction of 8. The HI absorption spectrum of CTB 109 (Fig. 2) excludes a large distance. Eg for 6 kpc the radial velocity of FM06 model is about $-$80 km s$^{-1}$, but no absorption is seen beyond $-$56km s$^{-1}$. Our distance measurement of 4.0 kpc to AXP 1E 2259+586 puts the AXP at the far edge of the Perseus arm and would give a small extinction of $A_{V} \sim$ 3 for the AXP. The reddening stays at $A_{V} \sim$ 3 out to about 6 kpc (Fig. 7 of DvK06), implying excess local extinction for the AXP at 4 kpc. Using N$_{H}$ = $A_{V}$ $\times$ 2.3$\times$10$^{21}$ cm$^{-2}$, with inferred excess $\Delta{A_{V}}$ = 5, gives excess $\Delta{N_{H}}$ $\approx$ 10$^{22}$ cm$^{-2}$. This is quite plausible given that AXPs should be produced in SN explosions of massive stars and these SN also are known to produce large amount of dust in their ejecta.
The new distance of 4 kpc to the AXP leads to a larger explosion energy of 1.3$\times$10$^{51}$ ergs than that based on 3 kpc \citep*{Vinet06}. However this is not a large enough increase to argue that CTB 109 had an unusually large explosion energy. The same situation occurs for SNR Kes 73/AXP 1E 1841-045 system \citep*{Tiaet08b}. Both cases have larger distances than previously estimated but still give normal explosion energies (10$^{51}$ ergs). This adds support to \citet*{Vinet06}'s argument against the possibility that the magnetars are formed from rapidly rotating (less than a few millisecond) proto-neutron stars. We have applied a new technique for determining the distance to CTB 109, i.e. absence or presence of HISA and HINSA. This is a new useful tool for finding distances to other objects in the Galactic plane.
\section{Acknowledgments}
WWT acknowledges support of the CBP and the NSFC. DAL thanks support from the Natural Sciences and Engineering Research Council of Canada. Di Li's work is supported by the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. The DRAO is operated as a national facility by the National Research Council of Canada.
|
\section{WAS EINSTEIN'S KINEMATICS INCOMPLETE?}
Einstein, in his pioneering paper \cite{EinsteinA1905}, insisted that when talking about motion, we must give a \emph{physical} meaning to ``time''. He showed how this can be done by introducing his idea of ``synchronized clocks''. However, he took for granted the concept of ``position''. He wrote:
\begin{quote} Let us take a system of co-ordinates in which the equations of Newtonian mechanics hold good. \ldots If a material point is at rest relatively to this system of co-ordinates, its position can be defined relatively thereto by the employment of rigid standards of measurement and the methods of Euclidean geometry, and can be expressed in Cartesian co-ordinates.
\end{quote}
Thus, he did not describe how position could be defined in a general, physically meaningful way. Nobody subsequently has done so. A little further in \cite{EinsteinA1905}, he said:
\begin{quote}
Let us in ``stationary'' space take two systems of co-ordinates, i.e.,two systems, each of three rigid material lines, perpendicular to one another and issuing from a point.
\end{quote}
Are the co-ordinate axes, then, material bodies? Do we need to think of what might happen to them when they move? Or are they merely conceptual? In a sense, then, Einstein's Kinematics was incomplete.
\section{EINSTEIN'S KINEMATICS COMPLETED}
In a recent paper \cite{Agashe1}, we suggested how Einstein's Kinematics could be completed. We proposed that an ``observation system'' could have the following ingredients. An observer, $S$, say, capable of sending and receiving light signals, is equipped with a \emph{single} clock and \emph{three} passive ``reflecting'' stations, say, $S_1,S_2,S_3$. Using a radar-like approach, the observer could obtain data sets as follows. He sends a signal in all directions at a time $t_0$ in his clock, and then records the times of arrivals of the echoes of this signal by reflection in the following four different ways: time $t_0'$ of arrival after reflection at the place P, say, of an event being observed (path $SPS$); time $t_1$ of arrival after reflection at the event P first, followed by a reflection at the station $S1$ back to $S$ (path $SPS_1S$); similarly, time instants $t_2$,$t_3$. Thus, he would obtain 5-tuples of data items of time, ${t_0,t_0',t_1,t_2,t_3}$. From each 5-tuple, he is to decide \emph{by definition} what could be meaningfully called the ``time of occurrence'' and ``place'' of the event. Following Einstein, we chose $t=\frac{t_0+t_0'}{2}$ as the \emph{definition} of the time of occurrence.
There remained the problem of deciding what could meaningfully be called the place of the event. Here, we proposed to go beyond the classical 3-dimensional rectangular Cartesian co-ordinate system idea which was only conceptual; we chose instead to think of the place of an event as an element of a 3-dimensional vector space to be equipped with a suitable scalar or inner product. The place of an event could thus be thought of as a ``position vector''. Any 3-dimensional vector space, $V$, say, would do. (Today, we know the possible advantages of such ``abstract'' \emph{representation}.) Since the reflecting stations deserved ``places'' of their own, it was natural to represent them by vectors $s_1,s_2,s_3$, say, forming a basis of $V$. Again, any basis would do. Of course, $S$ itself would be assigned the zero vector. Now, the reflecting stations could not be allowed to be totally arbitrary; they had to remain at fixed ``distances'' from $S$ and from one another. But what are distances? These had to be physically determinable. $S$ has only a clock- no measuring rods. As in radar, one could define ``distance'' in terms of \emph{time interval} through a parameter called the velocity of light, $c$. By doing some further signalling, involving various reflections, $S$ could obtain a set of 6 time intervals that would correspond to 6 transition times, $SS_1,SS_2,SS_3,S_1S_2, S_2S_3,S_3S_1$. These multiplied by $c$ would be taken as the lengths or ``norms'' of the 6 vectors $s_1,s_2,s_3,s_1-s_2,s_2-s_3,s_3-s_1$. These 6 numbers would then uniquely determine the scalar or inner product on $V$, thus making it into an inner product space. Note that although $V$ and a basis for it were arbitrarily chosen, the scalar product was determined by the observation system itself. The problem of obtaining from the 5-tuple of an event a representing vector $p$, say, in $V$ was then a problem of linear algebra (see\cite{Agashe1} for details, with a slightly different notation). There was a ``technical'' hitch, however. Not any set of 6 numbers would do; this was explored in an ``addendum''\cite{Agashe2}.
The stage was set to admit another observation system, an observer $S'$, say, with a clock and reflecting stations $S_1',S_2',S_3'$. This observer could choose a vector space, $V'$, say, not necessarily the same as $V$ of $S$, basis vectors $s_1',s_2',s_3'$ to represent its stations, and using the same ``velocity of light'' constant $c$, determine a scalar product on it, and finally obtain the representing vector $p'$, say, and time $t'$, say, of the same event for which $S$ had obtained $p$ as position vector and $t$ as time. To relate $p,t$ with $p',t'$, one assumed, as usual, that the system $S'$ was in uniform motion relative to system $S$ with velocity $v$. This motion would be observed by $S$, and thus, $v$ would be a vector in $V$. $S$ would observe the motions of $S',S_1',S_2',S_3'$ to be given by the vectors $d_0+tv,d_0+tv+d_1,d_0+tv+d_2,d_0+tv+d_3$, say, $d_0,d_1,d_2,d_3$ being all vectors in $V$. To go further, one needed some relation between the clocks of $S$ and $S'$ in the following sense. Suppose that as $S'$ moves, the clocks of $S$ and $S'$ \emph{at} $S'$ show values $t$ and $t'$, respectively. We need some relation between these two ``times''. We assume, with Einstein, linearity of this relation: $t' = \beta_1t$, where $\beta_1$ is some constant. This is the only assumption of linearity that we make. We then prove(see\cite{Agashe1}) that the following linear relations hold between the times and places of the events in the two systems:
\begin{equation}
t' = \beta_1\left[t-\frac{(p-d_0-tv,v)_S}{c^2 - v^2}\right]
\end{equation}
where the symbol $(u,w)_S$ denotes the scalar product of the vectors $u$ and $w$ in $V$, and
\begin{equation}
p' = T (p-(d_0+tv))
\end{equation}
where $T$ is a linear transformation on $V$ onto $V'$ such that it maps each vector $d_i$ of $V$ to the vector $s_i'$ of $V'$, i.e., the vectors in $V$ representing the relative positions of the stations of $S'$ are mapped to the vectors in $V'$ representing the stations of $S'$. This completes a summary of our derivation of the Lorentz transformation in\cite{Agashe1}. We call it ``Einstein's Lorentz transformation'' because we have followed an Einsteinian approach - except with respect to the meaning of ``position''.
\section{ LORENTZ TRANSFORMATION IN GALILEAN FORM}
Although we allowed the possibility that the representation vector spaces $V$ and $V'$ could be different, they could be chosen to be the same. Further, the vectors $d_i$ representing the relative positions in $V$ of the stations of $S'$ could be chosen to be the basis vectors for $S'$. Thus, we could choose $s_i' = d_i$. Of course, the scalar products could be different, as they are dictated by the observational data. The transformation $T$ then becomes the identity transformation, and for the vectors representing the place of the event in $S$ and $S'$, we obtain the following simple \emph{Galilean} relation:
\begin{equation}
p' = p-(d_0 +tv)
\end{equation}
which can be seen as the Galilean position vector of $P$ relative to $S'$. It must be emphasized, however, that there are still two representations because there are possibly two different scalar products for $S$ and $S'$, and these scalar products relate to two different calculations of distances in the two systems. Both are \emph{Euclidean} in the sense they are both based on a scalar product. The normal ``Euclidean'' co-ordinate systems use the distance $\sqrt{x^2 + y^2 + z^2}$, which is related to a special scalar product.
Let us consider in this context the Einstein form of the Lorentz equations:
\begin{equation}
\label{NLT}
x' = \beta (x - vt), \qquad
y' = y, \qquad
z' = z, \qquad
t' = \beta (t-\frac{vx}{c^2}).
\end{equation}
One could argue that one could put them in the Galilean form by using new variables $\bar{x}, \bar{y}, \bar{z}$:%
\begin{equation}
\bar{x} = x -vt, \qquad
\bar{y}=y, \qquad
\bar{z} =z
\end{equation}
and explicitly defining a new distance for $S'$ given in terms of norm by
\begin{equation}
{||(p,q,r)|| _{S'}}^2 = \beta^2 p^2 + q^2 + r^2 ,
\end{equation}
choosing the constant $\beta_1$ equal to $1/ \beta$, but this would have looked like a mathematical ``trick''. In contrast, here we have envisaged the possibility of a different scalar product and distance in the very notion of representation by a vector.
\section{REMARKS}
The Galilean form for the position vector has the advantage that the Einsteinian factor $\beta$ does not appear in it, making manipulations easier. Also, length contraction and time dilatation disappear. However, $\beta$ does remain in a slightly different appearance in the equation relating the times. The additional factor $\beta_1$ could be chosen as unity. The main point of this paper is, however, that the concept of ``position'' has to be defined and that there is a choice in representing position. The present paper could be read as a postscript to the earlier papers \cite{Agashe1} and \cite{Agashe2}.
It would be interesting to apply the coordinate-free vector representation of position to Maxwell's equations and to see how the Galilean form of the Lorentz transformation works out.
|
\section{Introduction}
Denote by $\ga$ the normalised Gaussian measure on $\BR^d$ having
density $x\mapsto\pi^{-d/2}\, \e^{-\mod{x}^2}$
with respect to the Lebesgue measure.
The \OU\ operator $\cL$ is the closure in $\ld{\ga}$ of the operator $\cL_0$,
defined on $C^\infty_c(\BR^n)$ by
$$
\cL_0
= -\frac{1}{2}\, \Delta + x\cdot\nabla,
$$
where $\Delta$ and $\nabla$ denote the Euclidean
Laplacian and gradient, respectively.
The operator $\cL$ is self-adjoint, and its
spectral resolution is
$$
\cL
= \sum^{\infty}_{j=0} j\, \cP_j,
$$
where $\cP_j$ denotes the orthogonal projection onto the linear span
of Hermite polynomials of degree~$j$ in $d$ variables.
The \OU\ operator generates a diffusion semigroup,
which has been the object of many investigations
during the last two decades.
In particular, efforts have been made
to study operators related to the \OU\ semigroup, with emphasis
on maximal operators \cite{S,GU,MPS, GMMST2},
Riesz transforms
\cite{Mu, Gun, M, Pisier, Pe, Gut, GST, FGS, FoS, GMST1, PS, U, DV}
and functional calculus \cite{GMST2,GMMST1,MMS1,HMM}.
In this paper, we shall concentrate on the first-order Riesz transforms
associated with $\cL$. Since the eigenspace associated to the
zero eigenvalue is nontrivial,
the positive square root of $\cL$ is not invertible, and one must
define the Riesz transforms carefully.
Consider the sequence $M:\BN\rightarrow \BC$, defined by
$$
M(j)=
\begin{cases}
0 & {\rm if}\ j=0\\
j^{-1/2} & {\rm if}\ j=1,2,\ldots
\end{cases}
$$
The operator $M(\cL)$, spectrally defined on $\ld{\ga}$,
extends to a bounded operator on $\lp{\ga}$ for every $p$ in $(1,\infty)$
(see, for instance, \cite[Theorem~1.2]{MMS1}).
The operator $M(\cL)$ plays, in this setting, the
same role as the operator $(-\Delta)^{-1/2}$
in Euclidean harmonic analysis.
We denote by $\partial_i$ the differentiation operator with respect to
the variable $x_i$. The formal adjoint of $\partial_i$
in $\ld{\ga}$ is the operator $\partial_i^* = 2x_i - \partial_i$.
Notice that, at least formally,
$$
\sum_{i=1}^d \partial_i^* \partial_i = 2\, \cL.
$$
For each $i$ in $\{1,\ldots,d\}$,
we define the operators $R_i$, $S_i$, $R_i^*$ and $S_i^*$
on finite linear combinations of Hermite polynomials by
\begin{equation} \label{f: def Riesz}
\begin{array}{ll}
R_i
= \partial_i M(\cL)
& \qquad S_i
= M(\cL) \partial_i \\
R_i^*
= M(\cL) \partial_i^*
& \qquad S_i^*
= \partial_i^* M(\cL).
\end{array}
\end{equation}
It is straightforward to check that $R_i^*$ and $S_i^*$
are the \emph{formal} adjoints in $\ld{\ga}$ of $R_i$ and $S_i$,
respectively.
Recall the action of $\partial_i$ and $\partial_i^*$
on Hermite polynomials:
$$
\partial_i H_n
= 2n\, H_{n-1}
\qquad\hbox{and}\qquad
\partial_i^* H_n
= H_{n+1},
$$
where $H_j$ denotes the Hermite polynomials of degree $j$
in the $i^{\small\textrm{th}}$ variable.
A straightforward argument, which uses these formulae and
the expression for the $\ld{\ga}$ norm of a Hermite polynomial,
then shows that $R_i$, $R_i^*$, $S_i$ and $S_i^*$ extend to
bounded operators on $\ld{\ga}$, and that $R_i^*$ and $S_i^*$
are the Hilbert space adjoints of $R_i$ and $S_i$,
respectively.
Note that, in contrast with the Euclidean case, these transforms are
not antisymmetric.
The operators $R_i$, $S_i$, $R_i^*$
and $S_i^*$ are all bounded on $\lp{\ga}$ for each $p$ in $(1,\infty)$
(see \cite{M}).
This may be easily proved by using the commutation relations
between $\partial_i$ and $\cL$ and the spectral multiplier
result \cite[Theorem~1.2]{MMS1}.
It is straightforward to check that none of these operators
are bounded on $\lu{\ga}$ or on $\ly{\ga}$.
Weak type $1$ estimates for $R_i$ and for $S_i^*$ have been proved in \cite{FGS}
and \cite{AFS}, respectively.
Thus, it is natural to ask what further estimates
these operators satisfy at the endpoints $p=1$ and $p=\infty$.
In the Euclidean case there are
substitute results at $p=1$ and $p=\infty$ saying that
the operators are bounded from $\hu{\BR^d}$ to $\lu{\BR^d}$ and
from $\ly{\BR^d}$ to $\BMO{\BR^d}$.
In our setting, the spaces
$\hu{\ga}$ and $\BMO{\ga}$ were defined by Mauceri and Meda
in \cite{MM}, where these authors also developed a theory of singular
integral operators in the Gaussian setting (see also \cite{CMM}
for a related theory in a more general setting). As
applications, they proved that the imaginary
powers of the \OU\ operator are bounded from $\hu{\ga}$
to $\lu{\ga}$ and from $\ly{\ga}$ to $\BMO{\ga}$, and that
the operators $R_i$ are bounded from $\ly{\ga}$ to $\BMO{\ga}$.
The question to be studied below is whether the
Riesz transforms defined in (\ref{f: def Riesz})
are bounded from $\hu{\ga}$ to $\lu{\ga}$ and
from $\ly{\ga}$ to $\BMO{\ga}$.
We find that this boundedness always holds in dimension one.
But, surprisingly, in higher dimensions, each Riesz
transform has one of these boundedness properties but not the other.
Our results for $d\geq 2$ are summarised in the table
in Theorem~\ref{t: main}.
The paper is organised as follows. Section 2 contains
background material and a few preliminary results,
including the expressions for the kernels of $R_i$ and $S_i^*$.
Our main result, Theorem~\ref{t: main}, is stated
in Section~\ref{s: TMR}. Its proof constitutes Sections~\ref{R}-\ref{D}.
\section{Preliminaries}
We briefly recall the Hardy space $\hu{\ga}$ from \cite{MM}.
A Euclidean ball $B$ is called \emph{admissible} if
\begin{equation} \label{f: 1-admissible}
r_B \leq \min\bigl(1,1/\mod{c_B}\bigr);
\end{equation}
here and in the sequel
$r_B$ and $c_B$ denote the radius and the centre of $B$, respectively.
The collection of all admissible balls will be denoted by $\cB_1$.
When there is equality in (\ref{f: 1-admissible}), we say that $B$
is a \emph{maximal ball} in $\cB_1$.
A (Gaussian) \emph{atom} is either the constant function $1$ or
a function $a$ in $\ly{\ga}$,
supported in an admissible ball $B$ and such that
\begin{equation} \label{f: prop atom}
\norm{a}{\infty} \leq \ga(B)^{-1}
\qquad\hbox{and}\qquad
\int_{\BR^n} a \, \wrt \ga =0.
\end{equation}
The space $\hu{\ga}$ is then the vector space of all functions
$f$ in $\lu{\ga}$ that admit a decomposition of the form
$\sum_j \la_j \, a_j$, where the $a_j$ are atoms
and the sequence of complex numbers $\{\la_j\}$ is summable.
The norm of $f$ in $\hu{\ga}$ is defined as the infimum
of $\sum_j\mod{\la_j}$ over all representations of $f$ as above.
In \cite{MM} the space $\hu{\ga}$ is defined by means of $(1,p)$-atoms
with $1<p<\infty$, but in \cite{MMS2} it is verified
that the space obtained is the same as ours.
Note that $\hu{\ga}$ is defined much as the atomic space $H^1$
on spaces of homogeneous type in the sense
of R.R.~Coifman and G.~Weiss \cite{CW}, but with one difference.
Namely,
only the exceptional atom and atoms with ``small supports'',
i.e., with supports contained in admissible balls,
appear in the definition of $\hu{\ga}$.
This difference is quite significant and has important consequences.
It is known \cite[Theorem~5.2]{MM}
that the Banach dual of $\hu{\ga}$ is isomorphic
to the space $BMO(\ga)$ of all functions of ``bounded mean
oscillation'', i.e., of all functions in $\lu{\ga}$ such that
\begin{equation} \label{f: BMO}
\norm{f}{*}
= \sup_{B \in \cB_1} \frac{1}{\ga(B)} \int_B \mod{f-f_B} \wrt \ga
< \infty,
\end{equation}
where $f_B = \frac{1}{\ga(B)} \int_B f \wrt \ga$.
A convenient norm on $BMO(\ga)$ is the following
$$
\norm{f}{BMO}
= \norm{f}{1} + \norm{f}{*}.
$$
\begin{remark} \label{rem: eq norm BMO}
An equivalent norm on $BMO(\ga)$ is obtained by replacing in (\ref{f: BMO})
balls in $\cB_1$ by cubes of sidelength at most $2\, \min(1,1/\mod{c_B})$
(see \cite[Section~2]{MM}).
As a consequence, a function $f$ is in $\hu{\ga}$
if and only if it admits a decomposition of the form $\sum_j \la_j\, b_j$
where $\{\la_j\}$ is summable and the $b_j$ are either the constant
function $1$ or functions in $\ly{\ga}$
supported in cubes $Q$ of sidelength at most $2\, \min(1,1/\mod{c_B})$,
satisfying
$$
\norm{b}{\infty} \leq \ga(Q)^{-1}
\qquad\hbox{and}\qquad
\int_{\BR^n} b \, \wrt \ga =0.
$$
An equivalent norm on
$\hu{\ga}$ is then obtained by taking the infimum
of $\sum_j\mod{\la_j}$ over all representations of $f$ as above.
\end{remark}
Given a bounded operator $T$ on $\ld{\ga}$,
we denote by $K_T$ the Schwartz kernel of $T$ and by $k_T$
the kernel of $T$ with respect to the Gaussian measure,
defined by
\begin{equation} \label{f: kT and KT}
k_T(x,y)
= \pi^{d/2} \, \e^{\mod{y}^2}\, K_T(x,y),
\end{equation}
in the sense of distributions in $\BR^d\times \BR^d$.
The reason for introducing $k_T$ is that if $K_T$ is locally integrable
in $\BR^d\times \BR^d$, then
$T$ is an integral operator with kernel $k_T$ with
respect to $\ga$, i.e.,
$$
Tf(x)
= \int_{\BR^d} k_T(x,y) \, f(y) \wrt \ga(y)
\quant x \in \BR^d \quant f \in C_c(\BR^d).
$$
One of the results in \cite{MM} says that if $T$
is bounded on $\ld{\ga}$ and $k_T$ is locally integrable
off the diagonal of $\BR^d \times \BR^d$, and satisfies
\begin{equation} \label{f: HIC H1L1}
\sup_{B\in \cB_1} \, r_B \, \sup_{y\in B} \int_{(2B)^c}
\bigmod{\nabla_y k_T(x,y)} \wrt\ga(x)
< \infty,
\end{equation}
then $T$ is bounded from $\hu{\ga}$ to $\lu{\ga}$,
and, consequently, on $\lp{\ga}$ for all $p$ in $(1,2)$.
Notice that (\ref{f: HIC H1L1})
is a local H\"ormander integral condition.
Similarly, if
\begin{equation} \label{f: HIC LIBMO}
\sup_{B\in \cB_1} \, r_B \, \sup_{x\in B} \int_{(2B)^c}
\bigmod{\nabla_x k_T(x,y)} \wrt\ga(y)
< \infty,
\end{equation}
then $T$ is bounded from $\ly{\ga}$ to $\BMO{\ga}$,
and, consequently, on $\lp{\ga}$ for all $p$ in $(2,\infty)$.
We shall use these criteria to prove the
boundedness results contained in Theorem~\ref{t: main},
and start by determining the kernels of $R_i$
and of $S_i^*$. The functions $\phi$ and $\psi$, defined~by
\begin{equation*} \label{fipsi}
\phi(r,x,y)
= \frac{ry-x}{\sqrt{1-r^2}} \qquad \mathrm{and} \qquad
\psi(r,x,y)
= \frac{rx-y}{\sqrt{1-r^2}}
\quant x,\,y \in \BRd \quant r \in (0,1),
\end{equation*}
will occur frequently.
Note that if $d\geq 2$, then $\phi$ and $\psi$ are vector-valued;
their components will be denoted by $\phi_i$ and $\psi_i$.
The arguments of these two functions will often be suppressed.
As shown in \cite[Lemma 2.2]{GMST1},
the Schwartz kernel of the operator $M(\cL)$, defined
in the introduction, is
\[
K_{M(\cL)}(x,y)
= \pi^{-(d+1)/2} \int_0^1 \Bigl[\frac{\e^{-|\psi|^2}}{(1-r^2)^{d/2}}
- \e^{-|y|^2}\Bigr]\wrt \rho(r),
\]
where the measure $\rho$, supported in $[0,1]$, is defined by
\begin{equation} \label{f: measure rho}
\wrt \rho(r)
= \frac{\wrt r}{r\, \sqrt{-\log r}}.
\end{equation}
Hence the Schwartz kernel $K_{R_i}$ of the operator $R_i$
agrees off the diagonal with the function
\[
\partial_{x_i} K_{M(\cL)}(x,y)
= - 2\pi^{-(d+1)/2} \int_0^1
\frac{r\, \psi_i\, \e^{-|\psi|^2}}{(1-r^2)^{(d+1)/2}} \wrt \rho(r).
\]
By using (\ref{f: kT and KT}) and the fact that
${|\phi|^2-|\psi|^2} =|x|^2-|y|^2 $, we get that
\begin{equation} \label{eq:kern}
k_{R_i}(x,y)
= -\frac2{\sqrt{\pi}} \, \e^{|x|^2} \int_0^1
\frac{r\, \psi_i\, \e^{-|\phi|^2}}{(1-r^2)^{(d+1)/2}} \wrt \rho(r)
\end{equation}
for all $x\neq y$.
To determine $k_{S_i^*}$, we argue similarly, observing that
\[
\partial_{x_i}^* \e^{-|\psi|^2}
= \frac{-2\phi_i}{\sqrt{1-r^2}} \,\, \e^{-|\psi|^2},
\]
and obtain
\begin{align} \label{eq:kerntilde}
k_{S_i^*}(x,y)
& = -\frac2{\sqrt{\pi}} \, \e^{|y|^2} \int_0^1
\Bigl[\frac{\phi_i \, \e^{-|\psi|^2} }{(1-r^2)^{(d+1)/2}}
+ x_i \, \e^{-|y|^2}\Bigr] \wrt \rho(r) \notag \\
& = -\frac2{\sqrt{\pi}} \, \e^{|x|^2} \int_0^1
\Bigl[\frac{\phi_i \, \e^{-|\phi|^2} }{(1-r^2)^{(d+1)/2}}
+ x_i \, \e^{-|x|^2}\Bigr] \wrt \rho(r)
\end{align}
for all $x\neq y$.
The constants $C<\infty$ and $c>0$ will depend only
on the dimension $d$, and they may vary from occurrence to
occurrence. By $f\sim g$ we mean $c < f/g < C$.
Lebesgue measure is denoted $\la$.
\section{Statements of results} \label{s: TMR}
The main result of this paper is the following.
\begin{theorem}\label{t: main}
In dimension one, the four operators $R_1$, $R_1^*$, $S_1$ and $S_1^*$
are bounded from $\hu{\ga}$ to $\lu{\ga}$ and from $\ly{\ga}$ to $BMO(\ga)$.
In dimension $d\ge2$, the boundedness properties of the operators
$R_i,R^*_i,S_i$ and $S^*_i$ are given by the following table,
where $B$ means ``bounded" and $U$ ``unbounded".
\medskip
\begin{center}
\begin{tabular}{|c||c|c|c|c|}
\hline
${\phantom a}$ & $R_i$ & $S_i$&$R_i^*$&$S_i^*$\\
\hline \hline
$H^1 \to L^1$ & $U$ & $U$ & $B$ & $B$ \\
\hline
$L^\infty \to BMO$ & $B$ & $B$ & $U$ & $U$ \\
\hline
\end{tabular}
\end{center}
\medskip
\end{theorem}
\begin{remark}
Note that all the operators $R_i$ and $S_i$, $i$ in $\{1,\ldots,d\}$,
have the same boundedness properties, and so do
the operators $R_i^*$ and $S_i^*$. Furthermore,
the boundedness properties of $R_i^*$ and $S_i^*$
are ``dual'' to those of $R_i$ and $S_i$. This is not
an accident, and the proofs of the statements concerning
the boundedness properties of
$R_i^*$ and $S_i$ will be obtained, by a duality argument,
from the properties of $R_i$ and~$S_i^*$.
\end{remark}
In addition to $R_i$, $R_i^*$, $S_i$ and
$S_i^*$ we also consider the operators
$M_i$ and $M_i^*$, defined on Hermite polynomials by
$$
M_i=x_i\, M(\cL) \qquad\qquad M^*_i=M(\cL)\, x_i.
$$
\begin{corollary}
The operators $M_i$ and $M_i^*$ extend to bounded operators on $\lp{\ga}$
for all $p$ in $(1, \infty)$ in any dimension.
For $d=1$, they are bounded from $\hu{\ga}$ to $\lu{\ga}$
and from $\ly{\ga}$ to $\BMO{\ga}$. For $d\geq 2$, they
are unbounded both from $\hu{\ga}$ to $\lu{\ga}$
and from $\ly{\ga}$ to $\BMO{\ga}$.
\end{corollary}
\begin{proof}
Since $M_i =(1/2) (R_i+S^*_i)$ and $M^*_i=(1/2) (R^*_i+S_i)$,
the $\lp{\ga}$ boundedness of $M_i$ and $M_i^*$
follows from that of $R_i$, $S_i$, $R_i^*$ and $S_i^*$
mentioned in the introduction. The remaining parts
of the corollary
are straightforward consequences of Theorem~\ref{t: main}.
\end{proof}
The proof of Theorem~\ref{t: main} occupies the remaining part of
the paper.
We first deal with $R_i$ and $S_i^*$. In Section~\ref{R},
it is proved that in dimension one,
$R_i$ is bounded from $\hu{\ga}$ to $\lu{\ga}$ and $S_i^*$
from $\ly{\ga}$ to $BMO(\ga)$.
Section~\ref{S} deals with the positive
statements concerning these two operators for $d\ge 2$, as indicated in the
table. The negative results claimed for $R_i$ and $S_i^*$ are
proved in Section~\ref{U}. Finally, a rather simple duality argument
in Section~\ref{D}
gives all the results for the remaining operators $R_i^*$ and $S_i.$
\section{Boundedness of $R_1$ and $S_1^*$ in one dimension} \label{R}
Let $d=1$. We first prove that $R_1$ is bounded
from $\hu{\ga}$ to $\lu{\ga}$.
Because of (\ref{f: HIC H1L1}) and (\ref{eq:kern}),
it suffices to show that there exists a constant $C$
such that for all balls $B$ in~$\cB_1$ and all points $y$ in $B$
\begin{equation}\label{basic2}
r_B \, \int_{(2B)^c} \!\!\! \wrt\la(x)\,
\Bigmod{\int_0^1 \frac{r\,\partial_y\bigl(\psi\,
\e^{-\phi^2}\bigr)}{1-r^2} \, \wrt {\rho(r)} }
\leq C.
\end{equation}
Next, observe that
\begin{align*}\notag
\partial_y\bigl(\psi\ \e^{-\phi^2}\bigr)
& = -\frac{1}{\sqrt{1-r^2}}\,\e^{-\phi^2}-2\phi\,
\psi \frac{r}{\sqrt{1-r^2}}\,\e^{-\phi^2} \\
& = F_1(r,x,y)+F_2(r,x,y).
\end{align*}
To deal with $F_2$, we first note that
$\partial \phi/\partial r = - \psi/(1-r^2)$, so that
$$
F_2(r,x,y) = -r\, \sqrt{1-r^2}\, \partial \e^{-\phi^2}/\partial r.
$$
Integrating by parts, we get
\[
\int_{0}^1 \frac{(-\log r)^{-1/2}}{1-r^2}\,F_2(r,x,y)\,\wrt r =
\int_{0}^1 \frac{\partial}{\partial r}
\frac{r(-\log r)^{-1/2}}{\sqrt{1-r^2}}\,\e^{-\phi^2}\,\wrt r.
\]
The integral to the right here is easily seen to be bounded by
\[
C\, \int_{0}^1 (1-r)^{-2}\, \e^{-\phi^2} \wrt r.
\]
Considering also $F_1$, we can then estimate the expression in
(\ref{basic2}) by
\begin{equation}
\label{eq:upper}
C\, r_B \int_{(2B)^c} \int_{0}^1 (1-r)^{-2}\, \e^{-\phi^2}\,\wrt r\wrt\la(x).
\end{equation}
Here we first integrate in $r$ only over $1-r_B/(2(1+|y|)) < r < 1$,
which implies
\[
|ry-x| = |(r-1)y+y-x| \geq |y-x| - \frac{r_B|y|}{2(1+|y|)} \geq
|y-x| - \frac{r_B}2 \geq \frac{|y-x|}2
\]
and thus
$|\phi| \geq |y-x|/(2\sqrt{1-r})$.
This part of (\ref{eq:upper})
can thus be estimated by the following expression, where
we make the change of variables $t = (1-r)/|x-y|^2$,
\begin{align*}
C\, r_B \int_{(2B)^c} \int_{1-r_B/(2(1+|y|))}^1
\!\! &\frac{\e^{-c|x-y|^2/(1-r)}\,}{(1-r)^2} \wrt r \wrt\la(x) \\
&\leq r_B \int_{(2B)^c} \frac{\wrt\la(x)}{|x-y|^2}
\int_0^\infty t^{-2}\e^{-c/t} \wrt t \leq C.
\end{align*}
In the remaining part of (\ref{eq:upper}), we switch
the order of integration and find
\begin{align*}
r_B \int_{(2B)^c}\!\!\! \wrt\la(x)\, &\int_{0}^{1-r_B/(2(1+|y|))}
\frac{\e^{-\phi(r,x,y)^2}}{(1-r)^2} \wrt r \\
&\leq r_B \int_{0}^{1-r_B/(2(1+|y|))} \!\!\frac{\wrt r}{(1-r)^2}
\int_{\BR^d} \e^{-\phi(r,x,y)^2} \wrt\la(x) \\
&\leq C \, r_B \int_{0}^{1-r_B/(2(1+|y|))}\frac{\wrt r}{(1-r)^{3/2}},
\end{align*}
which is easily seen to be bounded by a constant independent
of $y$ in $B$ and of $B$ in $\cB_1$ . The proof of the
boundedness of $R_1$ from $\hu{\ga}$ to $\lu{\ga}$ is complete.
\medskip
Now we prove that if $d=1$, then $S_1^*$ is bounded from $\ly{\ga}$
to $\BMO{\ga}$.
Recall that the functions $\phi$ and $\psi$ are now scalar valued.
By (\ref{f: HIC LIBMO}) and (\ref{eq:kerntilde}) it suffices to show that
there exists a constant $C$, independent of $x$ in $B$ and of
$B$ in $\cB_1$, such that
\begin{equation} \label{equiv}
r_B \,
\int_{(2B)^c} \!\! \wrt \la(y)\, \Bigmod{\int_0^1 \frac{1}{1-r^2} \,
\partial_x \bigl(\phi\, \e^{-\psi^2}
+{(1-r^2)} \, x \e^{-y^2}\bigr) \wrt \rho(r)}
\leq C.
\end{equation}
This is analogous to (\ref{basic2}), with the following
three differences. The variables $x$ and $y$, and thus also $\phi$
and $\psi$, are swapped; there is an extra factor
$1/r$; there is an extra term ${(1-r^2)} x \e^{-y^2}$.
For that part of the expression in (\ref{equiv}) obtained by integrating
only over $1/2 < r < 1$, we can copy the argument used to verify
(\ref{basic2}) above. The extra factor $1/r$ is harmless
in this interval, and so is the term ${(1-r^2)} x \e^{-y^2}$.
In the integration by parts with respect to $r$, one will now get
an integrated term at $r=1/2$, equal to $C\,\e^{-\psi(1/2,x,y)^2}$ and
easy to handle.
Consider then the integral over $0 < r < 1/2$. We have
\begin{equation} \label{eq:derivative}
\partial_x \bigl(\phi \,\e^{-\psi^2} +{(1-r^2)} \,x \,\e^{-y^2}\bigr)
= -(1-r^2)\, \Bigl[\frac{\e^{-\psi^2} }{(1-r^2)^{3/2}} - \e^{-y^2}\Bigr]
-\frac{2\phi\psi r}{(1-r^2)^{1/2}} \, \e^{-\psi^2}.
\end{equation}
Notice first that $ry-x = r(y-rx) + (r^2-1)x$, so that
\begin{equation} \label{eq:little}
|\phi|
\le |\psi| + |x|.
\end{equation}
This allows us to estimate the last term in (\ref{eq:derivative}) by
$C\,r(1+|x|)\e^{-\psi^2/2}$ for $0 < r < 1/2$.
The corresponding part of the expression in
(\ref{equiv}) can then be seen not to be larger than
$C\, r_B (1+|x|) \le C$, since $x \in B$.
For the first term in the right-hand side of (\ref{eq:derivative}), we write
\[
\frac{ \e^{-\psi^2} }{(1-r^2)^{3/2}} - \e^{-y^2}
= \frac{ \e^{-\psi^2} - \e^{-y^2} }{(1-r^2)^{3/2}}
+\Bigl[\frac{ 1 }{(1-r^2)^{3/2}} - 1\Bigr] \, \e^{-y^2}.
\]
The last term here is easily seen to be harmless and will be neglected.
For the first term to the right, we observe that
$\e^{-y^2} = \e^{-\psi(0,x,y)^2}$, so that
\[
\e^{-\psi(r,x,y)^2} - \e^{-y^2} = \int_0^r \frac{\partial}{\partial s}
\e^{-\psi(s,x,y)^2}\wrt s = \int_0^r
\frac{2 \phi(s,x,y) \psi(s,x,y)}{1-s^2}\e^{-\psi(s,x,y)^2}\wrt s.
\]
Since $s<r<1/2$
here, (\ref{eq:little}) implies that the last integral is at most
\begin{equation*}
\label{eq:s-integral}
C\, \int_0^r (1 + |x|)\, \e^{-\psi(s,x,y)^2/2}\wrt s.
\end{equation*}
Integrating this expression with respect to $y$, we get at most
$C\,r(1+|x|)$. This will give a contribution to the expression in
(\ref{equiv}) which is at most
$C\,r_B (1+|x|) \int_0^{1/2} (- \log r)^{-1/2} \wrt r$, and this is
uniformly bounded with respect to $x$ in $B$ and $B$ in $\cB_1$.
This concludes the proof that $S_1^*$ is bounded from
$\ly{\ga}$ to $\BMO{\ga}$.
\section{Boundedness of $R_i$ and $S_i^*$ in higher dimensions}
\label{S}
The boundedness of $R_i$ from $\ly{\ga}$ to $\BMO{\ga}$ for any $d$
was proved in \cite[Theorem 7.2~\rmii]{MM}.
To prove that $S_i^*$ is bounded from $\hu{\ga}$ to $\lu{\ga}$ in
all dimensions,
we follow the lines of the proof of \cite[Theorem 7.2]{MM}.
By (\ref{f: HIC H1L1}) and (\ref{eq:kerntilde}),
it is enough to verify that there exists a constant $C$, independent
of $y$ in $B$ and of $B$ in $\cB_1$ such that
\begin{equation}\label{basic3}
r_B \, \int_{(2B)^c} \Bigmod{\int_0^1 \nabla_y
\Bigl(\frac{\phi_i \, \e^{-|\phi|^2}}{(1-r^2)^{(d+1)/2}}
+ x_i \, \e^{-|x|^2}\Bigr) \wrt \rho(r)} \wrt\la(x)
\leq C.
\end{equation}
It is straightforward to verify that all the components of the gradient
with respect to $y$ that appear here are bounded in absolute value by
\[
C\, \frac{r\, (1+|\phi|^2)}{(1-r^2)^{(d+2)/2}} \, \e^{-|\phi|^2}
\le C\,\frac{r}{(1-r^2)^{(d+2)/2}} \, \e^{-|\phi|^2/2}.
\]
Thus the left-hand side of (\ref{basic3}) is no larger than
\begin{equation} \label{eq:doubleint}
C\, r_B
\int_{(2B)^c} \wrt\la(x) \int_0^1
\frac{r\, \e^{-|\phi|^2/2}}{(1-r^2)^{(d+2)/2}} \wrt \rho(r).
\end{equation}
We split the inner integral into integrals over
$(0, r_0)$ and $(r_0, 1)$, with $r_0 = 1 - r_B^2/4> 1/2$.
In each of the two resulting double integrals,
we switch the order of integration. That part of (\ref{eq:doubleint})
corresponding to $(0, r_0)$ will be less than
\begin{align*}
\quad C\, r_B \int_0^{r_0} \frac{(-\log r)^{-1/2}}{(1-r^2)^{(d+2)/2}}
\wrt r \, \int_{\BR^d} \e^{-|\phi|^2/2}
& \wrt\la(x) \\
& \le C\, r_B \int_0^{r_0} (-\log r)^{-1/2} (1-r^2)^{-1} \wrt r \\
& \le C \, r_B \, (1-r_0)^{-1/2} \\
& \le C.
\end{align*}
To deal with the integral over $(r_0, 1)$,
observe that $y \in B$ implies $|y| \le c_B + r_B
\le 2 r_B^{-1}$, in view of the definition of $\cB_1$. For $x \notin 2B$
and $r_0 < r < 1$ we then have
\[
|ry-x| = |y-x-(1-r)y| > r_B -|y|r_B^2/4 \ge r_B/2.
\]
The remaining part of (\ref{eq:doubleint}) is thus at most
\begin{equation} \label{eq:switch}
C\, r_B \int_{r_0}^1 \frac{r}{(1-r^2)^{(d+2)/2}} \wrt \rho(r) \,
\int_{|ry-x|>r_B/2} \e^{-|\phi|^2/2} \wrt\la(x).
\end{equation}
The inner integral in (\ref{eq:switch}) is dominated by
\[
(1-r^2)^{d/2}\, \int_{|z|>r_B/(2\sqrt{1-r^2})} \e^{-|z|^2} \wrt z
< C\, (1-r^2)^{d/2}\,\left(r_B/\sqrt{1-r^2}\right)^{-2},
\]
and the whole expression (\ref{eq:switch}) by
\[
C\, r_B^{-1} \int_{r_0}^1 (1-r)^{-1/2}\wrt r \le C\, r_B^{-1} (1-r_0)^{1/2} \le C.
\]
This concludes the proof of the
boundedness of $S_i^*$ from $\hu{\ga}$ to $\lu{\ga}$.
\section{Unboundedness results}
\label{U}
In this section, we shall prove that if $d\geq 2$, then
$R_1$ is unbounded from
$\hu{\ga}$ to $\lu{\ga}$ and $S_1^*$ is unbounded from
$\ly{\ga}$ to $\BMO{\ga}$.
The proofs for $R_2,\ldots,R_d$ and $S_2^*, \ldots, S_d^*$
are similar and are omitted.
Consider first $R_1$.
Let $Q$ be the cube with centre $(\xi,0,...,0)$ and
sidelength $2\xi^{-1}$, for some large $\xi>0$.
Denote by $Q_+$ and $Q_-$ those halves of $Q$
where $y_2>0$ and $y_2<0$, respectively.
{Set $b=\ga(Q)^{-1}(\1_{Q_+}-\1_{Q_-})$.
By Remark~\ref{rem: eq norm BMO}, $\norm{b}{H^1(\gamma)}$ is bounded
independently of $\xi$.}
Writing $\tilde y=(y_1,-y_2,y_3,\ldots,y_d)$, one can easily see from
(\ref{eq:kern}) that
\begin{align}\label{Ra}
\e^{-\mod{x}^2}\, R_1 a(x)
& = \frac2{\sqrt{\pi}\,\ga(Q)}\int_{Q_+} \int_{0}^1
\frac{r\, (y_1-rx_1)}{(1-r^2)^{(d+2)/2}}\,
\big(\e^{-|\phi(r, x, y)|^2}-\e^{-|\phi(r, x, \tilde y)|^2}\big)
\wrt \rho(r) \wrt\ga(y)\notag \\
& = \frac2{\sqrt{\pi}\,\ga(Q)}\int_{Q_+} \int_{0}^1
\frac{r\, (y_1-rx_1)}{(1-r^2)^{(d+2)/2}} \,
\e^{-|\phi(r, x, y)|^2}\, \big(1-\e^{-\tau(r,x_2,y_2)}\big)
\wrt \rho(r) \wrt\ga(y);
\end{align}
here and in the rest of the proof we write
$\tau(r,x_2,y_2)$ in place of $4rx_2y_2/(1-r^2)$.
Similarly, we write $\eta(\xi,x_1)$ instead of $\sqrt{(\xi-x_1)/\xi}$.
We shall evaluate $R_1a$ at points $x$ in the set
$$
A_Q =
\Bigl\{x\in\BR^d: x_1\in \Bigl(\xi - 1, \xi-\frac4\xi\Bigr), \,
x_2 \in \Bigl[\frac{\eta(\xi,x_1)}{2} , \eta(\xi,x_1)\Bigr],
|x_k| \leq \eta(\xi,x_1), k=3,...,d\Bigl\}.
$$
Observe that if $x\in A_Q$, $y\in Q_+$ and $r\in [0,1]$, then $y_1-rx_1>0$
and the integrands in (\ref{Ra}) are positive.
With $\xi$ large and $x \in A_Q$, the integration in $r$ will be
restricted to the interval
\begin{equation} \label{f: Int}
I(x_1) = \Bigset{r\in(0,1): \Bigmod{r-\frac{x_1}{\xi}}
<\frac{\eta(\xi,x_1)}{2\xi} }\subset (1/2, 1).
\end{equation}
For $x \in A_Q$ a lower bound for $\e^{-\mod{x}^2}\,R_1a(x)$ is
$$
\frac{c}{\ga(Q)}\int_{Q_+} \int_{I(x_1)} \frac{y_1-rx_1}{(1-r^2)^{(d+3)/2}}\,
\e^{-|\phi|^2}\ \big(1-\e^{-\tau(r,x_2,y_2)}\big) \wrt r\wrt\ga(y).
$$
Therefore,
\begin{multline}\label{difference}
\norm{R_1a}{1}
\ge \int_{A_Q} R_1a(x) \, \e^{-\mod{x}^2}\wrt \lambda(x) \\
\ge \frac{c}{\ga(Q)}\int_{A_Q}\int_{Q_+} \int_{I(x_1)}
\frac{y_1-rx_1}{(1-r^2)^{(d+3)/2}}\, \e^{-|\phi|^2}\,
\big(1-\e^{-\tau(r,x_2,y_2)}\big)\wrt r\wrt\ga(y) \,\wrt\lambda(x).
\end{multline}
To prove that the operator $R_1$ is unbounded from $\hu{\ga}$ to $\lu{\ga}$,
we only need to show that the last expression here
tends to infinity with $\xi$.
We claim that if $\xi$ is sufficiently large,
one has for $x$ in $A_Q$, $y$ in $Q_+$ and $r$ in $I(x_1)$
\begin{align}
1-r^2
& \sim \eta(\xi,x_1)^2 \label{dis1} \\
y_1-rx_1
&\ge c\ (\xi-x_1) \label{dis2}\\
\e^{-|\phi(r, x, y)|^2}
&\ge c \label{dis3}\\
1-\e^{-\tau(r,x_2,y_2)}
&\sim y_2/\eta(\xi,x_1). \label{dis4}
\end{align}
Deferring momentarily their proofs,
we show how these inequalities yield the desired conclusion.
Indeed, applying them and observing that
$$
\frac{1}{\ga(Q)}\int_{Q_+}y_2\wrt\ga(y) \ge\, \frac{c}\xi,
$$
we see from (\ref{difference}) that
\begin{align*}
\norm{R_1a}{1}
& \ge \frac{c}{\ga(Q)}\int_{A_Q}\int_{Q_+}
\eta(\xi,x_1)^{-(d+2)}\, \xi \, y_2\, \mod{I(x_1)}
\wrt\ga(y)\,\wrt\lambda(x) \\
& \ge c\,\int_{A_Q}\eta(\xi,x_1)^{-(d+1)}\, \xi^{-1} \wrt\lambda(x) \\
& = c\,\int_{\xi - 1}^{\xi-4/\xi} (\xi-x_1)^{-1} \wrt x_1\\
& \ge c\,\log \xi,
\end{align*}
which tends to infinity with $\xi$ as desired.\par
It remains to prove the inequalities (\ref{dis1})--(\ref{dis4}).
Observe first that for $x \in A_Q$ one has $4/\xi < \xi - x_1$.
Considering also the geometric mean of these two quantities, we have
\begin{equation}
\label{eq:star}
4/\xi < 2\, \eta(\xi,x_1) <\xi - x_1.
\end{equation}
To verify (\ref{dis1}), write for $r \in I(x_1)$
$$
1-r^2
\sim 1-r
=\eta(\xi,x_1)^2 -\Bigl(r-\frac{x_1}\xi\Bigr).
$$
Observe that $|r-x_1/\xi| < \eta(\xi,x_1)^2/4$, because of
the definition of $I(x_1)$ and (\ref{eq:star}), and (\ref{dis1})
follows.
For (\ref{dis2}), notice that similarly
$$
y_1-rx_1 > \xi - \frac1\xi
- \Bigl(\frac{x_1}\xi + \frac{\xi-x_1}{4\xi}\Bigr) x_1
> \xi - \frac{\xi-x_1}4 -\frac{x_1^2}{\xi} - \frac{\xi-x_1}4
> \frac{\xi^2-x_1^2}{\xi} - \frac{\xi-x_1}2
> \frac{\xi-x_1}2.
$$
Aiming at (\ref{dis3}), we write
\[
\mod{x_1-ry_1} \le \mod{r(y_1-\xi)} + \mod{\left(r - \frac{x_1}\xi\right)\xi}
< \frac1\xi + \frac{1}{2\xi}\eta(\xi,x_1)\,\xi <
\eta(\xi,x_1) \le C\,\sqrt{1-r^2},
\]
the last two inequalities in view of (\ref{eq:star}) and (\ref{dis1}),
respectively. Since $x\in A_Q$ and $y\in Q_+$, we similarly get
for $k=2,... ,d$
$$
\mod{x_k-ry_k}
\le \mod{x_k}+\mod{y_k}
\le \eta(\xi,x_1) +\xi^{-1}
\le C\,\sqrt{1-r^2}.
$$
Thus $\mod{x-ry}^2/(1-r^2)\le C$, which implies (\ref{dis3}).
Finally, we prove (\ref{dis4}). Because of (\ref{dis1}) and the facts that
$x\in A_Q$ and $y\in Q_+$, one has
$$
\tau(r,x_2,y_2) \sim \frac{\xi}{\xi-x_1}\, \eta(\xi,x_1)\, y_2
= \eta(\xi,x_1)\, y_2 < \frac1{\sqrt{\xi(\xi-x_1)}} < 1,
$$
and (\ref{dis4}) follows.
This concludes the proof that $R_1$ is unbounded
from $\hu{\ga}$ to $\lu{\ga}$.
\medskip
Next, we prove that in higher dimension, $S_1^*$ is unbounded from $\ly{\ga}$
to $\BMO{\ga}$.
We shall modify slightly the counterexample
constructed above.
The kernel is
\[
k_{S_1^*}(x,y)= \frac2{\sqrt{\pi}}\, \e^{|x|^2} \int_0^1
\Bigl[\frac{x_1 - ry_1}{(1-r^2)^{(d+2)/2}} \,
\e^{-|\phi|^2} + x_1 \e^{-|x|^2}\Bigr] \wrt \rho(r).
\]
With $\xi > 0$ large, we let the cube $Q$, its subset $Q_+$ and the region
$A_Q$ be as above. But now $f$ will be the characteristic
function of $A_Q$, and $S_1^* f$ will be evaluated in $Q$.
For $x \in Q_+$ we write $\tilde{x} = (x_1, -x_2, x_3, ..., x_d)$.
We shall verify that
\begin{equation}
\label{nonbmo}
\frac1{\ga(Q)} \int_{Q_+} (S_1^* f({x}) - S_1^* f(\tilde x))
\wrt \ga(x)
\end{equation}
tends to $+\infty$ as $\xi$ tends to $+\infty$. This would clearly mean that
$\norm{S_1^* f}{BMO}$ also tends to infinity, although
$\norm{f}\infty = 1$ for all $\xi$. Notice that for $x$ in $Q_+$
\[
S_1^* f({x}) - S_1^* f(\tilde x) = \frac2{\sqrt{\pi}}\, \int_{A_Q} \int_0^1
\frac{x_1-ry_1}{(1-r^2)^{(d+3)/2}}
\e^{-\mod{\psi}^2}\, \big(1-\e^{-\tau(r,x_2,y_2)}\big)
\wrt \rho(r) \wrt\la(y);
\]
recall that we write $\tau(r,x_2,y_2)$ in place of $4rx_2y_2/(1-r^2)$.
Here the integrand is positive. We now integrate with respect to
$\wrt \ga(x)$ over $x \in Q_+$, and restrict the integral in $r$
to the interval $I(y_1) \subset (1/2, 1)$, defined like $I(x_1)$
(see (\ref{f: Int})).
For such $r$, one can replace the measure $\wrt \rho(r)$ by
$(1-r^2)^{-1/2} \wrt r$.
This leads to a lower estimate of the quantity
(\ref{nonbmo}) in terms of a triple integral rather similar to that
in (\ref{difference}). The only differences are that $x$ and $y$
have switched roles and that we now have also an insignificant factor $1/r$.
Thus the integral will tend to
$\infty$ with $\xi$, like that in (\ref{difference}).
This concludes the proof that $S_1^*$ is unbounded from $\ly{\ga}$
to $\BMO{\ga}$.
\section{Duality arguments}
\label{D}
All the statements contained in Theorem~\ref{t: main}
that have not been proved yet will follow from those
proved above by a simple duality argument,
based on the following elementary lemma.
We point out that by a result of \cite{MMS2},
$\ld{\ga}$ is a subspace of $\hu{\ga}$,
which is dense because it contains all the atoms.
\begin{lemma} \label{l: duality}
Suppose that $T$ is a bounded operator on $\ld{\ga}$. Then
$T$ extends to a bounded operator from $\hu{\ga}$ to $\lu{\ga}$
if and only if its (Hilbert space) adjoint operator $T^*$
is bounded from $\ly{\ga}$ to $\BMO{\ga}$.
\end{lemma}
\begin{proof}
Suppose that $T$ extends to a bounded operator from $\hu{\ga}$
to $\lu{\ga}$.
Then for $g$ in~$\ly{\ga}$
\begin{equation} \label{f: norm T*f in BMO}
\norm{T^*g}{BMO}
= \sup\,\{ \mod{\prodo{f}{T^*g}}: f \in \ld{\ga}, \
\norm{f}{H^1}\le 1\},
\end{equation}
and
$$
\begin{aligned}
\mod{\prodo{f}{T^*g}}
& = \mod{\prodo{Tf}{g}} \\
& \le \norm{Tf}{1}\, \norm{g}{\infty} \\
& \le \norm{T}{H^1,L^1} \, \norm{f}{H^1}\, \norm{g}{\infty}.
\end{aligned}
$$
Thus $T^*$ is bounded from $\ly{\ga}$ to $BMO(\ga)$.
The converse implication is proved similarly. We omit the details.
\end{proof}
Now, the boundedness in one dimension of $R_1^*$ from $\ly{\ga}$ to $\BMO{\ga}$
follows by Lemma~\ref{l: duality}
from that of $R_1$ from $\hu{\ga}$ to $\lu{\ga}$, proved in
Section~\ref{R}. Similarly, the
unboundedness in dimension $d\geq 2$ of $R_i^*$ from $\ly{\ga}$ to $\BMO{\ga}$
follows from the unboundedness of $R_i$ from $\hu{\ga}$ to $\lu{\ga}$
proved in Section~\ref{U},
and the boundedness of $R_i^*$ from $\hu{\ga}$ to $\lu{\ga}$ for any $d$
follows from that of $R_i$ from $\ly{\ga}$ to $\BMO{\ga}$
proved in Section~\ref{S}.
A similar argument gives the boundedness of $S_i$
from $\hu{\ga}$ to $\lu{\ga}$ when $d=1$,
that of $S_i$ from $\ly{\ga}$ to $\BMO{\ga}$ for any $d$,
and the unboundednes of $S_i$ from $\hu{\ga}$ to $\lu{\ga}$
when $d\geq 2$.
The theorem is completely proved.
|
\section{Introduction and preliminaries}
An adaptive approximation scheme of the Wiener process is considered. The discretization points are placed at times when the value of the true process differs from the approximation by some amount, here denoted by $\eta$. This can be seen as a control problem where we want to track the true value of the process with our approximation, and where both the process and its approximation are fully observable. The approximation strategy presented here may be feasible when discretization is associated with some cost that should be kept low. Examples of related problems is that of discrete time hedging of derivative contracts in financial markets (see e.g. \citet{Geiss_Geiss:2006}) and certain space-time discretization schemes of stochastic differential equations (see e.g. \citet{Milstein_Tretyakov:1999}).
Let $X^{}$ be a diffusion process defined by $X^{}_t=\sigma W_t$, where $W$ denotes a one dimensional standard Wiener process. Define, for some $\eta>0$, a sequence of stopping times $\{t^{ \eta}_{i} \}_{i \geq 0}$ by
\begin{align*}
t^{ \eta}_{i+1} = \inf \{t>t^{ \eta}_{i} \, | \, |X^{}_t-X^{}_{t^{ \eta}_{i}}|=\eta \} \, ,
\end{align*}
where $t^{ \eta}_{0}=0$. The components of the sequence $t^{ \eta}$ may be seen as epochs of the renewal process $N^{ \eta}$ defined by $N^{ \eta}_t=\sup\{i: t^{ \eta}_{i} \leq t \}$. Furthermore, let the sequence $\{\tau^{ \eta}_{i} \}_{i \geq 1}$ of interarrival times be defined by $\tau_i^{ \eta}=t_{i}^{ \eta}-t_{i-1}^{ \eta}$, and define the renewal-reward process $\varphi$ by $\varphi^{ \eta}_t:=\sum_{i=1}^{N^{ \eta}_t} \tau_i^{ \eta}$. The process $X^{}_{\varphi^{ \eta}_t}$ may also be seen as a renewal-reward process, but with a reward that takes the values $-\eta$ and $\eta$ with equal probability.
The aim of this work is to investigate the asymptotic behavior of $(X^{}_t-X^{}_{\varphi^{ \eta}_t})/\eta$ as $\eta$ approaches $0$. It will be seen that this quantity converges, pointwise for each $t>0$, in distribution to a stochastic variable which is triagularly distributed.
Before we end this section we will state some resluts regarding barrier crossings and renewal processes. The main result is presented in Section~\ref{sec:main_res}. In Section~\ref{sec:num_res} we perform a simulation study and investigate the transition to the limiting distribution.
\subsection{The Wiener process with two absorbing barriers}
Since the components of the sequence $\{\tau^{ \eta}_{i} \}_{i \geq 1}$ are independent and identically distributed, we will let $\tau^{ \eta}$ denote a stochastic variable with the same properties as these $\tau^{ \eta}_{i}$'s, and which may be characterized by
\begin{align*}
\tau^{ \eta} = \inf \{ t>0 \, | \, |X^{}_t| = \eta \} \, .
\end{align*}
Now, consider the process $X^{}$ absorbed in $-\eta$ and $\eta$, that is $X^{}_{t \wedge \tau}$. The transition density of this process, from $X_0=0$, may be represented by (see \citet{Cox_Miller:1965})
\begin{align}
p^{ \eta}(t,x) = \sum_{k=1}^{\infty} \frac{1}{\eta^2} e^{-\frac{1}{2} \left( \frac{k \sigma \pi}{2 \eta} \right)^2 t } \sin \left( \frac{k \pi}{2} \right) \sin \left( \frac{k \pi (x+\eta)}{2 \eta} \right) , \label{eqn:pfct}
\end{align}
for all $(t,x) \in (0,\infty) \times [-\eta,\eta]$. This transition density may also be expressed as an infinite sum over Gaussian kernels (see \citet{Cox_Miller:1965})
\begin{align}
p^{ \eta}(t,x) = \sum_{k=-\infty}^{\infty} \frac{1}{\sqrt{2 \pi \sigma^2 t}} \left(e^{-\frac{(x-4 k \eta)^2}{2 \sigma^2 t}}-e^{-\frac{(x-2\eta+4 k \eta)^2}{2 \sigma^2 t}} \right) . \label{eqn:pfctG}
\end{align}
for all $(t,x) \in [0,\infty) \times [-\eta,\eta]$.
\begin{lemma} \label{lem:ptrip}
The integral of $p^{1}(t,x)$
\begin{enumerate}
\item[a)] with respect to $t$ over the interval $[a,b] \subset [0,\infty)$ may be represented as
\begin{align*}
\int_a^b p^{ \eta}(t,x) \text{d} t = \sum_{k=1}^{\infty} \int_a^b \frac{1}{\eta^2} e^{-\frac{1}{2} \left( \frac{k \sigma \pi}{2 \eta} \right)^2 t } \sin \left( \frac{k \pi}{2} \right) \sin \left( \frac{k \pi (x+\eta)}{2 \eta} \right) \text{d} t ,
\end{align*}
for all $x \in [-\eta,\eta]$.
\item[b)] with respect to $x$ over the interval $[a,b] \subset [-\eta,\eta]$ may be represented as
\begin{align}
\int_a^b p^{ \eta}(t,x) \text{d} x = \sum_{k=-\infty}^{\infty} \int_a^b \frac{1}{\eta^2} e^{-\frac{1}{2} \left( \frac{k \sigma \pi}{2 \eta} \right)^2 t } \sin \left( \frac{k \pi}{2} \right) \sin \left( \frac{k \pi (x+\eta)}{2 \eta} \right) \text{d} x , \label{eqn:repF}
\end{align}
for all $t \in (0,\infty)$, or as
\begin{align}
\int_a^b p^{ \eta}(t,x) \text{d} x = \sum_{k=-\infty}^{\infty} \int_a^b \frac{1}{\sqrt{2 \pi \sigma^2 t}} \left(e^{-\frac{(x-4 k \eta)^2}{2 \sigma^2 t}}-e^{-\frac{(x-2+4 k \eta)^2}{2 \sigma^2 t}} \right) \text{d} x , \label{eqn:repG}
\end{align}
for all $t \in [0,\infty)$.
\end{enumerate}
\end{lemma}
\begin{proof} \textit{a)} Define the functions $g^F_k$ and $G^F_n$by
\begin{align*}
g^F_k(t,x) = \frac{e^{-\frac{1}{2} \left( \frac{k \sigma \pi}{2 \eta} \right)^2 t }}{\eta^2} \sin \left( \frac{k \pi}{2} \right) \sin \left( \frac{k \pi (x+\eta)}{2 \eta} \right),
\end{align*}
and $G^F_n(t,x)=\sum_{k=1}^n g^F_k(t,x)$ then $\lim_{n \uparrow \infty}G^F_n(t,x)=p^{\eta}(t,x)$. Since $g_k^F(t,0) \geq 0$ it follows that
$$
0 \leq G_{n}^F(t,0) \leq G_{n+1}^F(t,0),
$$
and consequently by Lebesgues monotone convergence thorem
$$
\int_a^b \lim_{n \uparrow \infty} G_n(t,0) \text{d} t = \lim_{n \uparrow \infty} \int_a^b G_n(t,0) \text{d} t.
$$
Extending the integral we get
$$
\lim_{n \uparrow \infty} \int_a^b G_n(t,0) \text{d} t \leq \lim_{n \uparrow \infty} \int_0^\infty G_n(t,0) \text{d} t.
$$
Moving the integral inside of the sum in $G_n(t,0)$ and performing the integration over $\mathbb{R}_+$ we get the sum $\lim_{n \uparrow \infty} \sum_{k=1}^n 8/(k^2 \pi^2 \sigma^2)=4/(3 \sigma^2)$, and hence $\lim_{n \uparrow \infty} \int_0^\infty G_n(t,0) \text{d} t < \infty$. Since $|\sin(k \pi/2) \sin(k \pi (x+\eta)/(2 \eta))| \leq 1$ it holds that $|g_k^F(t,x)| \leq g_k^F(t,0)$ which implies that $|G_n^F(t,x)| \leq G_n^F(t,0)$. Since $G_n^F(t,0)$ is bounded by $\lim_{n \uparrow \infty} G_n^F(t,0)$ the function $G_n^F(t,x)$ is dominated by the integrable function $\lim_{n \uparrow \infty}$ $G_n^F(t,0)$ and by the dominated convergence theorem it follows that
$$
\int_a^b \lim_{n \uparrow \infty} G_n(t,x) \text{d} t = \lim_{n \uparrow \infty} \int_a^b G_n(t,x) \text{d} t.
$$
Moving the integral inside of the sum on the right hand side the claim is proved.
\textit{b)} \textit{Eqn. \eqref{eqn:repF}} From the proof of \textit{a)} we know that $G_n^F(t,x) \leq p^\eta(t,0) = \lim_{n \uparrow \infty}$ $G_n(t,0)$ which is bounded for every $t>0$. Since the set $[a,b]$ is bounded (i.e. $[a,b] \subset [-\eta,\eta]$), the claim now follows from the bounded convergence thorem.
\textit{Eqn. \eqref{eqn:repG}} Define the functions $g_k^G$ and $G_n^G$ by
\begin{align*}
g_k^G(t,x) = \frac{e^{-\frac{(x-k)^2}{2 \sigma^2 t}}}{\sqrt{2 \pi \sigma^2 t}} && \text{and} && G^G_n(t,x)=\sum_{k=-n}^n (g^G_{4k\eta}(t,x)-g^G_{2-4k\eta}(t,x)),
\end{align*}
then $\lim_{n \uparrow \infty} G^G_n(t,x)=$ $p^\eta(t,x)$. The function $G_n^G$ may be decomposed as
$$
G_n^G(t,x)=G_n^{G,1}(t,x)+G_n^{G,2}(t,x),
$$
where
$$
G_n^{G,1}(t,x)=\sum_{k=0}^n (g^G_{4k\eta}(t,x)-g^G_{4k\eta+2}(t,x))
$$
and
$$
G_n^{G,2}(t,x)=\sum_{k=-1}^{-n}(g^G_{4k\eta}(t,x)-g^G_{4k\eta+2}(t,x)).
$$
Since each term in $G_n^{G,1}$ is positive and each term in $G_n^{G,2}$ is negative it holds that
\begin{align*}
0 \leq G_n^{G,1}(t,x) \leq G_{n+1}^{G,1}(t,x) && \text{and} && 0 \geq G_{n}^{G,2}(t,x) \geq G_{n+1}^{G,2}(t,x).
\end{align*}
The claim now follows by Lebesgues monotone convergence theorem.
\end{proof}
\begin{lemma} \label{lem:ptri}
It holds that
\begin{equation*}
\sigma^2 \int_0^{\infty} p^{ 1}(t,x) dt=(1-|x|)^+ \, .
\end{equation*}
\end{lemma}
\begin{proof} From Lemma \ref{lem:ptrip} \textit{a)} we have that
\begin{equation*}
\int_0^{\infty} p^{ 1}(t,x) \text{d}t = \frac{8}{\pi^2 \sigma^2} \sum_{k=1}^{\infty} \frac{1}{k} \sin \left( \frac{k \pi}{2} \right) \frac{1}{k} \sin \left( \frac{ k \pi (x+1)}{2} \right) \, .
\end{equation*}
The idea is to find a function that can be expressed as a series which corresponds to the above sum. Let $s_1=1/2$ and $s_2=(x+1)/2$, then
\begin{equation*}
\frac{\pi^2 \sigma^2}{8} \int_0^{\infty} p^{ 1}(t,x) \text{d}t = \sum_{k=1}^{\infty} \frac{1}{k} \sin \left( k \pi s_1 \right) \frac{1}{k} \sin \left( k \pi s_2 \right) \, .
\end{equation*}
Define the function $h_s$ by
\begin{equation*}
h_s(x) = \left\{
\begin{array}{ll}
0 \, , & 0 \leq |x| \leq s \, , \\
1 \, , & s < |x| \leq 1 \, .
\end{array} \right.
\end{equation*}
The Fourier Cosine coefficients of $h_s$ are given by
\begin{equation*}
\begin{split}
c_0 & = \int_0^1 h_s(x) \text{d}x = 1-s \, , \\
a_k & = 2 \int_0^1 \cos(k \pi x) h_s(x) \text{d}x = -\frac{2 \sin(\pi k s)}{\pi k} \, .
\end{split}
\end{equation*}
Applying Parseval's formula yields
\begin{equation*}
\int_0^1 h_{s_1}(x) h_{s_2}(x) \text{d}x = 2 \sum_{k=1}^{\infty} \frac{\sin(\pi k s_1)}{\pi k} \frac{\sin(\pi k s_2)}{\pi k} + (1-s_1)(1-s_2) \, .
\end{equation*}
Assume that $x \in [0,1]$, then $0 \leq s_1 \leq s_2 \leq 1$ and
\begin{align*}
\sum_{k=1}^{\infty} \frac{2}{\pi^2 k^2} \sin(\pi k s_1) \sin(\pi k s_2) & = 1-s_2-(1-s_1)(1-s_2) \\
& = s_1(1-s_2) \, .
\end{align*}
Thus
\begin{align*}
\sigma^2 \int_0^{\infty} p^{ 1}(t,x) \text{d}t & = 4 \sum_{k=1}^{\infty} \frac{2}{k^2 \pi^2} \sin \left( \frac{k \pi}{2} \right) \sin \left( \frac{k \pi (x+1)}{2} \right) \\
& = (1-x) \, .
\end{align*}
Repeating the argument with $x \in [-1,0]$ yields the result.
\end{proof}
One important property in the theory of renewal processes is that of direct Riemann integrability of a function. A function function $H(\cdot)$ is said to be \textit{directly Riemann integrable} over $[0,\infty)$ if for any $h>0$, the normalized sums
\begin{align*}
h \sum_{n=1}^\infty \inf_{0 \leq \delta \leq h} H(nh-\delta) && \text{and} && h \sum_{n=1}^\infty \sup_{0 \leq \delta \leq h} H(nh-\delta),
\end{align*}
converge to a common finite limit as $h \downarrow 0$ (see chapter 4.4 in \citet{Daley_VereJones:1988}).
\begin{lemma} \label{lem:pdri}
The function $p^1(t,x)$ is directly Riemann integrable with respect to $t$ for each $x \in [-1,1]$.
\end{lemma}
\begin{proof}
We will start by considering the case when $x=0$. The function $p^1(t,0)$ is directly Riemann integrable if $p^1(t,0)$ is nonegative, monotonically decreasing and Lebesgue integrable (see chapter 4.4 in \citet{Daley_VereJones:1988}). Since each term in the representation \eqref{eqn:repF} is nonegative and monotonically decreasing for $x=0$ so is $p^1(t,0)$, and by Lemma \ref{lem:ptri} the integral of $p^1(t,0)$ over $[0,\infty]$ is given by $\int_0^\infty p^1(t,0) \text{d}t=1$ and thus $p^1(t,0)$ is Lebesgue integrable which proves that $p^1(t,0)$ is directly Riemann integrable.
Next let $x \in [-1,1] \setminus \{0\}$. The function $p(t,x)$ is directly Riemann integrable with respect to $t$ if $p^1(t,x) \geq 0$, $p(t,x)$ is uniformly continuous in $t$ and bounded from above by a monotonically decreasing integrable function (see chapter 4.4 in \citet{Daley_VereJones:1988}). Since $p(t,x)$ is a probability distribution for each $t$ it is clear that $p(t,x) \geq 0$. To show uniform continuity we will split the interval $[0,\infty)$ into two parts, say $[0,1]$ and $[1,\infty)$, and show that $p^1(t,x)$ is uniformly continuous on each part. For the interval $[0,1]$ we will use the representation \eqref{eqn:repG}. Let $g_k^G$ and $G_n^G$ be defined as in the proof of Lemma \ref{lem:ptrip}. It is clear that each $g_k^G$ is uniformly continuous in $t$ and thus also $G_n^G$ is uniformly continuous for each $n<\infty$. If we can shown that $G_n^G(t,x)$ for each $x \in [-1,1] \setminus \{0\}$ converges uniformly with respect to $t$ over $[0,1]$ as $n \uparrow \infty$, then also the limit $p(t,x)$ will be uniformly continuous. Rewrite $G_n^G$ as $G_n^G(t,x)=\sum_{k=0}^{n} \tilde{g}_k^G(t,x)$ where $\tilde{g}_0^G(t,x)=g^G_{0}(t,x)-g^G_{2}(t,x)$ and
\begin{align*}
\tilde{g}_k^G(t,x) = g^G_{4k}(t,x)-g^G_{2-4k}(t,x)+g^G_{-4k}(t,x)-g^G_{2+4k}(t,x), \text{ for } k \geq 1.
\end{align*}
According to Weierstrass M-test, if there is a series of constants $M_k$ such that $\sum_{k=0}^{\infty} M_k$ is convergent and $|\tilde{g}_k^G(t,x)| \leq M_k$ for all $t \in [0,1]$ then $G_n^G$ converges uniformly in $[0,1]$ as $n \uparrow \infty$. The functions $g_k(t,x)$ attains its maximum at $t=(x-k)^2/\sigma^2 \wedge 1$ for $t \in [0,1]$, and thus $g_k(t,x) \leq g_k((x-k)^2/\sigma^2 \wedge 1,x)$. The function $g_0(x^2/\sigma^2 \wedge 1,x)$ is bounded and it is easily seen that the functions $g_k^G$ may be bounded by $C/(1+k^2)$, for some bounded constant $C$, and which is clearly convergent. Hence, for each $x \in [-1,1] \setminus \{0\}$, $p(\cdot,x)$ is uniformly continuous in $[0,1]$. To show uniform continuity in $[1,\infty)$ we will use the representation \eqref{eqn:repF}. Let $t \geq 1$, then
\begin{align*}
|p^1(t+\delta,x)-p^1(t,x)| & \leq \sum_{k=1}^{\infty} e^{-\frac{k^2 \sigma^2 \pi^2}{8}t} |e^{-\frac{k^2 \sigma^2 \pi^2}{8}\delta}-1| \\
&\leq \sum_{k=1}^{\infty} \frac{8^2}{k^4 \sigma^4 \pi^4} \frac{k^2 \sigma^2 \pi^2}{8} \delta = \delta \frac{3}{4 \sigma^2}
\end{align*}
where we used the inequalites $e^{-y} \leq y^{-2}$ and $|e^{-y}-1|\leq y$ which holds for $y \geq 0$. Hence for every $\epsilon>0$ we may chose $\delta$ such that $\delta < 4 \sigma^2 \epsilon/3$ which holds for every $t$ in $[1,\infty)$. Hence $p^1(\cdot,x)$ is also uniformly continuous in $[1,\infty)$, which together with the previous result yields that $p(\cdot,x)$ is uniformly continuous in $[0,\infty)$. In the proof of Lemma \ref{lem:ptrip} we showed that $p(t,x) \leq p(t,0)$, and that $p(t,0)$ is a monotonically decreasing Lebesgue integrable function. Hence, $p(t,x)$ is also directly Riemann integrable with respect to $t$ for $x \in [-1,1] \setminus \{0\}$, which together with the first result of this proof yeilds that $p(t,x)$ is directly Riemann integrable for $x \in [-1,1]$.
\end{proof}
The next two lemmas regards properties of the random variable $\tau^{\eta}$ defined earlier in this section. Let $F_{\tau^{\eta}}$ denote the distribution function of $\tau^{\eta}$. Lemma \ref{lem:taudist} states that that $\tau^{\eta}$ has a density, which we will denote by $f_{\tau^{\eta}}$.
\begin{lemma} \label{lem:Etau}
The expectation of $\tau^{ \eta}$ is given by $E[\tau^{ \eta}] = \eta^2/\sigma^2$ .
\end{lemma}
\begin{proof} Let $g(x_0)=E[\tau^{ \eta}]$, where $x_0$ denotes the initial point of the process. The function $g$ satisfies the following ordinary differential equation (see \citet{Cox_Miller:1965})
\begin{align*}
\frac{\sigma^2}{2} \frac{d^2 g}{d x_0^2}(x_0) = -1 , \quad m_1(-\eta)=m_1(\eta)=0 \, .
\end{align*}
The solution to this problem, with $x_0=0$, is given by $g(0)=\eta^2/\sigma^2$, as was to be shown.
\end{proof}
\begin{lemma} \label{lem:taudist}
The random variable $\tau^\eta$ has a density, denoted by $f_{\tau^\eta}$, that may be represented as
\begin{multline*}
f_{\tau^\eta}(t) = \sum_{k=-\infty}^{\infty} \frac{1}{2 t \sqrt{2 \pi \sigma^2 t}} \Biggl( (\eta+4k\eta) e^{-\frac{(\eta+4k\eta)^2}{2 \sigma^2 t}}-(\eta+2-4k\eta)e^{-\frac{(\eta+2-4k\eta)^2}{2 \sigma^2 t}} \\
+(\eta-4k\eta) e^{-\frac{(\eta-4k\eta)^2}{2 \sigma^2 t}}-(\eta-2+4k\eta) e^{-\frac{(\eta-2+4k\eta)^2}{2 \sigma^2 t}} \Biggr),
\end{multline*}
for all $t \in [0,\infty)$.
\end{lemma}
\begin{proof}
In this proof we will use the representation \eqref{eqn:repG}. Let $g_k^G$ and $G_n^G$ be defined as in the proof of Lemma \ref{lem:ptrip}. By the use of Lemma \ref{lem:ptrip} for $t \in [0,\infty)$
\begin{align*}
P(\tau^{\eta} \leq t) = 1-\sum_{k=-\infty}^{\infty} \int_{-\eta}^{\eta} (g_{4k\eta}^G(t,x)-g_{2-4k\eta}^G(t,x)) \text{d} x .
\end{align*}
If each term in the sum above is differentiable on $[0,\infty)$ and
\begin{align}
\sum_{k=-\infty}^{\infty} \frac{d}{d t} \int_{-\eta}^{\eta} (g_{4k\eta}^G(t,x)-g_{2-4k\eta}^G(t,x)) \text{d} x \label{eqn:dPdtsum}
\end{align}
converges uniformly on $[0,\infty)$ then
\begin{align*}
\frac{d}{dt} P(\tau^{\eta} \leq t) = -\sum_{k=-\infty}^{\infty} \frac{d}{d t} \int_{-\eta}^{\eta} (g_{4k\eta}^G(t,x)-g_{2-4k\eta}^G(t,x)) \text{d} x .
\end{align*}
Calculating the integral and differentiating with respect to $t$ we get for each term in \eqref{eqn:dPdtsum}
\begin{equation}
\begin{split}
& \frac{d}{d t} \int_{-\eta}^{\eta} (g_{4k\eta}^G(t,x)-g_{2-4k\eta}^G(t,x)) \text{d} x \\
& = \frac{1}{2 t \sqrt{2 \pi \sigma^2 t}} \Biggl( (\eta+4k\eta) e^{-\frac{(\eta+4k\eta)^2}{2 \sigma^2 t}}-(\eta+2-4k\eta)e^{-\frac{(\eta+2-4k\eta)^2}{2 \sigma^2 t}} \\
& \qquad \qquad \qquad +(\eta-4k\eta) e^{-\frac{(\eta-4k\eta)^2}{2 \sigma^2 t}}-(\eta-2+4k\eta) e^{-\frac{(\eta-2+4k\eta)^2}{2 \sigma^2 t}} \Biggr), \label{eqn:dPdtterm}
\end{split}
\end{equation}
The maximum of the function $e^{-\frac{(x-k)^2}{2 \sigma^2 t}}/t^{3/2}$ in $[0,\infty)$ is attained at $t=(x-k)^2/(3 \sigma^2)$. For the first term in the expression above we get that
\begin{align*}
\frac{(\eta+4k\eta)}{2 t^{3/2} \sqrt{2 \pi \sigma^2}}e^{-\frac{(\eta+4k\eta)^2}{2 \sigma^2 t}} \leq \left(\frac{3}{2}\right)^{3/2} \frac{\sigma^2 e^{-\frac{3}{2}}}{\eta^2 \sqrt{\pi}} \frac{1}{(1+4k)^{2}},
\end{align*}
which may be bounded by $C/(1+k^2)$, where $C$ is a bounded constant. In a similar manner it can be shown that the rest of the terms in \eqref{eqn:dPdtterm} may also be bounded by $C/(1+k^2)$, and thus
\begin{equation}
\left| \frac{d}{d t} \int_{-\eta}^{\eta} (g_{4k\eta}^G(t,x)-g_{2-4k\eta}^G(t,x)) \text{d} x \right| \leq \frac{4C}{1+k^2}
\end{equation}
Since $\sum_{k=-\infty}^{\infty} 4C/(1+k^2)$ is a convergent series by Wierstrass M-test the sum \eqref{eqn:dPdtsum} converges uniformly on $[0,\infty)$, and hence, the density, $f_{\tau^{\eta}}$, may be represented by the sum \eqref{eqn:dPdtterm}. Since the terms in the sum of \eqref{eqn:dPdtterm} could be bounded by $4C/(1+k^2)$ we have that $|f_{\tau^{\eta}}(t)| \leq 4 C \sum_{k=-\infty}^\infty 1/(1+k^2)<\infty$ which shows that $f_{\tau^{\eta}}(t)$ is bounded in $[0,\infty)$.
\end{proof}
\subsection{Renewal processes}
In this paragraph we will focus on a renewal process denoted by $N$ with idenpendent and identically distributed interarrival times $\{ \tau_i \}_{i \geq 1}$. Define the renewal function $M$ by $M_t=E[N_t]$, and let $\mu$ denote the mean time between renewals, that is $\mu = E[\tau_i]$, which holds for all $i \geq 1$. Next, we will state the key renewal theorem that will be needed later on.
\begin{lemma}[Key renewal theorem] \label{lem:krt}
If $H(\cdot)$ is a directly Riemann-integrable function then
\begin{align*}
\lim_{t \rightarrow \infty} \int_0^t H(t-x) dM(x) = \frac{1}{\mu} \int_0^{\infty} H(x) dx \, .
\end{align*}
\end{lemma}
\begin{proof}
See e.g. \citet{Daley_VereJones:1988}.
\end{proof}
Let $F_{\tau}$ denote the common distribution function of the stochastic variables $\tau_i$. Since the components of $\{ \tau_i \}_{i\geq 1}$ are idependent and identically distributed the distribution function of the sum $\sum_{i=1}^k \tau_i$ may be represented by the $k$-fold convolution of $F_{\tau}$ (here denoted $F^{*k}_{\tau}$), i.e.
\begin{align*}
P\left(\sum_{i=1}^k \tau_i < t \right) = F^{*k}_{\tau}(t) \, .
\end{align*}
\begin{lemma}[Theorem 5.4 in \citet{Heyman_Sobel:1982}]
There exists a one-to-one correspondence between $F_{\tau}$ and $M$, and $M$ has the representation
\begin{align*}
M_t = \sum_{k=1}^{\infty} F^{*k}_{\tau}(t) \, .
\end{align*}
\end{lemma}
Under the assumption that $F_{\tau}$ has a density (here denoted $f_{\tau}$) we have that
\begin{align*}
f^{*k}_{\tau}(t) = \frac{d}{dt} F^{*k}_{\tau}(t) \, ,
\end{align*}
where $f^{*k}_{\tau}$ is the $k$-th convolution of the density function $f_{\tau}$. We may now define the renewal density $m$ by
\begin{align}
m_t := \frac{d}{dt} M_t = \sum_{k=1}^{\infty} f^{*k}_{\tau}(t) \, . \label{eqn:m_expr}
\end{align}
\section{Main result} \label{sec:main_res}
In this section we state and prove the main result of this paper. To ease the notation in the proof we will let $Z^{ \eta}_t=X^{ \eta}_t-X^{ \eta}_{\varphi^{ \eta}_t}$.
\begin{theorem} \label{thm:main}
Fix a point $t>0$, then
\begin{align*}
\frac{1}{\eta} \left( X^{}_t-X^{}_{\varphi^{ \eta}_t} \right) \stackrel{d}{\longrightarrow} \Lambda \quad \text{as } \eta \rightarrow 0 \, ,
\end{align*}
where $\Lambda$ is a stochastic variable with density function given by
\begin{equation*}
f_{\Lambda}(z) = (1-|z|)^+ \, .
\end{equation*}
\end{theorem}
\begin{proof}
Denote by $Y^{ \eta}_t(u)$ the quantity
\begin{equation*}
Y^{ \eta}_t(u)=X^{}_t-X^{}_{\varphi^{ \eta}_t}| \{ t-\varphi^{ \eta}_t=u \} \, .
\end{equation*}
Because of the time homogeneity of the process $X^{}$ the following equality in distribution holds
\begin{equation*}
X^{}_t-X^{}_{\varphi^{ \eta}_t}|\{ t-\varphi^{ \eta}_t=u \} \stackrel{d}{=} X^{}_u|\{ | X^{}_s | < \eta , \, 0 \leq s \leq u \} \, .
\end{equation*}
Consequently the density function of $Y^{ \eta}_t(u)$ can be expressed as
\begin{equation*}
f_{Y^{ \eta}_t(u)}(y)=\frac{p^{ \eta}(u,y)}{P(\tau^{ \eta} > u)} \, ,
\end{equation*}
The distribution function of $Z^{ \eta}_t$ is given by
\begin{equation*}
f_{Z^{ \eta}_t}(z) = \int_0^t f_{Y^{ \eta}_t(u)}(z) \text{d}F_{t-\varphi^{ \eta}_t}(u) \, ,
\end{equation*}
where
\begin{align*}
& \text{d}F_{t-\varphi^{ \eta}_t}(u) = \left\{ \delta(u-t) P(\tau^{ \eta}>t)+\sum_{k=1}^{\infty} \frac{\partial}{\partial u} P(t-\varphi^{ \eta}_t \leq u, \, N^{ \eta}_{t}=k) \right\} \text{d}u \, .
\end{align*}
The probability in the last term of the above expression can be rewritten as
\begin{equation*}
\begin{split}
P(t-\varphi^{ \eta}_t \leq u, \, N^{ \eta}_{t}=k) & = P \left( t - \sum_{j=1}^{k} \tau^{ \eta}_{j} \leq u , \, \sum_{j=1}^{k} \tau^{ \eta}_{j} < t < \sum_{j=1}^{k} \tau^{ \eta}_{j} + \tau^{ \eta}_{k+1} \right) \\
& = P \left( t - \sum_{j=1}^{k} \tau^{ \eta}_{j} \leq u , \, 0 < t - \sum_{j=1}^{k} \tau^{ \eta}_{j} < \tau^{ \eta}_{k+1} \right) \\
& = \int_{t-u}^{\infty} \int_{t-v}^{\infty} f_{\tau^{ \eta}}^{* k} (v) f_{\tau^{ \eta}}(z) \text{d}z \, \text{d}v \, ,
\end{split}
\end{equation*}
where $f_{\tau^{ \eta}}$ (which exists due to Lemma \ref{lem:taudist}) is the density function of $\tau^{ \eta}$, and $f^{*k}_{\tau^{ \eta}}$ denotes the $k$-th convolution of $f_{\tau^{ \eta}}$. Differentiating the above expression with respect to $u$ yields
\begin{equation*}
\begin{split}
\frac{\partial}{\partial u} \left( \int_{t-u}^{\infty} \int_{t-v}^{\infty} f_{\tau^{ \eta}}^{* k} (v) f_{\tau^{ \eta}}(z) dz \, dv \right) & = \int_{u}^{\infty} f_{\tau^{ \eta}}^{* k} (t-u) f_{\tau^{ \eta}}(z) \text{d}z \\
& = f_{\tau^{ \eta}}^{* k} (t-u) P(\tau^{ \eta}>u) \, .
\end{split}
\end{equation*}
This gives us that
\begin{equation*}
\text{d}F_{t-\varphi^{ \eta}_t}(u) = \left\{ \delta(u-t) P(\tau^{ \eta}>t)+\sum_{k=1}^{\infty} f_{\tau^{ \eta}}^{* k} (t-u) P(\tau^{ \eta}>u) \right\} \text{d}u \, .
\end{equation*}
Using the scaling property of the Brownian motion the following two relations are easily deduced
\begin{align*}
P(\tau^{ \eta}>t)=P(\tau^{ 1}>t/\eta^2) \quad \text{and} \quad \frac{1}{\eta} Y^{ \eta}_t(u) \stackrel{d}{=} Y^{ 1}_t(u/\eta^2) \, .
\end{align*}
The first of the two relations above yields
\begin{align*}
f_{\tau^{ \eta}}(t)=-\frac{\text{d}}{\text{d}t}P(\tau^{ \eta}>t)=-\frac{\text{d}}{\text{d}t}P(\tau^{ 1}>t/\eta^2) = \frac{1}{\eta^2}f_{\tau^{ 1}}(t/\eta^2) \, ,
\end{align*}
and consequently
\begin{align*}
& \text{d}F_{t-\varphi^{ \eta}_t}(u) = \left\{ \delta(u-t) P(\tau^{ 1}>t/\eta^2)+\sum_{k=1}^{\infty} \frac{1}{\eta^2} f_{\tau^{ 1}}^{* k} \left( \frac{t-u}{\eta^2} \right) P(\tau^{ 1}>u/\eta^2) \right\} \text{d}u \, .
\end{align*}
The relation $Y^{ \eta}_t(u)/\eta \stackrel{d}{=} Y^{ 1}_t(u/\eta^2)$ yields
\begin{align*}
& \int_0^t f_{Y^{ \eta}_t(u)/\eta}(y) \text{d}F_{t-\varphi^{ \eta}_t}(u) = \int_0^t f_{Y^{ 1}_t(u/\eta^2)}(y) \text{d}F_{t-\varphi^{ \eta}_t}(u) \, ,
\end{align*}
and thus
\begin{align*}
f_{Z^{ \eta}_t/\eta}(z) & = \int_0^t \frac{p^{ 1}(u/\eta^2,z)}{P(\tau^{ 1}>u/\eta^2)} \delta(u-t) P(\tau^{ 1}>t/\eta^2) \text{d}u \\
& \quad + \int_0^t \frac{p^{ 1}(u/\eta^2,z)}{P(\tau^{ 1}>u/\eta^2)} \sum_{k=1}^{\infty} \frac{1}{\eta^2} f_{\tau^{ 1}}^{* k} \left( \frac{t-u}{\eta^2} \right) P(\tau^{ 1}>u/\eta^2) \text{d}u \\
& = p^{ 1}(t/\eta^2,z) + \int_0^t p^{ 1}(u/\eta^2,z) \sum_{k=1}^{\infty} \frac{1}{\eta^2} f_{\tau^{ 1}}^{* k} \left( \frac{t-u}{\eta^2} \right) \text{d}u \, .
\end{align*}
Now, by a change of variables ($v=(t-u)/\eta^2$)
\begin{align*}
f_{Z^{ \eta}_t/\eta}(z) = p^{ 1}(t/\eta^2,z) + \int_0^{t/\eta^2} p^{ 1} \left( \frac{t}{\eta^2}-v , z \right) \sum_{k=1}^{\infty} f_{\tau^{ 1}}^{* k} \left( v \right) \text{d}v \, .
\end{align*}
Since
\begin{align*}
|p^1(t/\eta^2,x)| \leq \sum_{k=1}^{\infty} \frac{8 \eta^2}{k^2 \sigma^2 \pi^2}=\frac{4 \eta^2}{3 \sigma^2},
\end{align*}
we have that $\lim_{\eta \rightarrow 0} p^{ \eta}(y,t/\eta^2) = 0$. For the second term we have using \eqref{eqn:m_expr}, Lemma \ref{lem:Etau} and Lemma \ref{lem:krt} (which applicable since $p(t,x)$ is a directly Riemann integrable function due to Lemma \ref{lem:pdri})
\begin{align*}
\lim_{\eta \rightarrow 0} \int_0^{t/\eta^2} p^{ 1} \left( y,\frac{t}{\eta^2}-v \right) \sum_{k=1}^{\infty} f_{\tau^{ 1}}^{* k} \left( v \right) dv = \sigma^2 \int_0^{\infty} p^{ 1}(y,u) \text{d}u\, .
\end{align*}
Now by Lemma \ref{lem:ptri}
\begin{equation*}
\lim_{\eta \rightarrow 0} f_{Z^{ \eta}_t/\eta}(z) = (1-|z|)^+ \, ,
\end{equation*}
as was to be shown.
\end{proof}
\begin{remark}
Note that the limiting distribution does not depend on $\sigma$. This is unlike the case when discretization takes place on an equidistant grid, where $\sigma$ affects the variance of the limiting distribution. Instead, in the case of adaptive approximation, $\sigma$ is related to the expected number of discretization points.
\end{remark}
\begin{remark}
In the proof above all interarrival times $\tau_i^{ \eta}$ up to the time $t$ is used in order to characterize the distribution of $t-\varphi_t^{ \eta}$. However, we belive that $t-\varphi_t^{ \eta}$ may be characterized by the dynamics of the process $X$ in a small region around $X_t$, which would imply that the result of Theorem \ref{thm:main} is a local result. From this and the fact that the diffusion coefficient $\sigma$ is scaled away in the limiting expression of the distribution we conjecture that Theorem \ref{thm:main} would hold for a larger class of stochastic processes such as SDE's. We plan to address this in future research.
\end{remark}
\section{Numerical results} \label{sec:num_res}
In this section the transition of $f_{Z^{ \eta}_t/\eta}$ as $\eta$ goes from some large value towards zero is investigated. We will argue that for large values of $\eta$ the stochastic variable $Z^{ \eta}_t/\eta$ is approximately normally distributed, and thus as $\eta$ approaches zero we will see that $f_{Z^{ \eta}_t/\eta}$ goes from the density of a normally distributed random variable to the density of a triangularly distributed random variable.
A total of $50000$ trajectories of the process $X$ was simulated, over a period from $t=0$ to $t=0.5$, with $\sigma=1$, on a time grid with $200001$ equally spaced points. Trajectories of the approximation $X_{\varphi^{ \eta}_t}$ were calculated for a number of different values of $\eta$ in the range $[0.5,4.0]$.
Recall, from the proof of Theorem \ref{thm:main}, the expression of the density
\begin{align}
f_{Z^{ \eta}_t/\eta}(z) = p^{ 1}(t/\eta^2,z) + \int_0^{t/\eta^2} p^{ 1} \left( \frac{t}{\eta^2}-v,z \right) \sum_{k=1}^{\infty} f_{\tau^{ 1}}^{* k} \left( v \right) dv \, . \label{eqn:pdf_Zeta}
\end{align}
It is clear that for large values of $\eta$ it is the first term in \eqref{eqn:pdf_Zeta} that is the dominant one. Thus, in this case the density is approximately the same as the absorbed Wiener process. Furthermore, since $\eta$ was assumed to be large the density of the absorbed Wiener process is approximately the same as the Wiener process without absorbing barriers. Hence, for large $\eta$ we have that
\begin{align}
f_{Z^{ \eta}_t/\eta}(z) \approx \frac{\eta}{ \sigma \sqrt{t}} \phi\left( z \frac{\eta}{ \sigma \sqrt{t} } \right) \, , \label{eqn:fnorm}
\end{align}
where $\phi$ denotes the standard normal density function.
In Figure \ref{fig:pdf} the density of $f_{Z^{ \eta}_t/\eta}$, at $t=0.5$, as we let $\eta$ go from $4.0$ to $0.5$ is depicted. It is seen that when $\eta=4.0$ the distribution is quite close to the normal distribution. For $\eta=0.5$ the distribution on the other hand is quite close to the triangular distribution.
\begin{figure}
\centerline{
{\psfrag{x}[bc][bc]{$z$}
\psfrag{pdf}[bc][bc]{$\text{pdf}$}
\includegraphics[width=0.5\textwidth]{FigPdf2.eps}\includegraphics[width=0.5\textwidth]{FigPdf1.eps}}}
\caption{Left (large values of $\eta$): kernel estimates of $f_{Z^{ \eta}_t/\eta}(z)$ where $\eta=4.0$ (dotted line), $\eta=3.25$ (dash-dotted line) and $\eta=2.5$ (dashed line), and the Gaussian distribution (solid line). Right (small values of $\eta$): kernel estimates of $f_{Z^{ \eta}_t/\eta}(z)$ where $\eta=2.5$ (dashed line), $\eta=2.0$ (dash-dotted line) and $\eta=0.5$ (dotted line), and the triangular distribution (solid line). }
\label{fig:pdf}
\end{figure}
To further illustrate the transition from the normal distribution to the triangular distribution we measured the distance in therms of the Wasserstein metric between the, from the Monte Carlo simulation, estimated distribution and these two distributions. The distance between two distributions, with distribution functions $F$ and $G$, in terms of the Wasserstein metric is defined by
\begin{align*}
d_W(F,G) = \int_{\mathbb{R}} |F(x)-G(x)| dx \, .
\end{align*}
In Figure \ref{fig:dist} the Wasserstein distance between the empirical distribution and the triangular distribution as well as the distance between the empirical distribution and the normal distribution \eqref{eqn:fnorm}, at $t=0.5$, as a function of $\eta$ is depicted. Note that in the case of the normal distribution \eqref{eqn:fnorm} not only the empirical distribution but also the normal distribution that we compare with is dependent of $\eta$. It is seen that for $\eta$ smaller than $1.25$ the empirical distribution is relatively close to the triangular distribution whereas for values over $2.25$ it is close to the normal distribution \eqref{eqn:fnorm}. For $\eta$ in the interval $(1.25,2.25)$ the distribution is probably better explained by a mixture of the two distributions. The small offset from zero for small values of the distance is due to the variance of the monte carlo simulation.
\begin{figure}
\centerline{
{\psfrag{eta}[bc][bc]{$\eta$}
\psfrag{dW}[bc][bc]{$d_W$}
\includegraphics[width=0.5\textwidth]{FigDist.eps}}}
\caption{Distance in terms of the Wasserstein metric between the triangular distribution and the empirical distribution (squares), and the normal distribution \eqref{eqn:fnorm} and the empirical distribution (circles).}
\label{fig:dist}
\end{figure}
\begin{figure}
\centerline{
{\psfrag{time}[bc][bc]{Time}
\psfrag{variance}[bc][bc]{Variance}
\includegraphics[width=0.5\textwidth]{FigVar.eps}}}
\caption{The variance of $Z^{ \eta}_t/\eta$ as a function of time where $\eta=0.50$ (dotted line), $\eta=0.75$ (thin dash-dotted line), $\eta=1.00$ (thin dashed line), $\eta=1.50$ (thick dash-dotted line) and $\eta=2.25$ (thick dashed line), together with the function $(t/0.5^2)\wedge(1/6)$ (solid line).}
\label{fig:var}
\end{figure}
From \eqref{eqn:pdf_Zeta} it is clear that it is possible to fix $\eta$ and instead of letting $\eta$ approach zero let $t$ approach infinity. To capture this we have plotted the variance of $Z^{ \eta}_t/\eta$ as a function of $t$ for a couple of different values of $\eta$ (see Figure \ref{fig:var}). The constant $1/6$, that is the value of the variance of the triangularly distributed random variable, is also plotted in the figure. As expected it is seen that for low values of $\eta$ the limiting variance of $1/6$ is attained much faster than for higher values of $\eta$. From the argumentation above regarding high values of $\eta$ it is also clear that for low values of $t$ the distribution is approximately normal. Hence, the slope of the lines near zero is given by $1/\eta^2$, as is seen in the figure.
|
\section{Introduction}
Important progress has been made in the last few years to unveil the origin of the soft X-ray emission in obscured Active Galactic Nuclei (AGN). The first breakthrough was represented by high resolution spectra made available thanks to the gratings aboard \textit{Chandra} and XMM-\textit{Newton}. The `soft excess' observed in CCD spectra was found to be due to the blending of bright emission lines, mainly from He- and H-like transitions of light metals and L transitions of Fe, with low or no continuum, in most Seyfert 2 galaxies \citep[see e.g.][]{sako00b,Sambruna01b,kin02,gb07}. Spectral diagnostic tools agree that the observed lines should be produced in a gas photoionised by the AGN, with little contribution from any collisionally ionised plasma. A second breakthrough was made possible thanks to the unrivaled spatial resolution of \textit{Chandra}. The soft X-ray emission of Seyfert 2 galaxies appears to be morphologically correlated with that of the Narrow Line Region (NLR), as mapped by the [{O\,\textsc{iii}}] $\lambda 5007$ \textit{HST} images \citep[e.g.][]{yws01,iwa03,bianchi06}. Since the NLR is also believed to be a gas photoionised by the AGN, it was shown that a very simple model where the soft X-ray emission and the NLR emission are produced in the same material is possible \citep[e.g.][]{bianchi06}.
However, this scenario is clearly oversimplified. In particular, it is not clear whether different components of the same medium are spatially separated, possibly radially distributed, or co-exists at each radius, as a result of stratification. Moreover, the role of radio ejecta, which are often found to be strongly correlated with the morphology of the NLR, is still unclear. An exciting possibility to shed some more light on this issue is provided by spatially resolved X-ray spectroscopy, which is, unfortunately, limited by the compact structures of this class of objects (at most some arcsec). Indeed, such a study was performed with high resolution spectroscopy on only one source, NGC~1068, being very bright and extended \citep{brink02}. The results are in agreement with the expectations from a cone of plasma, irradiated by the central AGN. On the other hand, a similar analysis, but with CCD resolution, was performed on NGC~7582, showing that there are regions with a further source of ionisation, or lower density \citep{bianchi07b}.
Mrk~573 (a.k.a. UGC~1214, z=0.0172) is the third [{O\,\textsc{iii}}] brightest source in the \citet{schm03} sample of nearby Seyfert galaxies observed by \textit{HST}, only fainter than NGC~1068 and Mrk~3. Moreover, the NLR extension of Mrk~573 is by far the largest of the sample, reaching a total extent of 3 kpc, with a projected extension almost 9 arcsec wide. A triple radio source is associated to the galaxy, composed by a central core and two spots \citep{uw84}. A detailed analysis of the NLR of this source was performed, among others, by \citet{ferr99} and \citet{schl09}. They conclude that photoionisation by the central AGN is likely the dominant process, although the interaction with the radio jets must be taken into account in the overall scenario, introducing kinematic disturbances and shaping the NLR morphology.
Despite the brightness and the interesting features of Mrk~573, the source has never been studied in detail in X-rays. After the detection by \textit{Einstein}, Mrk~573 was observed by \textit{ROSAT}, which clearly detected an emission extended over about 10 arcsec, in agreement with the NLR dimensions. The source was then only observed by XMM-\textit{Newton} with a short exposure time of about 10 ks. It confirmed to be rather bright in the 0.5-2 keV band ($\simeq3\times10^{-13}$ cgs) and the detection of a very strong iron line, with an EW larger than 1 keV, revealed its nature as a Compton-thick source \citep{gua05b}. But the most important piece of information comes from the RGS high resolution soft X-ray spectrum, which clearly appears dominated by strong emission lines. The predominance of K lines, the ratio of the components of the {O\,\textsc{vii}} triplet and the detection of strong, narrow, radiative recombination continua (RRC) features all contribute to an interpretation in terms of photoionised gas \citep{gb07}.
In this paper, we re-analysed in detail the XMM-\textit{Newton} RGS spectrum of Mrk~573, adopting a self-consistent photoionisation model. Moreover, we present for the first time the imaging and spectral analysis of a \textit{Chandra} observation.
\section{Observations and data reduction}
In the following, errors correspond to the 90 per cent confidence level for one interesting parameter ($\Delta \chi^2 =2.71$), where not otherwise stated. The adopted cosmological parameters are $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_\Lambda=0.73$ and $\Omega_m=0.27$ \citep[i.e. the default ones in \textsc{xspec 12.5.1}:][]{xspec}. At the distance of Mrk~573, 1 arcsec corresponds to 360 pc. In all the fits, the Galactic column density along the line of sight to Mrk~573 is included \citep[$\mathrm{N_H}=2.96\times10^{20}$ cm$^{-2}$:][]{dl90}.
\subsection{X-rays: \textit{Chandra} and \textit{XMM-Newton}}
Mrk~573 was observed by \textit{Chandra} on 2006-11-18 for a total exposure time of 40 ks (obsid 7745), with the Advanced CCD Imaging Spectrometer \citep[ACIS:][]{acis}. Data were reduced with the Chandra Interactive Analysis of Observations \citep[CIAO:][]{ciao} 4.1 and the Chandra Calibration Data Base (CALDB) 4.1.2 software, adopting standard procedures. Images were corrected for known aspect offsets, reaching a nominal astrometric accuracy of 0.6 arcsec (at the 90 per cent confidence level). Spectra were extracted from three different regions: a circular region of 16 arcsec of radius (the default one analysed in Sect.~\ref{chandrafit}); the same region, but excluding the inner 1 arcsec; the nuclear region, i.e. the inner 1 arcsec. The imaging analysis was performed on event files without the pixel randomization and treated with the SER procedure \citep{li03}, in order to improve the positional accuracy. This allowed us to use a pixel size of 0.246 arcsec.
Mrk~573 was also observed by XMM-\textit{Newton} on 2004-01-15 for a total exposure time of $\mathbf{\simeq12}$ ks (obsid 0200430701), with the EPIC CCD cameras, the pn \citep{struder01} and two MOS \citep{turner01}, operated in Prime Full Window and Medium Filter. These data were already presented by \citet{gua05b} and \citet{gb07}. In this paper, we present a new and more detailed analysis of the RGS spectra, which were extracted with standard procedures, using SAS 8.0.1 \citep[last described in][]{sas610} and the most updated calibration files available at the time the data reduction was performed (October 2009). Background spectra were generated using blank field event lists, accumulated from different positions on the sky vault along the mission.
As shown in Fig.~\ref{mrk573field}, in the field there are several bright sources in the soft X-ray band. However, the extraction region (corresponding to the default 90 per cent of the point spread function in the cross-dispersion direction) includes only one contaminating source, named [TUM93]~J014401.5+022106 \citep{turner93}. We extracted its pn spectrum, and calculated a 0.5-0.8 keV flux of $1.6\times10^{-14}$ erg cm$^{-2}$ s$^{-1}$, i.e. a factor of 8 less than Mrk~573. It is unlikely to contaminate the line emission spectrum of Mrk~573, since any line, if present, would be displaced by the order of 1 \AA\ (without considering the effect of its unknown redshift), given its distance of around 1 arcmin from the nucleus. Such lines would therefore be at unidentified wavelengths, but none of them are detected (see Sect.~\ref{rgsanalysis}).
\begin{figure}
\begin{center}
\epsfig{file=mrk573_field.ps, width=0.95\columnwidth}
\end{center}
\caption{\label{mrk573field}XMM-\textit{Newton} EPIC pn image (0.3-2 keV). The RGS extraction region is shown in green.}
\end{figure}
We also extracted the pn spectrum of Mrk~573, already analysed by \citet{gua05b}. However, since the quality of the data of this short observation is lower than that of the \textit{Chandra} ACIS spectrum, we will not discuss it further in this paper.
\subsection{Radio: VLA}
We downloaded from the \textit{VLA} archive the observation of Mrk~573 performed on 1985-03-03 at 6 cm (4860 MHz) in A array configuration. Data reduction was performed with the NRAO software package \textsc{AIPS}, following standard procedures, and very little flagging of bad data points. The phases of the calibrator 0146+056 were interpolated and applied to the object, and 3C~48 was used as the primary flux calibrator. The final map was produced using the \textsc{AIPS} task \textsc{IMAGR} with a beam size of 0$\farcs$5$\times$0$\farcs$4, cleaning depths of several thousand iterations and a CLEAN gain factor of 0.1.
\subsection{\label{hst}Optical and IR: HST}
We retrieved the \textit{HST} observations we use in this paper from the Multimission Archive at STScI (MAST). The images were processed through the standard on-the-fly reprocessing system. Mrk~573 was observed with \textit{HST} with WFPC2 and the FR533N ramp filter as part of the GO program 6332 on 12/11/1995. The target was located in the WF2 camera to image the [{O\,\textsc{iii}}]$\lambda$5007 emission line. The total exposure time
for the two CR-SPLITted exposures is 600s. We combine the two images using the {\it crrej} task on {\it IRAF} to allow for cosmic-ray rejection. The image is clearly dominated by the emission line gas, therefore for the purposes of this work, a careful subtraction the continuum emission is not needed. The object was also observed in the near-IR with HST/NICMOS, as part of GO7867. Images were taken with both the F110W and F160W filters, which are similar to the J and H-bands, respectively, using the NIC1 camera. The exposure time is 1023s for both images. The data were originally presented by \citet{mp99}. The F110W filter includes the Pa$\beta$ emission line, while the F160W filter is relatively free of strong emission lines, at the redshift of Mrk~573. The image at longer wavelengths is clearly
dominated by the unresolved AGN emission. In order to assess the possible presence of structures in the central regions of the galaxy (e.g. starburst regions, dust), we subtract the stellar emission of the host galaxy using a model obtained by fitting ellipses to the galaxy isophotes. The result shows that no extra nuclear structures are
present. This image was therefore used as a reference for all the other images, by aligning the IR nucleus to the soft X-ray and [{O\,\textsc{iii}}] brightest pixels.
\section{Spectral analysis}
\subsection{\label{rgsanalysis}The soft X-ray RGS spectrum}
Following the procedure described in \citet{gb07}, the simultaneous fits on the spectra of the two RGS cameras were performed on $\simeq100$-channel-wide segments, adopting the Cash statistics \citep{cash76}. All the fits include a power law component ($\Gamma$ fixed to 1\footnote{Given the very limited band of these fits, the modellisation of the continuum is insensitive to the power law photon index, which can be frozen at any value.}) and as many emission lines as required (at the 90 per cent confidence level). A total of ten emission lines and two RRC were detected and identified securely (see Table \ref{rgslines}). The widths of all the features were consistent with being unresolved. The RRC are then properly modelled with the \textsc{redge} model, which fully takes into account the recombination edge profile, giving a direct estimate of the electron temperature of the gas.
\begin{table}
\caption{\label{rgslines}Mrk~573: detected emission lines and RRC in the XMM-\textit{Newton} RGS spectrum.}
\begin{center}
\begin{tabular}{lllll}
$\mathbf{E_{o}}$ & \textbf{Line id.} & $\mathbf{E_{th}}$ & kT & \textbf{Flux}\\
&&&\\
$0.5007\pm0.0011$ & {N\,\textsc{vii}} K$\alpha$ & 0.5003 & -- & $1.3^{+1.2}_{-0.9}$\\
$0.5614\pm0.0004$ &{O\,\textsc{vii}} K$\alpha$ (f) & 0.5610 & -- & $5.4^{+2.9}_{-2.1}$\\
$0.5739^{+0.0003}_{-0.0007}$ & {O\,\textsc{vii}} K$\alpha$ (r) & 0.5739 & -- & $3.1^{+2.2}_{-1.5}$\\
$0.6541^{+0.0007}_{-0.0006}$ & {O\,\textsc{viii}} K$\alpha$ & 0.6536 & -- & $2.2^{+1.2}_{-1.0}$\\
$0.6964^{+0.0015}_{-0.0010}$ & {O\,\textsc{vii}} K$\gamma$ & 0.6978 & -- & $0.9^{+0.7}_{-0.5}$\\
\multirow{2}{*}{$0.7266^{+0.0013}_{-0.0015}$} & {Fe\,\textsc{xvii}} M2 & 0.7252 & \multirow{2}{*}{--} & \multirow{2}{*}{$1.2^{+0.8}_{-0.6}$}\\
& {Fe\,\textsc{xvii}} 3G & 0.7272 & &\\
\multirow{2}{*}{$0.7389\pm0.0015$} & {O\,\textsc{vii}} RRC & 0.7393 & $<14$ & \multirow{2}{*}{$1.4^{+1.0}_{-0.7}$}\\
& {Fe\,\textsc{xvii}} 3F & 0.7390 & -- &\\
\multirow{2}{*}{$0.776^{+0.002}_{-0.003}$} & {O\,\textsc{viii}} K$\beta$ & 0.7746 & \multirow{2}{*}{--} & \multirow{2}{*}{$0.6^{+0.6}_{-0.4}$}\\
& {Fe\,\textsc{xviii}} L & 0.7747 & &\\
$0.826\pm0.002$ & {Fe\,\textsc{xvii}} 3C & 0.8257 & -- & $0.4^{+0.5}_{-0.3}$\\
\multirow{3}{*}{$0.8705^{+0.0017}_{-0.002}$} & {Fe\,\textsc{xviii}} L & 0.8626 & -- & \multirow{3}{*}{$1.0^{+0.7}_{-0.5}$}\\
& {O\,\textsc{viii}} RRC & 0.8714 & $<15$ &\\
& {Fe\,\textsc{xviii}} L & 0.8728 & -- &\\
\end{tabular}
\end{center}
Energies are in units of keV, fluxes of $10^{-5}$ ph cm$^{-2}$ s$^{-1}$, kT in eV. Theoretical energies are from CHIANTI \citep{dere97,dere09}. The labelling for {Fe\,\textsc{xvii}} lines follows that of \citet{brown98}.
\end{table}
\begin{figure*}
\begin{center}
\epsfig{file=nvii_contour.ps, width=0.8\columnwidth}
\epsfig{file=ovii_contour.ps, width=0.8\columnwidth}
\epsfig{file=oviii_contour.ps, width=0.8\columnwidth}
\epsfig{file=ovii_gamma_contour.ps, width=0.8\columnwidth}
\epsfig{file=oviirrc_contour.ps, width=0.8\columnwidth}
\epsfig{file=oviii_beta_contour.ps, width=0.8\columnwidth}
\epsfig{file=oviii_rrc_contour.ps, width=0.8\columnwidth}
\end{center}
\caption{\label{rgs}Mrk~573: centroid energy vs. flux contour plots for the detected emission lines in the XMM-\textit{Newton} RGS spectrum (see Table \ref{rgslines}). The contours refers to $\Delta$C=2.30, 4.61 and 9.21, i.e. confidence levels of 68, 90 and 99 per cent for two interesting parameters. Upper limits for the closest lines are also plotted ($\Delta$C=1.00, 2.71 and 4.00).}
\end{figure*}
The resulting temperature ($<15$ eV) is much less than would be expected from the same species, if the gas were in collisional equilibrium. This is a strong piece of evidence in favour of a dominant photoionised phase in the emitting region. However, the {O\,\textsc{vii}} RRC (0.7393 keV) may be significantly contaminated by the 3F component of {Fe\,\textsc{xvii}} L emission (0.7390 keV). This is suggested by the detection of the blend of the 3G and M2 components of the same species, at $\simeq0.7266$ keV. Our simulations with the \textsc{apec} model in \textsc{xspec} showed that the 3F/(3G+M2) ratio never exceeds $\simeq0.44$, for a wide range of temperatures. On the other hand, the observed ratio between the line detected at 0.7389 keV and that at 0.7266 keV is larger, being $1.2^{+0.7}_{-0.5}$. This means that a large part of the observed flux must be due to {O\,\textsc{vii}} RRC. The {O\,\textsc{viii}} RRC may also be contaminated by {Fe\,\textsc{xviii}} L lines (see Table~\ref{rgs}), but no other emission lines from that ion are detected in the spectrum.
In principle, diagnostics on the {O\,\textsc{vii}} triplet may also be a good indicator of the ionisation mechanism of the gas. Indeed, the predominance of the forbidden component is a sign of photoionisation, and the significant detection of the resonant transition a hint that pure recombination is not the only mechanism to produce the emission lines, but photoexcitation has also an important role. This, in turns, suggests that the gas column density should be not too large, in order not to suppress this process. However, given the large uncertainties on the line fluxes, it is impossible to exclude other solutions only on these grounds, like a contribution to the resonant line from a collisionally excited plasma.
Another hint for photoionisation comes from the {O\,\textsc{vii}} K$\alpha$ forbidden line to {O\,\textsc{viii}} K$\alpha$ ratio, which is larger than 1 ($2.5^{+1.9}_{-1.5}$). Together with a large total luminosity of the K$\alpha$ oxygen lines of $6\times10^{40}$ erg s$^{-1}$, this puts Mrk~573 in the photoionisation-dominated locus of the empirical diagnostic plot presented in \citet{gua09}.
Since the phenomenological analysis of the RGS spectrum favours an origin in a photoionised gas, we tried a more physical approach, performing a self-consistent fit on the whole spectrum (limited by the S/N to the $15-26$ \AA\ band).
We produced a grid model for \textsc{xspec} using \textsc{cloudy} 08.00 \citep[last described by][]{cloudy}. The main ingredients are: plane parallel geometry, with the flux of photons striking the illuminated face of the cloud given in terms of ionisation parameter $U$ \citep{of06}; incident continuum modelled as in \citet{korista97}; constant electron density $\mathrm{n_e}=10^5$ cm$^{-3}$; elemental abundances as in Table 9 of \textsc{cloudy} documentation\footnote{Hazy 1 version 08, p. 47: \url{http://viewvc.nublado.org/index.cgi/tags/release/c08.00/docs/hazy1_08.pdf?revision=2342&root=cloudy}}; grid parameters are $\log U=[-0.25:2.00]$, step 0.25, and $\log N_\mathrm{H}=[19.0:23.5]$, step 0.1. Only the reflected spectrum, arising from the illuminated face of the cloud, is taken into account in our model. We also produced tables with different densities ($\mathrm{n_e}=10^3-10^4$ cm$^{-3}$): all the fits presented in this paper resulted insensitive to this parameter, as expected since we are always treating density regimes where line ratios of He-like triplets are insensitive to density \citep{pd00}.
The fit with a single photoionised phase is presented in the upper panel of Fig.~\ref{rgsfit}. With $\log\mathrm{U}=0.6\pm0.3$ and $\log\mathrm{N_H}=20.6\pm0.7$, a good fit is obtained, with most of the lines detected in our phenomenological analysis reasonably modelled, including, notably, the RRCs. However, there are clear residuals at the {Fe\,\textsc{xvii}} 3G+M2 wavelength: these lines are completely missing from the photoionised model\footnote{The \textsc{cloudy} line database was rather inaccurate for the {Fe\,\textsc{xvii}} transition wavelengths and relative atomic parameters. We therefore modified it according to the \textsc{chianti} database.}. This is not surprising, given the very low ionic fraction of {Fe\,\textsc{xvii}} ($\simeq4\times10^{-4}$) at this ionisation parameter.
We therefore tried to add another photoionised phase, with a larger U. However, any attempt to reproduce the observed {Fe\,\textsc{xvii}} lines largely overpredicts the {O\,\textsc{viii}} K$\alpha$ and K$\beta$ lines. This forced us to abandon this approach and to try an alternative scenario, where the iron L lines are mainly produced in a collisional gas. Indeed, the addiction of a collisional phase ($\mathrm{kT}=0.30^{+0.10}_{-0.06}$ keV) perfectly takes into account the {Fe\,\textsc{xvii}} 3G+M2 lines, together with the 3C component at 0.826 keV, without affecting too much the rest of the spectrum. The improvement of the Cash statistics for the addition of the collisional gas component is $\Delta C=18$, with two less degrees of freedom. No further improvement is achieved allowing the elemental abundances to vary. The parameters of the photoionised phase change to $\log\mathrm{U}=0.1^{+0.5}_{-0.7}$ and $\log\mathrm{N_H}=20.5^{+0.9}_{-1.0}$. The residuals of this hybrid (photoionised+collisional) gas are presented in the lower panel of Fig.~\ref{rgsfit} and are rather good. The only exception is represented by some positive residuals for the {O\,\textsc{vii}} K$\gamma$, but they are only significant in the RGS1 spectrum. Indeed, a separate analysis of the two RGS spectra leads to a line detection only in the RGS1 ($2.5\pm1.4\times10^{-5}$ ph cm$^{-2}$ s$^{-1}$ at $0.6995\pm0.0013$ keV, formally inconsistent with {O\,\textsc{vii}} K$\gamma$), while in the RGS2 only an upper limit is measured if the line energy is fixed at the theoretical one ($<0.8\times10^{-5}$ ph cm$^{-2}$ s$^{-1}$). We therefore conclude that the RGS1 detection is insecure, and therefore cannot draw any conclusions based on that only\footnote{We note here that the line wavelength does not correspond to any known defective pixel in neither of the RGS cameras (\url{http://xmm.esac.esa.int/external/xmm_user_support/documentation/uhb/node59.html#3177}).}.
\begin{figure*}
\begin{center}
\epsfig{file=mrk573_rgs12_photofit.ps,width=2\columnwidth}
\epsfig{file=mrk573_rgs12_photo_apec.ps,width=2\columnwidth}
\end{center}
\caption{\label{rgsfit}XMM-\textit{Newton} RGS spectra ($15-26$ \AA) of Mrk~573, rebinned for displaying purposes only. \textit{Top}: Best fit with a single photoionised phase, whose parameters are reported above. \textit{Bottom}: Best fit with a photoionised and a collisional phase, whose parameters are reported above. The brightest emission lines for the adopted models are labelled.}
\end{figure*}
The total flux in the 0.5-0.8 keV band is $1.3\times10^{-13}$ erg cm$^{-2}$ s$^{-1}$. The contribution of the collisional phase is $4.6\times10^{-14}$ erg cm$^{-2}$ s$^{-1}$, i.e. roughly 1/3 of the total flux in this band.
\subsection{\label{chandrafit}The \textit{Chandra} broadband spectrum}
The broadband (0.4-8 keV) \textit{Chandra} X-ray spectrum of Mrk~573 appears dominated by a strong emission in the soft band, plus strong neutral iron K$\alpha$ emission, typical signatures of an highly obscured AGN. The high energy part of the spectrum is well fitted by a pure neutral reflection component plus the iron line. The properties of the reflection component are unconstrained, due to the low statistics, so we fixed the photon index to 1.7 \citep[as in typical Seyfert galaxies, see e.g.][]{bianchi09} and the cosine of the inclination angle to 0.45. The largest normalization compatible with the data leads to an iron EW of $\simeq1.8$ keV, perfectly in agreement with the expectations for a Compton-thick AGN \citep{mbf96}. Since the normalization of the reflection component is basically unconstrained in the following fits, where the soft X-ray component may contribute also to the high energy part of the spectrum (see below), we decided to fix it. The 2-10 keV flux and the iron line flux are in perfect agreement with the values measured with XMM-\textit{Newton} \citep{gua05b}.
To fit the soft X-ray band, at first we adopted the same model which successfully reproduces the RGS spectrum. This model is a good representation of the ACIS spectrum, when limited to the same band, but clearly fails to fit the remaining part of the data, because it cannot reproduce the observed emission from higher Z metals (see left panel of Fig.~\ref{broadband}). We added another photoionised phase to the fit, with a larger ionisation parameter, and allowed also the parameters of the other phases to vary. The resulting fit is very good ($\chi^2=74/79$ d.o.f., see right panel Fig~\ref{broadband} and Table~\ref{broadbandfit}). The best fit model includes two further emission lines, which can be readily identified with neutral Si K$\alpha$ and S K$\alpha$ at 1.740 keV and 2.308 keV, respectively \citep{house69}, likely arising from the same Compton-thick material responsible for the production of the features which dominate the high energy part of the spectrum, that is the Compton reflection component and the neutral iron K$\alpha$ line.
\begin{figure*}
\begin{center}
\epsfig{file=mrk573_acis_rgsfit.ps, width=6cm, angle=-90}
\epsfig{file=mrk573_bestfit.ps, width=6cm, angle=-90}
\end{center}
\caption{\label{broadband}Left: Mrk~573 \textit{Chandra} ACIS-S broadband spectrum, with the best fit for the RGS data in the soft X-ray part, and $\Delta\chi^2$ residuals. Right: The same, but with the best fit.}
\end{figure*}
The best fit parameters for the two photoionised phases are $\log\mathrm{U}=0.3^{+0.3}_{-0.6}$ and $\log\mathrm{N_H}=21.7^{+0.6}_{-0.8}$, and $\log\mathrm{U}=1.81^{+0.15}_{-0.12}$ and $\log\mathrm{N_H}=21.8^{+0.7}_{-0.8}$. The temperature of the collisional gas is $\mathrm{kT}=0.56\pm0.12$ keV.
If the three-phase model is re-applied to the RGS spectra, the fit is visually comparable to the one obtained above directly on these data, with only two phases. Therefore, although only 2 phases (one photoionised, the other collisional) is required by the RGS data, the \textit{Chandra} ACIS data reveals the presence of a further photoionised component, still allowed by the XMM-\textit{Newton} high-resolution spectrum.
The total 0.5-2 keV flux is $2.7\times10^{-13}$ erg cm$^{-2}$ s$^{-1}$, in agreement with the one measured with XMM-\textit{Newton} \citep{gua05b}. The contribution from the collisional phase is $\simeq20\%$, corresponding to an unabsorbed luminosity of $4.0\times10^{40}$ erg s$^{-1}$ in the same band. It is interesting to note that, of the remaining photoionised phases (which contribute roughly equally to the the soft X-ray flux), $\simeq15\%$ is constituted by a continuum component (which includes RRCs and the Thomson-scattered power law), while the remaining flux is in emission lines.
\begin{table}
\caption{\label{broadbandfit}Best fit parameters for the \textit{Chandra} spectrum of Mrk~573. Fluxes are in units of $10^{-13}$ erg cm$^{-2}$ s$^{-1}$, line fluxes in units of $10^{-6}$ ph cm$^{-2}$ s$^{-1}$, energies and kT in keV. See text for details.}
\begin{center}
\begin{tabular}{ll}
& \\
$\log\mathrm{U_1}$& $0.3^{+0.3}_{-0.6}$\\
$\log\mathrm{N_{H1}}$ & $21.7^{+0.6}_{-0.8}$ \\
$\log\mathrm{U_2}$ & $1.81^{+0.15}_{-0.12}$ \\
$\log\mathrm{N_{H2}}$ & $21.8^{+0.7}_{-0.8}$ \\
kT & $0.56\pm0.12$\\
E$_\mathrm{Si K\alpha}$ & 1.740$^*$ \\
F$_\mathrm{Si K\alpha}$ & $1.0\pm0.7$ \\
E$_\mathrm{S K\alpha}$ & 2.308$^*$ \\
F$_\mathrm{S K\alpha}$ & $1.2\pm0.8$ \\
E$_\mathrm{Fe K\alpha}$ & $6.35\pm0.03$ \\
F$_\mathrm{Fe K\alpha}$ & $5.5\pm1.7$ \\
$\chi^2$/dof & 74/79 \\
& \\
$F_{0.5-2}$ & $2.7\pm0.3$ \\
$F_{2-10}$ & $2.8\pm0.5$ \\
& \\
\end{tabular}
\end{center}
$^*$ fixed
\end{table}
\section{Imaging analysis}
The soft X-ray (0.2-2 keV) emission of Mrk~573 is clearly extended and closely resembles the morphology of the NLR, as mapped by the [{O\,\textsc{iii}}] \textit{HST} image (see left panel of Fig.~\ref{o3_softx}). This is often observed in Seyfert 2 galaxies, and clearly suggests a common physical origin for the two emissions \citep[e.g.][]{bianchi06}. Given the lower angular resolution of \textit{Chandra}, it is hard to say how good the correspondence between the X-rays and the NLR is at the smallest scale. However, there is no evidence that significant deviations are present in these data.
\begin{figure*}
\begin{center}
\epsfig{file=oiii_softx.ps, width=\columnwidth}
\epsfig{file=radio_softx.ps, width=\columnwidth}
\end{center}
\caption{\label{o3_softx}Left: \textit{Chandra} soft X-ray (0.2-2 keV) contours superimposed on the \textit{HST} {O\,\textsc{iii}} image. The contours refers to 0.001, 0.1, 0.2, 0.4 and 0.95 levels with respect to the brightest pixel. North is up, east to the left. Right: VLA radio (6 cm) contours superimposed on the \textit{Chandra} soft X-ray (0.2-2 keV) image.}
\end{figure*}
On the other hand, the right panel of Fig~\ref{o3_softx} shows the \textit{VLA} radio emission superimposed on the \textit{Chandra} data. As already reported by \citet{uw84} and \citet{fal98}, Mrk~573 has a unresolved radio core of a diameter size of $\sim$0.32 kpc (1 arcsec) and two-sided jets: the northwest and the southeast bubbles have a similar size of $\sim$0.48$\times$0.32 kpc$^{2}$ (1$\farcs$5$\times$1$\farcs$0). Table \ref{radiotable} reports flux and luminosities for these three radio components. It is clear that, although the inclination of the radio jets is aligned with that of the soft X-ray emission, the overall extension of the radio emission is far more compact.
\begin{table}
\caption{\label{radiotable}Flux (mJy) and luminosity (erg s$^{-1}$ Hz$^{-1}$) at 6 cm for the three radio components detected in Mrk~573.}
\begin{center}
\begin{tabular}{lll}
& & \\
& Flux & $\mathrm{\log L}$ \\
core & $0.95\pm0.03$ & 27.72\\
NW & $2.22\pm0.03$ & 28.09 \\
SE & $4.54\pm0.05$ & 28.40 \\
& & \\
\end{tabular}
\end{center}
\end{table}
We also tried to do some spatially resolved spectroscopy, to look for variations of the properties of the soft X-ray emitting gas along the distance from the nucleus. Following the analysis performed on the RGS high resolution spectra, we chose two narrow energy bands, which we know are dominated by emission from {O\,\textsc{vii}} (0.5-0.6 keV) and {O\,\textsc{viii}} (0.6-0.7 keV)\footnote{By means of simulations with the observed RGS fluxes of the emission lines and the spectral response of the \textit{Chandra} observation, we estimated that the 0.5-0.6 keV band is contaminated by less than 10\% from {O\,\textsc{viii}} photons, while the 0.6-0.7 keV band by around 25\% from {O\,\textsc{vii}} photons.}. As shown in Fig.~\ref{o7_o8}, there is an hint that the {O\,\textsc{vii}} emission is more extended than the {O\,\textsc{viii}} one, but the quality of the data does not allow us to quantify this difference.
\begin{figure*}
\begin{center}
\epsfig{file=o7_o8.ps, width=2\columnwidth}
\end{center}
\caption{\label{o7_o8}Mrk~573: \textit{Chandra} images in the 0.5-0.6 keV ({O\,\textsc{vii}}) and 0.6-0.7 keV ({O\,\textsc{viii}}) band. The images were smoothed with a 3x3 FWHM Gaussian filter and normalised to the brightest pixels. They are re-scaled logarithmically in the same way (see colourbar at the top). The black cross is the brightest pixel of the hard X-ray image. North is up, east to the left.}
\end{figure*}
We also extracted two spectra, one only of the inner 1 arcsec region, the other excluding it, while keeping all the remaining X-ray emitting region. While practically all the flux above 2 keV is concentrated in the nuclear region, approximately 15\% of the 0.5-2 keV flux is produced farther than 1 arcsec from the nucleus. This percentage is significantly lower than what is generally found in Seyfert 2s observed by \textit{Chandra} \citep[see e.g.][]{bianchi06}. The best fit parameters for the three phases of the soft X-ray emission in the nuclear region are consistent with those found for the outer region. There is an hint of an higher relative flux between the collisional and the photoionisation gas in the outer region with respect to the nucleus, but it is not statistically significant.
\section{Discussion}
\subsection{Fluorescence from Compton-thick material}
The high energy spectrum of Mrk~573 is typical of a Compton-thick Seyfert 2 galaxy. While the Compton reflection component and the strong iron K$\alpha$ line are ubiquitous signatures of reflection from a Compton-thick material, the presence of fluorescent K$\alpha$ lines from neutral silicon and sulphur are less common in spectra of obscured AGN \citep[but see e.g.][]{Sambruna01b,bianchi05b,til08}. This is due to the fact that the fluorescence yield is a strong function of the atomic number, and is thus much lower for low-Z metals with respect to iron. Using the \citet{basko78} formul\ae, valid in the case of a semi-infinite plane-parallel slab isotropically illuminated, we calculated the expected ratios between the fluxes of the silicon, sulphur and iron K$\alpha$ fluorescent lines. We adopted fluorescence yields of 0.042, 0.077 and 0.304, respectively \citep{km93}, and the \citet{ag89} abundances. For all the lines, we considered both the unscattered and the once-scattered photons, since with the quality of our data we cannot disentangle the two components even in the iron line. The calculated ratios are Si/Fe=$0.010-0.028$, S/Fe=$0.014-0.032$, Si/S=$0.70-0.88$, where the ranges take into account the dependencies on the inclination angle and the incident power law index. These values, once transformed in EWs with respect to the incident continuum, are very similar to the ones calculated by \citet{mfr97}.
The observed ratios are Si/Fe=$0.18\pm0.14$, S/Fe=$0.22\pm0.16$, Si/S=$0.8\pm0.8$. The fluxes of the observed silicon and sulphur fluorescent lines are significantly higher than predicted, although the uncertainties are quite large. It is difficult to recover the observed fluxes with elemental overabundances, which would be very high, and apparently not required by the ionised plasma dominating the soft X-ray emission. It is likely that the observed fluxes of these two lines are rather unreliable, because of the low statistics and resolution of the spectra. Indeed, the neutral Si K$\alpha$ is probably contaminated by the {Mg\,\textsc{xii}} K$\beta$ line, at 1.745 keV, which could be underpredicted by our photoionisation model. We note that the fluxes of neutral Si and S K$\alpha$ are also large in other sources where they are tentatively detected \citep[e.g.][]{Sambruna01b,bianchi05b,til08}.
\subsection{The soft X-ray emitting region}
The soft X-ray emission of Mrk~573 appears dominated by a photoionised gas, morphologically coincident with the optical NLR, as commonly found in Seyfert 2 galaxies \citep[e.g.][]{bianchi06,gb07}. From the \textit{Chandra} images, we can see that the two, symmetric, emitting regions have an inner radius $r_i<90 pc$ (pixel size of 0.246 arcsec), and an outer radius $r_o\simeq1$ kpc. The projected opening angle of the cones is roughly 50 degrees, in agreement with the 45 degrees measured by \citet{wt94}. Since the intrinsic luminosity of this source can only be indirectly estimated, because the nucleus is obscured by a Compton-thick material, it is difficult to derive a better guess for the inner radius of the emitting region, from the ionisation parameters of the fits with photoionisation models.
On the other hand, from the luminosity of the strongest line, the {O\,\textsc{vii}} forbidden line (which is not significantly contaminated by the collisional phase), we can calculate its emission measure (EM), and estimate the density of the material producing it. The luminosity (ph s$^{-1}$) of a recombination line of charge state $i$ is:
\begin{equation}
L_i =\int_V \! A_Z f_{i+1} \eta \alpha(T) n_e n_H \, dV
\end{equation}
\noindent where $A_Z$ is the abundance of the element, $f_{i+1}$ is the charge state fraction of the recombining ion, $\alpha(T)$ is the radiative recombination coefficient for the recombining ion at temperature $T$, $\eta$ is the fraction of recombinations leading to the relevant transition, and $n_e$ and $n_H$ are the electron and hydrogen density in the volume V of the emitting region \citep[e.g.][]{lied99}. The EM is defined as:
\begin{equation}
EM\equiv\int_V \! n_e n_H \, dV = \frac{L_i }{A_Z f_{i+1} \eta \alpha(T)}
\end{equation}
\noindent where the last equivalence assumes that all the relevant parameters are constant in the gas (or their average values are used). Adopting the best-fit parameters of the RGS fit, \textsc{cloudy} gives the following values: $A_Z=4.9\times10^{-4}$, $f_{i+1}=0.13$, $\alpha(T)=1.3\times10^{-11}$ cm$^3$ s$^{-1}$ (for an average temperature of $4.7\times10^4$ K), and $\eta=0.48$. From the observed luminosity of the {O\,\textsc{vii}} forbidden line, we get an EM of $9\times10^{64}$ cm$^{-3}$. The volume of the emitting region may be approximated ($r_i \ll r_o$) to $V\simeq g [1-\cos(\pi/4)] \pi r_o^3/6 $, where $g$ is the filling factor of the emitting gas within the spherical bi-cone with opening angle of 45 degrees. By assuming $n_e\simeq1.2 n_H$, we derive:
\begin{equation}
g n_e^2 \simeq 25 \, \mathrm{cm^{-6}}
\end{equation}
Therefore, the filling factor $g$ is likely to be much smaller than 1. If we take the average densities measured in the optical NLR \citep[of the order of $10^2-10^3$ cm$^{-3}$:][]{cap96,ferr99,schl09}, we derive $g\simeq10^{-3}-10^{-5}$, i.e. the emitting gas fills a very small fraction of the whole volume of the bi-conical region. Such a small filling factor is also required in order to recover a column density of the order of $3\times10^{20}-6\times10^{21}$ cm$^{-2}$, measured from the RGS and ACIS spectra, along the observed $\simeq1$ kpc. This result is in agreement with the very low mean density ($\simeq1$ cm$^{-3}$), derived from the EM of H$\beta$, if a unity filling factor is assumed \citep{schl09}.
\subsection{The collisional phase and the radio emission}
The detailed spectral analysis of the RGS spectra of Mrk~573 pinpointed the presence of a gas phase in collisional equilibrium, which contributes by $\simeq20-30\%$ to the overall soft X-ray luminosity. The observed thermal emission could be in the hot gas surrounding a starburst region. If all the 0.5-2 keV luminosity emitted by the collisional phase is associated to star formation, we can estimate a star forming rate (SFR) of $\simeq9$ M$_\odot$ yr$^{-1}$ \citep{ranalli03}. A SFR of the same order ($\simeq5$ M$_\odot$ yr$^{-1}$) can be estimated from the total radio luminosity of the two spots, if related completely to the star formation, assuming the relation found by \citet{bell03}, and an average spectral index of $\alpha=-0.75$ \citep{fal98}. A lower SFR ($\simeq2$ M$_\odot$ yr$^{-1}$) is derived from the far infrared (FIR) emission \citep[where the starburst emission largely dominates over the AGN, e.g.][]{fritz06} adopting the relation found by \citet{ken98}, and the FIR luminosity calculated from the \textit{IRAS} infrared fluxes \citep{mosh90} and the method proposed by \citet{hel88}. Moreover, as already reported in Sect.~\ref{hst}, there is no clear evidence of star-forming regions in the inner few kpc in the \textit{HST} near-IR images.
On the other hand, the collisional phase may be directly connected to the observed radio emission, both arising as free-free emission of a hot gas. The X-rays to radio ratio for thermal bremsstrahlung emission of a plasma at temperature $T$ is:
\begin{equation}
\frac {g\left( \nu_x, T \right) e^{-\frac{h\nu_x}{kT}}} {g\left( \nu_r, T \right) e^{-\frac{h\nu_r}{kT}}}
\end{equation}
\noindent where $g\left( \nu, T \right)$ are the Gaunt factors appropriate for the X-rays ($\nu_x$) and radio ($\nu_r$) frequency, at that temperature \citep[e.g.][]{longair92}. Adopting a temperature of $5\times10^6$ K, of the order of the one measured for the collisional plasma in the RGS and ACIS spectra, the Gaunt factor at 5 GHz is around 11, while at 1 keV is approximately 1.5 \citep{longair92}. Therefore, the luminosity ratio between the X-rays and the radio free-free emission should be $\simeq0.01$. For a radio luminosity of $10^{28}$ erg s$^{-1}$ Hz$^{-1}$ (representative of each of the regions detected in the VLA data, see Table \ref{radiotable}), we would expect an X-ray luminosity of $10^{26}$ erg s$^{-1}$ Hz$^{-1}$ at 1 keV, i.e. $10^{43.4}$ erg s$^{-1}$ keV$^{-1}$, several orders of magnitude larger than what is observed.
Note that this does not exclude that all the radio emission that we observe in Mrk~573 is due to free-free emission of a hot plasma. The nuclear core, for example, is likely to be absorbed by a substantial column density of neutral gas, like the X-ray nucleus, so that the X-ray emission possibly associated with the radio emission would be completely suppressed. On the other hand, the radio spots, extended on larger scales, should not be affected by such a large obscuration, and cannot be due to thermal bremsstrahlung emission, without overproducing the observed X-ray emission by a large factor. In any case, it is clear that the collisional plasma we detect in the X-rays cannot be produced by a thermal plasma which gives rise also to the radio emission.
The radio ejecta may still be responsible for the heating of the X-ray-emitting plasma, though not emitting in X-rays themselves. In this scenario, the radio luminosity of the two jets is mainly due to non-thermal synchrotron, while the steep power law index \citep{fal98} makes their X-ray luminosity negligible. Indeed, \citet{cap96} assigned a crucial role to the radio ejecta, because they compress the line-emitting gas, enhancing the emission where this interaction occurs. This interaction could also be the heating source of the gas at $\simeq5\times10^6$ K that we observe through its signatures of plasma in collisional equilibrium. This scenario could be tested by making a comparison between the pressure of the hot X-ray emitting gas and the minimum pressure of the radio jets. The latter can be estimated by assuming equipartition between the particles and the magnetic field. We model each jet lobe as a sphere of radius 0.2 kpc. We assume the electron energy distribution extends from a Lorentz factor $\gamma_{\rm min}=2$ up to $\gamma_{\rm max}=10^{5}$, with an electron energy index $p=2.4$ \citep[the results have only a weak dependence on the choice of $\gamma$; see][]{hard04}. Under the simple assumption that the jet is in the plane of the sky, and that relativistic beaming is unimportant, the minimum pressure of each lobe is $\sim 3\times10^{-11}$ barye. On the other hand, the pressure of the hot gas can be estimated from the density directly derived from the \textsc{APEC} normalisation of our \textit{Chandra} best fit. Assuming the same conical emitting volume as in the previous Section, but with an outer radius of 500 pc, similar to the one observed for the radio emission, we get a pressure of $\simeq1.4\times10^{-9}$ barye (a very similar value, $\simeq1.1\times10^{-9}$ barye, is derived if we use the \textsc{APEC} normalisation of our RGS best fit).
The pressure of the hot gas is therefore about 50 times larger than the minimal internal pressure in the radio jets. For the shocks to be the source of heating of the gas, the two pressures should be comparable. In order for the jet not to be suppressed (unless we are not seeing it at some special time), there must be some additional pressure. Indeed, it was suggested that departures from equipartition may characterise the weak jets and lobes observed in Seyfert galaxies and FR I-type radio galaxies \citep[see e.g.][for the case of NGC~2110]{evans06}, differently from the powerful jets observed in FR II-type radio galaxies. However, other studies of the NLR in Mrk~573 have shown that the the optical spectrum is well reproduced by nuclear photoionisation, and there is no evidence of shocks produced by the radio outflows \citep{ferr99,schl09}. The latter do not likely have any strong influence on the NLR medium, apart from some kinematic effects due to their expansion into the gas, eventually influencing the formation of the observed arcs \citep{schl09}.
\section{Conclusions}
We presented a self-consistent analysis of the XMM-\textit{Newton} RGS spectra of the Seyfert 2 galaxy, Mrk~573. Several pieces of evidence suggest that the dominant ionisation process of the soft X-ray emitting gas is photoionisation: the clear detection of RRC from {O\,\textsc{vii}} and {O\,\textsc{viii}}, and the prominence of the {O\,\textsc{vii}} forbidden line. A photoionisation model fully takes into account all the brightest emission lines, but a collisional phase is also required, in order to reproduce the {Fe\,\textsc{xvii}} lines. This component accounts for about 1/3 of the total luminosity in the 15-26 \AA\ band.
The broadband \textit{Chandra} ACIS spectrum confirms the Compton-thick nature of the source, dominated by a Compton reflection component and a strong neutral iron K$\alpha$ line. The soft X-ray emission needs a further photoionisation component with respect to the RGS spectrum, with a larger ionisation parameter, in order to reproduce emission from higher Z metals.
The \textit{Chandra} soft X-ray image closely follow the NLR morphology mapped by the [{O\,\textsc{iii}}] emission. On the other hand, the radio emission is far more compact, although clearly aligned with the NLR. It could be directly related to the collisional phase found in the X-ray spectra, in a plasma heated by the interaction with the radio ejecta, but the estimated pressure of the hot gas is much larger than the pressure of the radio jets, assuming equipartition and under reasonable physical parameters. Alternatively, the gas in collisional equilibrium may originate in a starburst region, requiring a star formation rate of $\simeq5-9$ M$_\odot$ yr$^{-1}$, but there is no clear evidence of this kind of activity from other wavelengths. Deeper X-ray observations are needed in order to confirm the presence of a gas in collisional equilibrium in Mrk~573, and understand its nature.
\section*{Acknowledgements}
SB, EP and GM acknowledge financial support from ASI (grant I/088/06/0). We would like to thank Craig Gordon for support on XSPEC and HEADAS software, Peter Young for support on CHIANTI, Peter van Hoof and Gary Ferland for support on CLOUDY, and Joel H. Kastner for support on the SER procedure. We also thank Alessandro Caccianiga for useful discussions. SB thanks the INAF-OAB for hospitality. CHIANTI is a collaborative project involving the NRL (USA), the Universities of Florence (Italy) and Cambridge (UK), and George Mason University (USA).
\bibliographystyle{mn2e}
|
\section{Introduction}
Star formation histories (SFHs) are a key measurement for understanding galaxy evolution in a $\Lambda$ Cold Dark Matter universe. Two main approaches have been used to constrain the SFHs of galaxies. One is to track the properties of different galaxy classes over cosmic time, such as star formation rate as a function of stellar mass \citep[e.g.][]{Noeske07}. The other approach is to observe galaxies in the nearby universe and estimate their SFHs based on their current properties \citep[e.g.][]{Heavens04}. Results from both approaches suggest a ``downsizing'' in star formation activity whereby the most massive galaxies form the bulk of their stars early (before $z\sim1$), while less massive, late type galaxies form the majority of their stars later \citep[e.g.][]{Brinchmann00}. \citet{Noeske07} suggests that both the onset and duration of star formation may be correlated with mass, so called ``staged'' star formation. For gas-rich dwarf systems, this scenario should result in relatively constant low-level star formation over much of a galaxy's lifetime.
In this paper we derive the stellar populations for the local dwarf irregular galaxy KKH~98 as determined by optical/IR color-magnitude diagrams (CMDs) of the individual member stars.
\begin{figure*}
\center
\includegraphics[scale=0.8]{f1.ps}
\caption{\label{fig:KKH98} $40 \arcsec \times 40 \arcsec$ images of KKH~98 in the F475W, F814W, and $K'$-band (top row; left, middle, and right respectively). Also shown are the three color composite (bottom left), and the full ACS field of the {\emph{HST}}\ F814W-band image (bottom right). North is up and East to the left in all images. The spatial resolution of the Keck AO $K'$-band image is a good match to the resolution of the optical {\emph{HST}}\ images. The F475W and F814W band {\emph{HST}}\ images reach depths of F475W$\sim27.6$ and F814W$\sim27.0$ (12\% photometric uncertainties). Photometry of these images recovered 2539 stars within the AO field. At this depth there is significant crowding in the {\emph{HST}}\ images. In contrast, the Keck AO $K'$-band image probes the brightest 592 near-IR sources, and shows no significant crowding. The very bright star at the top of the AO image was used as a tip-tilt guide star for the Keck AO system.}
\end{figure*}
For the most nearby galaxies ($d \la 4$ Mpc), color-magnitude diagrams of individual stars
can be used to estimate SFHs. Optical Hubble Space Telescope ({\emph{HST}}) imaging has been used extensively to study the stellar populations of nearby galaxies \citep[e.g.][]{Sarajedini05, Weisz08, Gogarten09, McQuinn09, Williams09a, Williams09b,Rejkuba09}. For instance, deep {\emph{HST}}\ and ground based imaging of Local Group dwarf galaxies has revealed a subset of dwarf spheroidals and ellipticals with extended star formation histories and star formation episodes as recent as 1 Gyr ago \citep[e.g.][]{Aparicio01,Dolphin05}. The ACS Nearby Galaxy Survey Treasury \citep[ANGST;][]{Dalcanton09} has expanded the Local Group efforts to include a large sample of more distant galaxies out to a distance, $d<4$ Mpc. Using the {\emph{HST}}\ Advanced Camera for Surveys (ACS) and archival {\emph{HST}}\ imaging, ANGST has released photometry for deep two-color optical images of nearly 70 nearby galaxies. Matching the CMDs of these galaxies to model CMDs allows reconstruction of the star formation histories of a wide variety of galaxies in both group and field environments.
Significantly less has been done to reconstruct the star formation histories from near-infrared (near-IR) stellar photometry. Several programs have used the SFHs of local group galaxies, as measured from optical data, to test stellar evolution models in the near-IR. For instance, \citet{Gullieuszik08} use the star formation history of Leo II dSph, obtained from optical CMDs, to estimate the expected production of AGB stars. They find that the models over-predict the observed numbers of AGB stars by as much as a factor of six. Likewise, \citet{Gullieuszik07} demonstrate how the near-IR colors of red giant branch (RGB) stars can be used to estimate the metallicity distribution within a galaxy. Other studies have used the near-IR luminosity functions of the AGB and RGB populations to estimate SFHs of local group galaxies \citep{Olsen06,Davidge09}
In spite of the limited analysis to date, the near-IR has several features that make it attractive for probing SFH. First, for most galaxies the bulk of the stellar luminosity is emitted in the near-IR \citep{Sawicki02}. Second, intermediate aged asymptotic giant branch (AGB) populations are both extremely luminous in the IR \citep{Aaronson85, Frogel90, Maraston06}, and lie in tight sequences in near-IR color-magnitude space. Third, the effects of reddening from dust are significantly reduced at near-IR wavelengths. Fourth, RGB age/metallicity degeneracies are reduced for optical-IR colors compared to optical or IR data alone \citep{Gullieuszik07}. Fifth, future missions such as the James Webb Space Telescope (JWST) and adaptive optics on the Thirty Meter Telescope (TMT) will be IR optimized. These missions will potentially resolve individual stars in galaxies out to the Virgo Cluster (18 Mpc) and beyond. In order to take full advantage of these missions a clear understanding of the IR properties of the CMD are required.
To test the efficacy of near-IR photometry for reconstructing the SFHs of nearby galaxies, we have begun a campaign to image ANGST galaxies with Keck Adaptive Optics (AO) and {\emph{HST}}\ WFC3. In this paper, we present results from Keck AO imaging of nearby dwarf irregular galaxy KKH~98. This galaxy was chosen because of its extended SFH and suitability of nearby AO guide stars. We obtained $K'$-band (2.12 $\mu$m) Keck laser guide star (LGS) AO imaging of KKH~98, as part of the Center for Adaptive Optics Treasury Survey \citep[CATS][]{Melbourne05a,Melbourne08a}. The spatial resolution of the AO image ($\sim0.1 \arcsec$) matched the spatial resolution of the {\emph{HST}}\ F475W and F814W images from ANGST. 592 individual IR bright stars ($K<23.5$) were resolved in the $K$-band, while 2539 stars were resolved in the optical (F475W $ <27.6$). We used the photometry of these stars to estimate the distance, metallicity, and star formation history of the galaxy.
This effort is the first to use resolved IR photometry of individual stars to reconstruct the complete star formation history (SFH) of a galaxy beyond the Local Group. We compare the IR results to the well constrained, optically derived SFH of KKH~98, providing an excellent opportunity to test these techniques at longer wavelengths. By refining SFH recovery methods in the IR, this work will help prepare for programs with JWST and TMT on more distant galaxies.
Throughout, we report Vega magnitudes and assume a $\Lambda$ cold-dark-matter cosmology: a flat universe with $H_0= 70$ km/s/Mpc, $\Omega_m=0.3$, and $\Omega_{\Lambda}=0.7$.
\section{{\emph{HST}}\ and Keck AO Observations of Dwarf Irregular Galaxy, KKH~98}
This study uses high-spatial resolution imaging data from both {\emph{HST}}\ and Keck AO to estimate the stellar populations within the dwarf Irr galaxy KKH~98. KKH~98 was first identified by \citet{Karachentsev01} as a low surface brightness galaxy in the Palomar Observatory Sky Survey II plates. It is a gas-rich, isolated dwarf irregular galaxy with an HI line width of $W_{50}=26$ km/s \citep{Karachentsev01}. \citet{Karachentsev02} find a distance of 2.45 Mpc based on the tip of the red giant branch. This distance gives an absolute magnitude of $M_B=-10.78$. Its properties are fairly typical of the local dwarf irregular population identified by \citet{Karachentsev01}.
\subsection{{\emph{HST}}\ Observations}
{\emph{HST}}\ observations of KKH~98 were obtained in both the F475W and F814W filters of ACS as part of the ANGST survey \citep[see][for the details]{Dalcanton09}. Three observations of the galaxy were made in each filter. The combined images have an effective exposure time of 2265 seconds and 2280 seconds, respectively, and are shown in Figure \ref{fig:KKH98}. The ACS field of view ($202\arcsec \times 202\arcsec$) encompasses the bulk of the galaxy (See Figure \ref{fig:KKH98}).
\subsection{Keck AO Observations}
On the night of November 24, 2007, we obtained $K'$-band (2.12 $\mu$m) Keck LGS AO imaging of KKH~98. We used the wide-field ($40\arcsec \times 40\arcsec$) camera of the NIRC2 instrument to achieve the largest field of view. With a pixel scale of $0.04 \arcsec$/pix, the wide-field camera under-samples the AO PSF ($0.06 \arcsec$ resolution), limiting the effective resolution to $\sim0.1 \arcsec$. However, as is discussed below, the loss of resolution did not affect our ability to measure the photometry of the uncrowded IR-bright stars in KKH~98.
Individual exposures were one minute, with a dither applied after two successive exposures. During the observations, the laser was fixed to the center of the field. Thus, as the telescope dithered, the laser dithered as well. This method was preferred over fixing the laser to the sky because of significantly lower overhead. In addition, as was demonstrated in \citet{Steinbring08}, this method provides a more uniform AO correction over the field. An $R=14.2$ tip/tilt star was located $\sim20 \arcsec$ north of the galaxy center (Figure \ref{fig:KKH98}). The AO system used this star to track and correct for image motion, while the laser spot was used for higher order wavefront corrections.
We followed the image reduction method outlined previously in \citet{Melbourne05a}. Sky and flatfield frames were created from the individual science frames after masking sources. Frames were then flatfielded and sky subtracted. We corrected frames for known NIRC2 camera distortions. We aligned images by centroiding on sources, and combined them with a clipped mean algorithm. The final reduced image has an effective exposure time of 45 minutes. Figure \ref{fig:KKH98} shows the resulting AO image of KKH~98. It covers the central $40 \arcsec \times 40 \arcsec$ region of the galaxy.
We observed the UKIRT photometric standard star FS3 immediately after the galaxy observations. However, the night was not photometric. Therefore we set the zeropoint for the AO data based on two bright unsaturated stars in the actual science frame which have $Ks$-band magnitudes from 2MASS \citep{Skrutskie06}. Ultimately, we compared our photometry to stellar isochrones produced in the $K$-band system \citep{Bessell88}. However the transformation to the $K$-band system is negligible, typically $\la0.02$ mag for red stars (including the transformation from $Ks$ to $K$ \citep{Carpenter01} and the color terms to transform from $K'$ to $Ks$ \citep{Wainscoat92}). Therefore, for simplicity, we heretofore report all AO magnitudes in the $K$-band Vega system.
The standard star observations of FS3 were obtained with tip-tilt AO correction to stabilize image motion, but without higher order LGS AO correction from the laser spot. This strategy produced a high signal-to-noise ratio (S/N) observation of the profile of the wings of the AO PSF (which are primarily set by seeing and tip-tilt). We used these standard star observations in the reconstruction of the AO PSF as outlined in the following section.
\begin{figure}
\center
\includegraphics[trim=20 0 0 0,scale=0.6]{f2.ps}
\caption{\label{fig:PSFsplice} Bottom: the azimuthally averaged stellar intensity as a function of radius (scaled to unity at a radius of $0.4\arcsec$) for the standard star FS 3 (solid line), and 3 stars in the AO image (black, green, and red diamonds). FS3 was observed with tip-tilt correction on but no higher order AO correction. The shape of this profile is therefore set by the seeing and tip-tilt correction. This figure shows that for radii larger than $\sim0.2 \arcsec$, the halo of the AO corrected PSF has a similar profile as the tip-tilt corrected standard star PSF. We therefore can use the high S/N PSF halo profile from our standard star (FS3) to track the profile of the our AO PSF out to large radii.The profiles of the 3 stars from the AO frame, which have higher order wavefront corrections, are significantly more peaked in their core than the standard star which only had tip-tilt correction. Top: the fraction of light within a given radius for one of the AO corrected stars. A vertical dashed line denoting a radius of $0.16\arcsec$ is shown. This radius is the size of the aperture used in our photometry, and it contains more than 40\% of the stellar flux. }
\end{figure}
\section{Photometry \label{sec:photometry}}
\citet{Williams09a} and \citet{Dalcanton09} describe the photometry of the {\emph{HST}}\ ANGST images. Stars in the ACS images were identified and measured with the photometry routine DOLPHOT \citep{Dolphin00} using PSF fitting of individual sources with model {\emph{HST}}\ PSFs from TinyTim\footnote{http://www.stsci.edu/software/tinytim/}. Aperture corrections were measured for uncrowded stars and then applied to PSF fitting results. Photometry lists were culled to remove extended objects and low S/N stars. Stars with S/N $>6$ in both filters were included in the final catalogue. The matched F475W and F814W-band catalogue contains 11324 stars. Twelve percent photometric uncertainties were measured at F475W$=27.6$ and F814W$=27.0$. Cutting the catalogue at these flux levels, and limiting to the field of view of the AO image results in a final set of 2539 stars.
The Keck AO image is significantly shallower than the {\emph{HST}}\ frames (Figure \ref{fig:KKH98}). However, the very deep {\emph{HST}}\ images already pinpoint the locations of stars in KKH~98. We take advantage of the {\emph{HST}}\ positions when measuring stars in the AO image, performing photometry on the sources already identified in the {\emph{HST}}\ frames. The AO PSF varies significantly across the field; thus we choose to use aperture photometry with aperture corrections calculated across a grid of sub-regions in the image. Discussions of the AO photometry --- including the PSF, aperture corrections, and aperture photometry --- are given below.
\begin{figure}
\includegraphics[trim=20 0 0 0,scale=0.6]{f3.ps}
\caption{\label{fig:apcor} The aperture correction for an $0.16\arcsec$ aperture as a function of separation from the center of the AO image. The correction increases linearly with distance to the image center. We fit all data points with a line, giving a unique correction for each star in the image. This fit also improves the aperture correction for stars on the outskirts of the image where there are few stars to make a good estimate of the PSF.
}
\end{figure}
\subsection{AO PSF}
To perform accurate photometry, we need a model of the AO PSF. Unfortunately, the AO correction degrades spatially with distance from the laser spot (anisoplantism) and the tip-tilt star (anisokinetisism), producing a spatially varying PSF. We model the PSF using two components, (1) an AO corrected core with a FWHM of roughly the diffraction limit of the telescope, and (2) a halo with a profile width set by the atmospheric seeing and AO tip-tilt correction. While the shape of the PSF halo is relatively stable across the image, the core varies significantly with position.
To get an accurate representation of the profile of the PSF core, we identify luminous, isolated stars within 25 equal-area sub-regions of the image. After subtracting neighbors, we register and coadd the stars in each sub-region, and trace the PSF profile of the coadded image.
Unfortunately, the PSF halo is spread over hundreds of pixels, and is much lower S/N than the diffraction limited core. Therefore its profile cannot be accurately traced even for the stacked PSFs described above. Instead we used images of the very bright flux standard star, FS 3, observed with tip-tilt AO correction on, and high order AO correction off. With high order correction off, light that would have been corrected into the PSF core by the AO system, remained in the halo. The resulting PSF has a very high S/N profile out to large distance ($\sim2\arcsec$) from the PSF center, and is a good match to the halo of the AO PSF (Figure \ref{fig:PSFsplice}).
To produce a final PSF for each of the 25 sub-regions, we splice together the halo and core images, scaling the halo image to map smoothly onto the image of the core. As was done in \citet{Melbourne08a}, we select the splice point to be the radius at which the flux in the azimuthally averaged core profile is 9 sigma above the sky noise, as this is a good match to bright AO corrected stars where the halo is easily observed. Figure \ref{fig:PSFsplice} shows that the PSF profile is relatively independent of the particular choice of splice point, for splice points larger than several times the diffraction limit (e.g. $\sim 0.2 \arcsec$). This figure also shows that the halo PSF profile from the standard star is a good match to the few AO corrected stars in the science frame that are bright enough to adequately detect the halo.
\subsection{AO Aperture Corrections}
The stellar photometry of the AO image was performed with small apertures on the high S/N core of the AO PSF. We chose a 4 pixel radius aperture ($0.16 \arcsec$). This aperture is 50\% larger than the typical FWHM of the stellar PSF, and maximizes the S/N of the aperture photometry. We convert the small aperture photometry to total flux with aperture corrections calculated from the 25 model PSFs, one for each sub-region of the image. Total flux for the model PSFs was measured with a curve of growth technique out to $2\arcsec$.
The aperture correction varies from 0.6 mags at the center of the field to $\sim1$ magnitude near the edge, where a larger fraction of the light remains uncorrected. Figure \ref{fig:apcor} shows the variation of the aperture correction across the field. A linear fit to the data gives aperture correction as a function of distance from the image center for an arbitrary stellar position. The fit provides a more finely sampled measure of the aperture correction across the field and improves the estimates in the outskirts of the image where the PSF is poorly measured due to a lack of stars. We use this functional form for the aperture corrections in the final photometry.
\begin{figure}
\center
\includegraphics[trim=20 0 0 0, scale=0.6]{f4.ps}
\caption{\label{fig:model_phot} Photometry of 500,000 artificial stars embedded at random locations in the AO image. The top panel shows the difference of input vs. measured magnitude while the bottom panel shows the standard deviation of the difference as a function of input magnitude. The thick dashed line represents the standard deviation calculated from all of the stars. The thin lines represent the standard deviations of groups of stars with consecutively larger separations from the image center (6, 12, 18, and 24 arcsec from the image center, as marked). Because the PSF is tighter near the center of the AO image, the S/N is higher for the photometry done there. The photometry shows low photometric uncertainty (better than 15\%) for stars brighter than $K=22$. }
\end{figure}
\subsection{AO Photometry \label{sec:phot}}
The deep {\emph{HST}}\ images of KKH~98 identify the locations of stars in the galaxy. To translate the {\emph{HST}}\ positions onto the Keck AO image, we first selected a set of 157 bright stars that are easily visible in both the {\emph{HST}}\ and AO images. We then used the IRAF CCTRAN program to transform the {\emph{HST}}\ coordinates into the AO image coordinates, using. a quadratic transformation. Since the geometric distortions in the AO image were removed during image processing, this transformation worked well across the entire frame.
We measured the $K$-band magnitude for each star in the {\emph{HST}}\ star list that was located on the AO frame. For each star, bright neighbors were identified with Sextractor. Scaled model PSFs were used to subtract the bright neighbors using a PSF fitting technique. We then measured aperture photometry on the star of interest, applying the aperture correction appropriate for the radial separation of the star from the center of the image (Figure \ref{fig:apcor}). A correction for the local sky was made, using a median sky value measured in an annulus of radius of $1.0 -1.2$ arcsec. When making the sky measurement, stars identified in the sky annulus by Sextractor segmentation maps and the {\emph{HST}}\ star lists were masked out.
Two types of uncertainty dominate the error budget in these measurements, (1) the photometric uncertainty, set primarily by the Poisson sky noise, and (2) the uncertainty introduced by the spatially varying PSF and aperture corrections. The photometric Poisson uncertainty can be measured directly from the images. For each star, we measure the standard deviation of pixels in the nearby sky to estimate this uncertainty. The uncertainty introduced by the aperture corrections can be measured from the scatter in Figure \ref{fig:apcor}. While the scatter is small near the center of the image, it increases significantly towards the edges of the image.
To estimate any systematics in the measurement methods, we also performed a Monte Carlo simulation, embedding 500,000 artificial stars into the image at random locations with sub-pixel sampling. Artificial stars ranged in magnitude from $K=17$ to 24, but were all assumed to be identified in the corresponding {\emph{HST}}\ image. We measured the photometry of the artificial stars in the exact same manner as the real stars in the {\emph{HST}}\ catalogue. Figure \ref{fig:model_phot} shows the standard deviation of the photometry as a function of input $K$-band magnitude. We find to $K=22$, the photometry is better than 0.2 mags. At $K=23.5$ the photometric uncertainty is about a magnitude.
An additional 22 IR-bright stars were identified in the AO image that were not included in the final {\emph{HST}}\ photometry lists, usually because they fell below a blending or S/N threshold in the F475W image. We found that these stars do not alter the conclusions of this paper and therefore do not include them in the final analysis.
\begin{figure*}
\centering
\includegraphics[trim=60 0 0 0, scale=0.8]{f5.ps}
\caption{\label{fig:CMD} Apparent (uncorrected for reddening) color-magnitude diagrams of stars in KKH~98. F475W $-$ F814W color is plotted on left, and F814W$-K$ color is plotted on the right. Suspected AGB stars are shown as filled red circles. While both the optical and IR CMDs show a tight red giant branch, the IR color shows a much tighter AGB sequence, resulting in a powerful age indicator for intermediate aged populations. The 50\% completeness limits (blue curve), and the cuts applied to the data, for modeling purposes (green lines) are shown for the optical data. The data cuts applied to the IR data are also shown. Typical photometric uncertainties at the faint end of each CMD are shown. At the bright end the uncertainties are roughly the size of the data-points. Data at the faint end of the IR CMD ($K<23$) have photometric uncertainties greater than 0.6 magnitudes. Thus the large scatter in the IR CMD at faint magnitudes is not a feature of stellar evolution but rather photometric error.}
\end{figure*}
\section{Results}
Figure \ref{fig:CMD} shows the F475W - F814W vs. $F814W$ CMD and the F814W$-K$ vs. $K$ CMD for KKH~98. Both CMDs show a tight red giant sequence with a tip at F814W$=23.2\pm0.1$ and $K\sim21.2 \pm 0.1$ (based on a visual fit to the CMD). The optical CMD exhibits a significant blue upper main sequence ($F475W - F814W\sim0.0$) suggesting the presence of very young stars in this galaxy. The young red helium burning sequence is poorly populated, although there do appear to be some young blue core helium burning stars. The main sequence is not obvious in the F814W$-K$ CMD, because of the flux limits of the $K$-band data.
In the IR CMD, a well-defined asymptotic giant branch (AGB) emerges from the tip of the red giant branch (TRGB) and extends to $K=18.5$. Compared with the optical CMD, the AGB stars in the F814W$-K$ CMD form a much tighter sequence. Tighter AGB sequences are expected at redder wavebands for several reasons: (1) lower amounts of self-extinction from dusty AGB envelopes, (2) a flatter color vs.\ stellar T$_{eff}$ relation \citep[e.g.][]{Gullieuszik07}, and (3) the thermally-pulsing AGB stars have much smaller pulsation amplitudes \citep{Whitelock06, Marigo08}. Because the AGB produces tight sequences in near-IR color-magnitude space, the AGB should prove a powerful indicator of star formation rate for intermediate aged populations, i.e. older than $\sim0.1$ Gyr.
\begin{figure*}
\centering
\includegraphics[trim=60 0 0 20, scale=0.75]{f6.ps}
\caption{\label{fig:CMD2} Same as Figure \ref{fig:CMD} only now with model isochrones overplotted \citep[from][]{Marigo08}. Top: metal poor isochrones, 1/10th solar, spanning ages from $10^7 - 10^{10}$ years (colored lines) nicely span the data. Bottom: more metal rich isochrones (dashed lines) with an age of $10^{9.5}$ years, the age by which roughly 80\% of the stars had formed, are also shown and are significantly redder than the data. This suggests that the stars of KKH~98 are metal poor. Note: the very red colors measured for some faint stars in the IR CMD (below the horizontal line) can be explained by photometric uncertainty (greater than 1 mag.), rather than some usual stellar population. }
\end{figure*}
At the faint end of the IR CMD (i.e. $K>23$), photometric uncertainties in the $K$-band produce significant scatter in the photometry. Thus the large scatter to red colors in Figure \ref{fig:CMD} (and Figure \ref{fig:CMD2}) is not a physical feature in the data (such as the red clump), just photometric noise.
Because of the spatially varying AO PSF, we investigated the possibility that the AO photometry might vary systematically across the image. However, CMDs generated from multiple sub regions across the image show no systematic deviations.
\begin{figure}
\center
\includegraphics[trim=10 0 0 0,scale=0.5]{f7.ps}
\caption{\label{fig:CMDresid} TOP: the optical CMD now displayed as a density Hess Diagram. The true data (top left), and the best-fit MATCH-derived model (top right) are shown. Also shown are the residual differences between the two (bottom left), and the residual difference normalized by Poisson statistics in each bin (bottom right). In the residual maps, darker regions denote an under-production of model stars. Lighter bins denote an over-production of model stars. The residuals (bottom left) vary from +10 stars per bin (white) to -11 stars per bin (black). The normalized residuals (bottom right) vary from +8 (white) to -5 (black) and can be thought of as sigma deviations per bin. The models reproduce the data, except at the faint end, and for the AGB, where the models over-produce AGB stars by roughly a factor of three (F475W $\sim25.7$ and F475W - F814W $\sim3$, see Figure \ref{fig:bi_fixbi} for more details). }
\end{figure}
\begin{figure}
\center
\includegraphics[trim=10 0 0 0,scale=0.5]{f8.ps}
\caption{\label{fig:CMDresid2} Same as Figure \ref{fig:CMDresid}, only now for the IR data. The residuals (bottom left) vary from +10 (white) stars to -9 stars (black), while the normalized residuals vary from +5 sigma to -5 sigma. The fit to the AGB is better for the IR CMDs than the optical CMDs, although the models still over-predict the the AGB by roughly a factor of two (see Figure \ref{fig:ik_fixik} for details).
}
\end{figure}
\subsection{Modeling the Color-Magnitude Diagrams \label{sec:sfh}}
We estimate the star formation history (SFH) of KKH~98 by comparing the observed CMDs with CMDs produced from the stellar isochrones of \citet{Girardi02}, with updated AGB models from \citet{Marigo08}. These isochrones were used together with the bolometric corrections and color transformations from \citet{Girardi08}. Sample isochrones, overlayed on the CMDs, are shown in Figure \ref{fig:CMD2}. Metal poor (1/10th solar) isochrones (top row Figure \ref{fig:CMD2}) span the photometric data. In contrast, more metal rich isochrones (bottom row Figure \ref{fig:CMD2}) tend to be redder than the data, suggesting that the stars in KKH~98 are metal poor.
We use the MATCH package \citep{Dolphin02} to fit the CMD for a star formation history, metallicity history, dust reddening, and distance modulus (DM), for an assumed Salpeter IMF and binary fraction of 0.35. MATCH uses a Poisson Maximum Likelihood statistic to determine the best fit model and the 1 sigma uncertainties on that model. For a complete description of the MATCH routine, see \citet{Dolphin02}.
\citet{Williams07} showed that the IMF choice does not affect the relative star formation rates (as a function of time), but can change the overall normalization. Our primary goals are to compare the SFH measured from the optical data to the SFH measured from the IR data, and to determine if the SFH is ``bursty'' or continuous. As these are relative measurements, neither will be significantly affected by choice of IMF or binary fraction.
Additional inputs to MATCH include: (1) a range of acceptable distances and reddening values; (2) time step sizes; and (3) ranges and binning sizes in color-magnitude space. In general we, attempted to reproduce the choices made in previous ANGST work \citep{Williams07, Williams09a}. Distance ranges were allowed within $\pm0.5$ mags of the distance modulus in the literature of $DM=26.95 \pm0.11$ \citep{Karachentsev02}. Reddening was allowed to vary from $A_V=0.1 - 0.7$, a reasonable range for a dwarf galaxy with foreground reddening $A_I=0.24$ as measured in the \citet{Schlegel98} Galactic dust maps. We use a series of 12 logarithmically spaced time-steps which allow us to explore the star formation history in both the recent (Myrs) and distant (Gyrs) past. The minimum sizes for color-magnitude bins are set by the uncertainties of the stellar isochrones and are roughly 0.05 mags in color and 0.1 mags in magnitude. However, when setting bin sizes, additional considerations include allowing for reasonable number statistics in each bin while providing the resolution necessary to resolve features in the CMDs. For the optical data we used color bins of size 0.1 mags and magnitude bins of size 0.15 mags. We doubled the sizes of these bins in the NIR to offset the smaller number statistics in the NIR CMDs. Small changes in the bin sizes were not found to alter the results.
We used the latest version of MATCH (version 2.3) which handles the isochrone color transformations for Carbon and Oxygen AGB stars independently. In older versions of MATCH, C/O ratios were not tracked and the color transformations produced AGB stars significantly redder (2-3 mags) than expected. The colour transformations of C stars in the updated version of MATCH were derived from \citet{Loidl01}. Because the information content of the IR CMD is biased towards the AGB populations, it was important for us to use the most realistic prescriptions for AGB photometry. In the optical CMDs these considerations were less vital because the information content at young and intermediate ages primarily arises from the Main Sequence turn-offs.
For the optical data, we chose to model the CMD with photometry brighter than the 12\% uncertainty limits and 50\% completeness limits (F475W$<27.6$ and F814W$<27.0$) to avoid fitting areas of the CMD with poor photometry or large completeness corrections. These limits result in photometry for 2539 stars in the area overlapping the AO data. For the IR CMD, we chose stars with $K < 23.5$, the point at which the $K$-band photometric uncertainty is about 1 magnitude. These cuts provide 592 stars. While the photometric precision at the fainter end of $K$ is poor, the long wavelength baseline between the F814W and $K$-bands means that there is still statistically interesting information at these fainter limits. Table \ref{tab:results} gives a summary of the MATCH derived results which are discussed in detail below.
Figures \ref{fig:CMDresid} and \ref{fig:CMDresid2} show Hess diagrams of the CMDs in the optical and IR (top left). Also shown are the best fit model CMDs derived by MATCH (top right), the residual difference between the model and the data (bottom left), and the residuals scaled by the statistical significance of bin (bottom right). The residual differences between the data and MATCH derived models are small for both the optical and IR CMDs, except for AGB stars, especially in the optical data where the models predict 130 AGB stars compared to 42 observed (with the caveat that the AGB must be brighter than the TRGB, and redder than $F475W - F814W = 1.75$). In the IR, the models predict 92 AGB stars (brighter than the TRGB). We return to these discrepancies in the following section and elaborate in the discussion section. In contrast, the numbers of stars in other evolutionary stages, such as the red giant branch and main sequence are better reproduced by the models. Table \ref{tab:cmdnum} gives the measured and predicted numbers of stars in different regions of the KKH~98 CMD.
\begin{figure}
\center
\includegraphics[scale=0.4]{f9.ps}
\caption{\label{fig:bi_fixbi} A realization of the best fit optical model CMD shown broken into different time bins (blue diamonds). The observational data are also shown (black points), with the AGB stars circled in red. The models shown that the main sequence is populated by young stars (log(age) $< 8.8$), whereas the red giant branch is primarily populated by old stars (log(age) $>9.5$). The model predicted numbers of AGB stars are given in the upper right of each CMD. For AGB stars above the TRGB (and $F475W - F814W > 1.5$), the data show 42 stars, while the models predict 130 stars. The lower left plot shows the magnitude limits provided to MATCH (green lines). }
\end{figure}
\begin{figure}
\center
\includegraphics[scale=0.4]{f10.ps}
\caption{\label{fig:ik_fixik} Same as Figure \ref{fig:bi_fixbi}, only now a realization of the best fit IR model CMD (blue diamonds) shown broken into different time bins. The model predicted numbers of AGB stars are given in the upper right of each CMD. For AGB stars above the TRGB, the data show 42 stars, while the models predict 93 stars. In this case we see that the bulk of the predicted AGB stars are old (log(age) $<9.5$). The lower left plot shows the magnitude limits provided to MATCH (green lines). }
\end{figure}
\subsubsection{Star Formation Histories}
The optical and IR derived SFHs of KKH~98 are plotted in Figure \ref{fig:sfh}. For the bulk of cosmic time, the SFRs from the two CMDs agree to within the uncertinties. For ages older than 3 Gyrs, both data sets show evidence for a modest, roughly constant star formation rate of $5\times10^{-4} M_{\odot}$ yr$^{-1}$, or $2 \times10^{-4} M_{\odot}$ yr$^{-1}$ kpc$^{-2}$. From $1 - 3$ Gyrs ago, both show a drop in star formation to roughly negligible rates. More recently, both show renewed star formation for ages $ < 0.5$ Gyrs ago. However, in these youngest time bins, there are differences in the details. At the youngest ages (age $< 0.1$ Gyr), the optical photometry suggests ongoing star formation, while the IR data set does not. This difference is not alarming, as the near-IR data are not very sensitive to hot (blue) main sequence stars produced by recent star formation.
Another difference is that the optical data set suggests a burst of star formation, at age $\sim0.5$ Gyrs (Figure \ref{fig:sfh}). In contrast, the IR data suggest SFRs roughly a factor of four smaller at that time. The evidence for a burst in the optical data comes from stars at the well-populated main sequence turn off (MSTO) at the faint end of the optical CMD. This is shown in Figure \ref{fig:bi_fixbi} where a realization of the predicted model CMD divided into different time bins. At log(age) $= 8.6 - 8.8$ [yr], large numbers of stars are predicted at the MSTO. In contrast the IR data do not reach deep into the main sequence, and therefore the IR-derived SFH is not sensitive to the main sequence turn-off feature. However, the IR data should contain information on populations of that age from the AGB sequence (see Figure \ref{fig:ik_fixik}). Therefore, if the burst is real, evidence for it should exist in the AGB population, assuming the AGB models are accurate. In the discussion section we examine possible explanations for the measured differences in the SFR at these ages. Here we discuss the effect of time binning.
\begin{figure}
\center
\includegraphics[trim=50 150 0 20,scale=0.6]{f11.ps}
\caption{\label{fig:sfh} The MATCH derived star formation history of KKH~98 from the optical CMD (blue) and the IR CMD (red). At early times, both show a relatively constant star formation rate of $5\times10^{-4} M_{\odot}$ yr$^{-1}$, or $2 \times10^{-4} M_{\odot}$ yr$^{-1}$ kpc$^{-2}$. From $1-3$ Gyrs ago, both show a significant drop in the star formation to negligible rates. More recently, (age $< 1$ Gyr) the SFR has roughly returned to its past average. However, the optical data suggest a possible burst of star formation at 0.5 Gyrs ago, not reproduced in the IR data. The IR CMD is also insensitive to the most recent star formation (i.e. age $<$ 60 Myr). Uncertainty estimates are shown for each time bin, and combine in quadrature (1) the systematic uncertainty between the models and the data, and (2) the uncertainty associated with small number statistics in a given time bin. The bottom panel shows the data rebinned to longer time periods.
}
\end{figure}
Time binning may play a role in the differences between the optical and IR derived SFHs. Logarithmic time bins have been selected to fully explore the SFH on both recent and older time scales. However, especially for the IR data, which has only 592 stars, the time binning may be more fine than the data warrant. The third panel of Figure \ref{fig:sfh} shows the IR results now binned with larger time steps. These bins were selected by first binning the data on small time scales and then tracking the goodness of fit parameters as bins were dropped from the sample. If the fits did not change by more than the measured uncertainties, a dropped bin would be incorporated with its neighboring bin. Ultimately, this resulted in 4 bins. We caution, however, that this method is likely to miss the importance of bins with little star formation such as those from $1 - 3$ Gyrs ago. For more on optimal binning techniques, see Williams et al. (2009 in press). The optical data have been binned to match the IR bins. Binned this way, there are still some differences between the two measured star formation histories at the youngest ages.
The SFR uncertainties shown in Figure \ref{fig:sfh} are a combination of (1) the systematic uncertainty between the models and the KKH~98 photometry, and (2) the random uncertainty associated with small number statistics of stars in any given color-magnitude bin. The first uncertainty is estimated by the MATCH routine from the maximum likelihood statistic as it attempts to fit models parameters within the adopted ranges. The second uncertainty is estimated by a Monte Carlo bootstrap simulation of 100 realizations, where we rerun MATCH but alter the input photometry list, so that it is composed of random draws from the actual photometry list. The systematic uncertainties are roughly equivalent for both the optical and IR data, and are similar in scale to the random uncertainties associated by small number statistics. The error bars shown in Figure \ref{fig:sfh} are the two uncertainties added in quadrature.
Surprisingly, the optical and IR model uncertainties are similar in scale despite the fact that the optical data contain contain a factor of 5 more stars. Several factors conspire to lead to this result. First, the color-magnitude bins used in the IR analysis are twice as large as the optical analysis, and thus typically contain as many or more stars as the optical bins. Second, a lack of stars in a feature can be just as important to producing a significant result as the presence of stars. For example the lack of AGB stars in the data forces the IR models to select against a burst of star formation at 0.5 Gyrs ago (see Figure \ref{fig:ik_fixik}), resulting in small numbers of AGB stars in both the models and data, and a low uncertainty assigned to this time bin. Whereas in the optical data, the main-sequence turn off suggests a burst at that time, and the lack of AGB stars acts to increase the uncertainty of the high SFR measured in this time bin.
\begin{figure}
\center
\includegraphics[trim=30 0 0 0,scale=0.55]{f12.ps}
\caption{\label{fig:met} The MATCH derived metallicity of new stars at each epoch from the optical (blue asterix) and IR (red diamonds) CMDs. Both show a preference for metal poor stars from roughly 1/30th to 1/10th solar. There may be a slight increase in metallicity over time. For timebins with negligible star formation estimated in the IR CMDs, no metallicity is plotted. }
\end{figure}
\subsubsection{Metallicity}
Figure \ref{fig:met} shows the MATCH measured metallicity of new stars at every epoch. Again results from both the optical (blue) and IR (red) CMDs are shown, however, in time bins where no star formation was measured, no metallicities are shown (several IR time bins). Both suggest that the chemical history of KKH~98 has been dominated by low metallicity star formation, producing stars with typical $[M/H]\sim -1.5$ to $-1.0$. There may be a trend to more metal rich stars at more recent times. MATCH can also be run in a mode where the metallicity is constrained to continually increase, as would be expected in a closed box galaxy model. When run in this mode, the MATCH results do not change appreciably.
\begin{figure}
\includegraphics[trim=40 0 0 0,scale=0.55]{f13.ps}
\caption{\label{fig:trgb2} The magnitude of the tip of the red giant branch (TRGB) as a function of metallicity and age in the F814W (top) and $K$-bands (bottom) as derived by the isochrones of \citet[][colored lines]{Marigo08}, the same isochrones used by MATCH. Also shown is the metallicity dependence on the $K$-band TRGB as estimated from old Galactic globular clusters \citep[dashed black line from][]{Valenti04}. The MATCH-derived metallicity of KKH98 of $[M/H]=-1.4$ is indicated by a hatched vertical region. For metal poor populations, $M_{F814W}^{TRGB} $ is roughly constant with metallicity, and for the metallicity of KKH~98, $M_{F814W}^{TRGB} =-4.0$. For metal rich systems, the $M_K^{TRGB}$ is roughly constant with metallicity. At the metallicity of KKH~98, the \citet{Valenti04} relation suggests $M_K^{TRGB}=-6.05$, whereas the \citet{Marigo08} model isochrones suggest a $K$-band tip about 0.2 mags fainter. The latter estimate gives a more consistent distance to KKH~98. }
\end{figure}
\subsubsection{Reddening and Distance}
MATCH provides an estimate of the foreground dust obscuration and distance to KKH~98. By varying these two parameters, MATCH effectively normalizes the models in color magnitude space so that they best match the observational data. Because different stellar evolution models will have different normalizations, the best fit reddening and distances reported by MATCH are model-dependent and therefore should not be considered measurements of the actual distance and foreground reddening. However, a comparison of the reddening and distance results from the two CMDs presented in this paper will help inform whether consistent results are possible from the optical and IR data sets.
For the optical data, MATCH infers a global $A_V= 0.52 \pm 0.06$. Using the \citet{Cardelli89} reddening law, this translates to an $A_I=0.23$. To within the uncertainties, the IR data gives a similar dust absorption of $A_V=0.42\pm0.08$. Using the optical CMD, MATCH finds a distance modulus of $DM=27.07\pm0.09$. The IR CMD produces a similar distance modulus of $DM=27.10\pm0.07$. To within the uncertainties, both estimates are similar to distances quoted in the literature --- e.g. $DM=26.95\pm0.11$\citep{Karachentsev02} and $DM=27.023$ \citep{Dalcanton09}. More on the distance to KKH~98 is given below.
\subsection{Estimating Distance from the Tip of the Red Giant Branch \label{sec:TRGB}}
While the MATCH-derived distances arise from normalizing the entire CMD to models, specific features in the CMD can also be used to estimate the distance. One of the best distance indicators available in the KKH~98 data set is the F814W-band magnitude of TRGB. Figure \ref{fig:trgb2} shows the F814W-band magnitude of the TRGB as a function of metallicity and age, using isochrones from \citet{Marigo08}\footnote{Models from CMD 2.1 http://stev.oapd.inaf.it/cmd}. For old, metal-poor stellar populations the TRGB has an absolute magnitude of $M_{F814W}^{TRGB}=-4.0$ (Figure \ref{fig:trgb2}). We estimate the apparent magnitude of the TRGB in KKH~98 to be at F814W$=23.2 \pm 0.1$. Correcting for reddening using $A_I=0.23$, this gives a distance modulus (DM) of $DM=26.97 \pm 0.10$, in good agreement with \citet{Karachentsev02} who found $DM=26.95\pm0.11$ from the $I$-band TRGB.
The apparent $K$-band magnitude of the TRGB ($K=21.2\pm0.1$ for KKH~98) can also be used to estimate the distance of KKH~98. In Galactic globular clusters with old stars, the $K$-band magnitude of the TRGB has been shown to be a function of metallicity \citep{Valenti04}. \citet{Gullieuszik07} parameterize this empirical function as:
\begin{equation}
\label{eqn:met}
M_K^{TRGB}=-6.92-0.62\;[M/H].
\end{equation}
\noindent
Assuming the MATCH-estimated KKH~98 metallicity for old stars of $[M/H]=-1.4$, equation \ref{eqn:met} gives $M_K^{TRGB}=-6.05$. Given the apparent TRGB magnitude of $K=21.2\pm0.1$ and negligible dust absorption in the $K$-band, we calculate a KKH~98 distance modulus of $DM=27.25\pm 0.10$. This distance measurement is somewhat farther than the distance estimated from the apparent F814W-band magnitude of the TRGB.
However, another estimate of the absolute $K$-band TRGB can be made from the \citet{Marigo08} models shown in Figure \ref{fig:trgb2}, which are also the models used by MATCH. These models suggest that, at low metallicities and old ages, the $K$-band magnitude of the TRGB is 0.2 magnitudes fainter than the value from Equation \ref{eqn:met}. Using the \citet{Marigo08} models, we find a distance modulus of $DM=27.05\pm 0.10$, which is in very good agreement with the DM derived from optical CMD.
\section{Discussion \label{sec:discuss}}
Using high spatial resolution optical ({\emph{HST}}) and near-IR (Keck AO) images, we have constructed CMDs of the stellar populations within dwarf irregular galaxy, KKH~98. From these data, we reconstructed the full star formation history of the galaxy, the first time this has been done in the IR for a galaxy beyond the Local Group. Below we compare the optical and IR results and demonstrate how these techniques may be useful for galaxies at larger distances.
\begin{figure}
\center
\includegraphics[scale=0.4]{f14.ps}
\caption{\label{fig:ik_fixbi} Same as Figure \ref{fig:ik_fixik}, only now the models (blue diamonds) are created using the optically-derived SFH. The over-production of the AGB in the old time bins is basically identical to that for the IR derived star formation history. However, in this case, significant numbers of AGB stars are also predicted at younger ages. The young model AGB stars are not only over-produced, but especially for log(age) $=$ 8.6 to 8.8, their colors are also typically 0.5 to 1 magnitude redder than the actual AGB populations in KKH~98. }
\end{figure}
\subsection{Comparison of Optical and IR Results}
Although the number of stars available for analysis in the IR was an order of magnitude smaller than for the deep {\emph{HST}}\ optical data, the bulk of measured properties --- including the broad trends in star formation history and metal enrichment --- agree between the two data sets. There are some differences however. The most obvious is that the IR data set is not sensitive to the youngest stars (i.e. age $< 0.1$ Gyr), because the bluest main sequence stars emit the bulk of their radiation in the UV and optical. Therefore, {\emph{HST}}\ optical observations are required to constrain the star formation histories at the youngest ages.
There are also differences in the details of the star formation histories at ages of $\sim0.1-1$ Gyrs ago, precisely when AGB stars are expected to dominate the information content of the IR CMD. For instance, 0.5 Gyrs ago the optical data suggest a burst with a star formation rate four times the prediction of the IR data. The evidence for this burst comes from the well-populated MS turn-off of the optical CMD at that age (Figure \ref{fig:bi_fixbi}).
A burst of this age should produce AGB stars observable in our optical and IR CMDs. Unfortunately, MATCH predicts that this burst should produce roughly the entire observed AGB population, which would only make sense if there were no AGB production at older times (Figures \ref{fig:bi_fixbi}). The over-production of AGB stars by the models is even more striking if we model the burst in the IR. Figure \ref{fig:ik_fixbi} shows a realization of the model IR CMD created with the optically-derived star formation history. Not only does the model predict large numbers of AGB stars at 0.5 Gyrs ago, it also suggests that their colors should be $0.5-1$ magnitude redder than the AGB populations seen in the data. As the SFH derived from the IR data relies heavily on the AGB sequences, it is now clear why the IR data under predicts the star formation in this time bin; there are few AGB stars in the KKH~98 data that match the luminous, red AGB stars predicted by the models for a burst of that age. Figure \ref{fig:ik_fixbi} also suggests that age/metallicity degeneracies are not producing the discrepancy; the predicted AGB stars are well separated from the observed AGB stars in color space.
Not only do the models over-produce young metal poor AGB populations (and predict redder colors than are seen), they also over-produce old metal poor AGB populations (Figures \ref{fig:bi_fixbi} and \ref{fig:ik_fixik}). For populations older the 3 Gyrs, the models predict $\sim57 \pm 8$ (optical, assuming Poisson uncertainties) and $\sim79 \pm9$ (IR) AGB stars in KKH~98. These predictions are both larger than the total number of 42 observed AGB stars, which should presumably include some younger AGB stars as well. Taken together, these results suggest issues with models of AGB stars especially for metal poor populations such as those found in KKH~98.
The modelling of AGB depends crucially on the prescriptions for two difficult-to-model phenomena: mass loss and the efficiency of third dredge up episodes. To better address the uncertainties related to these processes, modellers resort to synthetic codes in which a few parameters are tuned to fit observational data, typically from the Magellanic Clouds \citep[e.g.][]{Marigo07}.
While synthetic AGB codes have successfully reproduced the numbers, luminosities, and colors of AGB stars in the Magellanic Clouds \citep{Cioni99, Nikolaev00, Marigo07}, the situation in other galaxies has been less encouraging. For instance, \citet{Gullieuszik08} showed that models calibrated on the LMC and SMC over-predict the AGB lifetimes of the old, metal poor dwarf spheroidal, Leo~II, by as much as a factor of six. Girardi et al. (in preparation) reaches a similar conclusion based on a larger sample of ANGST galaxies.
Perhaps more relevant to the present paper are the conclusions from a similar work from \citet{Held10}: In Leo I dSph, a metal poor galaxy containing young stellar populations (therefore more similar to KKH~98), they find that the carbon stars predicted by the models are about twice those observed. We may be seeing a similar issue in KKH~98, where we would need to roughly triple the AGB population in the data to match the models, and the difficulties with the AGB seem to be particularly problematic at younger ages. We will potentially be able to better constrain AGB models in this vital time range with observations of the AGB populations of a wide variety of galaxies, as will be possible with the ANGST WFC3 survey.
\subsection{Application of These Techniques in Other Systems}
In KKH~98, the IR bright stars are well separated, with typical separations of $~\sim1\arcsec$. Assuming these separations are common within other galaxies, crowding should not limit the technique until much larger distances (up to 10 Mpc for Keck). The primary limiting factor for future ground based observations with AO will therefore be sensitivity above the thermal background. Our study was completed with a 45 minute exposure. At larger distances, significantly more exposure time, or better AO correction will be required to reach the depth of our study ($\sim2$ magnitudes below the TRGB). For instance, to reach comparable S/N in a galaxy twice as distant would require 16 times the total exposure time, or 12 hours. There are, however, several ways in which this time requirement could be reduced.
1) Improved AO performance. The observing conditions during this run were sub-optimal with only $\sim20$\% Strehl ratio (ratio of peak luminosity in the PSF to theoretical maximum peak luminosity for a diffraction limited image). With the Keck AO upgrade, typical Strehl ratio's of 30\% or more are common. A higher Strehl ratio will dramatically increase the sensitivity above background. In addition, if the Strehl ratio were more uniform across the image larger numbers of faint stars would be well photometered towards the edges of the image. More uniform AO performance will be possible with upcoming Multi-Conjugate AO systems such as the one being developed for Gemini south, which will have an isoplanatic patch 4 times the size of the Keck AO system.
2) At shorter wavelengths, the thermal IR background is reduced. In $H$-band, \citet{Melbourne07a} showed that it was possible to photometer an $H=24.0$ [Vega] point source to a photometric precision of 17\% in an hour long Keck AO exposure. The $H$-band sensitivity is significantly deeper than the $K$-band limit. The drawbacks for using this strategy are, (1) the AO system typically achieves smaller Strehl ratios at shorter wavelengths, and (2) there is a smaller wavelength lever-arm in the CMD when using $H$-band (1.6 $\mu$m) data compared with $K$-band (2.2 $\mu$m). Despite these limitations $H$-band may be a good strategy for galaxies at larger distances than KKH~98.
3) With near-IR WFC3 observations from {\emph{HST}}, both the thermal background and field of view issues can be solved. We expect the planned WFC3 observations of the ANGST sample to reach significantly deeper than our Keck AO observations. These observations will be made in both the $H$ and $J$-bands. The roughly factor of 3 loss in resolution should be relatively unimportant at the distances of the ANGST sample, where we have shown that the IR luminous stars are typically not crowding limited. In more distant galaxies, crowding will be a problem for {\emph{HST}} at these wavelengths.
4) Future instruments such as AO on a Thirty Meter Telescope or JWST should also overcome the limitation of our current data set. These telescopes will potentially resolve individual stars in galaxies out to the Virgo Cluster (18 Mpc) and beyond.
\section{Conclusions}
We obtained high spatial resolution $K$-band imaging of dwarf irregular galaxy KKH~98 using the Keck LGS AO system. The resolution of the AO images ($\sim0.1\arcsec$) was a good match to the existing {\emph{HST}}\ optical images. In addition, the AO PSFs had a significantly higher contrast above the background compared with seeing limited images, resulting in 15\% photometry at $K\sim22$ [Vega] for our 45 minute exposure \citep[similar to the Keck AO photometric uncertainty found by][in the Local Group dwarf IC 10]{Vacca07}.
From these images, we created optical and near-IR CMDs of the stars in KKH~98, and estimated the star formation and metal enrichment history of the galaxy. On average, KKH~98 has experienced a low level ($\sim5 \times10^{-4} M_{\odot}$ yr$^{-1}$) star formation rate over its lifetime, and has maintained a metallicity of roughly 1/30th --- 1/10th solar. In addition, there is evidence for a prolonged period of quiescence (1-3 Gyrs ago) where little star formation occurred. Despite the relatively small number of IR bright stars (592 compared to 2539 in the optical), the results from the IR data match those from the optical for the bulk of cosmic time.
However, there are some discrepancies in the optical and IR derived SFHs for ages $<1$ Gyr. In the optical CMD, younger populations were primarily constrained by the main sequence turn-offs, whereas the IR CMD constrained younger populations via the asymptotic giant branch. Discrepancies between the two can be explained by an over-production of AGB stars in the SFH models. We estimate that the models over-predict the numbers of AGB stars in KKH~98 by as much as a factor of three. The over-production of AGB stars in SFH models has been noted previously in studies of low metallicity galaxies \citep{Gullieuszik08}, suggesting that AGB lifetimes (and mass loss rates) may be metallicity dependent.
Because the IR bright stars were not crowded, these techniques may be useful for measuring stellar populations at much larger distances, especially with the James Webb Space Telescope and AO on future Thirty-Meter class telescopes. However, improved models of AGB evolution may be required to take full advantage of these techniques in the IR.
\acknowledgments
The adaptive optics data presented herein were obtained at the Keck Observatory, which is operated as a scientific partnership among Caltech, UC, and NASA. The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. The laser guide star adaptive optics system was funded by the W. M. Keck Foundation. This work has been supported in part by the NSF Science and Technology Center for Adaptive Optics, managed by the University of California (UC) at Santa Cruz under the cooperative agreement No. AST-9876783. S. M. Ammons acknowledges support from the Bachmann instrumentation program through UCO/Lick. L. Girardi was partially supported by contract ASI-INAF I/016/07/0.
\bibliographystyle{apj}
|
\section{Introduction}
\label{introduction} Theory of the so-called $\mathcal{PT}$-symmetric Hamiltonians has been a very active research
area in recent years \cite{Bender, Bender-1,CGM,Dorey,Znojil2,Ali1,Shin,Shin2,Znojil1}. While they are not
self-adjoint in general, many $\mathcal{PT}$-symmetric Hamiltonians have real spectra
\cite{Bender,CGM,Dorey,Shin}. Self-adjointness of Hamiltonians is not physical requirement, but guarantees reality
of spectrum and conservation of probability. So one searches for non-self-adjoint Hamiltonians that have real
spectra. In this paper, we study the Schr\"odinger eigenvalue problems in the complex plane with complex
polynomial potentials under decaying boundary conditions along two f\/ixed rays to inf\/inity. We will prove that the
Schr\"odinger eigenvalue problem has inf\/initely many real eigenvalues if and only if the eigenvalue problem is
$\mathcal{PT}$-symmetric or it is a translation of a $\mathcal{PT}$-symmetric problem.
For an integer $m\geq3$ we consider the Schr\"odinger eigenvalue problem
\begin{equation}\label{ptsym}
H u(z):=\left[-\frac{d^2}{dz^2}-(iz)^m-P(iz)\right]u(z)=\lambda u(z),\quad\text{for some $\lambda\in\mathbb C$},
\end{equation}
with the boundary condition that
\begin{equation}\label{bdcond}
\text{$u(z)\rightarrow 0$ exponentially, as $z\rightarrow \infty$ along the two rays} \ \arg
(z)=-\frac{\pi}{2}\pm \frac{2\pi}{m+2},
\end{equation}
where $P$ is a polynomial of degree at most $m-1$ of the form
\[
P(z)=a_1z^{m-1}+a_2z^{m-2}+\cdots+a_{m},\qquad a:=(a_1,a_2,\ldots, a_m)\in\mathbb C^m.
\]
The existence of inf\/initely many eigenvalues is known. Sibuya \cite{Sibuya} showed that the eigenvalues of $H$ are
zeros of an entire function of order $\rho:=\frac{1}{2}+\frac{1}{m}\in(0, 1)$ and hence, by the Hadamard
factorization theorem (see, e.g., \cite{Conway}), there are inf\/initely many eigenvalues.
The Hamiltonian $H$ in \eqref{ptsym} with the potential $V(z)$ under the boundary condition \eqref{bdcond} is
called {\it $\mathcal{PT}$-symmetric} if $\overline{V(-\overline{z})}=V(z)$, $z\in\mathbb C$. Note that
$V(z)=-(iz)^m-P(iz)$ is a $\mathcal{PT}$-symmetric potential if and only if $a\in\mathbb R^m$.
In this paper, we will derive the following asymptotic expansion of the eigenvalues $\lambda_n$, ordered in the
order of their magnitude.
\begin{theorem}\label{the_main}
There exists $N=N(m)\in\mathbb Z$ such that the eigenvalues $\left\{\lambda_n\right\}_{n=N}^{\infty}$ of $H$ satisfy
that
\begin{equation}\label{asymp_eqn1}
\sum_{j=0}^{m+1}c_j(
{a})\lambda_n^{\rho-\frac{j}{m}}+O\left(\lambda_n^{-\rho}\right)=n+\frac{1}{2}\quad \text{as $n\to+\infty$},
\end{equation}
where $c_0( {a})= \pi^{-1}\sin\left(\frac{\pi}{m}\right)B\left(\frac{1}{2},\,1+\frac{1}{m}\right)$ and $c_1(a)=0$
and $c_j( {a})$, $2\leq j\leq m+1,$ are explicit non-zero polynomials in the coefficients $ {a}$ of the
polynomial potential, defined in~\eqref{c_def}.
\end{theorem}
As a consequence of \eqref{asymp_eqn1}, we obtain a necessary and suf\/f\/icient condition for the anharmonic
oscillator $H$ to have inf\/initely many real eigenvalues as follows.
\begin{theorem}\label{equiv-for}
The anharmonic oscillator $H$ with a potential $V(z)$ has infinitely many real eigenvalues if and only if $H$
with the potential $V(z-z_0)$ is $\mathcal{PT}$-symmetric for some $z_0\in\mathbb C$.
Moreover, no $H$ has infinitely many real and infinitely many non-real eigenvalues.
\end{theorem}
If $H$ is $\mathcal{PT}$-symmetric, then $u(z)$ is an
eigenfunction associated with the eigenvalue $\lambda$ if and only
if $\overline{u(-\overline{z})}$ is an eigenfunction associated
with the eigenvalue $\overline{\lambda}$. Thus, the eigenvalues of
a~$\mathcal{PT}$-symmetric Hamiltonian $H$ either are real or come
in complex conjugates. Also,
magnitude of the large eigenvalues is strictly increasing~\cite{Shin2}. Thus,
if $H$ is $\mathcal{PT}$-symmetric, then
eigenvalues are all real with at most f\/initely many exceptions.
This paper is organized as follows. In Section~\ref{prop_sect},
we will introduce results of Hille~\cite{Hille},
Sibuya~\cite{Sibuya}, and Shin~\cite{Shin2,Shin3} on properties of the solutions of~\eqref{ptsym}. We will present asymptotics of solutions of
\eqref{ptsym}, introduce the spectral determinant of $H$, and provide an asymptotic expansion for the spectral
determinant in the sector where all but f\/initely many eigenvalues lie. In Section \ref{asymp_eigen}, we will
prove Theorem~\ref{the_main} where we use the asymptotic expansion of the spectral determinant. In Section~\ref{last_sect}, we will prove Theorem~\ref{equiv-for} and provide other spectral results; expres\-sing~$\lambda_n$ as an asymptotic formula in $n$ and obtaining an asymptotic formula for nearest neighbor spacing of
eigenvalues, monotonicity of their magnitude, and more.
\section{Properties of the solutions}
\label{prop_sect}
In this section, we will introduce results of Hille \cite{Hille}, Sibuya \cite{Sibuya}, and Shin
\cite{Shin2,Shin3} on the properties of solutions of \eqref{ptsym}.
First, let $u$ be a solution of (\ref{ptsym}) and let $v(z)=u(-iz)$. Then $v$ solves
\begin{equation}\label{rotated}
-v^{\prime \prime}(z)+[z^m+P(z)+\lambda]v(z)=0.
\end{equation}
Solutions of \eqref{rotated} have simple asymptotic behavior near inf\/inity in the complex plane and in order to
describe the asymptotic behavior, we will use the Stokes sectors. The Stokes sectors $S_k$ of the equation
(\ref{rotated}) are
\[
S_k=\left\{z\in \mathbb C:\left|\arg (z)-\frac{2k\pi}{m+2}\right|<\frac{\pi}{m+2}\right\}\quad\text{for}\ k\in \mathbb Z.
\]
Hille \cite[\S~7.4]{Hille} showed that every nonconstant solution of (\ref{rotated}) either decays to zero or
blows up exponentially in each Stokes sector. Moreover, if a nonconstant solution $v$ of \eqref{rotated} decays in
a~Stokes sector, it must blow up in its two adjacent Stokes sectors. However, when $v$ blows up in a Stokes
sector, $v$ need not be decaying in its adjacent Stokes sectors. Thus, the decaying boundary condition
\eqref{bdcond} becomes that $v(z)\to 0$ as $z\to \infty$ in the sectors $S_1\cup S_{-1}$.
We will also use the following functions of the coef\/f\/icients $a\in\mathbb C^m$ of the polynomial $P$. Let
\begin{gather*}
\text{$b_{j,k}( {a})$ be the coef\/f\/icient of $z^{mk-j}$ in ${\frac{1}{2}\choose{k}}\left(P(z)\right)^k$\,\,for
$1\leq k\leq j$,\, and}
\\
b_j( {a})=\sum_{k=1}^j b_{j,k}( {a}),\quad j\in\mathbb N.
\end{gather*}
We also let $r_m=-\frac{m}{4}-\nu(a)$ and $\mu( {a})=\frac{m}{4}-\nu( {a})$, where
\begin{gather*}
\nu( {a})=\left\{
\begin{array}{ll}
0 &\text{if $m$ is odd,}\\
b_{\frac{m}{2}+1}( {a}) &\text{if $m$ is even.}
\end{array}
\right.
\end{gather*}
Now we are ready to introduce results of Sibuya \cite{Sibuya} and Shin \cite{Shin3} on existence of a solution
that has a f\/ixed asymptotic behavior in $S_{-1}\cup S_0\cup S_1$, and on its properties.
\begin{theorem}\label{prop}
Equation \eqref{rotated}, with $a\in \mathbb C^m$, has a solution $f(z,a,\lambda)$ with the following properties.
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $f(z,a,\lambda)$ is an entire function of $z$, $a $ and $\lambda$.
\item[$(ii)$] $f(z,a,\lambda)$ and $f^\d(z,a,\lambda)=\frac{\partial}{\partial z}f(z,a,\lambda)$ admit the following asymptotic
expansions. For each fixed $\varepsilon>0$,
\begin{gather*}
f(z,a,\lambda)= z^{r_m}\big(1+O\big(z^{-1/2}\big)\big)\exp\left[-F(z,a,\lambda)\right],\\
f^\d(z,a,\lambda)= -z^{r_m+\frac{m}{2}}\big(1+O\big(z^{-1/2}\big)\big)\exp\left[-F(z,a,\lambda) \right],
\end{gather*}
as $z$ tends to infinity in the sector $|\arg z|\leq \frac{3\pi}{m+2}-\varepsilon$, uniformly on each compact set
of $(a,\lambda)$-values. Here
\[
F(z,a,\lambda)=\frac{2}{m+2}z^{\frac{m}{2}+1}+\sum_{1\leq j<\frac{m}{2}+1}\frac{2}{m+2-2j}b_j(a) z^{\frac{1}{2}(m+2-2j)}.
\]
\item[$(iii)$] For each $\delta>0$, $f$ and $f^\d$ also admit the asymptotic expansions,
\begin{gather}
f(0,a,\lambda)= \left[1+O\left(\lambda^{-\rho}\right)\right]\lambda^{-1/4}\exp\left[L(a,\lambda)\right],\label{eq1}\\
f^\d(0,a,\lambda)= -\left[1+O\left(\lambda^{-\rho}\right)\right]\lambda^{1/4}\exp\left[L(a,\lambda)\right],\label{eq2}
\end{gather}
as $\lambda\to\infty$ in the sector $|\arg(\lambda)|\leq\pi-\delta$, uniformly on each compact set of $a\in\mathbb C^m$,
where $\rho:=\frac{1}{2}+\frac{1}{m}$
\[
L(a,\lambda)=\int_0^{+\infty}\left(\sqrt{t^m+P(t)+\lambda}-
t^{\frac{m}{2}}-\sum_{j=1}^{\lfloor\frac{m+1}{2}\rfloor}b_j(a)t^{\frac{m}{2}-j}
-\frac{\nu(a)}{t+1}\right)dt.
\]
\item[$(iv)$] The entire functions $\lambda\mapsto f(0,a,\lambda)$
and $\lambda\mapsto f^\d(0,a,\lambda)$ have orders
$\rho=\frac{1}{2}+\frac{1}{m}$.
\end{enumerate}
\end{theorem}
\begin{proof}
See Theorems 6.1, 7.2, 19.1, and 20.1 in \cite{Sibuya} for proof. Sibuya showed \eqref{eq1} and \eqref{eq2} with
error terms $o(1)$ and I improved them to $O\left(\lambda^{-\rho}\right)$ in \cite{Shin3}.
\end{proof}
\subsection{Spectral determinant}
In this subsection, we will introduce the function whose zeros are the eigenvalues of $H$.
Let
\[
\omega=\exp\left[\frac{2\pi i}{m+2}\right]
\]
and def\/ine
\[
G^k(a):=\big(\omega^{-k}a_1, \omega^{-2k}a_2,\ldots,\omega^{-mk}a_m\big)\quad \text{for} \ k\in \mathbb Z.
\]
Then one can see from the def\/inition of $b_{j,k}( {a})$ that
\begin{equation}\label{bjk}
b_{j,k}\big(G^{\ell}( {a})\big)=\omega^{-j\ell}b_{j,k}( {a})\qquad \text{and}\qquad b_j\big(G^{\ell}( {a})\big)=\omega^{-j\ell}b_j(
{a})\quad \text{for} \ \ell\in\mathbb Z.
\end{equation}
Also, by Theorem \ref{prop} there exists $f(z,G^k(a),\omega^{-mk}\lambda)$ that decays in $S_0$ and we let
\[
f_k(z,a,\lambda):=f\big(\omega^{-k}z,G^k(a),\omega^{-mk}\lambda\big).
\]
So $f_k$ decays in $S_k$ and blows up in its two adjacent Stokes sectors. Since $f_0$ decays in $S_0$ while~$f_1$
blows up in the same sector, we conclude that $f_0$ and $f_1$ are linearly independent. Thus, any solution can be
expressed as a linear combination of these two. In particular,
\[
f_{-1}(z,a,\lambda)=C(a,\lambda)f_0(z,a,\lambda)+\widetilde{C}(a,\lambda)f_{1}(z,a,\lambda)\quad \text{for some
$C(a,\lambda)$ and $\widetilde{C}(a,\lambda)$.}
\]
One can write these coef\/f\/icients in terms of the Wronskians of solutions. That is,
\[
C(a,\lambda)=\frac{\mathcal{W}_{-1,1}(a,\lambda)}{\mathcal{W}_{0,1}(a,\lambda)}\qquad\text{and}\qquad
\widetilde{C}(a,\lambda)= -\frac{\mathcal{W}_{-1,0}(a,\lambda)}{\mathcal{W}_{0,1}(a,\lambda)},
\]
where $\mathcal{W}_{j,k}=f_jf_k^\d -f_j^\d f_k$ is the Wronskian of $f_j$ and $f_k$. Since $f_0$ and $f_1$ are
linearly independent, $\mathcal{W}_{0,1}(a,\lambda)\not=0$ and likewise $\mathcal{W}_{-1,0}(a,\lambda)\not=0$.
Then by Theorem \ref{prop}$(i)$, for each f\/ixed $a\in\mathbb C^m$, the function $\lambda\mapsto C(a,\lambda)$ is entire.
Moreover, $\lambda$ is an eigenvalue of $H$ if and only if $C(a,\lambda)=0$. We call $C(a,\lambda)$ the spectral
determinant of $H$.
\subsection[Asymptotic expansions of $C(a,\lambda)$]{Asymptotic expansions of $\boldsymbol{C(a,\lambda)}$}\label{asymp_sect}
In \cite{Shin2}, the asymptotics of $C(a,\lambda)$ is provided and it is showed that all but f\/initely many zeros
$\lambda$ of $C(a,\lambda)$ lie in a small sector containing the positive real $\lambda$-axis. Here we will
improve this asymptotics of $C(a,\lambda)$ as $\lambda\to\infty$ in a sector containing the positive real
$\lambda$-axis as follows.
\begin{theorem}\label{thm_sector2}
Suppose that $m\geq 4$. Then for each f\/ixed $a\in\mathbb C^m$ and $0<\delta<\frac{\pi}{m+2},$
\begin{gather}
C(a,\lambda)= \big[\omega^{\frac{1}{2}}+O\left(\lambda^{-\rho}\right)\big]\exp\left[L\big(G^{-1}(a),\omega^{-2}\lambda\big)-L(a,\lambda)\right]\nonumber\\
\phantom{C(a,\lambda)=}{} +\big[\omega^{\frac{1}{2}+2\nu(a)}+O\left(\lambda^{-\rho}\right)\big]\exp\left[L(G(a),\omega^{2}\lambda)-L(a,\lambda)\right],\label{asymp_1}
\end{gather}
as $\lambda\to\infty$ in the sector
\begin{equation}\label{sector4}
\pi-\frac{4\lfloor\frac{m}{2}\rfloor\pi}{m+2}+\delta \leq \arg(\lambda)\leq \pi-\frac{4\pi}{m+2}-\delta.
\end{equation}
If $m=3$ then
\begin{gather}
C(a,\lambda)= \big[-\omega^{-2}+O\left(\lambda^{-\rho}\right)\big]\exp\left[L\big(G^{4}(a),\omega^{-2}\lambda\big)-L(a,\lambda)\right]\nonumber\\
\phantom{C(a,\lambda)=}{} -\big[i\omega^{\frac{7}{4}}+O\left(\lambda^{-\rho}\right)\big]\exp\left[-L(G^2(a),\omega^{-1}\lambda)-L(a,\lambda)\right],\label{asymp_2}
\end{gather}
as $\lambda\to\infty$ in the sector
\begin{equation}\label{sector2-1}
-\frac{\pi}{5}+\delta \leq \arg(\lambda)\leq \pi-\delta.
\end{equation}
\end{theorem}
\begin{proof}
In Theorems 13 and 14 of Shin \cite{Shin2}, similar asymptotics
of $C(a,\lambda)$ are proved with the error terms $o(1)$ instead
of $O\left(\lambda^{-\rho}\right)$. With the improved asymptotics
\eqref{eq1} and \eqref{eq2}, one can closely follow proofs of
Theorems 13 and 14 in \cite{Shin2} to complete the proof.
\end{proof}
Next, in order to further examine the improved asymptotics of $C(a,\lambda)$ in Theorem \ref{thm_sector2}, we will
use the following asymptotics of $L(a,\lambda)$.
\begin{lemma}\label{asy_lemma}
Let $m\geq 3$ and $a\in\mathbb C^m$ be fixed. Then
\[
L(a,\lambda)=\sum_{j=0}^{\infty}K_{m,j}(a)\lambda^{\frac{1}{2}+\frac{1-j}{m}}-\frac{\nu(a)}{m}\ln(\lambda)
\]
as $\lambda\to\infty$ in the sector $|\arg(\lambda)|\leq \pi-\delta$ with
$K_{m,0}(a)=\frac{B\left(\frac{1}{2},1+\frac{1}{m}\right)}{2\cos\left(\frac{\pi}{m}\right)}>0$ and for $j\geq 1$,
\begin{equation}\label{Kmj=}
K_{m,j}( {a})=\sum_{k=\lfloor\frac{j-1}{m}\rfloor+1}^{j}K_{m,j,k}\,b_{j,k}( {a}),
\end{equation}
where for $\frac{j}{m}\leq k\leq j$,
\begin{gather*}
K_{m,j,k} :=\left\{
\begin{array}{ll}
\displaystyle -\frac{2}{m}
\quad &\text{if $j=1$},\vspace{2mm}\\
\displaystyle \frac{2}{m}\left(\ln 2-\sum\limits_{s=1}^{k-1}\frac{1}{2s-1}\right) \, &\displaystyle \text{if $j=\frac{m}{2}+1$, $m$ even,}\vspace{2mm}\\
\displaystyle \frac{1}{m}B\left(k-\frac{j-1}{m},\,\frac{j-1}{m}-\frac{1}{2}\right) &\displaystyle
\text{if $j\not=1$ or $j\not=\frac{m}{2}+1$.}
\end{array}\right.
\end{gather*}
\end{lemma}
\begin{proof}
See the Appendix in \cite{Shin3} for a proof.
\end{proof}
\begin{remark}\label{remark1}\sloppy
The coef\/f\/icients $K_{m,j}( {a})$ have the following properties that can be deduced from~\eqref{Kmj=} and the
def\/inition of $b_{j,k}(a)$:
\begin{itemize}\itemsep=0pt
\item[$(i)$] The $K_{m,j}( {a})$ are all real polynomials in terms
of the coef\/f\/icients $ {a}$ of $P$.
\item[$(ii)$] For each
$1\leq j\leq m$, the polynomial $K_{m,j}( {a})$ depends only on
$a_1,a_2,\dots,a_{j}$. Furthermore, it is a non-constant linear
function of $a_j$.
\end{itemize}
\end{remark}
\section{Proof of Theorem \ref{the_main}}\label{asymp_eigen}
In this section, we will prove Theorem \ref{the_main}.
\begin{proof}[Proof of Theorem~\ref{the_main}]
In order to get asymptotics for the eigenvalues $\lambda$, we
examine asymptotics of $C(a,\lambda)$ since the zeros of
$C(a,\lambda)$ are the eigenvalues of $H$. Since all but f\/initely
many eigenvalues lie near the positive real $\lambda$-axis, the
asymptotics in Theorem \ref{thm_sector2} would be suf\/f\/icient for
our purpose.
For $m\geq 4$ and $a\in\mathbb C^m$ f\/ixed, we set $C(a,\lambda)=0$ in
\eqref{asymp_1}, rearrange the asymptotic equation to get
\[
\big[1+O\left(\lambda^{-\rho}\right)\big]\exp\left[L\big(G(a),\omega^2\lambda\big)-L\big(G^{-1}(a),\omega^{-2}\lambda\big)\right]
=-\omega^{-2\nu(a)},
\]
and absorb the term $[1+O\left(\lambda^{-\rho}\right)]$
into the exponential term to obtain
\[
\exp\left[L\big(G(a),\omega^2\lambda\big)-L\big(G^{-1}(a),\omega^{-2}\lambda\big)+O\left(\lambda^{-\rho}\right)\right]
=-\omega^{-2\nu(a)},
\]
or equivalently
\begin{equation}\label{main_eq}
L\big(G(a),\omega^2\lambda\big)-L\big(G^{-1}(a),\omega^{-2}\lambda\big)+O\left(\lambda^{-\rho}\right) =(2n+1)\pi i
-\frac{4\nu(a)\pi}{m+2}i,\quad \text{for some $n\in\mathbb Z$}.
\end{equation}
Since $L(a,\lambda)=K_{m,0}(a)\lambda^{\rho}(1+o(1))$ by Lemma \ref{asy_lemma}, one gets
\begin{equation}\label{L-estmat}
L\big(G(a),\omega^2\lambda\big)-L\big(G^{-1}(a),\omega^{-2}\lambda\big)+O\left(\lambda^{-\rho}\right)
=2iK_{m,0}(a)\sin\left(\frac{2\pi}{m}\right)\lambda^{\rho}(1+o(1)),
\end{equation}
as $\lambda\to\infty$ in the sector \eqref{sector4} containing the
positive real $\lambda$-axis. Thus, the function
\[
\lambda\mapsto H(a,\lambda):=L(G(a),\omega^2\lambda)
-L(G^{-1}(a),\omega^{-2}\lambda)+O\left(\lambda^{-\rho}\right)
\]
maps the region $|\lambda|\geq M$ and $|\arg(\lambda)|\leq
\varepsilon$ for some large $M>0$ and small $\varepsilon>0$ onto a region that contains the positive imaginary
axis near inf\/inity. Also, $H(a,\cdot)$ is analytic in the region $|\lambda|\geq M$ and $|\arg(\lambda)|\leq
\varepsilon$. Then, Sibuya \cite[pp.~131--133]{Sibuya} showed that for every closed ball $\{a\in\mathbb C^m: |a|\leq
r\}$, there exists $N_0=N_0(m, r)\in\mathbb N$ such that
for every integer $n\geq N_0$ there exists exactly one $\lambda_n$ for which \eqref{main_eq} holds, that is,
\begin{equation}\label{L-eq}
H(a,\lambda_n)=(2n+1)\pi i -\frac{4\nu(a)\pi}{m+2}i.
\end{equation}
(Sibuya showed the existence of such an $N_0$ with the estimate \eqref{L-estmat}.)
Hence, from Lemma \ref{asy_lemma} we infer that
\begin{gather}
H(a,\lambda_n)= L\big(G(a),\omega^2\lambda_n\big)-L\big(G^{-1}(a),\omega^{-2}\lambda_n\big)+O\left(\lambda_n^{-\rho}\right)\nonumber\\
\phantom{H(a,\lambda_n)}{} = \sum_{j=0}^{m+1}\left[ K_{m,j}(G(a))\left(\omega^{2}\lambda_n\right)^{\frac{1}{2}+\frac{1-j}{m}}
-K_{m,j}(G^{-1}(a))(\omega^{-2}\lambda_n)^{\frac{1}{2}+\frac{1-j}{m}}\right]\nonumber\\
\phantom{H(a,\lambda_n)=}{} +\frac{\nu\left(G^{-1}(a)\right)}{m}\ln\left(\omega^{-2}\lambda_n\right)-\frac{\nu\left(G(a)\right)}{m}\ln\left(\omega^{2}\lambda_n\right)
+O\left(\lambda_n^{-\rho}\right)\nonumber\\
\phantom{H(a,\lambda_n)}{}= 2i\sum_{j=0}^{m+1}K_{m,j}(a)\sin\left(\frac{2(1-j)\pi}{m}\right)\lambda_n^{\frac{1}{2}+\frac{1-j}{m}}+\frac{\nu(a)}{m}\frac{8\pi
i}{m+2}+O\left(\lambda_n^{-\rho}\right),\label{K-eq}
\end{gather}
where we used $\nu\left(G^{\pm1}(a)\right)=-\nu(a)$ and by \eqref{bjk} and \eqref{Kmj=}, for $1\leq j\leq m+1$,
\begin{gather}
K_{m,j}(G(a))\big(\omega^{2}\big)^{\frac{1}{2}+\frac{1-j}{m}}-K_{m,j}\big(G^{-1}(a)\big)\big(\omega^{-2}\big)^{\frac{1}{2}+\frac{1-j}{m}}\nonumber\\
\qquad{}=K_{m,j}(a)\omega^{-j}\big(\omega^{2}\big)^{\frac{1}{2}+\frac{1-j}{m}}-K_{m,j}(a)\omega^j\big(\omega^{-2}\big)^{\frac{1}{2}+\frac{1-j}{m}}\nonumber\\
\qquad{}=K_{m,j}(a)\big(e^{(1-j)\frac{2\pi i}{m}}-e^{-(1-j)\frac{2\pi i}{m}}\big)\nonumber\\
\qquad{}=2iK_{m,j}(a)\sin\left(\frac{2(1-j)\pi}{m}\right).\label{K_eq}
\end{gather}
Next, one can combine \eqref{L-eq} and \eqref{K-eq} and rearrange the resulting equation to obtain
\eqref{asymp_eqn1}. That~is,
\[
\sum_{j=0}^{m+1}c_j(a)\lambda_n^{\frac{1}{2}+\frac{1-j}{m}}
+O\left(\lambda_n^{-\rho}\right)=n+\frac{1}{2}\quad \text{as
$n\to\infty$},
\]
where for $0\leq j\leq m+1$,
\begin{equation}\label{c_def}
c_j(a):=\left\{
\begin{array}{ll}
\displaystyle \frac{1}{\pi}K_{m,j}(a)\sin\left(\frac{2(1-j)\pi}{m}\right)
&\displaystyle\text{if $j\not=\frac{m}{2}+1$},\\
&\\
\displaystyle \frac{2\nu(a)}{m} &\displaystyle \text{if $j=\frac{m}{2}+1$, $m$ even.}
\end{array}\right.
\end{equation}
We still need to examine how many eigenvalues of $H$ are not in the set $\{\lambda_n\}_{n\ge N_0}$ to complete our
proof. To do this we will use Hurwitz's theorem in complex analysis.
Hurwitz's theorem \cite{Conway} says that if a sequence of analytic functions $\varphi_k$ converges uniformly to an analytic
function $\varphi$ on any compact subsets of their common domain, then for all large~$k$ functions~$\varphi_k$ and
$\varphi$ have the same number of zeros in any open subset whose boundary does not contain any zeros of~$\varphi$.
Since the eigenvalues are the zeros of $C(a,\lambda)$ that is an analytic function in variables $a$ and $\lambda$,
Hurwitz's theorem implies that eigenvalues varies continuously as~$a$ varies. That is, $\lambda_n(a)$ varies continuously from
$\lambda_n(\widehat{a})$ to $\lambda_n(\widetilde{a})$ as $a$ varies from $\widehat{a}$ to $\widetilde{a}$ in the
closed ball, where we used $\lambda_n(a)$ in order to clearly indicate its dependence on $a$.
When $a=0$, there exists an integer $N=N(m)\leq N_0$ such that the eigenvalues of $H$ that are not
in $\{\lambda_n(0)\}_{n\geq N_0}$ is exactly $(N_0-N)$.
Also, Hurwitz's theorem implies that when eigenvalues collide as $a$ varies, the number of eigenvalues before and after the
collision remains the same. So there is no sudden appearance or disappearance of the eigenvalues as $a$ varies and
hence, there are exactly $(N_0-N)$-eigenvalues that are not in the set $\{\lambda_n(a)\}_{n\geq N_0}$. Therefore,
the eigenvalues $\{\lambda_n(a)\}_{n\geq N}$ satisfy \eqref{asymp_eqn1}. This completes the proof for $m\geq 4$.
For $m=3$, proof is very similar to the case $m\geq 4$ above and we will use \eqref{asymp_2} in the place of~\eqref{asymp_1}. If $C(a,\lambda)=0$ then we rearrange the asymptotic formula \eqref{asymp_2} with
$C(a,\lambda)=0$ to obtain
\[
\big[1+O\left(\lambda^{-\rho}\right)\big]\exp\left[-L\big(G^{4}(a),\omega^{-2}\lambda\big)-L\big(G^{2}(a),\omega^{-1}\lambda\big)\right]=e^{\pi
i},
\]
as $\lambda\to\infty$ in the sector~\eqref{sector2-1}. Next, like
in \eqref{K_eq} we examine
\begin{gather*}
-K_{3,j}\big(G^4(a)\big)\left(\omega^{-2}\lambda\right)^{\frac{1}{2}+\frac{1-j}{3}}-K_{3,j}\big(G^2(a)\big)
\left(\omega^{-1}\lambda\right)^{\frac{1}{2}+\frac{1-j}{3}}\nonumber\\
\qquad{}=-\big[\omega^{-4j}\omega^{-2\left(\frac{1}{2}+\frac{1-j}{3}\right)}+\omega^{-2j}\omega^{-\left(\frac{1}{2}+\frac{1-j}{3}\right)}\big]K_{3,j}(a)\lambda^{\frac{1}{2}+\frac{1-j}{3}}\nonumber\\
\qquad{} =-\big[e^{-\frac{4(j-1)}{3}\pi i}-e^{-\frac{2(j-1)}{3}\pi i}\big]K_{3,j}(a)\lambda^{\frac{1}{2}+\frac{1-j}{3}}\nonumber\\
\qquad{} =2i\sin\left(\frac{2(1-j)\pi}{3}\right)K_{3,j}(a)\lambda^{\frac{1}{2}+\frac{1-j}{3}},\nonumber
\end{gather*}
where we used \eqref{bjk}, \eqref{Kmj=}, and $\omega^{\frac{5}{2}}=-1$. Then one can complete proof for $m=3$ just like in the case of $m\geq 4$.
\end{proof}
\begin{remark}
{\rm We order the eigenvalues in the order of their magnitude and we showed that eigenvalues vary continuously as
the coef\/f\/icients $a=(a_1,\dots, a_m)$ vary. For the large eigenvalues, $\lambda_n(a)$ are continuous functions of $a\in\mathbb C^m$ (c.f. \eqref{mono}). However,
for small eigenvalues, even though eigenvalues varies continuously as $a$ varies, we do not claim that
$\lambda_n(a)$ ordered in their order of magnitude are continuous. While these small eigenvalues are continuously
moving in the complex plane, their indices can be switched and how we number these small eigenvalues does not
af\/fect the asymptotic result in Theorem~\ref{the_main}.}
\end{remark}
\section{Corollaries and proof of Theorem \ref{equiv-for}}\label{last_sect}
In this section, we will introduce some results deduced from Theorem \ref{the_main}.
First, we invert \eqref{asymp_eqn1}, expressing $\lambda_n$ as an asymptotic formula in $n$. Also, we obtain an
asymptotic formula for nearest neighbor spacing of eigenvalues and monotonicity of their magnitude.
\begin{corollary}
One can compute numbers $d_j( {a})$ explicitly such that
\begin{equation}\label{somm_eq}
\lambda_n=\sum_{j=0}^{m+1}d_j( {a})\cdot
\left(n+\frac{1}{2}\right)^{\frac{2m}{m+2}\left(1-\frac{j}{m}\right)}+O\big(n^{-\frac{4}{m+2}}\big)\quad \text{as
$n\to+\infty$}.
\end{equation}
Also, the space between successive eigenvalues is
\[
\lambda_{n+1}-\lambda_n
\underset{n\to+\infty}{=}\frac{2m}{m+2}\left(\frac{\pi}{B\left(\frac{1}{2},1+\frac{1}{m}\right)}\right)^{\frac{2m}{m+2}}\cdot
\left(n+\frac{1}{2}\right)^{\frac{m-2}{m+2}}+o\big(n^{\frac{m-2}{m+2}}\big)\quad \text{as $n\to+\infty$.}
\]
In particular, $\lim\limits_{n\to\infty}\left|\lambda_{n+1}-\lambda_n\right|=\infty$ and
$\lim\limits_{n\to+\infty}\arg(\lambda_n)=0.$ Hence:
\begin{equation}\label{mono}
\left|\lambda_n\right|<\left|\lambda_{n+1}\right|\quad \text{for all large $n$}.
\end{equation}
\end{corollary}
\begin{proof}
Since \eqref{asymp_eqn1} is an asymptotic equation, one can solve it for $\lambda_n$ to get \eqref{somm_eq}, where
$d_0(a)=\left(c_0\right)^{-\frac{2m}{m+2}}$ and $d_j( {a})$ for $1\leq j\leq m+1$ can be computed recursively and
explicitly. Then since $d_0>0$, one can deduce
\begin{equation}\label{eq_111}
\lim_{n\to0}\arg(\lambda_n)=0\qquad \text{and}\qquad \lim_{n\to\infty}|\lambda_n|=\infty.
\end{equation}
Also, we obtain
\begin{equation}\label{mono_11}
\lambda_{n+1}
\underset{n\to+\infty}{=}\lambda_n+\frac{2m}{m+2}\left(d_0(a)\right)^{\frac{2m}{m+2}}\cdot
\left(n+\frac{1}{2}\right)^{\frac{m-2}{m+2}}+o\big(n^{\frac{m-2}{m+2}}\big)\quad \text{as
$n\to+\infty$,}
\end{equation}
where we use the generalized binomial expansion to $\big(1+\big(n+\frac{1}{2}\big)^{-1}\big)^{\alpha}$ for
some $\alpha\in\mathbb R$. Since $d_0>0$, equations~\eqref{eq_111} and~\eqref{mono_11} imply
$\lim_{n\to0}\left|\lambda_{n+1}-\lambda_n\right|=\infty$ and hence, \eqref{mono} holds.
\end{proof}
When $H$ is $\mathcal{PT}$-symmetric, the eigenvalues either are real or else appear in complex conjugates. Thus,
if $H$ is $\mathcal{PT}$-symmetric, then by monotonicity of magnitude of the large eigen\-values~\eqref{mono}, $H$
cannot have an inf\/inite pairs of complex conjugate eigenvalues and hence all but f\/initely many eigenvalues are
real.
Next, we will prove Theorem \ref{equiv-for} on a necessary and suf\/f\/icient condition for an anharmonic oscillator
$H$ having inf\/initely many real eigenvalues.
\begin{proof}[Proof of Theorem~\ref{equiv-for}]
Suppose that $H$ or its translation is $\mathcal{PT}$-symmetric. It is easy to check that the potentials $V(z)$
and $V(z-z_0)$ generate the same set of eigenvalues for any $z_0\in\mathbb C$. Thus, since $H$ or its translation is
$\mathcal{PT}$-symmetric, eigenvalues of $H$ either are real or come in complex conjugates. Then by~\eqref{mono},
we conclude that all but f\/initely many eigenvalues are real. Since there are inf\/initely many eigenvalues
\cite{Sibuya} as we noted in the Introduction, there are inf\/initely many real eigenvalues.
{\sloppy
Suppose that $H$ has inf\/initely many real eigenvalues. Then
f\/irst, from the properties of~$K_{m,j}(a)$ in
Remark~\ref{remark1}
and \eqref{c_def}, one can derive
the following properties of $c_j(a)$:
\begin{itemize}\itemsep=0pt
\item[$(i)$] The $c_j( {a})$ are all real-valued polynomials in
terms of the coef\/f\/icients $ {a}$ of $P$.
\item[$(ii)$] For
$2\leq j\leq m$, the polynomial $c_j( {a})$ depends only on $a_1,a_2,\dots,a_{j}$. Furthermore, it is a~non-constant linear function of $a_j$.
\end{itemize}
}
Notice that if $c_j(a)\not\in\mathbb R$ for some $0\leq j\leq m+1$, then \eqref{asymp_eqn1} implies that all but f\/initely
many eigenvalues are non-real. Thus, if $H$ has inf\/initely many real eigenvalues, then $c_j(a)\in\mathbb R$ for all
$0\leq j\leq m+1$.
Recall that $c_0(a)$ is real and depends only on $m$, and $c_1(a)=0$. Next, if $a_1\not=0$, then we can choose
$z_0=-\frac{a_1}{m}i$ and we consider the translation of $H$, replacing the potential $V(z)$ by $V(z-z_0)$. The
translated $H$ has its potential with a~vanishing $z^{m-1}$-term (i.e., $a_1=0$). Then by the above properties $(i)$
and $(ii)$ of $c_2(a), c_2(a)\in\mathbb R$ is a real polynomial in $a_1$ and $a_2$, and it is a nonconstant linear
function in $a_2$. Thus, since $a_1=0\in\mathbb R$ and $c_2(a)\in\mathbb R$, we obtain $a_2\in\mathbb R$. Likewise, since $a_1, a_2 ,
c_3(a)\in\mathbb R$, by the above properties of $c_3(a), a_3\in\mathbb R$ and hence, recursively we conclude that $a\in\mathbb R^m$.
Therefore, the translated $H$ is $\mathcal{PT}$-symmetric.
\end{proof}
\pdfbookmark[1]{References}{ref}
|
\section{Introduction}
\par The issue of controlling chaos has attracted great interest in the
last two decades; much work has been done in the case of dissipative
systems. Ott, Grebogi and York (OGY)~\cite{ott}developed a method by
which chaos can be suppressed by making small time-dependent
perturbations in order to shadow one of the infinitely many periodic
orbits embedded in the chaotic attractor. This method has been
extended by many following
studies~\cite{kapitaniak,dressler,pyragas} and successfully applied
to experimental systems~\cite{ditto}.
As a result of the absence of attractors in conservative systems and
the complex nature of the phase space combining regular and
irregular regions, constructing an effective method has remained a
challenging question. Lai Ding and Grebogi (LDG) extended the OGY
method to Hamiltonian systems by incorporating the notion of stable
and unstable directions at each periodic point.
In relation to the above question, Chandre \emph{et
al.}~\cite{chandre,chandre1,vittot} have developed a method for
constructing barriers in phase space to subdue chaotic behavior for
large times. This method is based on the introduction of a
specifically designed small control term which changes the dynamics
from chaotic to regular behavior. Zhang \emph{et al.}~\cite{zhang}
have introduced a method for controlling chaos in two-dimensional
Hamiltonian system which they called adaptive integrable mode
coupling based on the separation of the system into two coupled
subsystems one of which is stable and the other unstable; when the
unstable system comes into the vicinity of the integrable system the
conditions are reset resulting in an effective adaptive control.
Cartwright \emph{et al.}~\cite{cartwright} have constructed a method
which permits stabilization of KAM islands through forward iteration
of the orbits, and transforming them into global attractors of the
embedded system. Ciraolo \emph{et al.}~\cite{ciraolo} have used a
method of adding a small perturbation to modify systems, such as
nonlinear plasmas, to follow more regular motion. Additional
efficient methods have been proposed
in~\cite{Kulp,bolotin,ding,yugui,zhihua}; some of these methods
require tracking the trajectories while others involve the addition
of interacting terms for which the criteria are somewhat sensitive.
Geometric approaches for the analysis of the stability of a given
Hamiltonian have been widely
discussed~\cite{jacobi,hadamard,casettiReport,casettiPRE,cipriani,BenZionPRL,BenZionPRE1,BenZionPRE2,Kawabe,szyd1,szyd2,szyd3,szyd4},
for which curvature of a manifold is associated with stability.
The pioneering work of Oloumi and Teychenn\'e~\cite{oloumi} proposes
a stabilization method by considering the Gaussian curvature of the
potential energy as a source of chaos. Even if the condition of
negative curvature of the potential and instability of the dynamics
are not completely equivalent they were able to successfully control
the instability by avoidance of negative curvature regions of the
potential energy (ANCRP).
In this work we make use of a recently developed geometrical
criterion~\cite{BenZionPRL} which is highly sensitive to instability
in Hamiltonian systems, and has been shown to be in complete
agreement with the results of the numerical technique of surface of
section (Poincar\'e plot) and more effective than other geometric
methods ~\cite{BenZionPRE2}. This criterion is based on an
equivalence between motions generated by Hamiltonian in the standard
form with quadratic kinematic terms and additive potentials (which
we shall call the {\it Hamilton} description) and a Hamiltonian in
which the dynamics is described by a metric-type function of
coordinates multiplying the momenta in bilinear form (which we shall
call the {\it Gutzwiller}
description)~\cite{gutzwiller,curtis,moser,eisenhardt,arnold}. The
mapping between these equivalent descriptions was first introduced
by Appel~\cite{appell,cartan,landau,zerzion}. The criterion for
stability developed in ~\cite{BenZionPRL} establishes an inverse
mapping of the motion described completely geometrically in the
Gutzwiller space back to the Hamilton description, carrying with it
the covariance under diffeomorphisms that is a property of the
Gutwiller dynamics. The resulting orbits in the Hamilton
description, with the special choice of coordinates serving as the
basis for the Appel type relation, reduce to the standard
description of the Hamilton orbits under the Hamilton equations. In
general geometric form, as geodesic equations, however, they are
subject to analysis in terms of geodesic deviation, and the
resulting formula can be reduced to a computation in the special
coordinates of a new (symmetric) matrix valued criterion for
stability~\cite{BenZionPRL} which has been shown to be very
effective in a wide range of
examples~\cite{BenZionPRE1,BenZionPRE2,yahalom}.
\par In the analysis of these systems, it was shown that the presence
of negative eigenvalues for the stability matrix in the admissible
physical region (for which $E>V$) results in a chaotic type
Poincar\'e plot. Since the condition is local in coordinate space,
the result necessarily implies some degree of ergodicity. There has,
however, been no proof that the existence of negative eigenvalues is
not strongly associated with the structure of the dynamics more
globally, as is often the case, for example, for analytic function
theory, where the presence of a complex pole implies a distortion of
a relatively large region of the complex plane. The results of our
application to the control of dynamical systems, however, indicates
that the regions of negative eigenvalue correspond to a very
localized phenomena. The observed instability they induce appears
to be due to the passage of orbits through these regions, and they
are not strongly correlated with the more global structure of the
dynamics.
\par What we have done here is to identify the regions of negative
eigenvalue in the coordinate space, and modify the Hamiltonian in
these regions locally, removing the source of instability either by
removing the instability inducing terms or varying the coupling to
these terms to bring the Hamiltonian closer to an integrable form in
these regions. The remaining part of the space is left to develop
according to the full, nonlinear, symmetry breaking, evolution. The
effect on the Poincar\'e plots, as we show here, is dramatic. This
result provides strong evidence that chaotic Hamiltonian systems
derive their properties from sources that may be thought of as
highly localized in the regions of negative eigenvalues for the
stability matrix. In addition to giving an interesting insight into
the nature of Hamiltonian chaos, it also provides an effective
control method in which the Hamiltonian dynamics is left to exercise
its full function in a large domain of configurations, but is only
constrained in prescribed local regions of the configuration space.
\par In the next section, we review the basic ideas underlying the
geometrical criterion, and in the following section we give
numerical results for the application of our control procedure for
the cases of an oscillator with broken symmetry and a potential
which may be derived from the Toda form.
\begin{figure*}
\centering
$(a_{I})$ \subfigure{ \includegraphics[width=0.25\textwidth]{Cdana0_2.eps}}
$(a_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Dana0_2.eps}}
$(a_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlDana0_2.eps}}
\\
$(b_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{Cdana0_25.eps}}
$(b_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Dana0_25.eps}}
$(b_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlDana0_25.eps}}
\\
$(c_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{Cdana0_33333.eps}}
$(c_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Dana0_33333.eps}}
$(c_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlDana0_33333.eps}}
\\
$(d_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{Cdana0_5.eps}}
$(d_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Dana0_5.eps}}
$(d_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlDana0_5.eps}}
\caption{Effect of control on the model of Eq. \eqref{harmonic}. The first column shows the physical region (closed black curves) corresponding to the sequence of energies,
$\frac{1}{5},\frac{1}{4},\frac{1}{3},\frac{1}{2}$ corresponding to $a,b,c,d$. The regions of negative eigenvalues (instability) are shown in gray. The second column shows the Poincar\'e
plots for the uncontrolled system. The third column shows the Poincar\'e plots for the controlled system (Eq. \eqref{harmonicControl} with $r=0.4$ for all cases). The (red) dashed line in column one
shows the boundary of the control modification.}
\end{figure*}
\begin{figure*}
\centering
$(a_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{0_5.eps}}
$(a_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Toda0_5.eps}}
$(a_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlToda0_5.eps}}
\\
$(b_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{0_6.eps}}
$(b_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Toda0_6.eps}}
$(b_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlToda0_6.eps}}
\\
$(c_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{0_8.eps}}
$(c_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Toda0_8.eps}}
$(c_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlToda0_8.eps}}
\\
$(d_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{1.eps}}
$(d_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Toda1.eps}}
$(d_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlToda1.eps}}
\\
$(e_{I})$\subfigure{\includegraphics[width=0.25\textwidth]{1_5.eps}}
$(e_{II})$\subfigure{\includegraphics[width=0.25\textwidth]{Toda1_5.eps}}
$(e_{III})$\subfigure{\includegraphics[width=0.25\textwidth]{controlToda1_5.eps}}
\caption{Effect of control on the model of Eq. \eqref{toda}. The first column shows the physical region (closed black curves) corresponding to the sequence of energies,
$0.5,0.6,0.8,1.0,1.5$ corresponding to $a,b,c,d,e$. The regions of negative eigenvalues (instability) are shown in gray. The second column shows the Poincar\'e
plots for the uncontrolled system. The third column shows the Poincar\'e plots for the controlled system (Eq. \eqref{todaControl} with $\alpha=0.5$ for all cases). The (red) dashed line in column one
shows the boundary of the control modification.}
\end{figure*}
\section{Theory}
\par It has been shown ~\cite{appell,BenZionPRL} that a Hamiltonian system of the
form (we use the summation convention)
\begin{equation}\label{q1}
H= {{p^i}^2 \over 2M} + V(x),
\end{equation}
where $V$ is a function of space variables alone, can be put into
the equivalent form
\begin{equation}\label{q2}
H_G = {1 \over 2M} g_{ij}p^i p^j,
\end{equation}
where $g_{ij}$ is conformal and is a function of the coordinates
alone. One can easily see that the orbits described by the Hamilton
equations for \eqref{q2} coincide with the geodesics on a Riemannian
space associated with the metric $g_{ij}$, i.e., it follows directly
from the Hamilton equations associated with \eqref{q1}
that~\cite{gutzwiller}
\begin{equation}\label{q3}
{\ddot x}_\ell = -\Gamma_\ell^{mn} {\dot x}_m {\dot x}_n,
\end{equation}
where the connection form $\Gamma_\ell^{mn}$ is given by
\begin{equation}\label{q4}
\Gamma_\ell^{mn} = {1\over 2} g_{\ell k} \bigl\{ {\partial g^{km}
\over
\partial x_n}+ {\partial g^{k n} \over
\partial x_m}- {\partial g^{n m} \over
\partial x_k} \bigr\},
\end{equation}
and $g^{ij}$ is the inverse of $g_{ij}$.
\par For a metric of conformal form
\begin{equation}\label{q5}
g_{ij} = \varphi \delta_{ij},
\end{equation}
with inverse $g^{ij}= \varphi^{-1}\delta_{ij}$,on the hypersurface
defined by $H_G=H=E=constant$, and assuming that the observable
momenta are the same at every $t$, the requirement of equivalence
implies that ~\cite{BenZionPRL}
\begin{equation}\label{q6}
\varphi = {E \over E-V(x)}.
\end{equation}
\par To see that the Hamilton equations obtained from \eqref{q2} can,
be put into correspondence with those obtained from the Hamiltonian
of the potential model \eqref{q1}, we first note, from the Hamilton
equations for \eqref{q2}, that
\begin{equation}\label{q7}
{\dot x}_i = {\partial H_G \over \partial p^i}= {1 \over M}g_{ij}
p^j.
\end{equation}
We then define the {\it velocity field}
\begin{equation}\label{q8}
{\dot x}^j \equiv g^{ji}{\dot x}_i = {1 \over M}p^j,
\end{equation}
coinciding formally with one of the Hamilton equations implied by
\eqref{q1}, for which we label the coordinates ${x^j}$.
\par To complete our correspondence with the dynamics induced by
\eqref{q1}, consider the Hamilton equation for ${\dot p}^i$
generated by $H_G$,
\begin{equation}\label{q9}
{\dot p}^\ell = -{\partial H_G \over \partial x_\ell} = -{1 \over
2M}{\partial g_{ij} \over
\partial x_\ell} p^i p^j .
\end{equation}
With \eqref{q8} this becomes
\begin{equation}\label{q16}
{\ddot x}^\ell = -M^\ell_{mn}{\dot x}^m {\dot x}^n,
\end{equation}
where
\begin{equation}\label{q17}
M^\ell_{mn}\equiv {1 \over 2}g^{\ell k}{\partial g_{n m} \over
\partial x^k}.
\end{equation}
\par Eq. \eqref{q16} has the form of a geodesic equation, with a
truncated connection form~\cite{BenZionPRL}. Note that performing
parallel transport on the local flat tangent space of the Gutzwiller
manifold (for which $\Gamma_\ell^{mn}$ and $g_{ij}$ are compatible),
the resulting connection, after
raising the tensor index to reach the Hamilton manifold, is
exactly the ``truncated'' connection \eqref{q17}.
\par Substituting \eqref{q5} and \eqref{q6} into \eqref{q16} and \eqref{q17}, the Kronecker
deltas identify the indices of ${\dot x}^m$ and ${\dot x}^n$; the
resulting square of the velocity cancels a factor of $(E-V)^{-1}$,
leaving the Hamilton-Newton law derived from the Hamilton equations
directly from \eqref{q1}. Eq. \eqref{q16} is therefore a
geometrically covariant form of the Hamilton-Newton law, exhibiting
what can be considered an underlying geometry of standard
Hamiltonian motion.
\par Since the coefficients $M^\ell_{mn}$ constitute a connection
form, they can be used to construct a covariant derivative. It is
this
covariant derivative which must be used to
compute the rate of transport of the geodesic deviation $\xi^\ell =
x'^\ell - x^\ell$ along
the (approximately common) motion of neighboring orbits in the
Hamilton manifold, since it follows the geometrical structure of the
geodesic curves on $\{x^\ell\}$.
\par The relation
\begin{equation}\label{q18}
{\ddot \xi}^\ell = -2 M^\ell_{mn}{\dot x}^m {\dot \xi}^n -
{\partial M^\ell_{mn}\over \partial x^q} {\dot x}^m{\dot x}^n \xi^q,
\end{equation}
obtained from \eqref{q16}, can be
factorized in terms of the covariant derivative
\begin{equation}\label{q19}
\xi^\ell_{;n} = {\partial \xi^\ell \over \partial x^n} +
M^\ell_{nm}\xi^m.
\end{equation}
One obtains
\begin{equation}\label{q20}
{{D_M}^2 \over {D_M} t^2} \xi^\ell = {{R_M}^\ell}_{qmn} {\dot
x}^q{\dot x}^n \xi^m,
\end{equation}
where the index $M$ refers to the connection \eqref{q17}, and
\begin{equation}\label{q21}
{{{R_M} ^\ell}_{qmn} = {\partial M^\ell_{qm} \over \partial x^n}
-{\partial M^\ell_{qn} \over \partial x^m} + M^k_{qm}M^\ell_{nk} -
M^k_{qn}M^\ell_{mk}}
\end{equation}
corresponds to the curvature associated with the connection form
$M^\ell_{mn}$
This
expression does not coincide with the curvature of the Gutzwiller
manifold (given by this formula with $\Gamma_{qm}^\ell$ in place of
$M_{qm}^\ell$), but is a {\it dynamical curvature} which is
appropriate for geodesic motion in $\{x^\ell\}$
\par With the conformal metric in noncovariant form \eqref{q5},\eqref{q6} (in the
coordinate system in which \eqref{q6} is defined), the dynamical
curvature \eqref{q21} can be written in terms of derivatives of the
potential $V$, and the geodesic deviation equation \eqref{q20}
becomes
\begin{equation}\label{q22}
{D^2{\bf \xi} \over Dt^2} = - {\cal V}P{\bf \xi},
\end{equation}
where the matrix ${\cal V}$ is given by
\begin{equation}\label{q23}
{\cal V}_{\ell i} = \bigl\{ {3 \over M^2v^2} {\partial V \over
\partial x^\ell} {\partial V \over \partial x^i} + {1 \over M}
{\partial^2 V \over \partial x^\ell \partial x^i} \bigr\}.
\end{equation}
and
\begin{equation}\label{q24}
P^{ij} = \delta^{ij} - {v^i v^j
\over v^2},
\end{equation}
with $ v^i \equiv {\dot x}^i$, defining a projection into a
direction orthogonal to $v^i$.
\par Instability should occur if at least one of the eigenvalues of
$P{\cal V}P$ is negative,
in terms of the second covariant derivatives of the
transverse component of the geodesic deviation. This condition is
easily seen to be equivalent to the same condition imposed on the
spectrum of ${\cal V}$, and is thus independent of the direction of
the motion on the orbit.
\par The condition implied by the geodesic deviation
equation \eqref{q22}, in terms of
covariant derivatives, in which the orbits are viewed geometrically
as geodesic motion, is a new condition for instability
~\cite{BenZionPRL}, based on the underlying geometry, for a
Hamiltonian system of the form \eqref{q1}, providing new insight
into the structure of the unstable and chaotic behavior of
Hamiltonian dynamical systems. The method has been successfully
applied to many potential
models~\cite{BenZionPRL,BenZionPRE1,BenZionPRE2,yahalom}. We now
apply the method, which is simple and straightforward, since the
instability criterion is defined locally in coordinate space, to
extract from a chaotic Hamiltonian the part of the dynamics in the
regions for which ${\cal V}$ has negative eigenvalues which lead to
unstable motion . We show by direct simulation that this procedure
is remarkably effective.
\par The results confirm the notion of the locality of our criterion,
since the (local) removal of parts of the Hamiltonian inducing
instability by this criterion strongly affects the global stability
of the motion.
\section{Results and discussion}
\par In the following we give some examples of Hamiltonian motion
which are unstable and exhibit chaotic behavior. In two dimensions,
the terms in the potential that break rotational symmetry have a
form, in these examples, that induces chaotic behavior. The matrix
${\cal V}$ computed over the physically accessible region exhibits
certain regions with negative eigenvalues. We have arranged our
simulation to replace the Hamiltonian {\it in the regions of
negative eigenvalues} by a Hamiltonian in which the symmetry
breaking terms are either removed or for which the coupling
coefficient is decreased, but in regions of positive eigenvalues,
the Hamiltonian remains in its original form. The changes are
carried in relatively small regions of the configuration space in
the neighborhood of the regions of negative eigenvalues.
We take for illustration here a simple and important case of coupled
harmonic oscillators with perturbation
\begin{equation}\label{harmonic}
V(x,y)=\frac{1}{2}(x^2+y^2)+6x^2y^2
\end{equation}
and a generalization of the Toda potential
\begin{equation}\label{toda}
V(x,y)=\frac{1}{2}(x^2+y^2)+x^2y-\frac{1}{3}y^3+\frac{3}{2}x^4+\frac{1}{2}y^4.
\end{equation}
The first of these is known to generate chaotic behavior; as we have
shown previously, the transition to chaotic behavior as a function
of energy and of the coupling to the perturbation is well described
by our geometric criterion \cite{BenZionPRL}. In the present work,
we set the coupling to the chaos inducing perturbation to zero in
the regions for which the geometric criterion results in negative
eigenvalues.
In our previous study of the second example Eq. \eqref{toda} we
studied the sensitivity of the dynamical behavior to the choice of
energy \cite{BenZionPRE2}. In the present work, at the energies for
which the system exhibits chaotic behavior we have removed the chaos
inducing perturbation in the regions for which the geometric
criterion results in negative eigenvalues.
Fig. 1 shows the effects on the dynamical behavior of the change of
the coupling to a stable value in the regions of negative
eigenvalues for the perturbed oscillator potential \eqref{harmonic}.
The results for different values of energy are shown; these values
correspond to different sizes of the physically accessible region.
Note that the radius of the region of positive eigenvalues does not
change in this example.
The first row shows the physical region and the location of negative
eigenvalues. The interior of the circle corresponds to a region in
which no negative eigenvalue occur, chosen in this way to simplify
the computation. The second row shows the surface of section
Poincar\'e plot(surface of section) of the original uncontrolled
Hamiltonian indicating chaotic dynamics. The third row shows the
Poincar\'e plot generated by the controlled potential:
\begin{equation}\label{harmonicControl}
V(x,y)=\left\{\begin{matrix}
\frac{1}{2}(x^2+y^2)+6x^2y^2 & x^2+y^2<r^2\\
\frac{1}{2}(x^2+y^2) & x^2+y^2\geq r^2
\end{matrix}\right.
\end{equation}
where $r$ stands for the radius of the region of positive
eigenvalues.
Fig. 2 shows the effect of control on second system \eqref{toda},
using the controlled potential:
\begin{equation}\label{todaControl}
V(x,y)=\left\{\begin{matrix}
\frac{1}{2}(x^2+y^2)+x^2y-\frac{1}{3}y^3+\frac{3}{2}x^4+\frac{1}{2}y^4 & y > -\alpha\\
\frac{1}{2}(x^2+y^2) & y \leq -\alpha
\end{matrix}\right.
\end{equation}
where $\alpha$ stands for the limit (a horizontal dashed line) of
the region of positive eigenvalues.
In both cases, for sufficiently high energies the uncontrolled
systems become less stable and a chaotic signature appears. We
examine the method for different values of the energy, all
corresponding to chaotic motion of the uncontrolled system. One can
easily see that the Poincar\'e plots in controlled systems present
almost completely regular motion.
\section{Conclusion}
We see from these computations that in regions of positive eigenvalues
the Hamiltonian with chaos inducing terms do not generate
instability. If the chaos inducing terms are removed or decreased in
coupling in the regions of negative eigenvalues, the chaotic motion
of the system is reduced or disappears. We conclude that the chaotic
behavior of the system is associated with local properties of the
Hamiltonian, and that our local criterion is effective in
identifying these regions. Furthermore, our procedure provides an
effective method for control of chaotic systems of this type.
|
\section{Introduction}
The analysis of phase transitions and the associated microscopic structures is a well-developed scientific approach in physics. In real systems, the observation of phases and their different macroscopic behavior comes first, and a subsequent analysis reveals how the structure of one phase is transformed into the structure of the other phase. This transition is associated with the change of a so-called {\it control parameter}, such as the temperature. Most interesting are abrupt changes in functions measuring the macroscopic behavior, e.g., the density or heat capacity, that happen with small changes in the control parameter. The function showing the non-analytic behavior or {\it singularities} is the {\it order parameter} of the system and can be seen as a fingerprint of the underlying phase transition. Starting with the analysis of random graphs \cite{Bollobas2001} and simple percolation models \cite{Stauffer1994,Bollobas2006}, combinatorial objects came into the focus of statistical physicists. A thorough analysis revealed that these simple systems also show phase transitions.
Whereas in percolating systems the phases and their different behaviors are visually accessible, this is not the case for other combinatorial systems with a proposed phase transition. One of the most important of these systems is the so-called {\it satisfiability problem (SAT)}. Given some Boolean formula, it asks whether there exists an assignment of Boolean values to its variables such that it is satisfied, i.e., such that it evaluates to $true$. SAT problems belong to the set of NP-hard problems, i.e., so far there is no algorithm to solve them in polynomial time \cite{Garey1979}. As with many other NP-hard problems, satisfiability problems arise not only in theory but also in industry, e.g., in automotive configuration \cite{Sinz2003}, in software and hardware design \cite{k-itfhv-98}, biological sciences \cite{Elser2006}, and artificial intelligence \cite{Biere2009}.
Since satisfiability problems are so abundant, understanding when and why they are hard and developing better algorithms is crucial.
A classic family for analyzing the hardness is the random $k$-SAT family in which the $k$ variables of each clause are drawn uniformly at random and without repetition from the set of all variables. Each variable is negated with probability $0.5$. The ratio between the number of clauses $m$ and variables $n$ denoted by $\alpha\equiv m/n$ parameterizes the probability P[UNSAT] of finding an unsatisfiable instance at a given $\alpha$. It was observed early \cite{Cheeseman1991,Mitchell1992} that plotting P[UNSAT] against $\alpha$ shows a sharp threshold behavior at some critical $\alpha_c$. Furthermore, around this $\alpha_c$ it also takes various algorithms the longest time to solve random 3-SAT problems, i.e., the problems are {\it hard}.
To quantify the hardness, either the number of distinct steps of the solving algorithm is counted, or simply the time measured until the problem is solved. The divergence of the hardness together with the sudden jump of P[UNSAT] at some critical $\alpha$ resembles a phase transition-like behavior \cite{Hartmann2005,Percus2006}. The numerical analysis of this sharp threshold behavior resulted in $\alpha_c = 4.15\pm 0.05$ \cite{Kirkpatrick1994}.
Kirkpatrick and Selman could also show that there is a non-trivial finite size effect, i.e., that the width of the window in which the transition takes place is proportional to $n^{-2/3}$ for $3$-SAT. It is thought that this sharp threshold phenomenon is of first order, i.e., in the limit of infinite system size and for $\alpha < \alpha_c$ P[UNSAT] $= 0$, and for $\alpha > \alpha_c$ P[UNSAT] $= 1$. For 2-SAT, this could be rigorously shown \cite{Chvatal1992}, but for all $k \geq 3$ it is an open question. Note that $2$-SAT itself is not NP-hard \cite{Garey1979}.
To analyze the nature of this sharp threshold behavior, the $k-SAT$ problem can also be represented as a spin-glass model, and different theoretical analyzes have arisen from this approach \cite{Monasson1999,Mezard2002b,Mezard2002,Krakala2007}.
Since these theoretical analyzes rely on the thermodynamical limit whereas numerical approaches can only tackle system sizes of up to $100$ or even be restricted to system sizes below $40$ (depending on the specific question), it is not surprising that none of the theoretical approaches matches the numerical value of $\alpha_c = 4.15\pm 0.05$. The approach that comes closest is based on the analysis of survey propagation that results in $\alpha_c=4.267$, which is believed to be exact \cite{Mezard2002}. The applied order parameter is very technical and it is difficult to analyze how it relates to P[UNSAT].
To find out more about the behavior of $3$-SAT, we first repeated the experiment of Kirkpatrick and Selman, and increased the then available system size from $100$ to $200$. A subsequent finite size scaling is much more in accordance with the old value of $\alpha_c=4.15\pm 0.05$ than with $\alpha_c=4.267$. In the second step, we aim at understanding a different parameter, namely the {\it entropy} of the system, i.e., the logarithm of the number of solutions a satisfiable instance has. It was shown by an approach from statistical physics that the entropy is still finite at $\alpha_c$, i.e., the number of solutions is still exponential \cite{Monasson1996}. Monasson and Zecchina state that `` hence (...) the transition itself is due to the abrupt appearance of logical contradictions in all solutions and not to the progressive decreasing of the number of these solutions down to zero.'' Such a sudden emergence of logical contradictions on a macroscopic level would be a good sign of a genuine phase transition.
In this paper we give numerical evidence that the explanation for the finite entropy at $\alpha_c$ is far simpler, namely that the average number of solutions of satisfiable instances is universally described by a lognormal distribution over a range of different system sizes and $4.0 \leq \alpha \leq 4.5$. This means that, although many of the instances are already unsatisfied at $\alpha_c$, some of the satisfiable instances have a large number of solutions left, which accounts for the high average number of solutions. A lognormal distribution can be the result of the iterative application of a factor drawn from some distribution. This raises the question of whether the phase transition of P[UNSAT] may be only a sharp threshold phenomenon that is not based on the non-trivial restructuring of interacting entities. In the following we will first discuss our numerical findings regarding the average number of solutions, then give an alternative explanation for the rise of the hardness at $\alpha_c$ and finally discuss some simple models with different kinds of sharp threshold phenomena. The last model shows qualitatively the same behavior as P[UNSAT].
In summary, we do not attack the idea that $k$-SAT shows phase transitions in general but we put on display some simple explanations and models that raise doubt about whether the proposed phase transition of P[UNSAT] is more than a simple sharp threshold phenomenon. In general, the once obvious border between first order and continuous phase transitions and their respective properties has become so blurred that scientists from neighboring disciplines, e.g., computer scientists or chemists and even statistical physicists not specialized in spin-glasses, have difficulties to find out what are the properties that define a phase transition. Our main contribution in this paper are thus the above mentioned toy models that are so simple that they cannot be considered to have a genuine phase transition. Still, they mimic some important properties of the $3$-SAT system. With this we would like to open a discussion with the spin-glass community to understand what differentiates the simple models from $3$-SAT and what exactly makes a phase transition. The paper thus aims at starting a discussion of the difference between a mere sharp threshold phenomenon and a genuine phase transition. We hope that a discussion of what properties are required for acknowledging a phase transition will help to support the interdisciplinary discussion in this area.
The paper is organized as follows: After giving some definitions in Sec. \ref{definitions}, we will discuss in Sec. \ref{sat} the question of whether the sharp threshold phenomenon of P[UNSAT] is directly caused by a continuous phase transition of an order parameter related to P[UNSAT]. We will furthermore discuss whether there is any evidence at all for the existence of two different phases. Sec. \ref{models} finally introduces two simple statistical models that show similar phase transition-like behavior without any underlying interacting elements. The first one is clearly trivial, while the second shows a non-trivial finite size scaling effect. From these models, we develop a simple toy model that qualitatively shows the same properties as P[UNSAT] in random $3$-SAT. Finally, we discuss our findings in Sec. \ref{summary}.
\section{Definitions}
\label{definitions}
Let $V$ be a set of $n$ variables $\{v_1, \dots, v_n\}$. Each variable has two literals, a positive literal denoted by $v_i$ and a negated literal denoted by $-v_i$. A Boolean formula in {\it conjunctive normal form (CNF)} consists of $m$ subsets of {\bf and}-connected literals, called {\it clauses} or {\it constraints}. The clauses are {\bf or}-connected. An {\it assignment} is a function $a: V \rightarrow \{true,\ false\}^n$ that assigns each variable a Boolean value, i.e., {\it true} or {\it false}. With a given Boolean formula in CNF and a given assignment, the formula can be {\it evaluated}: a positive literal which is assigned $true$ evaluates to $true$, and to $false$ if it is assigned $false$. A negated literal which is assigned $false$ evaluates to $true$ and to $false$ otherwise. A {\it clause} evaluates to $true$ if at least one of its literals evaluates to $true$, and the whole formula evaluates to $true$ if all clauses evaluate to $true$. The {\it satisfiability problem}, or SAT problem for short, asks whether a given Boolean formula has at least one assignment such that it evaluates to $true$. Such an instance is called {\it satisfiable (sat)}, and one where no satisfying assignment can be found is called {\it unsatisfiable (unsat)}. If all clauses contain $k$ literals, we speak of $k$-SAT. If, moreover, the instance is created by choosing the $k$ literals uniformly at random without repetition, we speak of random $k$-SAT. $\alpha$ denotes the ratio between the number of clauses $m$ and the number of variables $n$ in a random $k$-SAT instance.
For each two assignments $a$ and $a'$, the {Hamming distance} $d(a, a')$ is defined as the number of different assignments to the variables.
The SAT problem can be solved by different algorithms, the most widely used being based on the following scheme, first proposed by Davis et al. \cite{Davis1962}. It is a kind of trial-and-error procedure in which a growing subset of variables is assigned Boolean values until we either find a solution or encounter a contradiction. In each step, take one of the variables that is yet unassigned and assign either $true$ or $false$ to it. Say, variable $v_i$ is assigned $true$. Now, the instance can be simplified by (temporarily) removing all clauses which contain the positive literal of $v_i$ since they are already satisfied. Furthermore, we can temporarily remove the negated literal from all clauses since it cannot contribute to the satisfaction of the clauses it is contained in. If after this step all clauses have been removed, we have found a solution to the problem. If we encounter an empty clause, all of its originally contained variables have been assigned the wrong value and thus we have found a contradiction. In this case, we have to backtrack and restore the instance up to the point where $v_i$ was unassigned. Then, the same procedure is tried, but assigning $false$ to $v_i$. As long as there is no solution and no contradiction in the simplified instance, we simply proceed with the partial assignment. If all decisions lead to contradictions, the instance is unsatisfiable. There are many improvements to this basic scheme, e.g., specifying an order in which the variables are assigned \cite{Marques-Silva1999} and learning \cite{Marques-Silva1996}. One basic improvement is {\it unit propagation}, i.e., whenever a clause has only one literal left, it can only be satisfied when the variable's assignment is set accordingly. Note that the assignment of such a variable is called a {\it dependent decision} while the assignment of Boolean values to all other so-called {\it free} variables is called {\it independent decision}.
\section{Random 3-SAT}
\label{sat}
It is well known that $3$-SAT belongs to the set of the so-called $NP$-hard problems, i.e., problems for which so far no algorithm with polynomial runtime has been found \cite{Garey1979}. In the worst case, finding a solution to these problems can take exponential time such that even relatively small instances cannot be solved within months. On the other hand, many real-world SAT problems can be solved in a short time despite their huge size. Since this behavior is not well understood, research has been dedicated to understanding why and how hard instances emerge and what their structure looks like.
It was observed early \cite{Mitchell1992, Cheeseman1991} that plotting P[UNSAT] against $\alpha$ shows a sudden jump at some value $\alpha_c$ independent of the system size $n$. Furthermore, around this value $\alpha_c$ it also takes various algorithms the longest time to solve random 3-SAT problems. This divergence of the hardness and the sudden jump of P[UNSAT] at some universal $\alpha$ resembles a phase transition-like behavior \cite{Hartmann2005,Percus2006}. In their classic paper from 1994, Kirkpatrick and Selman used the well-understood model of percolation in growing random graphs and the techniques deployed in this area for the identification of critical phenomena in random $3$-SAT: ``We use finite-size scaling, a method from statistical physics in which the observation of how the width of a transition narrows with increasing sample size gives direct evidence for critical behavior at a phase transition.'' They scaled the curves for different $k$ according to $n^{\nu}*(\alpha-\alpha_c)/\alpha_c$ and evaluated $\alpha_c$ to be $4.15\pm 0.05$ and the critical exponent $\nu$ for $k=3$ to be $2/3$ \cite{Kirkpatrick1994}.
Today a value of $\alpha_c=4.267$ is often cited for the P[UNSAT] threshold \cite{Mezard2002}, but plotting P[UNSAT] against the rescaled parameter $y = n^{0.66}(\alpha-4.12)/4.12$ yields a much better scaling than that for the rescaled parameter $y=n^{0.66}(\alpha-4.267)/4.267$ (s. Figure~\ref{finiteSize}). The reason for this mismatch is not totally clear. It could be due to the still quite small system size in our experiments.
In this paper we suggest that the observed threshold phenomenon of P[UNSAT] is not so much a sign of criticality but simply caused by the law of large numbers. In general it is not easy to prove that an observed sharp threshold behavior is not caused by the critical behavior associated with a phase transition since there are many possible interactions that could be causing it. In the next section we will first analyze the typical number of solutions, which is closely related to the entropy of the system.
\subsection{Number of solutions}
A first-order phase transition is deeply connected to a sudden increase in {\it order}. For example, when water freezes the molecules are fitted into a neat structure that shows high order. It is difficult to see intuitively what kind of order is measured by P[UNSAT]. However, when a continuous phase transition is
studied using an {\it existence parameter} instead of a
{\it quantitative parameter}, it may seem to be rather like a
first order transition, as we will exemplify in the case of site percolation in 2D.
Here, one can ask about the behavior of two different but related parameters: ``Is there a biggest connected component (BCC) of size O(n)?'' (this is the existence parameter) or ``What is the size of the BCC ?'' (and this is the quantitative parameter). Plotting the relative size of the BCC shows a continuous phase transition at some critical value, i.e., the first parameter is a quantitative one that reveals the complex behavior of the system. At the critical value, a finite fraction of all vertices is spanned by the BCC, i.e., it has size $O(n)$. Since the second parameter just asks for the existence of a BCC with size $O(n)$, it will trivially show a first-order phase transition-like behavior at the same value \cite{Stauffer1994}. Thus, in this system, the seemingly first-order phase transition-like behavior of the existence parameter is just a trivial implication from the true continuous phase transition concerning the quantitative
parameter.
Since P[UNSAT] asks whether there exists a solution or not, we first analyzed whether the seemingly first-order phase transition of P[UNSAT] also belongs to this type, i.e., whether it is an indicator of a more complex continuous phase transition of a related quantity like the behavior of the number of solutions.
An instance is $unsat$ if and only if it has no solution---this is a typical {\it existence parameter}. A possible quantitative parameter of which this existence parameter could be an indicator is the average number of solutions. The logarithm of this quantity is the entropy of the system at a given $\alpha$ \cite{Monasson1996}.
Figure~\ref{avgNumberSol} shows that the average number of solutions $<s>$ can be fitted to a simple exponential law, i.e.,
\begin{equation}
\label{avgNumb}
<s> = 2^n\left(\frac 7 8\right)^m.
\end{equation}
This simple behavior of the average number of solutions coincides with the so-called annealed estimate of the number of solutions \cite{Kirkpatrick1994}, which is based on the fact that any solution will be `killed' with a probability of $1/8$ by a clause drawn uniformly at random. But although this estimate has been used for a long time, it is surprising that the average number of solutions follows it so closely since it does not take into account that in reality the solutions' probability to be deleted are dependent: i.e., two very similar solutions have a higher probability to be killed by the same constraint whereas two solutions that assign the opposite values to variables can never be killed by the same constraint. Thus, it is still surprising that the average number of solutions universally follows this simple law for all system sizes. Furthermore, Figure~\ref{avgNumberSol} reveals that at $\alpha_c$ there is---on average---still an exponential number of solutions although we know that the probability of finding a satisfiable instance drops to zero for large system sizes. This has also been proven rigorously by \cite{Monasson1996}. It is clear that without the gap between the critical value of $\alpha_c$ and the point $\alpha = 5.19$ where the {\it average} number of solutions becomes $1$, there would not have been much interest in the seemingly critical behavior of P[UNSAT].
The only possibility to achieve an exponential average number of solutions at $\alpha_c$ and P[UNSAT] $\rightarrow 0$ for $n \rightarrow \infty$ is to have a strongly right-skewed distribution of the number of solutions an instance has. Indeed, as Figure~\ref{distNumbSol} shows, the distributions of satisfiable instances displays a universal behavior. Over an interval of $\alpha=4.0-4.5$ and different system sizes $n=30-100$, the cumulative distribution of the number of satisfiable instances, $P(s)$, can be fitted by the cumulative distribution of a lognormal distribution given as
\begin{equation}
P(s)=\frac{1}{2}+\frac{1}{2}{\rm erf}\left[\frac{\ln(s)-\mu}{\sigma\sqrt{2}}\right],
\label{eq:lognormal}
\end{equation}
where $\mu$ and $\sigma$ correspond to the mean and the standard deviation
of $\ln(s)$, and erf denotes the error function.
The lognormal distribution of $s$ explains that there is no need of a sudden drop of $<s>$ at $\alpha_c$ since the average is dominated by some instances with a high number of solutions, although most instances are already unsatisfiable. In summary, neither the typical number of solutions nor its distribution shows critical behavior around $\alpha_c$.
Since we know now that the distribution of $s$ is highly skewed, another intuitive measure is the {\it quenched average}, i.e., the average $<\log(s+1)>$ of the logarithm of the number of solutions, shown in Figure~\ref{quenchedAverage}. Note that also this does not show any interesting behavior around $\alpha_c=4.15$.
In summary, it does not seem to be the case that the sharp-threshold phenomenon of P[UNSAT] is the simple indicator of a related, continuous phase transition of a quantitative measure.
\subsection{Are there two different phases in $k$-SAT?}
This leads us back to the question of whether we really have two phases in this system, one consisting of satisfiable instances and one consisting of unsatisfiable instances. In $k$-SAT, the main problem is that we cannot observe two different phases by eye. In this special case, the sharp threshold behavior had been observed first. and this lead to the definition of the ``phases'' instead of observing and defining the phases first before analyzing the transition between them. This happened because the sharp threshold phenomenon divided the instances into two different groups that match our intuition. Maybe, however, an unsatisfiable instance is just an instance with $0$ solutions and not substantially different from an instance with exactly $1$ solution. The question is thus whether the two `phases' are just a differentiation that is convenient for computer scientists or whether they relate to a small structural change in some interaction on a microscopic scale that leads to a huge change in macroscopic behavior.
Hardness has been used to argue that there are two different phases, since it shows a diverging behavior around $\alpha_c$. Of course, hardness, measured as the number of independent decisions of a DPL-like algorithm \cite{Davis1962} or simply by the runtime, depends on the specific implementation. Nonetheless, the basic picture is always the same, namely that it peaks around $\alpha_c$\footnote{Note that the maximum itself is difficult to locate and might also shift with $n$.}. The question is whether this maximum is genuine or directly dependent on the definition of a satisfiable and an unsatisfiable instance. We will give evidence here that the occurrence of a maximal runtime around $\alpha_c$ is directly implied by the definition of a {\it decision algorithm}. The problem is that a decision algorithm does different things in the two cases: if it runs on a satisfiable instance, it stops after the first solution is encountered. Otherwise, a proof has to be given that no solution exists. For DPL-like algorithms \cite{Davis1962}, this means that in the first case only some fraction of the whole decision tree has to be searched while for unsatisfiable instances the whole tree has to be traversed. We can assume two things:
\begin{enumerate}
\item the decision trees of typical satisfiable and typical unsatisfiable instances at a given $\alpha$ are of approximately the same size;
\item the locations of the solutions in the leaves of the tree are uniform.
\end{enumerate}
Thus, let the size of a typical decision tree at a given $\alpha$ be denoted by $t(\alpha)$. Even if an instance has just one solution, we will on average traverse only half of the tree to find it. For an unsatisfiable instance at the same $\alpha$, we will on average take double the time to find the solution. Since at $\alpha_c$ there are more unsatisfiable than satisfiable instances, this is already an explanation for the increasing runtime at $\alpha_c$. Of course, the behavior of the average hardness is a bit more complicated than this. The average hardness $h(\alpha)$ can be dissected into $h_{sat}(\alpha)$ and $h_{unsat}(\alpha)$, the hardness of satisfiable and unsatisfiable instances at $\alpha$. With this,
\begin{equation}
h(\alpha) = (1-\mbox{P[UNSAT]})h_{sat}(\alpha)+ \mbox{P[UNSAT]}h_{unsat}(\alpha).
\end{equation}
Note that the hardness $h_{unsat}(\alpha)$ is simply given by the average size $t_{unsat}(\alpha)$ of the decision tree of unsatisfiable instances at $\alpha$. While $h_{unsat}(\alpha)=t_{unsat}(\alpha)$ seems to be a simple, exponentially decreasing function with $\alpha$ (s. Figure~\ref{hardness}a), $h_{sat}(\alpha)$ is at a maximum around $\alpha_c$ (s. Figure~\ref{hardness}b). $h_{sat}(\alpha)$ can be approximated as the product of $t_{sat}(\alpha)$, the size of the average decision tree of satisfiable instances at $\alpha$, and $\phi_T(\alpha)$, the average fraction of the decision tree that is traversed before a solution is found. While the first is decreasing with $\alpha$, the latter is increasing with $\alpha$. Thus, the maximum around $\alpha_c$ in $h_{sat}(\alpha)$ is introduced artificially by stopping after the first solution is encountered.
If we instead look at the runtime of an algorithm that counts the number of {\bf all solutions} an instance has, we see no singularity of the hardness around $\alpha_c$ as Figure~\ref{hardness}. shows. We thus conclude that the hardness supports the view that there are no two phases since the size of the decision tree decreases smoothly with growing $\alpha$, at least for the system sizes that could be computed.
Summarizing the results so far, we could not find a measure which is related to the existence question measured by P[UNSAT] and which shows a continuous phase transition. We also did not find any measure that is independent of P[UNSAT] and therefore proves that indeed an unsatisfiable instance is structurally different from an instance with $1$ solution. Instead, we will now present results from two very simple statistical systems that show a sharp threshold phenomenon. We will then use these systems to develop a simple toy model that shows qualitatively the same behavior as $3$-SAT and shows quite clearly that no phase transition is needed to produce a $3$-SAT-like system.
\section{Sharp threshold phenomena in simple statistical systems}
\label{models}
In this section we discuss two simple stochastic processes. The first one, is a simple coin tossing example that is discussed in Sec. \ref{coin} and the second is a statistical problem, called the {\it coupon collector's problem}, discussed in Sec. \ref{couponColl}.
\subsection{Throwing a Biased Coin}
\label{coin}
In the book {\it Computational complexity and statistical physics}, the editors briefly discuss the question of whether sharp thresholds are more than just an effect of the law of large numbers. They contrast $SAT$ with the following simple system \cite[p.8]{Percus2006}: a biased coin is tossed that shows heads with probability $\beta$ and tails with probability $1-\beta$. Let an instance consist of $\hat{n}$ tosses and let $\hat{n}$ define the system size. We expect the chance $P[\# heads > \# tails]$ to see more heads than tails in one of these instances to change from $0$ for $\beta < 0.5$ to $1$ for $\beta > 0.5$ with an ever-increasing sharpness with growing $\hat{n}$. With this example, Percus et al. indicate that sharp threshold phenomena {\it per se} are not so surprising, but they don't settle the question of whether this simple system will already show finite size scaling.
The question is thus whether the curve $P[\#heads > \#tails]$ for low $\hat{n}$ just fluctuates stronger or is indeed less steep than that of a larger system. This question is settled by Figure~\ref{FigCoin}.
Figure~\ref{FigCoin}a shows the fraction of $10,000$ instances of $\hat{n}$ tosses each where more heads than tails were shown. The curves meet approximately at $\beta = 0.5$. Plotting them against the rescaled parameter $y = \hat{n}^{0.5}(\beta-0.5)/0.5$ shows a perfect universal scaling. This model is especially interesting since here also the sharp threshold behavior results from asking a peculiar kind of question. Instead of looking at the more natural question of $P[heads]$ which is of course identical to $\beta$, the behavior artificially becomes a sharp threshold behavior by asking when it is more likely to see more heads than tails in any given system size. Moreover, this most simple system also displays a finite size scaling effect. Naturally, the corresponding exponent $\beta = 0.5$ is the one dictated by the law of large numbers. Thus, although a finite size scaling effect can be seen, nobody would regard it as the effect of a phase transition since the exponent is a trivial one. The next example is much more interesting since it shows a non-trivial exponent.
\subsection{The Coupon Collector's Problem}
\label{couponColl}
The simple system of coin tossing cannot easily be likened to $3$-SAT. We will thus introduce a second statistical problem called the {\it coupon collector's problem}: let there be a set of $n'$ distinguishable objects called {\it coupons}, identified by a coupon ID from $1$ to $n'$. Each coupon is contained multiple times in a large multi-set and collectors can purchase coupons from this multi-set by drawing one item uniformly at random. We will assume that each coupon ID has the same probability of being drawn. The coupon collector problem asks how many draws have to be made expectedly until each coupon ID is drawn at least once, i.e., the question of when the collection is completed. In essence, once the collector has collected $k$ different IDs, the chance of picking a new ID is $\frac{n'-k}{n'}$ and thus the expected time to find a new one is $\frac{n'}{n'-k}$. Summing over these expected times gives $\frac{n'}{n'}+\frac{n'}{n'-1}+\ldots+n'=n'\left(\frac{1}{1}+\frac 1 2 + \ldots+\frac 1 {n'} \right) = n' H_{n'}$. This can be approximated to be $n' \ln {n'}+ \Upsilon n' + \frac 1 2 + O(1)$, where $\Upsilon \simeq 0.57722$ denotes the Euler--Mascheroni constant. The variance is bound from above by $2n'^2$.
For a set of $x$ collections, we now define $P[full, t]$ to be the fraction of full collections after $t$ draws. Of course, the number of draws depends on the system's size. We thus define $\gamma := \frac{t}{n' \ln{n} + 0.577 n' + 0.5}$ and plot $P[full,\gamma]$ against $\gamma$. Figure~\ref{cc}a shows the result for different system sizes from $10$ to $1000$ in dependence of $\gamma$. Interestingly, this looks like a phase transition at a critical $\gamma_c = 1$. Furthermore, we define a rescaled parameter $z = n'^{0.17}\left(\gamma-\gamma_c\right)$ against which we plot the functions, as shown in Figure~\ref{cc}b.
Note that the critical exponent is far away from the trivially expected $0.5$. We can now define two phases: full collections and incomplete collections. With this, Figure~\ref{cc} shows clearly that there exists a first-order phase transition between the two phases. Or does it?
But of course, a system as simple as the coupon collector's problem does not meet the intuition about a system with a phase transition and it especially cannot exhibit any non-trivial collective behavior.
Just defining that one condition of a system, i.e., whether a collection is complete or not, represents two phases does not make them different phases.
Also, the finite size scaling effect cannot justify the notion of a phase transition since it seems to be mainly an effect of the law of large numbers.
In the following we will highlight the connection between the coupon collector's problem and the behavior of P[UNSAT] in $3$-SAT.
\subsubsection{Connection between random $k$-SAT and the coupon collector's problem}
When $\alpha = 0$ each random $k$-SAT instance has exactly $2^n$ solutions. Every added clause $C = \{l_1, l_2, \dots, l_k\}$ excludes all solutions in which all negated literals $l_i$ are assigned $true$ and all positive literals $l_j$ are assigned $false$. That is, each added clause extinguishes a fraction of $2^{-k}$ of all remaining solutions. Of course, some of the solutions might already have been extinguished by a clause added earlier. An instance becomes unsatisfiable when all of its possible assignments have been extinguished by some clause. Thus, the question is very similar to that of the coupon collector's problem: in each time step we draw uniformly at random $k$ literals that extinguish a $2^{-k}$th of all possible assignments and we want to know when all possible assignments are extinguished. Of course, there are two main differences: we draw more than one `coupon' at once, namely $2^{n-k}$, and moreover these are not independent of each other. The first condition alone would just reduce the expected completion time by some factor, but the effect of the second condition is harder to estimate.
Note that there is really no kind of interaction between the clauses. Given a set of solutions $S$ that are left for some instance $I$, adding a clause will lead to the following reduced set of solutions $S'$: let $s \in S$ be any solution that does not satisfy the newly added clause. This cannot be a solution of the new instance, and thus it is removed from $S$. Let now $s \in S$ be some solution that satisfies the newly added clause. Since it was contained in $S$, this means that the assignment given by $s$ satisfies at least one literal in all the clauses added so far plus at least one in the newly added clause. Thus, this solution is in $S'$. The clauses are independent of each other in the sense that the only solutions extinguished by a clause are those that don't satisfy it. There is no cumulative effect of the clauses such that after adding some of them a whole avalanche of solutions is extinguished. Note, however, that the solutions in $S$ are not independent of each other since if $s \in S$, other solutions $s'$ with a low Hamming distance to $s$ have a higher probability of being in $S$ than those with a large distance.
\subsection{A toy model for $3$-SAT}
Neither the coupon collector's problem nor coin tossing displays one of the main qualitative behaviors of $3$-SAT. The main point of interest is the gap between $\alpha = 5.19$ at which the average number of solutions meets $1$ and the point $\alpha_c$ at which most instances are already unsatisfiable.
In the following, we introduce a toy model that shows this more involved behavior but is still quite simple and not likely to have a real phase transition. The toy model is based on the following idea: an instance is represented by a number, starting with $2^n$. This represents the number of solutions left at a given $\alpha$. Adding a clause is mimicked by multiplying this number by some reduction factor.
Of course, simply multiplying the number by $7/8$ is already enough to produce the average number of solutions shown in Figure~\ref{avgNumberSol}, and also a sharp threshold behavior of P[UNSAT]. But, unfortunately, the latter takes place at $\alpha = 5.19$. Looking at the real reduction factor, it turns out that the distribution broadens with $\alpha$ and is shifted to the right. We used this observation for the toy model of random 3-SAT, in which we draw a multiplicative factor from a normal distribution with a standard deviation $\sigma = 0.0585*\alpha$ and an average of $\mu(\alpha)$ given by
\begin{equation}
\mu(\alpha)=\left\lbrace
\begin{matrix}
0.875+0.009*\alpha & \alpha<3.8 \cr
0.875+0.170*\alpha & \alpha\geq 3.8
\end{matrix}
\right. .
\end{equation}
If the drawn number is lower than 0 or higher than 1, we set it to 0 or 1, respectively. This factor is then multiplied with the current number of the toy model instance. An instance of the toy model represents an unsatisfiable instance if its number drops below 1. Thus, $P_{\mbox{toy}}[UNSAT, \alpha]$ gives the fraction of toy model instances at $\alpha$ whose number is below 1.
In Figures~\ref{toy} and \ref{toydist} we show our simulation results for the toy model defined above. According to Figure~\ref{toydist}a, the average number of solutions follows the same exponential behavior as expected from (\ref{avgNumb}), and $\left< s\right>$ drops below 1 at $\alpha=5.19$. Surprisingly, a sharp threshold behavior can be observed when plotting P[UNSAT] as a function of $\alpha$ as shown in Figure~\ref{toydist}b-c. Similar to $3$-SAT, the transition point of the threshold behavior at $\alpha=4.76$ is separated from the point where $\left< s\right>=1$ by a non-negligible gap. Furthermore, the distribution of the numbers $P_{\mbox{toy}}[s,\alpha ]$ is best described by a lognormal distribution and shows the same universal scaling behavior as the real $P[s,\alpha ]$ distribution, as displayed in Figure~\ref{toydist}.
In summary, this toy model shows the same qualitative properties as the real $3$-SAT system.
\section{Summary}
\label{summary}
In this article we have raised the question of whether or not the sharp threshold phenomenon displayed by P[UNSAT] around $\alpha = 4.2$ is a mere statistical event that does not relate to a phase transition in the classical sense. Our intuition is that there is no interaction of the elements of a Boolean instance, i.e., clauses, variables, or solutions, that leads to this phenomenon. We also see no principal difference between instances with at least $1$ solution and those with no solution. We thus believe that the sharp threshold behavior of P[UNSAT] can rather be likened to the sharp threshold phenomena in simpler systems, like the coupon collector's problem. Of course, it is obvious that approaches from statistical physics were successful in describing $3$-SAT and that some of these results lead to the most powerful SAT-solver based on survey propagation \cite{Monasson1999}. It is important that we not question the phase transition shown for other order parameters like backbone size \cite{Monasson1999b}, clustering of the solution space \cite{Krakala2007}, or the order parameter associated with the messages in survey propagation \cite{Mezard2002}, but only P[UNSAT] as an order parameter of a real phase transition.
We conclude by describing one of the possibly many examples where asking
somewhat different questions about the states of the same system may
easily lead to the conclusion that more than one observable
transition (and of different kinds) takes place in the system,
even though it is widely accepted that there is only a single relevant
transition in it.
Consider the Ising model on a face-centered cubic lattice. As
the system cools down from high temperatures, we ask two simple
questions (without loss of generality we can assume that for low
temperatures the up spins take over):
\begin{enumerate}
\item What is the total spontaneous magnetization of the system? (ratio
of up spins minus the ratio of down spins)
\item Is there a percolating cluster of down spins present?
\end{enumerate}
The (textbook level) answers are:
\begin{enumerate}
\item Below a critical temperature $T_c^I$, the spontaneous magnetization
sharply increases as the number of up spins starts to grow quickly.
The associated transition is a prototype of continuous phase
transitions (involving fluctuations, etc).
\item At a temperature $T_p^I < T_c^I$, the probability that a
percolating (connected infinite) cluster of down spins is present
suddenly drops from 1 to 0 (as if a first-order transition was taking
place).
\end{enumerate}
We suggest that the lesson from this analogy is the following: the
answer one gets depends very much on the question. Our conclusion is
that it remains to be demonstrated that asking ``What is
the probability of having a satisfiable instance in 3-SAT?'' is the right
question. We argue that this particular question (order parameter) is not closely related
to the variety of possible rich transitions
taking place in this paradigmatic satisfiability problem.
Has the question of whether or not P[UNSAT] actually undergoes a phase transition, more to it than just being a simple question of how to name something? In this interdisciplinary field it is very important to be careful with terms; a phase transition is more than just a sharp threshold phenomenon and requires proof that the supposed phases behave differently in some aspect that is independent of their definition. The simple stochastical systems presented here stress the point that a sharp threshold phenomenon, even if accompanied by a non-trivial finite size scaling effect, is not enough to show a genuine phase transition - an independent proof of two different phases is needed in addition. We hope that this article will trigger a discussion about the observations to be made in categorizing a sharp threshold phenomenon as a non-trivial phase transition, and thereby support ongoing interdisciplinary research in this field.
\section*{Acknowledgement}
The authors thank A. Hartmann and G. Istrate for their numerous helpful comments on the early versions of our manuscript.
KAZ was supported by a grant by the Deutsche Akademie der Naturforscher Leopoldina (BMBF LPD 9901/8-182). This work was supported by the Hungarian National Science Fund (OTKA K68669, NK77824), the National Research and Technological
Office (NKTH, Textrend) and the J\'anos Bolyai Research Scholarship of the Hungarian Academy of Sciences.
\section*{References}
|
\section*{Introduction}
The convergence of $U$-statistics has been intensively studied for estimators based on families of i.i.d. random variables and variants of them. In most cases, the independence assumption is crucial. When dealing with Feynman-Kac and other interacting particle systems of Monte-Carlo type, one faces a new type of problem. Namely, in a sample of $N$ particles obtained through the corresponding algorithms, the distributions of the particles are correlated -although any finite number of them is asymptotically independent with respect to the total number $N$ of particles. It happens so (and this is the main contribution of the present article to show) that this asymptotic independence is enough in practice to insure the convergence of $U$-statistics based on interacting particle systems. In the following, we prove therefore the convergence of $U$-statistics for different particle systems under mild assumption that are satisfied by Feynmann-Kac particle systems. The case of Bird and Nanbu systems also fits in this framework and will be treated elsewhere by the third Author \cite{rubenthaler-2009}. To study the asymptotics of Feynman-Kac systems, whose properties are crucial to ensure the convergence of the statistics, we will use a functional representation, as introduced in \cite{DPR-2006} in the framework of discrete interacting particle systems.
The article is organized as follows. To fix the notations and the general framework of interacting particle systems, we first recall the (Feynman-Kac, continuous) interacting particle model. The model first appeared in quantum physics, in the work of Feynman and Kac in the 1940-50's, as a way to encode the motion of a quantum particle evolving in a potential (e.g. the interaction potential of a quantum field theory, viewed as a perturbation of the free Hamiltonian) in terms of path-integral formulas. It was realized progressively that interacting particle systems could be used in incredibly many different settings in probability and statistics.
A detailed list of the (still expanding) application areas of these models is contained in \cite{del-moral-2004}, to which we refer for further informations.
Recall simply, since we focus here on its statistical features, that the model is mainly used, in applied statistics, as a Bayesian nonlinear filtering model: the motion of the particles is driven by a diffusion process and the potential encodes the likewood of the states with respect to observations or to some reference path.
We study then the associated empirical joint distributions of a finite number $k$ of particles and study the convergence of the distributions in terms of the total number $N$ of particles of the system. This is closely related to our previous joint work \cite{DPR-2006} on discrete Feynman-Kac models -although the continuous hypothesis we use in the present article leads to some simplification of the tricky combinatorics that showed up in the discrete framework.
We turn then to $U$-statistics for interacting particle systems and prove that under mild asumptions on the behavior of the system (satisfied e.g. by Feynman-Kac and Boltzmann systems) several asymptotic normality properties holds.
\section{Feynman-Kac particle systems}\label{Sec:1}
Let us consider a $E$-valued Markov process $X_t$, where $E=\mathbb{R}^d$ (or an arbitrary metric space) with a time-inhomogeneous infinitesimal generator $L_t$, continuous trajectories, and a positive bounded potential function $V_t$, $0\leq V_t(x)\leq V_\infty$. We assume that the distribution of $X_0$ is $\gamma_0= \eta_0$. Notice that these hypothesis are meaningful for most applications, but could be accomodated to more general ones, see \cite{del-moral-2004}. We are interested in the unnormalized (resp. normalized) distribution flows $\gamma_t$ and $\eta_t$ that are solutions, for sufficiently regular test functions $f$ and under appropriate regularity conditions, of the nonlinear equations:
$$\frac{d}{dt}\gamma_t(f)=\gamma_t(L_t(f))-\gamma_t(fV_t)$$
and
$$\frac{d}{dt}\eta_t(f)=\eta_t(L_t(f))+\eta_t(f(\eta_t(V_t)-V_t)).$$
In terms of $X_t$, we have:
$$
\gamma_t(f) = \mathbb{E}\left( f(X_t) \exp\left(-\int_0^t V_s(X_s) ds\right)\right),\ \eta_t(f)=\frac{\gamma_t(f)}{\gamma_t(1)}\ .
$$
\subsection{Definitions and notations}
Let us fix first of all some notations.
For $q\in \mathbb{N}^*$, we write $[q]:=\{1,\dots,q\}$. For $q,N \in \mathbb{N}^*$, we set $\langle q,N \rangle := \{s\in [N]^{[q]}, s\ \text{injective}\}$.
For $q$ even, we write $\mathcal{I}_q$ for the set of partitions of $[q]$ in pairs. We have $$\# \mathcal{I}_q = \frac{q!}{2^{q/2}\left( \frac{q}{2} \right)!}\ .$$
The set of smooth bounded (resp. smooth bounded symmetric) functions on $E^q$ is written $
\mathcal{B}_b ( E^q)$ (resp. $
\mathcal{B}_b^{\text{sym}}( E^q)$). We also write $\mathcal{B}^{\text{sym}}_0 ( E^q)$ for the set of symmetric functions:
$$\mathcal{B}^{\text{sym}}_0 ( E^q) := \left\{ F \in \mathcal{B}^{\text{sym}}_b ( E^q) : \int_{E} F(x_1, \dots, x_q) \gamma_t(dx_q) = 0\right\}.$$
Notice that the set of functions $\mathcal{B}^{\text{sym}}_0 ( E^q)$ depends on $t$, so that a better notation would be $\mathcal{B}^{\text{sym}}_{0,t} ( E^q)$. However, since in practice the abbreviated notation should not lead to confusion, we decided not emphasize this dependency for notational simplicity. We write simply $\mathcal{B}_0 ( E)$ for the centered functions: $F \in \mathcal{B}_b ( E) : \int_{E} F(x) \gamma_t(dx) = 0$
The empirical (possibly random) measure associated to a (possibly random) vector $x=(x_1,\dots,x_N )\in E^N$ is given by
$$
m(x)=\frac{1}{N} \sum_{i=1}^N \delta_{x_i}\ .
$$
We have, for all $q\in \mathbb{N}^*$, $F:E^q \rightarrow \mathbb{R}$:
$$
m(x)^{\otimes q}(F) = \frac{1}{N^q} \sum_{s\in [N]^{[q]}} F(x_{s(1)},\dots,x_{s(q)})\ .
$$
We also consider the corresponding $U$-statistics:
$$
m(x)^{\odot q}(F) = \frac{1}{(N)_q} \sum_{s \in \langle q , N\rangle } F(x_{s(1)},\dots,x_{s(q)})
$$
$$={{N}\choose{q}}^{-1}\sum\limits_{1\leq i_1<...<i_q\leq N}F_{sym}(x_{i_1},\dots,x_{i_q})\ ,
$$
where
$
(N)_q=\frac{N!}{(N-q)!}=\# \langle q, N\rangle\
$ and
$
(F)_{\text{sym}}(y_1,\dots,y_q) := \frac{1}{q!} \sum_{\sigma \in S_q} F(y_{\sigma(1)},\dots,y_{\sigma(q)})\ ,
$
with $S_q$ the symmetric group of order $q$.
The notion of differential for sequences of signed measures\footnote{From now on, \it measure \rm will have the more general meaning of \it signed measure\rm} will also be useful.
The total variation norm is written $||~~||_{tv}$, so that, for any linear operator $L$ on $\mathcal{B}_{b}(E^q)$,
$$||L||_{tv}:=\sup_{f\in\mathcal{B}_b(E^q):\Vert f\Vert\leq 1}~|L(f)|$$
Let $(\Theta^N)_{N\geq 1}$, be a uniformly bounded sequence of measures on $E^q$, in the sense that $\sup_{N\geq 1}{\|\Theta^N\|_{\rm{ } tv}}<\infty$.
The sequence
$\Theta^N$ is said to converge strongly to some measure $\Theta$,
as $N\uparrow\infty$ if and only if
$$
\forall f\in \mathcal{B}_b(E)\qquad
\lim_{N\uparrow\infty} \Theta^N(f)= \Theta(f)
$$
\begin{definition} Let us assume that $\Theta^N$ converges strongly to $\Theta$.
The discrete derivative of the sequence $(\Theta^N)_{N\geq 1}$ is the sequence of
measures $ (\partial \Theta^N)_{N\geq 1}$ defined by
$$
\partial \Theta^N:=N~\left[ \Theta^N-\Theta\right]
$$
We say that $\Theta^N$ is differentiable, if $ \partial \Theta^N$
is uniformly bounded, and if it
converges strongly to some measure $\partial \Theta$, called simply the derivative of $\Theta^N$,
as $N\uparrow\infty$.
\end{definition}
The discrete derivative $\partial\Theta^N$ of a differentiable
sequence can itself be differentiable. In this situation, the derivative of the discrete derivative is called the second derivative of $\Theta^N$ and it is denoted by $ \partial^2 \Theta=\partial \left(\partial\Theta \right)$, and so on.
A sequence $\Theta^N$ that is
differentiable up to order $(k+1)$, has the following representation
$$
\Theta^N=\sum_{0\leq l\leq k }\frac{1}{N^l}~\partial^l\Theta+\frac{1}{N^{k+1}}~\partial ^{k+1}\Theta^N
$$
with $\sup_{N\geq 1}{\|\partial ^{k+1}\Theta^N\|_{\rm{ } tv}}<\infty$
, and the convention $\partial^0\Theta=\Theta$, for $l=0$.
\subsection{A genetic particle model}
A particle system approaching the measures $\eta_t$ (and therefore also $\gamma_t$) is the following. At $t=0$, the random vector $\Xi_t=(\xi^1_t,\dots,\xi^N_t)$ is a family of i.i.d. random variables distributed according to $\gamma_0=\eta_0$. Each entry $\xi^i_t$ of the vector diffuses according to the generator $L_t$, independently of the other entries. Each entry $\xi_t^i$ has an exponential clock of parameter $V_\infty$ (independent of all the other variables defined up to now). When the clock of $\xi^i_t$ rings, say at $\tau$, it can
\begin{itemize}
\item jump to a randomly (and uniformly) chosen particle in the family, including itself, with a probability $\frac{V_\tau(\xi_t^i)}{V_\infty}$
\item stay where it is with probability $1- \frac{V_\tau(\xi_t^i)}{V_\infty}$.
\end{itemize}
Up to a renormalization of the jump type generators (which is convenient for our present purposes), this is the model described in \cite[Sect. 1.5.2]{del-moral-2004}.
Notice that the law of $\xi_t^i$ depends on $N$, so that a more consistent notation for $\xi_t^i$ would be $\xi_t^{N,i}$ -whenever we want to emphasize the dependency on $N$, we will switch to this second notation.
The corresponding empirical\footnote{The adjective \it empirical \rm refers, in the present article, to any measure, process or statistics obtained from a particle system approximation.} measures and empirical U-statistics are given (and related) by:
$$\eta_t^N =m(\Xi_t)=m(\xi^1_t,\dots, \xi^N_t),$$
$$\gamma_t^N =\gamma_t^N(1)\cdot\eta_t^N, \text{with} \ \gamma_t^N(1)=\exp\left( -\int_0^t \eta_s^N(V_s) ds \right)\ ,$$
$$(\gamma_t^N)^{\odot q}(F)=\gamma^N_t(1)^q\cdot(\eta_t^N)^{\odot q}(F).$$
These are the analogs in continuous time to the random measures and U-statistics defined and studied in \cite{DPR-2006} in a discrete time setting.
We recall that $\eta_t^N$ and $\gamma_t^N$ are known to converge as $N\to \infty$ to $\eta_t$ and $\gamma_t$ \cite{del-moral-2004}.
In oder to study the convergence of the (empirical) U-statistics $(\gamma_t^N)^{\odot q}$ and $(\eta_t^N)^{\odot q}$, we are going to rewrite them by means of functional expansions (actually, Laurent series in the parameter $N$). These expansions will allow us, later, to control the convergence and (among others) to extend to particle systems the classical central limits theorems for U-statistics.
Deriving these expansions is the main purpose of the first part of the article, together with first results of convergence.
For these purposes, it is useful to introduce an auxiliary $q$-particle system. The reasons for its introduction will become clear later and stem from a backward analysis of Feynman-Kac trajectories of families of $q$ particles.
\begin{definition}\label{Def:aux}
The \it auxiliary system \rm of $q$ particles, $\hat{\xi}^1_t,\dots,\hat{\xi}^q_t$ is defined as follows:
the random variables $\hat{\xi}^1_0,\dots,\hat{\xi}^q_0$ are independant of law $\eta_0$. Moreover, the particles
\begin{equation}
\label{Def:hatX1}
\hat{\xi}^1_t,\dots,\hat{\xi}^q_t \text{ diffuse according to } L_t
\end{equation}
(one more time, independently of one another)
and undergo the following jumps. For any $(i,j)\in [q]^2,\ i\not= j$, there is an exponential clock of parameter $V_\infty/N$ and a corresponding Poisson point process $T^{(i,j)}_1,T^{(i,j)}_2,\dots$ of parameter $V_\infty/N$. The $T^{(i,j)}_1,T^{(i,j)}_2,\dots$ are named the ringing times.
At a ringing time $t\in \{ T^{(i,j)}_1,T^{(i,j)}_2,\dots\}$,
\begin{equation}
\label{Def:hatX2}
\hat{\xi}^i_t
\begin{cases}
\leftarrow \hat{\xi}^j_t & \text{ with proba. } \frac{V_t(\hat{\xi}^i_t)}{V_\infty}\\
\leftarrow \hat{\xi}^j_t & \text{ with proba. } 1-\frac{V_t(\hat{\xi}^i_t)}{V_\infty}\ .
\end{cases}
\end{equation}
\end{definition}
The notation $\hat{\xi}^i_t \leftarrow \hat{\xi}^j_t$ means that $\hat{\xi}^i_t$ jumps to (or is substituted by) $\hat{\xi}^j_t$. When $q=N$, the particle systems $(\hat{\xi}^1_t,\dots,\hat{\xi}^N_t)_{t\geq 0}$ has the same law as $({\xi}^1_t,\dots,{\xi}^N_t)_{t\geq 0}$.
\
We set, for an arbitrary $F\in \mathcal{B}_b ( E^q)$,
$$
F^e((\hat{\xi}^1_s,\dots,\hat{\xi}^q_s)_{0\leq s \leq t}) = F(\hat{\xi}^1_t,\dots,\hat{\xi}^q_t) \exp\left( {-\int_0^t [V_s(\hat{\xi}^1_s) +\dots + V_s(\hat{\xi}^q_s)] ds }\right)\ .
$$
and
$$
E_{t,k}(F) := \mathbb{E}\left\{ F^e((\hat{\xi}^1_s,\dots,\hat{\xi}^q_s)_{0\leq s \leq t})| k\ \text{rings on} [0;t] \right\}
\ .$$
Notice, for further use, that $E_{t,k}(F)$ does not depend on $N$. This is because, as a consequence of the general properties of Poisson point processes, conditionally to the hypothesis that there are $k$ rings on $[0,t]$, the distribution of the ringing times is uniform on $[0,t]$ and therefore independent of the parameter $V_\infty/N$; since the other parameters of the process ($L_t$ and the jump probabilities $\frac{V_t(\hat{\xi}^i_t)}{V_\infty}$) are independent of $N$, the property follows.
\subsection{Expansion of the unnormalized measure}\label{unnorm}
Let us write for an arbitrary $F\in \mathcal{B}_b ( E^q)$,
$$\mathbb{Q}^N_{t,q}(F):=\mathbb{E} ((\gamma_t^N)^{\odot q}(F)).$$
We refer to $\mathbb{Q}^N_{t,q}$ as to the empirical unnormalized measure associated to the particle system $\xi_t^i$. Since the joint distribution of the sequence $\xi_t^1,\dots,\xi_t^N$ is invariant by permutations, $\mathbb{Q}^N_{t,q}(F)=\mathbb{Q}^N_{t,q}(F_{sym})$, and we can assume without restriction that $F$ is a symmetric function.
The first question we adress is the (exact) computation of the speed of convergence of the empirical unnormalized measure $\mathbb{Q}^N_{t,q}$ to $\gamma_t^{\otimes q}$.
\begin{theorem}
\label{Theo:FKunnormalized}
For $F\in \mathcal{B}_b^{sym} ( E^q)$, we have the Laurent expansion
\begin{eqnarray}
\label{Eq:FKunnormalized1}
\mathbb{Q}^N_{t,q}(F) &=&
\sum_{k=0}^\infty \frac{(\lambda t)^k e^{-\lambda t}}{k!} E_{t,k}(F)\\
\label{Eq:FKunnormalized2}
&=& \gamma_t^{\otimes q}(F) +\underset{k+i\geq 1}{\sum_{k,i\geq 0}^\infty} (-1)^i\frac{(\lambda t)^{k+i}}{k!i!} E_{t,k}(F)\\
&=& \gamma_t^{\otimes q}(F) +\sum_{r=1}^\infty \frac{1}{N^r}\sum_{k=0}^r\left( \frac{(-1)^{r-k}}{k!(r-k)!} \left( q(q-1) V_\infty t \right)^r E_{t,k}(F)\right)
\end{eqnarray}
where $\lambda:={q(q-1)}\frac{V_\infty}{N}$.
In particular, $\mathbb{Q}^N_{t,q}$ is differentiable up to any order with
$$
\partial^r \mathbb{Q}_{t,q}(F) = \sum_{k=0}^r \frac{(-1)^{r-k}}{k!(r-k)!} \left( {q(q-1)} V_\infty t \right)^r E_{t,k}(F)\ .
$$
\end{theorem}
Notice in particular that, although the $E_{t,k}(F)$s depend on the choice of the upper bound $V_\infty$ for the potential function $V$, the coefficients of the development do not (they are the derivatives of $\mathbb{Q}_{t,q}$, that do not depend on $V_\infty$).
\proof[Proof of Theorem \ref{Theo:FKunnormalized}]
\
We have first, for $f\in \mathcal{B}_b ( E)$
$$\frac{d}{dt} \mathbb{E} (\eta_t^N(f))=\mathbb{E}(\eta_t^N(L_t(f)))+\sum\limits_{i,j=1}^N\mathbb{E}\left(\frac{V_t(\xi_t^i)\left( f(\xi_t^j)-f(\xi_t^i) \right)}{N^2}\right)$$
and
$$\frac{d}{dt}\mathbb{E}( \gamma_t^N(f))=\mathbb{E}(\gamma_t^N(L_t(f)))-\mathbb{E}(\eta_t^N(V_t)\gamma_t^N(f))+\mathbb{E}\left(\gamma_t^N(1)\sum\limits_{i,j=1}^N\frac{V_t(\xi_t^i)\left( f(\xi_t^j)-f(\xi_t^i) \right)}{N^2}\right)$$
For $F\in \mathcal{B}_b^{sym} ( E^q)$, let us introduce the useful notation:
$F_i:E^{q-1}\longrightarrow \mathcal{B}_b ( E),$
$$F_i(x_1,...,x_{q-1})(y):=F(x_1,...,x_{i-1},y,x_i,...,x_{q-1})$$
and let us extend $L_t$ to functions in $\mathcal{B}_b^{sym}$:
$$L_t(F)(x_1,...,x_n):=\sum\limits_{i=1}^qL_t(F_i(x_1,...,x_{i-1},x_{i+1},...,x_n))(x_i).$$
We get:
\begin{eqnarray}
\label{Eq:deriv1}
\lefteqn{
\frac{d}{dt} \mathbb{E}((\gamma_t^N)^{\odot q} (F)) }
\\
\nonumber
& = & \mathbb{E}( (\gamma_t^N)^{\odot q} (L_t (F))) - \mathbb{E}(q \eta_t^N(V_t) (\gamma_t^N)^{\odot q }(F))\\
\nonumber
&& + \mathbb{E}\left\{ \gamma_t^N(1)^q\sum_{s\in <q,N>} \sum_{i=1}^q \sum_{k=1}^N V_t(\xi^{s(i)})\right. \\
\nonumber
&&\left. \times\left( \frac{F_i(\xi^{s(1)}_t,\dots,\xi^{s(i-1)}_t,\xi^{s(i+1)}_t,\dots, \xi^{s(q)}_t)(\xi_t^k) - F(\xi^{s(1)}_t,\dots,\xi^{s(i)}_t,\dots, \xi^{s(q)}_t)}{N\cdot(N)_q} \right) \right\}
\end{eqnarray}
The second term in the right hand side of (\ref{Eq:deriv1}) reads:
\begin{eqnarray*}
&&\mathbb{E}\left( -q\frac{\gamma_t^N(1)^q} {N\cdot(N)_q} \sum_{s\in <q,N>}\sum_{k=1}^N F(\xi^{s(1)}_t , \dots , \xi^{s(q)}_t) V_t(\xi^k_t) \right)=\\
&&\mathbb{E}\left( -q\frac{\gamma_t^N(1)^q} {N\cdot(N)_q} \sum_{s\in <q,N>}F(\xi^{s(1)}_t , \dots , \xi^{s(q)}_t) (V_t(\xi^{s(1)}_t)+\dots + V_t(\xi^{s(q)}_t)) \right)+\\
&&\mathbb{E}\left( -q\frac{\gamma_t^N(1)^q} {N\cdot(N)_q} \sum_{s\in <q+1,N>} F(\xi^{s(1)}_t , \dots , \xi^{s(q)}_t) V_t(\xi^{s(q+1)}_t) \right).
\end{eqnarray*}
Similarly, making use of the symmetry properties of $F$, the last term in the right hand side of (\ref{Eq:deriv1}) reads:
\begin{eqnarray*}
\mathbb{E}\left( q\frac{\gamma_t^N(1)^q}{N\cdot (N)_q} \sum_{s\in <q,N>} \sum_{k=1}^N V_t(\xi^{s(q)}_t) F(\xi^{s(1)}_t,\dots,\xi^{s(q-1)}_t,\xi^k_t)\right)-\\
\mathbb{E}\left( \frac{\gamma_t^N(1)^q}{(N)_q} \sum_{s\in <q,N>} F(\xi^{s(1)}_t,\dots,\xi^{s(q-1)}_t,\xi_t^{s(q)})(V_t(\xi^{s(1)}_t)+\dots+V_t(\xi^{s(q)}_t))\right)\\
\end{eqnarray*}
The first term in this last sum decomposes then into (making use once again of the symmetry properties of $F$):
\begin{eqnarray*}
\mathbb{E}\left(\frac{\gamma_t^N(1)^q}{N\cdot (N)_q} \sum_{s\in <q,N>} (V_t(\xi^{s(1)}_t)+\dots+V_t(\xi^{s(q)}_t)) F(\xi^{s(1)}_t,\dots,\xi^{s(q-1)}_t,\xi^{s(q)}_t)\right)+\\
\mathbb{E}\left( q\frac{\gamma_t^N(1)^q}{N\cdot (N)_q} \sum_{s\in <q,N>} \sum_{k=1}^{q-1} V_t(\xi^{s(q)}_t) F(\xi^{s(1)}_t,\dots,\xi^{s(q-1)}_t,\xi^{s(k)}_t)\right)+\\
\mathbb{E}\left(q\frac{\gamma_t^N(1)^q}{N\cdot (N)_q} \sum_{s\in <q+1,N>} F(\xi^{s(1)}_t,\dots,\xi^{s(q-1)}_t,\xi^{s(q)}_t)V_t(\xi^{s(q+1)}_t) \right)
\end{eqnarray*}
Reorganizing the summands in these expansions, we get finally that the two last terms of (\ref{Eq:deriv1}) sum up to :
\begin{eqnarray*}
&& - \mathbb{E}\left( \frac{\gamma_t^N(1)^q }{(N)_q} \sum_{s\in <q,N>} F(\xi^{s(1)}_t,\dots,\xi^{s(q)}_t) (V_t(\xi^{s(1)}_t)+\dots+V_t(\xi^{s(q)}_t)) \right)+\\
&&\mathbb{E} \Big{(} \frac{\gamma_t^N(1)^q}{N\cdot (N)_q} \sum_{s\in <q,N>} \sum_{i,r=1}^q V_t(\xi^{s(i)}_t) \left[F(\xi^{s(1)}_t,\dots,\xi^{s(i-1)}_t,\xi^{s(r)}_t,\xi^{s(i+1)}_t,\dots,\xi^{s(q)}_t)\right.\\
&& ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \left. - F(\xi^{s(1)}_t,\dots,\xi^{s(q)}_t)\right] \Big{)}\ .
\end{eqnarray*}
We set
\begin{eqnarray*}
F_{V_t}(x_1,\dots,x_q) &:=& (V_t(x_1)+\dots+V_t(x_q)) F(x_1,\dots,x_q) \\
F_{V_t}^{(i,r)}(x_1,\dots, x_q) &=& V_t(x_i) (F(x_1,\dots,x_{i-1},x_r,x_{i+1},\dots,x_q)-F(x_1,\dots,x_q)).
\end{eqnarray*}
Then the equations above give
$$
\frac{d}{dt}\mathbb{Q}^N_{t,q}(F) = \mathbb{Q}^N_{t,q} L_t (F) - \mathbb{Q}^N_{t,q}(F_{V_t}) + \sum_{1\leq i,r\leq q} \frac{1}{N} \mathbb{Q}^N_{t,q}(F_{V_t}^{(i,r)})\ .
$$
And so
\begin{eqnarray}\label{fundam}
\mathbb{Q}^N_{t,q}(F) = \mathbb{E}\left(F(\hat{\xi}^1_t,\dots,\hat{\xi}^q_t) \exp\left( -\int_0^t V_t(\hat{\xi}^1_s) +\dots + V_t(\hat{\xi}^q_s) ds\right)\right) \ .\end{eqnarray}
We then obtain Equation (\ref{Eq:FKunnormalized1}) of the theorem by noticing that
$$
\mathbb{P}(k \text{ rings on} [0,t] ) = \frac{(\lambda t)^k e^{-\lambda t}}{k!}\ .
$$
We obtain Equation (\ref{Eq:FKunnormalized2}) by developping the term $e^{-\lambda t}$ and by noticing that
$$
E_{t,0}(F) = \gamma_t^{\otimes q}(F)\ .
$$
\endproof
\subsection{Wick theorem for interacting particle systems}
We say that two particles of the auxiliary system $\hat\xi^i$, $\hat\xi^j$, $i\not=j$ interact at $t$ if and only if $\hat\xi^i_t$ jumps to $\hat\xi^j$ or $\hat\xi^j_t$ jumps to $\hat\xi^i$ at $t$.
We say that a trajectory of the auxiliary system of particles $\hat\xi^1_t,\dots,\hat\xi^q_t$ is exactly Wick-coupled on $[0,t]$ if and only each particle of the system has exactly one interaction with another particle between $0$ and $t$ (notice that the existence of such trajectories requires $q$ to be even). We write $\mathcal{W}_t$ for the set of Wick-coupled trajectories on $[0,t]$; the set $W_t$ embeds into the set of trajectories with exactly $q/2$ rings on $[0,t]$.
\begin{theorem}\label{Thm:Wick1}
If $F \in \mathcal{B}^{\text{sym}}_0 ( E^q)$ then, for $r< q/2$, $\partial^r \mathbb{Q}^N_{t,q}(F) =0$.
Moreover, for $q$ even,
\begin{eqnarray*}
\partial^{q/2} \mathbb{Q}_{t,q}(F) & =& \frac{({q(q-1)} V_\infty t)^{q/2}}{(q/2)!}E_{t,q/2}(F)
\end{eqnarray*}
In particular,
\begin{equation}
\label{Eq:majWick}
\limsup_{N\rightarrow +\infty} N^{q/2}|\mathbb{Q}^N_{t,q}(F)| \leq \frac{({q(q-1)} V_\infty t)^{q/2}}{(q/2)!} \Vert F\Vert_\infty .
\end{equation}
\end{theorem}
\begin{proof}[Proof of Thm \ref{Thm:Wick1}]
Recall that $F$ is symmetric and belongs to $\mathcal{B}^{\text{sym}}_0 ( E^q)$ so that, for any $i\leq q$, $\int_E F(x_1,...,x_q)\gamma_t(dx_i)=0$.
The proof follows then from Thm~\ref{Theo:FKunnormalized} together with the observation that, if there are strictly less than $q/2$ rings, then at least one of the particles does not interact with the others so that, if $k<\frac{q}{2}$, $E_{t,k}(F)=0$.
\end{proof}
\begin{corollary}
With the assumptions of Thm~\ref{Theo:FKunnormalized} and $q$ even, we have:
\begin{eqnarray*}
\partial^{q/2} \mathbb{Q}_{t,q}(F) & =& \frac{q!}{(q/2)!} (V_\infty t)^{q/2} \mathbb{E}((F^e(\hat{\xi}^1_s, \dots, \hat{\xi}^q_s)_{0\leq s \leq t}) | \mathcal{W}_t).
\end{eqnarray*}
If, furthermore, $F=(f^1 \otimes \dots \otimes f^q)_{\text{sym}}$ (with $f^1,\dots,f^q$ centered with respect to $\gamma_t$) then we have the Wick-type expansion
\begin{eqnarray*}
\mathbb{E}(F^e((\hat{\xi}^1_s, \dots, \hat{\xi}^q_s)_{0\leq s \leq t})| \mathcal{W}_t) = \frac{2^{q/2} (q/2)!}{q!} \sum_{I_q\in\mathcal{I}_q} \prod_{\{i,j\} \in I_q} E'_{t,1}(f^i \otimes f^j)
\end{eqnarray*}
where the sum is over the set ${\mathcal{I}_q}$ of partitions $I_q$ of $[q]$ into pairs, $i,j$ being paired if $\{i,j\} \in I_q$ and where, for a function $G$ of two variables
$$
E'_{t,1}(G) := \mathbb{E}(G(\hat{\mu}^1_t,\hat{\mu}^2_t)\exp\left(-\int_0^t [V_s(\hat{\mu}^1_s) + V_s(\hat{\mu}^2_s)] ds \right)|\text{one ring on } [0,t])
$$
with $\hat{\mu}^1_t,\hat{\mu}^2_t$ an auxiliary system of two particles (defined with the same rules as a general auxiliary system of $q$ particles).
\end{corollary}
\begin{proof}
Indeed, from the enumeration of possible interactions between particles and the general properties of Poisson processes, we have:
$$\mathbb{P}(\mathcal{W}_t|\text{$q/2$ rings on}\ [0,t])=\frac{q!}{(q(q-1))^{q/2}}.$$
The first part of the Corollary follows.
The last part follows from the observation that the distributions of two particles (or of blocks of particles) are independent conditionnally to the assumption that they do not interact (either directly or through interactions with other particles), so that quantities such as $ \mathbb{E}(F^e((\hat{\xi}^1_s, \dots, \hat{\xi}^q_s)_{0\leq s \leq t})| \mathcal{W}_t) $ may be computed by disjoint integrations over blocks (pairs in this particular case) of interacting particles. Since $Card(\mathcal{I}_q)=\frac{q!}{2^{q/2} (q/2)!}$, the Corollary follows.
\end{proof}
\subsection{Expansion of the normalized measure}
The purpose of the present section is to prove for the measures associated to the empirical $U$-statistics $(\eta^N_t)^{\odot q}$ properties similar to the ones obtained in the unnormalized case. We conclude the section with a Wick formula in this setting.
We set, for $F\in \mathcal{B}_b(E^q)$:
$$
\mathbb{P}_{t,q}^N(F) = \mathbb{E}( (\eta^N_t)^{\odot q} (F))\ .
$$
\begin{theorem}
\label{Theo:FKnormalized}
The sequence $(\mathbb{P}^N_{t,q})$ is differentiable up to any order.
\end{theorem}
Let us mention that an explicit formula (that we omit) for the derivatives $\partial^l \mathbb{P}_{t,q}$ follows immediately from Fla~\ref{formm} in our proof.
\proof
Let us expand first $(\eta_t^N)^{\odot q }$ in terms of $(\gamma^N_t)^{\odot q}$. Let, in the following, $F$ be a bounded symmetric function of $q$ variables. We first have:
\begin{eqnarray*}
(\eta_t^N)^{\odot q } (F) & = & (\gamma^N_t)^{\odot q} (F) \gamma^N_t(1)^{-q}\\
& = & (\gamma^N_t)^{\odot q} (F) \frac{1}{\gamma_t(1)^q} \frac{1}{\left(1-\left( 1- \frac{\gamma^N_t(1)}{\gamma_t(1)} \right) \right)^q}\ .
\end{eqnarray*}
We set $\check{F}_t:=\frac{F}{\gamma_t(1)^q}$ and get:
$$
(\eta_t^N)^{\odot q } (F) = (\gamma^N_t)^{\odot q} (\check{F}_t) \frac{1}{\left(1-\left( 1- \frac{\gamma^N_t(1)}{\gamma_t(1)} \right) \right)^q}=(\gamma^N_t)^{\odot q} (\check{F}_t)\frac{1}{(1-u^N_t)^q}\ ,
$$
with $$u^N_t=1-\frac{\gamma^N_t(1)}{\gamma_t(1)} \ .$$
Recall the decomposition (that holds for any $m$, see \cite[Lemma 4.11]{DPR-2006}):
\begin{eqnarray*}
\frac{1}{(1-u^N_t)^q} &=& \sum_{0\leq k \leq m} (q-1+k)_k \frac{(u^N_t)^k}{k!} + (u^N_t)^{m} \sum_{1\leq k \leq q} C_{q+m}^{k+m} \left( \frac{u^N_t}{1-u^N_t}\right)^k
\end{eqnarray*}
from which it follows that:
\begin{eqnarray*}
\nonumber
(\eta_t^N)^{\odot q } (F) & = & (\gamma^N_t)^{\odot q} (\check F_t) \sum_{0\leq k \leq m} (q-1+k)_k \frac{(u^N_t)^k}{k!} \\
\nonumber
&&~~~~~ + (\gamma^N_t)^{\odot q} (\check{F}_t) (u^N_t)^{m} \sum_{1\leq k \leq q} C_{q+m}^{k+m} \left( \frac{u^N_t}{1-u^N_t}\right)^k\\
&&~~~~=: (1)+(2)\ .
\end{eqnarray*}
For an arbitrary function $f$ on $E$, we set: $\theta_t(f):=\frac{f-\eta_t(f)}{\gamma_t(1)}$. Then, recalling the equality $\gamma_t(1) = \exp\left( -\int_0^t \eta_s(V_s) ds \right)$, we get
\begin{eqnarray*}
\frac{\gamma^N_t(1)}{\gamma_t(1)} &=& e^{ -\int_0^t [\eta_u^N(V_u)-\eta_u(V_u)] du }\\
&=&1- \int_0^t [\eta_u^N(V_u)-\eta_u(V_u)] \frac{\gamma^N_u(1)}{\gamma_u(1)}du\\
&=&1-\int_0^t\gamma_u^N(\theta_u(V))du \ ,
\end{eqnarray*}
where the last identity follows from the rewriting $$[\eta_u^N(V_u)-\eta_u(V_u)]\frac{\gamma^N_u(1)}{\gamma_u(1)}=\eta_u^N\left( \frac{V_u-\eta_u(V_u)}{\gamma_u(1)} \right)\gamma^N_u(1)=\gamma_u^N(\theta_u(V_u)).$$
In particular, we get, for an arbitrary $k>0$,
\begin{equation}
\label{Eq:decint}
\left(1-\frac{\gamma^N_t(1)}{\gamma_t(1)} \right)^k = \int_{s_1,\dots s_k \in [0,t] }\prod_{i=1}^k \gamma_{s_i}^N(\theta_{s_i}(V_{s_i})) ds_1 \dots ds_k
\end{equation}
$$=k!\ \int_{0\leq s_1\leq \dots \leq s_k\leq t }\prod_{i=1}^k \gamma_{s_i}^N(\theta_{s_i}(V_{s_i})) ds_1 \dots ds_k.$$
We therefore have
\begin{eqnarray}
\nonumber
(1) &=&
\sum_{0\leq k \leq m} (q-1+k)_k \int_{0\leq s_1 \leq \dots\leq s_k \leq t} \left[ \prod_{i=1}^k \gamma_{s_i}^N(\theta_{s_i}(V_{s_i}))\right](\gamma^N_t)^{\odot q} (\check{F}_t) ds_1 \dots ds_k \\
&=& \sum_{0\leq k \leq m} (q-1+k)_k \times \int_{0\leq s_1 \leq \dots\leq s_k \leq t} \gamma^{k,q,N}_{{\mathbf s},t}( S^{k,q}_{s_1,\dots,s_k,t}(F) )ds_1\dots ds_k \label{Eq:dev1}
\end{eqnarray}
with ${\mathbf s}:=\{s_1,\dots,s_k\}$, $\gamma^{k,q,N}_{{\mathbf s},t}= \gamma^N_{s_1}\otimes \dots \gamma^N_{s_k} \otimes (\gamma^N_t)^{\odot q} $, $S^{k,q}_{{\mathbf s},t}(F) = \theta_{s_1}(V_{s_1}) \otimes \dots \otimes \theta_{s_k}(V_{s_k}) \otimes \check{F}_t$.
We introduce the operators (for $G$ of $k+q$ variables, $j\leq k$)
$$
D_{j}G(x_1,\dots,x_{k+q}) = \frac{N-q-k+j}{N} G(x_{1},\dots,x_{k+q})+ \frac{1}{N}\sum_{j+1\leq i \leq k+q} G_j(x_1,\dots,x_{k+q})(x_i)
$$
We then have, for any empirical measure $m(x)=\frac{1}{N}\sum_{1\leq i \leq N} \delta_{x_i}$ and for any function $G$ of $k+q\leq N$ variables and any measure $\mu$ on $E^{j-1}$
$$
\mu\otimes (m(x) \otimes m(x)^{\odot (k+q-j)}) (G) =\mu \otimes m(x)^{\odot (k+q-j+1)} (D_{j}G)\ .
$$
For $G$ function of $q+k$ variables and $s\leq t$ and $j\leq k+q$, we define then the Markov operator
\begin{eqnarray}
\label{Eq:defRq}
\lefteqn{R^j_{s,t} G(x_1,\dots,x_{k+q}) =}\\
\nonumber &&\mathbb{E}(G(x_1,\dots,x_{k+q-j},\hat{\xi}^{1}_{t-s},\dots,\hat{\xi}^j_{t-s})e^{-\int_0^{t-s}(V_{s+u}(\hat\xi_u^1)+\dots +V_{s+u}(\hat\xi_u^j))du}\\
\nonumber
&& ~~~~~~~~~~~~~~~~~~~~~~~~~|\hat{\xi}^1_0=x_{k+q-j+1},\dots,\hat{\xi}^j_{0}=x_{k+q})\ ,
\end{eqnarray}
where the $\hat\xi^i$ are defined as before, excepted for the initial condition that reads now $\hat\xi_0^i=x_{k+i}$.
Reasoning as before (and using the auxiliary system $\hat\xi_s^i$ as in Sect.~\ref{unnorm}), we get for $s_k\leq t$ and $H$ a function of $k+q$ variables, $\mu$ a measure on $E^{j-1}$,
\begin{eqnarray*}
\mathbb{E}(\mu\otimes (\gamma^N_{s_k} \otimes (\gamma^N_t)^{\odot q}) (H)|{\cal F}_{s_k})
&=& \mu \otimes(\gamma^N_{s_k} \otimes (\gamma^N_{s_k})^{\odot q} )R^{q}_{s_k,t} (H)\\
&=& \mu \otimes (\gamma^N_{s_k})^{\odot (q+1)} D_{q+1} R^{q}_{s_k,t} (H)\ ,
\end{eqnarray*}
where ${\cal F}_{t}$ stands for the natural filtration on the probability space underlying the particle system. Here, we took advantage of the Markovian properties of the system and of Eq~\ref{fundam}.
So, by recurrence~:
\begin{eqnarray*}
\mathbb{E}(\gamma^{k,q,N}_{{\mathbf s},t}( G )) =
(\eta_0)^{\otimes(k+q)} R^{k+q}_{0,s_1} D_1 R^{k-1+q}_{s_1,s_2} \dots D_k R^{q}_{s_k,t}( G)\ .
\end{eqnarray*}
We proceed now as for the unnormalized measure and introduce still another more general auxiliary particle system $({\check{\xi}}^1_t,\dots,\check{\xi}^{k+q}_t)_{t\geq 0}$. The system depends on $k$ (and $q$), but we do not emphasize this dependency to abbreviate the notations. This should not create ambiguities since the dependency on $k$ should be obvious from the context in the following formulas.
Let $S_1,S_2,\dots,S_k$ be the order statistics of $k$ uniform random variables in $[0,t]$ (we take $k$ i.i.d. uniform variables in $[0,t]$ and sort them). The $S_i$ are independent of all the other variables. We set $S_i:=t$ for $k<i\leq k+q$. The system (which agrees with the auxiliary system $\hat\xi^i_t$ when $k=0$) is defined as follows:
\begin{itemize}
\item The law of $(\check{\xi}^1_0,\dots,\check{\xi}^{k+q}_0)$ is $\eta_0^{\otimes k+q}$
\item The particles $\check{\xi}^i_s$ diffuse according to $L_s$ and undergo the following jumps:
\item Each couple $(i,j)\in [q+k]^2$ has an exponential clock of parameter $V_\infty/N$. At a ringing time $s$:
\begin{itemize}
\item If $s>S_i$ or $s>S_j$, then nothing happens.
\item Else,
$$
\check{\xi}^i_s
\begin{cases}
\leftarrow \check{\xi}^j_s & \text{ with proba. } \frac{V_s(\check{\xi}^i)}{V_\infty}\\
\leftarrow \check{\xi}^i_s & \text{ with proba. } 1-\frac{V_s(\check{\xi}^i)}{V_\infty}
\end{cases}
$$
\item At $s=S_i$, we make the following replacement (with $U$ sampled uniformly in $\{i+1,\dots,k+q\}$)
$$
\check{\xi}^U_s \leftarrow \check{\xi}^i_s \text{ with proba. } \frac{k-i+q}{N}\ .
$$
We say that there has been a parasite at $s$, so that the system has at most $k$ parasites.
\end{itemize}
\end{itemize}
With the notation
\begin{eqnarray*}
\lefteqn{
\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{S_1},\dots,\check{\xi}^k_{S_k},\check{\xi}^{k+1}_t,\dots,\check{\xi}^{k+q}_t)
}
\\
&:=&S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{S_1},\dots,\check{\xi}^k_{S_k},\check{\xi}^{k+1}_t,\dots,\check{\xi}^{k+q}_t)e^{-\int_0^{S_1}\dots\int_0^t V_u(\check\xi_u^1)du\dots V_u(\check\xi_u^{k+q})du}
\ ,
\end{eqnarray*}
the equation (\ref{Eq:dev1}) gives
\begin{eqnarray*}
\mathbb{E}((1)) & = & \sum_{0\leq k \leq m} \frac{(k-1+q)_k}{k!}\mathbb{E}(\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{S_1},\dots,\check{\xi}^k_{S_k},\check{\xi}^{k+1}_t,\dots,\check{\xi}^{k+q}_t)) \\
&& =\sum_{0\leq k \leq m} \frac{(k-1+q)_k}{k!} \\
&&~~~~~~~~~~\times \mathbb{E}\left\{ \sum_{r=0}^{+\infty} \frac{\left( \Lambda_S V_\infty/N\right)^re^{-\Lambda_s V_\infty/N}}{r!} \mathbb{E}(\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{S_1},\dots)|r \text{ rings },S_1,\dots,S_k) \right\}
\end{eqnarray*}
with
$$
\Lambda_S := q(q-1)(t-S_k) +(q+1)q(S_k-S_{k-1})+\dots +(q+k)(q+k-1) (S_1-0)\ ,
$$
for which we will use the crude bound $\Lambda_S \leq (q+k)^2 t$.
We set
$$
\bar{E}_{k,r,l} = \mathbb{E}( \Lambda_S^l \mathbb{E}(\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{s_1},\dots)|r \text{ rings },S_1,\dots,S_k))\ ,
$$
and notice that the parasites are sampled independently of the $S_i$ and of the ringing times, so that:
$$
\bar{E}_{k,r,l} = \sum_{i=0}^k\mathbb{P}(i~\text{parasites})\mathbb{E}( \Lambda_S^l \mathbb{E}(\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{S_1},\dots)|i ~\text{parasites}, r~\text{rings},S_1,\dots,S_k))\ .$$
The probability $\mathbb{P}(i~ \text{parasites})$ depends on $t$, $q$ and $k$, but we omit the corresponding indices to simplify the notation.
The following lemma is crucial:
\begin{lemma}\label{crucial}
For $i+r<\frac{k}{2}$, we have:
$$\bar{D}_{k,r,l,i}:=\mathbb{E}( \Lambda_S^l \mathbb{E}(\check S_{{\mathbf S},t}^{k,q} (\check{\xi}^1_{s_1},\dots)|i ~\text{parasites}, r~\text{rings},S_1,\dots,S_k))=0\ ,$$ so that:
$$
\bar{E}_{k,r,l} = \sum_{i=(\left\lceil {\frac{k}{2}}\right\rceil-r)^+}^k\mathbb{P}(i~ \text{parasites})\bar{D}_{k,r,l,i}\ .$$
where $\left\lceil x\right\rceil$ stands for $\inf\{k\in \mathbb{N}:k\geq x\}$.
\end{lemma}
Indeed, for $i+r<\frac{k}{2}$, at least one of the $\check{\xi}^j$, $j\leq k$ does not interact with the other particles, so that, by Fubini's lemma, $\theta_{S_j}(V_{S_j})(\check{\xi}^j_{S_j})e^{-\int_0^{S_j}V_u(\check\xi_u^j)du}$ can be integrated independently.
However, conditionnally to the assumption that $\check{\xi}^j$ does not interact with the other particles, we have:
$$\mathbb{E} (\theta_{S_j}(V_{S_j})(\check\xi_{S_j}^j)e^{-\int_0^{S_j}V_u(\check\xi_u^j)du})=\eta_{S_j}(\theta_{S_j}(V_{S_j}))=0,$$
and the lemma follows.
We then have (by developping the exponential in $\mathbb{E}((1))$)
\begin{eqnarray}
\label{Eq:esp(1)}
\lefteqn{
\mathbb{E}((1)) = \sum_{0\leq k \leq m} \frac{(k+q-1)_k}{k!} \sum_{r=0}^{+\infty} \frac{V_\infty^r }{N^r r!} \sum_{i=0}^{+\infty} \frac{(-1)^i}{i!} \left(\frac{V_\infty }{N} \right)^i \bar{E}_{k,r,r+i}
}
\\
\label{formm}
&=&\sum_{0\leq k \leq m} \frac{(k+q-1)_k}{k!} \sum_{r=0}^{+\infty} \frac{V_\infty^r }{N^r r!} \sum_{i=0}^{+\infty} \frac{(-1)^i}{i!} \left(\frac{V_\infty }{N} \right)^i \\
\nonumber
&&~~~~~~~~~~~~~ \sum_{j=(\left\lceil \frac{k}{2}\right\rceil-r)^+}^k\mathbb{P}(j~ \text{parasites})\bar{D}_{k,r,r+i,j}\ .
\end{eqnarray}
By the construction of the auxiliary system of particles, $\mathbb{P}(j~ \text{parasites})=O(\frac{1}{N^j})$.
For further use, the explicit formula is:
\begin{eqnarray*}
\lefteqn{\mathbb{P}(j \text{ parasites })}\\
& =&
\sum_{1\leq k_1 \leq \dots \leq k_j \leq k} \left( \frac{(q+k_1-1)}{N} \dots \frac{q+k_j-1}{N} \prod_{l\in [k] \backslash \{k_1,\dots,k_j\}} \left( 1-\frac{q+l-1}{N} \right)\right) \ ,
\end{eqnarray*}
so that there exists coefficients $(a_{l})$ (depending on $k$ and $q$, as usual we omit the corresponding indices for notational simplicity) such that
$$
\mathbb{P}(j \text{ parasites }) = \sum_{l=j}^k \frac{a_{l}}{N^l}\ .
$$
Because of the condition $i+r<\frac{k}{2}$ in Lemma~\ref{crucial}, it follows that the coefficient of any $\frac{1}{N^p}$ in the expansion of $\mathbb{E}((1))$ is a finite sum of coefficients which are independent of $N$. This is essentially all what we need in order to prove the Thm.~\ref{Theo:FKnormalized}, provided we are able (Step 1 below) to give a suitable upper bound for the remainder terms in the expansion of $\mathbb{E}((1))$ at a given order, and (Step 2) to give a suitable upper bound to $\mathbb{E}((2))$. This is the (cumbersome) purpose of the following conclusion to the proof of the Thm.~\ref{Theo:FKnormalized}, that we include for completeness sake but that the reader may skip, if she/he wishes.
\underline{Step 1}: Upper bound for the remainder terms in the expansion of $E((1))$ at the order p, assuming that $m$ is even and $m=2p$.
Notice that for an arbitrary bounded function $F$,
$$
\Vert \check F \Vert_\infty \leq \Vert F \Vert_\infty e^{tV_\infty q}\ ,
$$
$$
\Vert S_{{\mathbf S},t}^{k,q}\Vert_\infty \leq 2^{k} V_\infty^{k} e^{t(k+q)V_\infty} \Vert F \Vert_\infty \ ,
$$
and so
$$
|\bar{E}_{k,r,r+i}| \leq (t (q+k)^2)^{r+i} 2^k V_\infty^k e^{t(k+q)V_\infty} \Vert F\Vert_\infty\ .
$$
In the Fla~\ref{Eq:esp(1)} for the expectation $\mathbb{E}((1))$, the sum of the terms with a factor $1/N^{r+i}$, $r+i\geq p$ is, in absolute value, equal to (we use here: $\forall x\geq 0$, $\sum_{i\geq j} \frac{x^i}{i!} \leq \frac{x^{j}}{j!} e^x$)
\begin{eqnarray*}
&& \left| \sum_{0\leq k \leq 2p} \frac{(k+q-1)_k}{k!} \sum_{r=0}^{+\infty} \frac{V_\infty^r }{N^r r!} \sum_{i\geq (p-r)_+} \frac{(-1)^i}{i!} \left(\frac{V_\infty }{N} \right)^i \bar{E}_{k,r,r+i} \right|\\
&&~~~~\leq \sum_{0\leq k \leq 2p} \frac{(k+q-1)_k}{k!} \left[ \sum_{r=0}^{p-1} \frac{V_\infty^r}{N^r r!}
2^{k} V_\infty^{k} e^{t(k+q)V_\infty} \Vert F \Vert_\infty \right.\\
&&~~~~ \times \frac{1}{(p-r)!} \left( \frac{V_\infty t(q+k)^2}{N}\right)^{p-r} \exp\left(\frac{V_\infty t(q+k)^2}{N} \right) (t(k+q)^2)^r \\
&&~~~~~~~~+ \left. \sum_{r=p}^{+\infty} \frac{V_\infty^r}{N^r r!} 2^{k} V_\infty^{k} e^{t(k+q) V_\infty} \Vert F \Vert_\infty \exp\left(\frac{V_\infty t(q+k)^2}{N} \right) (t(q+k)^2)^r \right]\\
&&~~~~~ \leq \frac{1}{N^p} \sum_{0\leq k \leq 2p} \frac{(k+q-1)_k}{k!} \\
&&~~~~~~~~~~\times \left[ \sum_{r=0}^{p-1} V_\infty^{p+k} 2^k e^{t(k+q)V_\infty} \Vert F \Vert_\infty \frac{1}{(p-r)!r!} (t(q+k)^2)^p e^{\frac{V_\infty t(q+k)^2}{N}} \right. \\
&&~~~~~~~~~ \left. + 2^k \frac{(t(k+q)^2)^pV_\infty^{k+p}}{p!}e^{V_\infty t (k+q)} \Vert F \Vert_\infty e^{2\frac{V_\infty t (q+k)^2}{N}} \right]=O(\frac{1}{N^p})\ .
\end{eqnarray*}
Since, by Fla~\ref{formm}, $\mathbb{E}((1))$ is the sum of this term with a polynomial in $\frac{1}{N}$, the proof of the Step 1 is concluded.
\underline{Step 2}: Upper bound to $E((2))$.
Since $\gamma_t(1)=e^{-\int_0^t\eta_u(V_u)du}$ and $\gamma_t^N(1)=e^{-\int_0^t\eta_u^N(V_u)du}$, we get:
$$
|\mathbb{E}((2))| \leq \Vert F \Vert_\infty e^{tq V_\infty} \left( \sum_{1\leq k \leq q} C_{q+2p}^{k+2p} (e^{tV_\infty}+1)^k\right) \mathbb{E}((u^N_t)^{2p})\ .
$$
We have, with the same auxiliary system of the $\check{\xi}$'s as before but with $q=0$, $k=2p$ and with $S_1,\dots,S_{2p}$ the order statistics of $2p$ uniform variables on $[0,t]$,
\begin{eqnarray*}
\lefteqn{ \mathbb{E}((u^N_t)^{2p})}\\ & = & \mathbb{E}(\theta_{S_1}(V_{S_1})\otimes \dots\otimes \theta_{S_{2p}}(V_{S_{2p}}) (\check{\xi}^1_{S_1},\dots,\check{\xi}^{2p}_{S_{2p}})e^{-\int_0^{S_1}\dots\int_0^{S_{2p}}V_{u_1}(\check\xi_{u_1}^1)\dots V_{u_p}(\check\xi_{u_{2p}}^{2p})du_1\dots du_{2p}})\ .
\end{eqnarray*}
We get (with, now, $\Lambda_S= (2p)(2p-1)S_1+(2p-1)(2p-2)(S_2-S_1)+\dots+1\times(S_{2p}-S_{2p-1})$)
\begin{eqnarray*}
\mathbb{E}((u^N_t)^{2p})&&= \mathbb{E}(\sum_{r=0}^{+\infty} \left( \frac{\Lambda_S V_\infty}{N} \right)^r \frac{1}{r!} e^{-\Lambda_S V_\infty/N} \mathbb{E}( \theta_{S_1}(V_{S_1})\otimes \dots\otimes \theta_{S_{2p}}(V_{S_{2p}})\\
&& (\check{\xi}^1_{S_1},\dots,\check{\xi}^{2p}_{S_{2p}})e^{-\int_0^{S_1}\dots\int_0^{S_{2p}}V_{u_1}(\check\xi_{u_1}^1)\dots V_{u_p}(\check\xi_{u_{2p}}^{2p})du_1\dots du_{2p}}|r \text{ rings},S_1,\dots,S_{2p}))\\
&& = \mathbb{E}(\sum_{r=0}^{+\infty} \left( \frac{\Lambda_S V_\infty}{N} \right)^r \frac{1}{r!} e^{-\Lambda_S V_\infty/N} \sum_{i=0}^{2p-1}\mathbb{P}(i \text{ parasites})\\
&&~~~~~~~ \times \mathbb{E}( \theta_{S_1}(V_{S_1})\otimes \dots\otimes \theta_{S_{2p}}(V_{S_{2p}}) (\check{\xi}^1_{S_1},\dots,\check{\xi}^{2p}_{S_{2p}})\\
&&e^{-\int_0^{S_1}\dots\int_0^{S_{2p}}V_{u_1}(\check\xi_{u_1}^1)\dots V_{u_p}(\check\xi_{u_{2p}}^{2p})du_1\dots du_{2p}}|r \text{ rings}, i \text{ parasites},S_1,\dots,S_{2p}))\ .
\end{eqnarray*}
We have
$$
\mathbb{P}( i \text{ parasites}) \leq C_{2p-1}^i \left( \frac{2p-1}{N} \right)^i
$$
and know that, if $r+i<p$
$$
\mathbb{E}( \theta_{S_1}(V_{S_1})\otimes \dots\otimes \theta_{S_{2p}}(V_{S_{2p}}) (\check{\xi}^1_{S_1},\dots,\check{\xi}^{2p}_{S_{2p}})$$
$$e^{-\int_0^{S_1}\dots\int_0^{S_{2p}}V_{u_1}(\check\xi_{u_1}^1)\dots V_{u_p}(\check\xi_{u_{2p}}^{2p})du_1\dots du_{2p}}|r \text{ rings}, i \text{ parasites},S_1,\dots,S_{2p})) = 0 \ .
$$
So that
\begin{eqnarray*}
\mathbb{E}((u^N_t)^{2p}) & \leq & \frac{1}{N^p} \sum_{r=0}^{+\infty} \frac{(t(2p)^2 V_\infty)^r}{r!} \sum_{i=(p-r)_+}^{2p-1} C_{2p-1}^i ( 2p-1)^i e^{2p t V_\infty}2^{2p}V_\infty^{2p}
\end{eqnarray*}
and
\begin{eqnarray*}
|\mathbb{E}((2))|& \leq & \frac{1}{N^p} \times \Vert F \Vert_\infty e^{tq V_\infty} \left(\sum_{1\leq k \leq q} C_{q+2p}^{k+2p} (e^{tV_\infty}+1)^k \right) \\
&& ~~ \times e^{t(2p)^2 V_\infty} \sum_{i=0}^{2p-1} C_{2p-1}^i ( 2p-1)^i e^{2p t V_\infty}2^{2p}V_\infty^{2p}=O(\frac{1}{N^p}),
\end{eqnarray*}
and the proof of the Thm is complete.
\endproof
\begin{corollary}
\label{Cor:FKWick}
If $F\in \mathcal{B}^{\text{sym}}_0 (E^q)$, then
\begin{itemize}
\item $\partial^l \mathbb{P}_{t,q} (F)=0$, $\forall l <q/2$
\item for $q$ even, $\partial^{q/2} \mathbb{P}_{t,q}(F)$ simplifies to
$$
\partial^{q/2} \mathbb{P}_{t,q}(F)=
\frac{V_\infty^{q/2}}{(q/2)!} (q(q-1) t)^{q/2} \mathbb{E}(\check F^e(\hat{\xi}^1_t,\dots,\hat{\xi}^q_t)|\frac{q}{2} \text{rings})\ ,
$$
or
$$\partial^{q/2} \mathbb{P}_{t,q}(F)=\frac{\partial^{q/2} \mathbb{Q}_{t,q}(F)}{\gamma_t(1)^q}.$$
\end{itemize}
\end{corollary}
\proof[Proof of Corollary \ref{Cor:FKWick}]
Indeed, the same argument as in the proof of Lemma~\ref{crucial} shows that, for a centered function $F\in{\cal B}_0^{sym}$, the terms $\overline{D}_{k,r,r+i,j}$ vanish for $j+r<\frac{k+q}{2}$ (this is because at least one of the $\check\xi^j,\ j\leq k+q$ does not interact with the other particles if there are less than $\frac{k+q}{2}$ rings and parasites).
It follows that, if $q$ is even (resp. odd), $\frac{q}{2}$ (resp. $\frac{q+1}{2}$) is the smallest power of $\frac{1}{N}$ appearing in the expansion of $\mathbb{E}((1))$ for $m>q$ and that the contribution of $(2)$ is negligible with respect to $1/N^{q/2}$, from which the first part of the Corollary follows.
The same reasoning shows that, for $q$ even, only the case $k=i=j=0$ and $r=\frac{q}{2}$ contributes to the coefficient of $(\frac{1}{N})^{\frac{q}{2}}$. As noticed before, for $k=0$, the systems $(\check{\xi})$ and $(\hat{\xi})$ have the same law.
The second statement follows.
\section{First results on asymptotic normality}
The empirical measure $\eta_t^N$ converges weakly to $\eta_t$: for all bounded $f$, $\eta^N_t(f) \underset{N \rightarrow + \infty }{\overset{\text{a.s.}}{\longrightarrow}} \eta_t(f)$. The Wick theorem (Cor. \ref{Cor:FKWick}) allowed us to improve on this convergence result. Namely, we know from the previous section that
\begin{itemize}
\item For $q$ odd and $F \in \mathcal{B}^{\text{sym}}_0(E^q)$: $N^{q/2} \mathbb{E}((\eta^N_t )^{\odot q} (F) ) \underset{N \rightarrow \infty}{\longrightarrow} 0$,
\item For $q$ even, $F \in \mathcal{B}^{\text{sym}}_0(E^q)$: $N^{q/2} \mathbb{E}((\eta^N_t )^{\odot q} (F) ) \underset{N \rightarrow \infty}{\longrightarrow} \Delta_{q/2}(F)$,
\end{itemize}
where $\Delta_{q/2}$ is a shortcut for the signed measure:
$$\Delta_{q/2}(F)=\frac{(q(q-1)tV_\infty)^{q/2}}{(q/2)!} \mathbb{E}(\check F^e(\hat{\xi}^1_t,\dots,\hat{\xi}^q_t)|\frac{q}{2} \text{rings})\ .$$
Furthermore, for $F=(f_1\otimes ...\otimes f_q)_{sym}$, with $f_i\in \mathcal{B}_0 (E)$, this formula simplifies to:
$$\Delta_{q/2}(F)=\frac{(2V_\infty t)^{q/2}}{\gamma_t(1)^q}\sum\limits_{I_q\in{\cal I}_q}\prod\limits_{\{i,j\}\in I_q}E_{t,1}'(f_if_j)=\sum\limits_{I_q\in{\cal I}_q}\prod\limits_{\{i,j\}\in I_q}W_t(f_i \otimes f_j),$$
where $W_t:=\frac{2V_\infty t}{\gamma_t(1)^2}E_{t,1}'$.
A striking consequence of this Wick theorem for interacting particle systems and their U-statistics is that it leads immediately to an explicit result of asymptotic normality for the vector-valued random measures $(\eta_t^N)^q$.
Notice that the same properties hold for the system studied in \cite{rubenthaler-2009} and so we have the same asymptotic normality result for this system.
More precisely:
\begin{theorem} \label{Theo:CLT}
$\forall f_1,\dots,f_q \in \mathcal{B}_0 (E)$,
$$
N^{1/2} (\eta^N_t(f_1),\dots,\eta^N_t(f_q)) \underset{N\rightarrow +\infty}{\overset{\text{law}}{\longrightarrow}} \mathcal{N}(0,K)
$$
with $K(i,j)= \eta_t(f_i f_j) + \mathbb{E}(W_t(f_i \otimes f_j)) $.
\end{theorem}
\begin{proof}
For any $u_1,\dots,u_q$, we have:
\begin{eqnarray*}
&& \mathbb{E}\left(\exp\left( N \eta^N_t \left(\log\left(1+\frac{iu_1 f_1 + \dots +iu_q f_q}{\sqrt{N}} \right)\right) \right)\right) \\
&& = \mathbb{E}\left( \exp\left( \sum_{k\geq 1} \frac{(-1)^{k+1}}{k} N^{1-k/2} \eta^N_t( (iu_1 f_1 + \dots iu_q f_q )^k )\right) \right)\\
&& \underset{N \rightarrow +\infty}{\sim} \mathbb{E}(\exp(\sqrt{N} (i u_1 \eta^N_t(f_1) + \dots + i u_q \eta^N_t (f_q)))) \\
&&~~~~~~~~~~~\times \exp\left( -\frac{1}{2} \eta_t ((u_1f_1 + \dots + u_q f_q)^2) ) \right) \ ,
\end{eqnarray*}
where the last equivalence follows from the differentiability of $\mathbb{P}_{t,1}^N$.
We also have
\begin{eqnarray*}
&& \mathbb{E}\left(\exp\left( N \eta^N_t \left(\log\left(1+\frac{iu_1 f_1 + \dots iu_q f_q}{\sqrt{N}} \right)\right) \right)\right) \\
&& ~~~ = \mathbb{E}\left( \prod_{j=1}^N \left( 1+ \frac{ i u_1 f_1(\xi^j_t) + \dots + i u_q f_q (\xi^j_t)}{\sqrt{N}}\right)\right) \\
&& ~~~ = \mathbb{E}\left(\sum_{0 \leq k \leq N} \frac{1}{N^{k/2}} \sum_{1\leq j_1 ,\dots,j_k \leq q } i^k u_{j_1} \dots u_{j_k} \sum_{1\leq i_1 < \dots < i_k \leq N } f_{j_1}(\xi^{i_1}_t) \dots f_{j_k}(\xi^{i_k}_t)\right)\\
&& ~~~ = \mathbb{E}\left(\sum_{0 \leq k \leq N} \frac{(N)_k}{N^{k/2}} \sum_{1\leq j_1, \dots , j_k \leq q } i^k u_{j_1} \dots u_{j_k} \frac{1}{k!} ( \eta^N_t )^{\odot k} (f_{j_1} \otimes \dots \otimes f_{j_k} )\right) \\
&& ~~~ \underset{N\rightarrow + \infty}{\longrightarrow} \sum_{k \geq 0, k \text{ even }} (-1)^{k/2} \sum_{1\leq j_1, \dots , j_k \leq q}\frac{ u_{j_1} \dots u_{j_k}}{k!} \sum\limits_{I_k\in{\cal I}_k}\prod\limits_{\{a,b\}\in I_k}W_t(f_a \otimes f_b)\\
&&~~~=\sum_{k \geq 0, k \text{ even }} \frac{(-1)^{k/2}}{2^{k/2} (k/2)!} \sum_{1\leq j_1, \dots , j_k \leq q} u_{j_1} \dots u_{j_k} \mathbb{E}(W_t(f_{j_1} \otimes f_{j_2})) \dots \mathbb{E}(W_t(f_{j_{k-1}} \otimes f_{j_k})) \\
&& ~~~ = \sum_{k \geq 0, k \text{ even }} \frac{(-1)^{k/2}}{2^{k/2} (k/2)!} \left( \sum_{1\leq j_1,j_2 \leq q} u_{j_1} u_{j_2} \mathbb{E}( W_t(f_{j_1}\otimes f_{j_2})) \right)^{k/2} \\
&& ~~~ = \exp\left( - \frac{1}{2} \sum_{ 1\leq j_1,j_2 \leq q} u_{j_1} u_{j_2} \mathbb{E}(W_t(f_{j_1}\otimes f_{j_2})) \right)
\end{eqnarray*}
The Theorem follows.
\end{proof}
\section{Convergence of empirical $U$-statistics.}
\subsection{Hoeffding's decomposition}
We want here to study the convergence under fairly general assumptions of a sequence of empirical U-statistics $(\eta_t^N)^{\odot q}(F)$ when $N\rightarrow +\infty$, where the $\eta_t^N$s are empirical measures on some space $E$ and $F $ is a bounded symmetrical function on $E^q$.
Let us first write an analog of Hoeffding's decomposition (cf. \cite{lee-1990}, \cite{delapena-gine-1999}). We fix some measure $\eta_t$.
For $1\leq k \leq q$ and any $F$ bounded symmetrical on $E^q$, we define
$$
F^{(k)}(x_1,\dots,x_q)= \int_{E^{q-k}} F(x_1,\dots,x_k,x_{k+1},\dots,x_q) \eta_t^{\otimes (q-k)}(dx_{k+1},\dots,dx_{q})\ .
$$
Notice that $F^{(q)}=F$ and that, for an arbitrary $k$, $F^{(k)}$ is symmetrical on $E^k$.
We define recursively (the function $F$ being fixed --for notational simplicity, we do not stress the dependency of $\theta$, $h^{(i)}$... in the following formulas):
\begin{eqnarray*}
\theta &=& \int_{E_n^q} F(x_1,\dots,x_q) \eta_t^{\otimes q}(dx_1,\dots,dx_q)\\
h^{(1)}(x_1) &=& F^{(1)}(x_1)-\theta\\
h^{(k)}(x_1,\dots,x_k) &=& F^{(k)}(x_1,\dots,x_k) - \sum_{j=1}^{k-1} \sum_{(k,j)}h^{(j)} -\theta\ ,
\end{eqnarray*}
where $\sum_{(k,j)}h^{(j)}$ is an abbreviation for
$\sum_{1\leq i_1 < \dots < i_j \leq k}[ h^{(j)}(x_{i_1},\dots,x_{i_j})]$.
The $h^{(k)}$ are bounded and symmetrical. They are constructed so as to be centered and to satisfy, for $l<k$, $(h^{(k)})_l=0$, see e.g. \cite[Sect. 1.6]{lee-1990}.
\subsection{Asymptotic normality for empirical U-statistics.}
As for classical U-statistics, the Hoeffding decomposition can be used to study the convergence of the empirical U-statistics $(\eta_t^N)^{\odot q}$.
Notice first that, $\forall N$,
\begin{equation}
\label{Eq:dec1}
(\eta_t^N)^{\odot q} (F) = (\eta_t^N)^{\odot q} (h^{(q)}) +(\eta_t^N)^{\odot q}(\sum_{j=1}^{q-1} \sum_{(q,j)} (h^{(j)})) +\theta\ .
\end{equation}
Notice that, $\forall 1\leq j \leq q-1$,
\begin{eqnarray*}
\sum_{(q,j)} h^{(j)}& = & \frac{1}{j!} \sum_{b\in \langle j,q \rangle } h^{(j)} (x_{b(1)},\dots,x_{b(j)})\ .
\end{eqnarray*}
So, with the notations of the first sections:
\begin{eqnarray}
\nonumber
(\eta_t^N)^{\odot q} (\sum_{(q,j)} h^{(j)}) &=& \frac{1}{(N)_q} \sum_{a \in \langle q,N \rangle} \frac{1}{j!} \sum_{b\in \langle j,q \rangle} h^{(j)} (\xi^{a\circ b (1)}_t,\dots,\xi^{a\circ b (j)}_t)\\
\nonumber
&=& \frac{1}{ j!(N)_q} \times (N-j)_{(q-j)} (q)_j \sum_{a' \in <j,N>} h^{(j)}(\xi^{a'(1)}_t,\dots,\xi^{a'(j)}_t)\\
\label{Eq:stat1}
&=& \frac{(q)_j}{j!} (\eta_t^N)^{\odot j}(h^{(j)})\ .
\end{eqnarray}
\begin{theorem}
Suppose we have, for all $ j \geq 2$ and $f \in \mathcal{B}^{\text{sym}}_0 (E^j)$,
\begin{equation}
\label{Eq:stat2}
\mathbb{E}( ((\eta_t^N)^{\odot j}(f))^2) \leq \frac{C}{N^{j}}\ ,
\end{equation}
for some constant $C$ depending on $t,N,j,\Vert f \Vert_\infty$. Assume further that, with $F$ as above,
\begin{equation}
\label{Eq:stat3}
\sqrt{N} \eta_t^N(h^{(1)}) = \sqrt{N}(\eta_t^N(F^{(1)}) - \theta) \overset{\text{law}}{\underset{N\rightarrow +\infty}{\longrightarrow}} \mathcal{N}(0,\sigma^2)\ ,
\end{equation}
then
\begin{equation}
\label{Eq:stat4}
\sqrt{N}((\eta_t^N)^{\odot q}(F)- \theta) \overset{\text{law}}{\underset{N\rightarrow +\infty}{\longrightarrow}} \mathcal{N}(0,q^2 \sigma^2)\ .
\end{equation}
\end{theorem}
The Theorem follows from Hoeffding's decomposition, together with the identity
$$(\eta_t^N)^{\odot q}(\sum\limits_{(q,1)}h^{(1)})=q\ \eta_t^N(h^{(1)}).$$
\begin{corollary}
For the Feynman-Kac particle systems of the first sections, we have:
$$\sqrt{N}((\eta_t^N)^{\odot q}(F)-\theta )\longrightarrow \mathcal{N}(0,q^2\eta_t((F^{(1)})^2)).$$
\end{corollary}
Indeed, notice first that the Eq~\ref{Eq:stat3} holds for Feynman-Kac particle systems e.g. as a Corollary of Th.~\ref{Theo:CLT}. The Corollary follows then from the following Lemma.
\begin{lemma}
For the particle system described in Section \ref{Sec:1}, we have, $\forall j$, $\forall f \in \mathcal{B}^{sym}_0(E^j)$,
$$
\mathbb{E}(((\eta^N_t)^{\odot j}(f))^2) \leq \frac{C}{N^j}\ ,
$$
for some constant $C$ (depending on $t$, $\Vert f \Vert_\infty$).
\end{lemma}
\begin{proof}
\begin{eqnarray*}
((\eta^N_t)^{\odot j}(f))^2 & = & \frac{1}{((N)_j)^2} \sum_{a,b \in \langle j,N \rangle } f(\xi^{a(1)}_t,\dots,\xi^{a(j)}_t)f(\xi^{b(1)}_t,\dots,\xi^{b(j)}_t) \\
& = & \frac{1}{((N)_j)^2} \sum_{k=0}^j \underset{ \# \text{Im}(a) \cap \text{Im}(b) = k}{\sum_{a,b \in \langle j,N \rangle }}f(\xi^{a(1)}_t,\dots,\xi^{a(j)}_t)f(\xi^{b(1)}_t,\dots,\xi^{b(j)}_t)\\
& = & \frac{1}{((N)_j)^2} \sum_{k=0}^j \sum_{a \in \langle 2j-k, N \rangle } C_j^k f(\xi_t^{a(1)},\dots,\xi_t^{a(j)}) \\
&& ~~~~~~~~~~~~~~~~~~~ \times f(\xi_t^{a(1)},\dots, \xi_t^{a(k)},\xi_t^{a(j+1)},\dots,\xi_t^{a(2j-k)}) \\
& = & \frac{(N)_{2j}}{((N)_j)^2} (\eta^N_{t})^{\odot 2j} (f\otimes f) \\
&&\quad + \sum_{k=1}^j \frac{C_j^k (N)_{2j-k}}{((N)_j)^2} \frac{1}{(N)_{2j-k}} \sum_{ a \in \langle 2j-k, N \rangle } f(\xi_t^{a(1)},\dots,\xi_t^{a(j)})\\
&& \qquad \qquad \qquad \qquad \times f(\xi_t^{a(1)},\dots, \xi_t^{a(k)},\xi_t^{a(j+1)},\dots,\xi_t^{a(2j-k)})
\end{eqnarray*}
Notice first that, because of the very definition of $(\eta^N_{t})^{\odot 2j}$ (as an average over arbitrary configurations of distinct systems of $2j$ particles), although $f\otimes f$ is not a symmetrical function, $(\eta^N_{t})^{\odot 2j} (f\otimes f)=(\eta^N_{t})^{\odot 2j}(Sym (f\otimes f))$, where $Sym$ stands for the canonical symmetrization map from ${\mathcal B}^{sym}_0(E^q)\otimes {\mathcal B}^{sym}_0(E^q)$ to ${\mathcal B}^{sym}_0(E^{2q})$.
Therefore, by Corollary \ref{Cor:FKWick}, $\exists C_0, \frac{(N)_{2j}}{((N)_j)^2} \mathbb{E}((\eta^N_{t})^{\odot 2j} (f\otimes f) ) \leq \frac{C_0}{N^j}$. For $k \in \{1,\dots,j\}$, we set $g_k(x_1,\dots,x_{2j-k})=f(x_1,\dots,x_j)f(x_1,\dots,x_k,x_{j+1},\dots,x_{2j-k})$. We then have
\begin{eqnarray*}
&&\frac{1}{(N)_{2j-k}} \sum_{ a \in \langle 2j-k, N \rangle } f(\xi_t^{a(1)},\dots,\xi_t^{a(j)}) \times f(\xi_t^{a(1)},\dots, \xi_t^{a(k)},\xi_t^{a(j+1)},\dots,\xi_t^{a(2j-k)}) \\
&& \quad = (\eta^N_t)^{\odot (2j-k)} (g_k) \ .
\end{eqnarray*}
This function $g_k$ has the particular feature that there are $\alpha=2j-2k$ indexes $i$ such that $\int_{E^{2j-k}} g_k(x_1,\dots,x_{2j-k}) \eta_t(dx_i) =0$ (namely $i=k+1,\dots,2j-k$).
As for the proof of Corollary \ref{Cor:FKWick}, we go back to the proof of Lemma \ref{crucial}. For the function $g_k$ defined above, the terms $\overline{D}_{k',r,r+i,j'}$ vanish for $j'+r < \frac{k'+\alpha}{2}$. So $j-k$ is the smallest power of $N$ appearing in the Laurent-type expansion of $\mathbb{E}((\eta^N_t)^{\odot (2j-k)} (g_k))$ and $\exists C_k, \mathbb{E}((\eta^N_t)^{\odot (2j-k)} (g_k)) \leq \frac{C_k}{N^{j-k}}$. This finishes the proof.
\end{proof}
\subsection{Wiener integrals expansions}
The convergence of symmetric statistics to multiple Wiener integrals has been studied intensively, see e.g. the seminal \cite{dynkin-mandelbaum-1983} or the account of the classical theory in \cite{lee-1990}.
We are interested here in proving that the main statements of the theory still hold for empirical U-statistics and focus on a particular example, namely the one of a symmetrical and bounded kernel constructed as a product of centered functions.
From the technical point of view, the key issue w.r. to the classical theory of limit distributions for U-statistics based on i.i.d. assumptions is related to the existence of interactions between the particles of the system. Taking these interactions into account amounts in practice to replace the central limit theorem by the analogous statement for interacting particle systems, namely Thm.~\ref{Theo:CLT}.
Let us consider a kernel (a symmetrical bounded fonction on $E^q$) $F$ of the form $F=(f_1\otimes ...\otimes f_q)_{sym}$ with $f_i\in {\mathcal B}_0(E)$. Then, $F^{(k)}=h^{(k)}=0$ for $k=1,...,q-1$, and we expect $N^{\frac{q}{2}}(\eta_t^N)^{\odot q}(F)$ to converge in law.
This is indeed the case.
Let us start with some classical results (see e.g. \cite{rota-wallstrom-1997} for a systematical approach by means of the poset of partitions).
For an empirical signed measure $m(x)=\sum\limits_{1\leq i\leq N}\delta_{x_i}$, we first have the formula of Rubin and Vitale (\cite{dynkin-mandelbaum-1983},\cite{lee-1990} p. 85):
$$m(x)^{\odot q}(F)=\sum\limits_{\cal P}\prod\limits_{V\in \cal P}(-1)^{|V|-1}(|V|-1)!m(x)(f_V)$$
where $f_V, V=\{v_1,...,v_k\}$, stands for $f_{v_1}...f_{v_k}$ and $\cal P$ runs over the set of partitions of $[q]$ into disjoint subsets.
Assume now that $m(x)=N\cdot \eta_t^N=:m_N$.
For $|V|=1$, $N^{-\frac{1}{2}}m_N(f_V)$ converges in law to ${\cal N}(0,\eta_t(f_V^2)+\mathbb{E} (W_t(f_V \otimes f_V))$. For $|V|\geq 2$, and since $\eta_t^N$ converges to $\eta_t$, $N^{-1}m_N(f_V)$ converges to $\eta_t(f_V)$.
Let us consider now a partition $\cal P$ of $[q]$ with $j_1$ sets with 1 element, $j_2$ sets with 2 elements, ..., $j_q$ sets with $q$ elements. We notice that $j_1+2j_2+2j_3+...+2j_p<j_1+2j_2+...+qj_q=q$ excepted if $j_3=...=j_q=0$ (and then the two terms are equal). In particular,
$N^{-\frac{q}{2}}\prod\limits_{V\in\cal P }(-1)^{|V|-1}(|V|-1)!m_N(f_V)$ converges in probability to $0$ if $j_3+...+j_q>0$.
By Slutsky's theorem
it follows that, if $N^{-\frac{q}{2}} m_N^{\odot q}(F)=N^{\frac{q}{2}} (\eta_t^N)^{\odot q}(F)$ converges in law, it converges to the same limit as $\sum\limits_{\overline{\cal P}}\prod\limits_{V\in \overline{\cal P}}(-1)^{|V|-1}(|V|-1)!m_N(f_V)$, where the sum is restricted now to partitions $\overline{\cal P}$ of $[q]$ with $j_3=...=j_q=0$.
The next theorem follows.
\begin{theorem}
With the above asumptions, we have the convergence in law:
$$N^{\frac{q}{2}} (\eta_t^N)^{\odot q}(F)\rightarrow \sum\limits_{k=0}^{\left\lceil \frac{q}{2}\right\rceil}(-1)^k
\sum\limits_{1\leq i_1<...<i_{q-2k}\leq q}I(f_{i_1})...I(f_{i_{q-2k}})\sum\limits_{J}\eta_t(f_{J_1})....\eta_t(f_{J_k}).$$
Here, $J$ runs over the partitions of $[q]-\{i_1,...,i_{q-2k}\}$ in ordered pairs $J_1=\{j_1,j_1'\},...,J_k=\{j_k,j_k'\}$ (where $j_i<j_i'$ and $j_1<...<j_k$).
Besides, $\left\lceil x\right\rceil$ stands for the integer part of a real number $x$ (the highest integer less or equal to $x$).
The $I(f_i), i=1,...,q$ are Wiener integrals of the functions $f_i$. They form a Gaussian family with $\mathbb{E}[I(f_i)]=0$ and $\mathbb{E}[I(f_i)I(f_j)]=\eta_t(f_if_j)+\mathbb{E}(W_t(f_i \otimes f_j))$.
\end{theorem}
\begin{corollary}
In the particular case $F=f^{\otimes q}$ with $\eta_t(f^2)=1$, we get the convergence in law:
$$N^{\frac{q}{2}} (\eta_t^N)^{\odot q}(F)\rightarrow H_k(I(f))$$
where $H_k$ stands for the $k$-th Hermite polynomial.
\end{corollary}
Indeed, in that case, since the number of partitions of $[q]$ with $q-2k$ singletons and $k$ pairs is
$\frac{q!}{2^k(q-2k)!k!}$, the Theorem simplifies to a convergence in law of $N^{-\frac{q}{2}} (\eta_t^N)^{\odot q}(F)$ to
$$\sum\limits_{k=0}^{\left\lceil \frac{q}{2}\right\rceil}(-1)^k\frac{q!}{2^k(q-2k)!k!}
I(f)^{q-2k},$$
where one recognizes the expansion of the $q$th (``probabilistic'') Hermite polynomial:
$$H_q(x)=\sum\limits_{k=0}^{\left\lceil \frac{q}{2}\right\rceil}(-1)^k\frac{q!}{2^k(q-2k)!k!}
x^{q-2k}.$$
\providecommand{\noopsort}[1]{}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
In the last decades, our understanding of the intricate phenomena in clusters of galaxies has advanced tremendously based on observations with the constantly improving generations of X-ray satellites. Clusters are now known to show complex morphologies and exciting physics on all scales and are thus among the most rewarding objects to target in the X-ray domain. Detailed spatial and spectral information about clusters can reveal important clues about the energetics of large-scale gas motions in the intracluster medium (ICM) and about the interaction between the ICM and the active galactic nucleus (AGN) in the central galaxy. Both AGN-ICM interactions in the form of shocks and bubbles \citep{Binney95,churazov2000,Boehringer02,Brueggen02,Birzan04,Forman05,Voit05,fabian2006,McNamara07} and gas "sloshing" in the cluster dark matter potential \citep{Markevitch01,Churazov03,markevitch2007} have been invoked, among other explanations, as possible ways to heat the gas in the centre of cool-core clusters, preventing the radiative cooling of the gas at high rates which should occur in the absence of heating but which is not observed \citep{Peterson01,Peterson03,Boehringer01}. Moreover, detailed spatial and spectral information about clusters is essential in constraining the chemical enrichment history of the ICM \citep{deplaa2007,Werner08rev,boehringer2009} and possible mechanisms responsible for the transport of heavy elements throughout the hot cluster medium \citep{Rebusco06,simionescu2008a,simionescu2008b}.
One of the best targets to perform in-depth studies of the ICM is M87 at the centre of the Virgo cluster. This is the nearest galaxy cluster ($\sim16$ Mpc), allowing us to resolve phenomena on small scales. M87 is among the brightest extragalactic X-ray sources in the sky, a guarantee for excellent spectral statistics and accurate temperature and metal abundance determinations within the technical capabilities of the detectors. M87 is among the best studied galaxies in the Universe and its central region has been the target of very deep observations with XMM-Newton, {\it Chandra}, and recently Suzaku. The hot gas atmosphere of M87 shows striking signs of AGN-ICM interaction, including an AGN-driven classical shock \citep{Forman05,Forman06,simionescu2007a} and clear enhancements in X-ray surface brightness associated with the radio lobes to the east and southwest of the core \citep{Feigelson87,bohringer1995,Belsole01,Young02,Forman05,Forman06}. Initial XMM-Newton observations showed that multi-temperature models including a cool gas component are needed to describe the spectra in these regions, also known as the E and SW X-ray arms \citep{Belsole01,Molendi02}. More recently, with deeper XMM-Newton data, a correlation was found between the percentage of cool gas and the metallicity in these regions, based on which \citet{simionescu2008a} concluded that the cool gas is metal-rich and that it was uplifted by the AGN \citep[following the scenario proposed by][]{Churazov01} from the central parts of the galaxy which had been enriched by stellar mass loss and supernova explosions. The energetics and metal-transport mechanisms associated with the AGN activity in the central parts of M87 have therefore been thoroughly investigated to date and have yielded important information about the physics of the AGN-ICM interaction. Here, we present the results of Suzaku and XMM-Newton observations extending the in-depth analysis to the outskirts of M87 to look for evidence for large-scale gas motions in the Virgo core and to bring gas sloshing and AGN-ICM interaction together in explaining the metal transport processes in the ICM.
We focus on a clearly observed surface brightness edge that lies 90 kpc (19$^\prime$) to the north of M87 and extends $140^{\circ}$ in azimuth in the archival ROSAT $\sim$30~ks image, its large scale making it an optimal target for Suzaku given the limitations related to its point-spread function. We show, using recent XMM-Newton and Suzaku data, that this edge is a ``cold front'', a contact discontinuity associated with gas sloshing in the ICM, typical of those found in many cool core clusters \citep{markevitch2007}.
\citet{simionescu2007a} have previously detected a weaker counterpart cold-front $\sim 33$ kpc to the south-east of the core. The opposite and staggered placement of these two features is a typical example of sloshing as seen in simulations \citep{Tittley05,Ascasibar06}. As the gas sloshes back and forth in the gravitational potential, cooler and denser parcels of gas from central parts of the cluster can move and come into contact with the hotter outskirts creating ``cold-front'' signatures, where the temperature on the denser side of the surface brightness edge is lower, which is the opposite of what one expects from a shock front \citep{Markevitch01}. Since heavy elements in cooling flow clusters are concentrated toward the centre \citep[e.g.][]{degrandi2001,Leccardi08}, the displaced parcels of cool central gas are also more metal-abundant. Thus, the abundance should be discontinuous across these fronts, as long as sloshing occurs within the region with a strong gradient. In general, sloshing should spread the heavy elements from the centre outwards, providing, beside the central AGN, a mechanism for distributing and transporting the metals produced in the central galaxy into the ICM.
In this paper, we discuss the abundance gradients for different heavy elements across the cold front to quantify the metal transport due to sloshing and to study the chemical enrichment histories of the gas on either side of the surface brightness edge.
\section{Data reduction}
\subsection{Suzaku}
On 2008 June 8, Suzaku observed two fields offset to the North by 19$^\prime$ and 33$^\prime$ with respect to the centre of M87. The effective exposure times were 59.7 ks for the pointing furthest away from the M87 centre (``M87Nfaint'') and 21.9 ks for the pointing with a smaller radial offset (``M87Nbright''). The distance between the Southern edge of the field-of-view of the M87Nfaint pointing and the surface brightness discontinuity was chosen to be $\sim 3^\prime$ (of the order of the Suzaku PSF) in order to avoid contamination from the bright side of the front into the faint-side spectra.
The data were reduced using the tools available in the HEAsoft package (version 6.7) to create a set of cleaned event lists with hot or flickering pixels removed and including only ASCA grades 0,2,3,4,6. The recommended filtering criteria were used to remove South Atlantic Anomaly (SAA) transits and times of low geomagnetic cut-off rigidity (COR), eliminating data with COR $\leq$ 6 GV.
In addition, all data taken within 20$^\circ$ of the sun-lit Earth limb or 5$^\circ$ of the dark Earth limb were excised to reduce contamination from scattered solar X-ray flux. Data obtained with 3x3 and 5x5 editing modes were merged into a single event file. The lightcurves for both data sets after the initial cleaning do not show anomalous count rates either in the full or in the soft energy bands. Ray-tracing simulations based on the available ROSAT data show that the expected contamination by out-of-field-of-view emission is below 1.5\% in both pointings. An exposure-corrected flux image of the two combined Suzaku pointings is shown in Fig. \ref{suzakuim} (no background subtraction has been performed for the image processing). The cold front is clearly visible in the image.
For the spectral analysis, we subtracted the non X-ray background (NXB) obtained with \texttt{xisnxbgen} and then modeled the cosmic X-ray background (CXB) with 3 components:
two thermal components to account for the local hot bubble (LHB) and for the Galactic halo (GH) emission, and a power-law to account for the integrated emission of unresolved point sources. The temperature of the LHB was frozen to 0.08~keV, that of the GH to 0.2~keV, and the power-law index to 1.41; unless otherwise noted, the spectrum normalisations were fixed based on the fluxes given by \citet{kuntz2000} for the thermal components and by \citet{deLuca04} for the power-law. We note that the region investigated is within approximately 0.2 of the virial radius of the Virgo cluster, thus the CXB level is still low compared to the emission from the ICM.
\begin{figure}
\includegraphics[width=\columnwidth, bb=49 36 564 757]{suzakumosaic.pdf}
\caption{Mosaic of the two Suzaku fields (exposure-corrected). The dashed line marks the approximate position of the outer cold front.}
\label{suzakuim}
\end{figure}
\subsection{XMM-Newton}
A series of 6 pointings with an average exposure time of 20~ks was performed with XMM-Newton in June 2008, with an additional seventh pointing of the same length completed on 2008 November 12. These seven pointings cover an annulus around the existing deep observations of the central part of M87 from 2000 June 19 and 2005 January 10, which had a combined effective exposure time of 120~ks. A combined exposure-corrected flux image of the MOS data from all the pointings is shown in Fig. \ref{extrreg}, together with the residuals obtained by dividing this image by a circularly symmetric beta model, with a best-fit core radius of $12.6^{\prime\prime}$ and $\beta=0.38$. Two surface brightness discontinuities, one at a radius of $\sim$33 kpc towards the SE and one at $\sim$90 kpc towards the NW are clearly seen.
For data reduction we used the 8.0.0 version of the XMM-Newton Science Analysis System (SAS). Out-of-time events were subtracted from the EPIC/pn data using the standard SAS prescription for the extended full frame mode.
To remove flaring from soft protons, we used a two-stage cleaning of the event files. First, we extracted a light-curve in the hard energy band (10-12 keV for MOS1 and MOS2 and 12-14 keV for pn) using 100 second time bins and excluded the time periods in each observation when the count rate exceeded the mean by more than 3$\sigma$. After this first cleaning, we used the resulting event files to extract a light-curve in a broad energy band (0.3-10 keV) with 10-s time bins and again excluded the time intervals with count rates larger than the average by more than 3$\sigma$.
We then examined the soft band (0.3--0.5 keV) images of each MOS observation to look for chips with anomalously high flux in this energy range \citep{snowden2007}, and found that none of our observations were affected by this instrumental problem.
For the background subtraction, we used a combination of blank-sky maps from which point sources have been excised \citep{ReadPonman} and closed-filter observations.
This is necessary because the instrumental background level of XMM-Newton is variable and increases with time. For each detector, we added to the corresponding blank sky background set a fraction of the closed filter data designed to compensate for the difference between the out of field of view (OoFoV) hard-band count rate in the observation and in the blank sky data \citep[for more details, see][]{simionescu2008b}.
\begin{figure*}
\includegraphics[width=\textwidth,bb=36 237 577 555]{xmm_pscorr.pdf}
\caption{XMM-MOS mosaic of M87. {\it Left:} Exposure corrected image divided by the best-fitting radially symmetric beta model. {\it Right:} Extraction regions used for the radial profiles in Sect. \ref{prof} and the two Suzaku fields overplotted on the exposure corrected image. Colorbar units are counts per second.}
\label{extrreg}
\end{figure*}
\section{Temperature and iron abundance profiles}\label{prof}
We first present radial profiles across and away from the northwestern cold front in order to establish the projected temperature and Fe abundance jumps associated with this feature and compare the different trends towards the NW and SE from the core.
For the XMM-Newton data, the spectral extraction regions used were 90$^\circ$-wide sectors of elliptical annuli towards the NW and SE, respectively. The thickness of these annuli increases logarithmically with distance from the M87 centre, in order to balance the decreasing surface brightness. The exact regions used are shown in Fig. \ref{extrreg}. For the Suzaku data, we chose elliptical annuli with the same ellipticity and position angle as for the XMM-Newton data. For illustration, the ellipse delimiting the two Suzaku annuli immediately outside and immediately inside the outer cold front is over-plotted in Fig. \ref{suzakuim}. No sector selection was performed since the opening angle of the area covered by Suzaku is already limited (see green and yellow squares overplotted in Fig. \ref{extrreg}) and all annuli were fixed to have a width of 3$^\prime$, due to constraints from the Suzaku PSF.
We use XSPEC 12.5.1 to model our spectra with a single-temperature plasma model in collisionally ionised equilibrium. Unless otherwise stated, we used the \texttt{apec} plasma model \citep{smith2001}. The determination of the Galactic absorption column density based on X-ray spectra can be easily biased by subtle underlying multi-temperature structure in the ICM \citep{simionescu2008b} or by uncertainties in the calibration of the oxygen edge around 0.5~keV and of Suzaku's contamination layer. We therefore opt to fix this parameter to $N_{\mathrm{H}}=2.0\times10^{20}$~cm$^{-2}$ \citep{Lieu96,kalberla2005}.
Point sources identified using the X-ray images were excluded from the spectral analysis for the XMM-Newton spectra. No unusually bright point sources are present within the field of view of Suzaku. The spectra obtained by different detectors of the same satellite were fit simultaneously with their relative normalisations left as free parameters. This is mainly done in order to account for effective area variations due to XMM-Newton chip gaps, the failed CCD6 on MOS1, and the different locations of the field of view edges for MOS and pn. The instrument normalisations of the three XIS detectors agree to within 5\% for each region analyzed.
The MOS, pn, and XIS1 data were fit in the 0.4--7.0 keV band while for XIS0 and XIS3 we used the 0.4--10 keV band. The spectra were binned with a minimum of 30 counts per bin before fitting.
The normalisation, temperature, and the abundances of O, Ne, Mg, Si, S, Ar, Ca, Fe, and Ni were left as free parameters. The abundances of the other elements heavier than He were fixed to 0.5 of the Solar value. For consistency with previous chemical enrichment studies of M87 with XMM-Newton by \citet{Matsushita03}, we adopt the solar abundances of \citet{Feldman92}.
Our spectral fitting results are shown in Fig. \ref{ktfeprof}. We find an excellent agreement between XMM-Newton and Suzaku in determining all the spectral parameters, despite the different background subtraction techniques employed.
The X-ray surface brightness profile, obtained by dividing the spectrum normalisation from XSPEC by the area of the corresponding extraction region, shows a very clear discontinuity at $19^\prime$ ($\sim$90~kpc), the radius of the outer cold front. At the radius of the inner cold front, the surface brightness profile towards the SE crosses from being systematically higher to being systematically lower than the corresponding profile towards the NW. The jump however is less pronounced, and is only evident once the strong radial trend is removed using a symmetric model, as was done in Fig. \ref{extrreg}.
The temperature profile towards the SE rises beyond the inner cold front at $\sim$33~kpc discussed by \citet{simionescu2007a} and then becomes flat, while the projected temperature to the NW stays flat or decreases slightly in the radial range beyond 30~kpc and only shows a sharp rise at the cold front at $19^\prime$ ($\sim$90~kpc).
Therefore, from the temperature profiles, we can conclude that the surface brightness discontinuities both at 33 and 90~kpc are cold fronts rather than shocks, exhibiting a lower temperature of the gas on the denser side. After increasing outside of the 90~kpc cold front, the temperature seems to decrease again beyond 30$^\prime$. This is however most likely a feature of the general Virgo temperature profile rather than being associated with the event that caused the cold front. A more detailed deprojection analysis will be presented by Forman et al. (in prep).
Typically two types of cold fronts have been discussed in the literature \citep[see][]{Owers09,markevitch2007} and two different mechanisms for producing them have been proposed. On one hand, cold fronts have been seen in merging clusters, where often the remnant core belonging to an infalling group or smaller cluster can still be identified. The best and most famous example here is the bullet cluster \citep{Markevitch02}. The cold front then exists at the interface between the cool, dense remnant core and the ICM of the cluster with which it is merging. This is clearly not the case for M87, which appears relatively relaxed apart from the X-ray arms and other small scale features in the very center (inner $\sim1^\prime$), all of which seem related to AGN-ICM interaction \citep[e.g.][]{Feigelson87,Matsushita02,Forman05,simionescu2008a}.
The most probable explanation for the opposite and radially staggered cold fronts in M87 thus is gas sloshing, similar to that proposed by \citet{Markevitch01} in A1795, where the gravitational potential of a cool core cluster can be perturbed by a past subcluster infall. Due to this disturbance, the central ICM is displaced and fronts are created where cooler and denser parcels of gas from the centre come into contact with the hotter outskirts. Numerical simulations show that the sloshing typically follows a spiral pattern, which can explain the placement of the two cold fronts in M87 at different radii on opposite sides from the centre \citep{Tittley05,Ascasibar06}. Using the mass profile of \citet{Matsushita02}, the Keplerian velocity is approximately 550 km/s for an orbit with a radius of 33~kpc and 850 km/s for a radius of 90~kpc. A rough estimate for the time needed for the two fronts to become out of phase by 180 degrees is thus $\pi/(\omega_{33}-\omega_{90})$ = 0.4 Gyr, where $\omega=v/r$ is the angular velocity at the corresponding radius. This estimate is likely to be low.
An interesting topic for further simulations in the case of M87 will be whether the large scale bulk motions due to ICM sloshing are strong enough to cause the observed bending of the AGN-inflated radio lobes and of the associated E and SW X-ray arms.
Apart from the marked increase in temperature beyond the surface brightness discontinuities at 33 and 90~kpc, another common feature is a sharp drop in Fe abundance outside each cold front. A discontinuity in the metal abundance profile is expected because the Fe abundance in Virgo, as in the vast majority of cooling core clusters, is centrally peaked \citep[Fig. \ref{ktfeprof}; see also e.g.][]{Leccardi08}. This means that the central gas displaced by sloshing is also richer in metals and thus sloshing contributes to transporting the heavy elements from the centre outwards.
To quantify this, we calculate the excess mass of Fe towards the SE behind the inner cold front (in an opening angle of $90^\circ$ between $\sim$14 and 33~kpc) and the excess mass of Fe towards the NW behind the outer cold front (in an opening angle of $90^\circ$ between $\sim$33 and 90 kpc). This is done simply by estimating the gas mass in each annulus based on the spectrum normalisation, multiplying by the Fe mass fraction determined from the Fe abundance, subtracting the corresponding value for the opposite sector, and finally summing over all annuli in the cited radial ranges. Note that this is likely an underestimate of the mass of Fe transported by the outer cold front. This is firstly because this front extends over an opening angle larger than $90^\circ$ (although it is less pronounced outside the region considered here). Secondly, the low Fe abundance outside the outer cold front suggests that the initial Fe profile to the NW before the metal transport was lower than estimated from taking the SE profile at corresponding radii.
We obtain an excess Fe mass of $4.5\times10^6$ M$_\odot$ behind the inner cold front and $13.7\times10^6$ M$_\odot$ behind the outer cold front.
This represents 31\% and 22\% of the average Fe mass per sector within the corresponding radial range for the inner (33~kpc) and outer (90~kpc) fronts, respectively. If the Fe mass averaged between the NW and SE sectors is representative also for each of the other two sectors towards the NE and SW, then the mass of Fe transported by sloshing represents $\sim$6--8\% of the total Fe mass within the considered radial ranges.
Alternatively, if we relate the mass of Fe transported by the two cold fronts to the total Fe masses between 14--33 and 33--90~kpc obtained using the cumulative Fe mass profile of M87 from \citep{bohringer2004}, we obtain very similar ratios of $\sim$6--9\%.
It is noteworthy that both the large-scale 90~kpc cold front and the inner cold front seem to transport a very similar fraction of the mass of Fe available within the radial region affected by the sloshing.
By comparison, the Fe mass uplifted by the AGN in the bright X-ray arms is only $1.5\times10^6$ M$_\odot$ \citep{simionescu2008a}. Note that this is only an estimate of the current amount of metals being displaced by either process - depending on the strength of the mixing within the ICM, a fraction of these metals may eventually fall back to their original position within the cluster.
While it was expected that the sloshing that causes cold fronts is an important means of transporting metals in the ICM and that this would cause the Fe abundance to be discontinuous at the surface brightness jump towards the NW, the strength of this feature is surprising. Immediately outside the cold front at 90 kpc, we reach a very low, flat abundance plateau. The average abundance here is $0.25\pm0.02$ solar (see also next section) relative to the solar units of \citet{Feldman92}, which translates to only $0.17\pm0.01$ solar in the units of \citet{Angr}. This is as low as the expected metallicity close to the virial radius in other clusters \citep[0.2 solar,][]{fujita2007}. Such low Fe abundances are rarely measured in the ICM. For a sample of nearby galaxy clusters observed with XMM-Newton, \citet{Leccardi08} find that the average metallicity profile keeps decreasing out to very large radii, reaching $\sim$0.2 solar only at around 0.3--0.4 r$_{180}$. A similar trend is shown by the Fe abundance profiles measured with Suzaku \citep[e.g.][]{Sato09}.
For an average temperature of 2.5~keV, the virial radius of the Virgo cluster based on the scaling relations of e.g. \cite{Arnaud05} is $r_{200}\sim$ 1.2~Mpc, which makes it very surprising to see gas with such a low metallicity outside a radius of only 90~kpc.
Towards the SE, the metallicities at corresponding radii beyond the outer cold front are systematically higher, closer to what one would expect at these relatively small distances from the cluster centre. The likely scenario is that sloshing in the more distant past has influenced the abundances of the gas towards the SE and, because Virgo is a relatively young and unrelaxed object, we are only now witnessing the dispersion of higher metallicity gas into the pristine region towards the NW, whose abundance is that typical of the cluster outskirts.
\begin{figure}
\includegraphics[width=\columnwidth,bb=40 144 380 690]{prof_Sx_kT_Fe_Suz_XMM_apec2.pdf}
\caption{X-ray surface brightness, temperature and Fe abundance profiles towards the NW and SE as a function of the average semi-major axis (SMA) of the elliptical annuli used as extraction regions. The positions of the two cold fronts at 33 and 90~kpc are marked with vertical dotted and dashed lines, respectively.}
\label{ktfeprof}
\end{figure}
\section{Metal abundance ratios and supernova enrichment history}
It is believed that the peak in Fe abundance associated with cool core clusters is produced by type Ia supernova ejecta in the brightest cluster galaxy (BCG) which is spatially coincident with this peak \citep{degrandi2004,bohringer2004}. While SN~Ia have a longer delay time and still occur in these BCGs, which have typically old stellar populations, core-collapse supernovae (SN$_{\rm CC}$) usually explode only within a short time after a star formation event and would thus not be expected to enhance abundances significantly in cool cores. Elements which are predominantly produced by SN$_{\rm CC}$, such as O, Ne and Mg, or even intermediate elements produced by both SN$_{\rm CC}$ and SN~Ia such as Si and S, are therefore expected to be less abundant compared to Fe in the vicinity of the BCG. Whether or not such trends in the e.g. O/Fe or Si/Fe ratios are in fact observed is, however, a matter of debate in recent literature. \citet{finoguenov2000} show an increase of Si/Fe with radius in a sample of clusters observed with ASCA. While the presence of such a gradient was confirmed, more recently, for the case of galaxy groups \citep{Rasmussen09}, the {\it constant} Si/Fe and increasing O/Fe between the centre and outskirts of M87 prompted \citet{finoguenov2002} to propose a diversity of SN~Ia: faint SN~Ia with less complete burning and higher Si/Fe yields would contribute to the enrichment of the central parts of the ICM, while brighter SN~Ia with lower Si/Fe yields would dominate the enrichment in the outskirts. A strong increase in O/Fe with radius claimed by \citet{tamura2004} has not been confirmed by recent analyses of deeper cluster observations, which show no significant increase in O/Fe within 0.1 r$_{200}$ \citep{simionescu2008b}. The latest Suzaku results also indicate constant Mg/Fe, Si/Fe, and S/Fe abundance ratios with radius and only a very weak, low-significance increase in O/Fe out to as far as 0.4 r$_{180}$ for several objects which have been studied in detail \citep{sato2007b,Sato09}.
Trends in abundance ratios across cold fronts are an important missing piece of the puzzle in this ongoing debate. This is because sloshing effectively brings into contact gas which has been near the BCG and ``pristine'' gas at large radii, which has not been influenced (or has been influenced much less) by the BCG at late times. If the BCG has a large impact on the metal budget of the gas in its vicinity, this should be reflected in marked differences between the chemical enrichment patterns of the gas on either side of the cold front.
To test this, we focus here on the northwestern cold front at 90~kpc, for which the good energy resolution of the Suzaku observations allows more precise determinations of the metal abundances than possible with other data sets \citep[see previous work by e. g. ][for the case of A496 with {\it Chandra}]{Dupke03,Dupke07}.
Since precise determinations of the metal abundances other than Fe require significantly more counts, we must substantially increase the sizes of the extraction regions. We thus can only determine one value of the O and Mg abundances for each side of the cold front, rather than being able to produce a radial profile. For this, we fitted in parallel the Suzaku spectra extracted from the first 3 and last 6 ellipses used to create the radial profiles in Fig. \ref{ktfeprof} to determine the metal abundances on the bright and faint side of the cold front, respectively. The temperature and normalisation of each ellipse spectrum was free in the fit, while all elemental abundances were coupled between the data sets in each case. This ensures that we account for variations in temperature across the large regions sampled by the spectra, which would otherwise bias the abundance measurements. Note that allowing the Fe abundance also to vary between the different elliptical extraction regions does not change the measured abundances for the other elements. The results are shown in Table \ref{abund}.
\begin{table}
\caption{Metal abundances on the bright and faint sides of the cold front.
The Suzaku spectra extracted from the first 3 and last 6 ellipses used to create the radial profiles were fit in parallel with the temperature and normalisations left free (3T/6T model) in order to determine the metal abundances on the bright and faint side of the cold front, respectively. The errors are given at the $\Delta \chi^2$=1 confidence level.}
\begin{center}
\begin{tabular}{l|cc}
\hline
\hline
Parameter & bright side 3T & faint side 6T \\
\hline
O & $0.30\pm0.07$ & $0.17\pm0.07$ \\
Ne & $0.42\pm0.07$ & $0.31\pm0.08$ \\
Mg & $0.34\pm0.07$ & $0.22\pm0.07$ \\
Si & $0.36\pm0.04$ & $0.16\pm0.04$ \\
S & $0.39\pm0.05$ & $0.23\pm0.05$\\
Ar & $0.55\pm0.10$ & $0.45\pm0.12$\\
Ca & $0.63\pm0.16$ & $0.60\pm0.18$ \\
Fe & $0.41\pm0.02$ & $0.25\pm0.02$\\
Ni & $0.35\pm0.14$ & $0.28\pm0.16$ \\
\hline
$\chi^2$/d.o.f. & 2730/2696 & 3848/3678 \\
\hline
\end{tabular}
\end{center}
\label{abund}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth, bb= 18 144 592 718]{omgfe.pdf}
\caption{O/Fe and Mg/Fe abundance profiles. In black, values determined for several central annuli by \citet{Matsushita03} with XMM-Newton. In blue, the results for the bright side and in red, for the faint side of the cold front.}
\label{omgfeprof}
\end{figure}
We caution that the abundance of any element produced predominantly by SN$_{\rm CC}$ is unfortunately very difficult to determine from the X-ray spectra. The Ne X line is blended with the strong Fe-L complex between 0.9--1.2 keV, which is very sensitive to the exact modeling of the temperature structure of the emitting gas. The O determination is sensitive to the normalisation of the 0.2~keV galactic halo component, which is difficult to determine exactly because the soft X-ray background around M87 is patchy. Raising or lowering this normalisation by a factor of 2 changes the best-fitting O abundance on the faint side of the cold front by as much as 40\%, while Mg changes only by 14\% and the temperature by less than 0.02~keV. The Mg line is also blended with Fe-L emission and, although the Fe-L line emission around the Mg line is much weaker than that surrounding the Ne line, different spectral codes disagree on its modeling. The current collisional ionization equilibrium (CIE) plasma model implemented in SPEX \citep{kaastra1996} gives Mg abundances lower by a factor of 2 compared to the \texttt{vapec} model, which provides a better fit to the Mg/Fe-L line blend \citep[e.g.][]{Matsushita03}, motivating our choice of the latter plasma model for the present work. An improved version of SPEX containing an updated Fe-L line library, which is yet to be released, provides better agreement with \texttt{vapec} (J. de Plaa, private communication; see Table \ref{mekalapec}). The temperature structure moreover plays an important role in determining the elemental abundances, including Mg.
If the metal abundances are obtained by fitting the integrated spectra of the regions in front of and beyond the cold front with single temperature models, there is an indication of a relative increase in Mg/Fe on the faint side compared to the bright side of the surface brightness jump, despite the systematic offsets in determining the absolute Mg/Fe between the different plasma models. However, accounting for the temperature structure by allowing the temperature to vary across several rings with fixed abundances shows that this trend is only an artefact.
\begin{table}
\caption{Systematics in determining the Mg/Fe ratio.
The 3T/6T model refers to fitting the first 3 and last 6 ellipses used to create the radial profiles in parallel with the temperature and normalisations left free.
}
\begin{center}
\begin{tabular}{l|cc}
\hline
\hline
Mg/Fe & bright side & faint side \\
\hline
vapec 3T/6T & $0.81\pm0.17$ & $0.88\pm0.29$ \\
vapec 1T & $0.77\pm0.15$ & $1.14\pm0.32$ \\
spex 1T & $0.42\pm0.15$ & $0.68\pm0.32$ \\
new spex 1T & $0.97\pm0.18$ & $1.28\pm0.40$\\
\hline
\end{tabular}
\end{center}
\label{mekalapec}
\end{table}
The results presented in Table \ref{abund} show a decrease both in the Fe abundance and in the abundances of all elements predominantly produced by SN$_{\mathrm{CC}}$ across the cold front. Although the individual significances of the changes in the O, Ne and Mg abundances are small due to the large statistical uncertainties, the fact that all three abundances decrease by a similar relative amount (26--43\%), which is approximately the same as the relative change in the Fe abundance (40\%), is a convincing indication that the measured differences are reliable. Adding in quadrature the significances of the variations in O, Ne and Mg across the cold front suggests a change in the abundance of SN$_{\mathrm{CC}}$ products at the 90\% confidence level. This change requires the presence of a centrally peaked distribution not only for Fe but also for O, Ne, and Mg, and rules out models in which SN$_{\mathrm{CC}}$ products are uniformly distributed throughout the ICM, while only SN~Ia products show a radial gradient.
Because all elemental abundances decrease by very similar relative amounts, we find no significant change in either Mg/Fe, Ne/Fe or O/Fe across the cold front.
Comparing the Mg/Fe and O/Fe ratios on either side of the cold front with those measured at smaller radii by \citet{Matsushita03} using XMM-Newton data (also using the \texttt{vapec} model), we find a very slow increase across the entire radial range.
The data are consistent with no increase in Mg/Fe from the very centre of the galaxy out to roughly 210 kpc (45$^\prime$), as shown in Fig. \ref{omgfeprof}.
\citet{Doherty09} show that the stellar halo of M87 truncates at an average radius of $\sim$150 kpc, therefore we are clearly probing the chemical enrichment history in regions both inside and beyond the sphere of influence of the galaxy. The very low Fe abundance outside the cold front (beyond $20^\prime$ in Fig. \ref{ktfeprof}) also strongly suggests that we are probing the chemical enrichment of the abundance floor at which the ICM metallicity is expected to flatten out at large radii.
The roughly constant Mg/Fe and O/Fe abundance ratios indicate similar chemical compositions and, consequently, enrichment histories in all these regions, although within the large statistical errors we cannot rule out a recent influence of SN~Ia in M87 on the metallicity budget of the Virgo ICM.
Assuming the core-collapse model yields of \citet{nomoto2006} with solar metallicity of the progenitor,
the best-fitting Mg/Fe ratio in the centre of M87 (based on the average of the three data points from \citealt{Matsushita03}) corresponds to 27\% of all supernova explosions contributing to the chemical enrichment being SN~Ia, while the best-fitting Mg/Fe ratio on the faint side of the cold front suggests a relative contribution of $N_{\rm SN~Ia}/(N_{\rm SN~Ia}+N_{\rm SNcc}) \sim$20\% (independent of the assumed SN~Ia yield model). An enrichment of the gas on the faint side of the cold front (beyond 90~kpc) purely by SN$_{\mathrm{CC}}$ without any contributions by SN~Ia would imply a Mg/Fe ratio of 2.6 solar and is ruled out at the 5.8$\sigma$ level.
The total mass of Fe contained within a 90~kpc radius from the centre of M87 is approximately $2.3\times10^8$ M$_\odot$ \citep{bohringer2004}. If 27\% of all supernovae in the core region are SN~Ia (based on the Mg/Fe ratio discussed above) and each SN~Ia produces $\sim$0.8 M$_\odot$ of Fe \citep{iwamoto1999}, while the rest are core collapse supernovae \citep[which, for a solar metallicity of the progenitor and assuming a Salpeter IMF, produce approximately 0.09 M$_\odot$ of Fe per explosion,][]{nomoto2006}, then $2.2\times10^8$ SN~Ia and $6\times10^8$ SN$_{\mathrm{CC}}$ are needed to produce all the metals observed in this region.
For a Salpeter initial mass function (IMF) between 0.1 and 100 M$_\odot$, approximately one star with $M>8$ M$_\odot$ which will explode as a SN$_{\mathrm{CC}}$ is expected per 100 solar masses of stars formed. For a B-band mass to light ratio of around 8 \citep{Kronawitter00} and a blue-band luminosity of $10^{11}L_{B\odot}$ corresponding to the absolute B magnitude of M87 of -22.14 cited by \citet{Peletier90}, the stellar mass of M87 is $8\times10^{11}$ M$_\odot$ and $8\times10^9$ SN$_{\mathrm{CC}}$ would have exploded as the galaxy was forming, 13 times more than the $6\times10^8$ SN$_{\mathrm{CC}}$ needed to produce the metals in the central 90~kpc from the M87 centre. The central Mg peak can thus be explained if only less than 10\% of the SN$_{\mathrm{CC}}$ products from the galaxy formation phase are retained rather than being uniformly distributed into a flat abundance profile.
The expected present SN~Ia rates in elliptical galaxies are still relatively uncertain, from $0.18\pm0.06$ SNU \citep{Cappellaro99} to $0.36^{+0.22}_{-0.14}\pm0.02$ SNU \citep{Sharon07} or $0.28^{+0.11}_{-0.08}$ SNU \citep{Mannucci08}, where 1 SNU is equal to one supernova explosion per century per $10^{10}L_{B\odot}$ of the galaxy.
Since redshift z=1, corresponding to a look-back time of 7.7 Gyr, between $10^8$ and $4.6\times10^8$ SN~Ia could have enriched the centre of M87 within the current uncertainties, between half to twice the amount needed in total. It is therefore possible that a fraction of the SN~Ia products in the central peak stem from the proto-cluster phase rather than recent SN~Ia in M87, although that requires the SN~Ia rates to be on the lower end of the measured confidence intervals and not to have been larger in the past, as predicted by e.g. \citet{Renzini93}; using 0.28 SNU, the most recent measurement of \citet{Mannucci08}, exactly the required number of SN~Ia, $2.2\times10^8$, would explode since z=1, thus all the Fe would be produced only by the central galaxy at late times. Another possibility to explain the constant Mg/Fe ratio in this case is that some of the Fe outside the metal peak was produced recently by SN~Ia explosions of intra-cluster stars. For a diffuse intra-cluster light with an average apparent magnitude of 27 mag/arcsec$^2$ \citep{Mihos05,Kormendy09}, 30\% of the Fe outside the 90~kpc cold front and within the radius covered by the XMM-Newton mosaic could have been produced by SN~Ia since z=1, for a SN~Ia rate of 0.28 SNU.
We plot in Fig. \ref{femgs} the measured Fe/S vs. Fe/Mg ratios for the M87 centre \citep{Matsushita03} and on either side of the observed cold front using our measured values. The Fe/Mg, as discussed above, is an important diagnostic of the relative contribution of SN~Ia to SN$_{\mathrm{CC}}$. Using this ratio and assuming the same core-collapse model yields as above, we can determine the amount of Si-group elements produced by the SN$_{\mathrm{CC}}$ and consequently the ratio of Si-group elements to Fe produced by SN~Ia, which is an important diagnostic for the explosion mechanism of the latter. We overplot in Fig. \ref{femgs} three different abundance ratio models based on the deflagration (W7) and delayed detonation (WDD1, WDD2) SN~Ia yields of \citet{iwamoto1999}. We used S as a representative element for the Si-group because the Si abundance is more easily biased by calibration issues around the Si edge and by modeling the underlying Si instrumental line. We find marginal evidence supporting the scenario proposed by \citet{finoguenov2002}, in that the data points at larger radii favor a W7 model whereas the abundance ratios at the centre of M87 favor the WDD1/WDD2 SN~Ia yields.
\begin{figure}
\includegraphics[width=\columnwidth,bb=18 144 592 718]{snplot_apec.pdf}
\caption{Fe/S vs. Fe/Mg abundance ratios on the two different sides of the cold front (this work) and in the centre of M87 \citep{Matsushita03}. Overplotted are supernova enrichment models obtained by combining various relative numbers of SN~Ia \citep[using the WDD1, WDD2 and W7 yield models of ][]{iwamoto1999} and SN$_{\mathrm{CC}}$ \citep[whose yields are modeled based on][assuming a solar metallicity of the progenitor and averaged over the Salpeter IMF]{nomoto2006}.}
\label{femgs}
\end{figure}
\section{Conclusions}
Two cold fronts are observed in M87: one towards the SE at a radius of $\sim$33~kpc, and one towards the NW with a radius of $\sim90$~kpc. The opposite and staggered placement of these cold fronts, as well as the absence of a remnant subcluster, make sloshing the most viable explanation for the presence of these cold fronts. The ICM distribution in and around M87 is very similar to results obtained from numerical simulations of sloshing in cool core clusters \citep{Tittley05,Ascasibar06}. For the mass profile of M87, the time needed for two fronts at these radii to become out of phase by 180 degrees is approximately 0.4 Gyr.
We present accurate projected temperature profiles which show that the temperature is higher on the X-ray fainter side of each cold front, opposite from what would be expected in the case of shock fronts. We also show that the metallicity drops sharply outside each cold front.
Such a discontinuity in the metal abundance profile is expected because sloshing displaces the central, cooler, more metal-rich gas and brings it into contact with hotter, more metal-poor ICM in the outskirts.
In this way, sloshing also contributes to transporting the heavy elements from the centre outwards.
We estimate that the inner cold front transports $4.5\times10^6$ M$_\odot$ of Fe, while for the outer cold front the total mass of Fe transported amounts to $13.7\times10^6$ M$_\odot$. These masses represent in each case the same fraction ($\sim$6--8\%) of the total Fe mass available within the radial range affected by sloshing, which is much larger than the amount of Fe transported by AGN-related processes, amounting to only $1.5\times10^6$ M$_\odot$ of Fe.
We find no strong indications for a change in the Mg/Fe and O/Fe ratios across a large radial range including the outer cold front, given a proper modeling of the temperature distribution within the extraction regions.
This requires the presence of a centrally peaked distribution not only for Fe but also for O, Ne, and Mg, and rules out models in which SN$_{\mathrm{CC}}$ products are uniformly distributed throughout the ICM, while only SN~Ia products show a radial gradient.
Since stellar mass loss in the central galaxy typically contributes only small amounts of heavy elements to the ICM metal budget,
the peaks in the distribution of SN$_{\mathrm{CC}}$ products are likely to stem from star formation, triggered as the BCG assembled during the protocluster phase, as proposed by \citet{simionescu2008b}. The central Mg peak can be explained if only less than 10\% of the SN$_{\mathrm{CC}}$ products from the galaxy formation phase are retained rather than being uniformly distributed into a flat abundance profile.
The very low Fe abundance of only $0.25\pm0.02$ solar \citep[$0.17\pm0.01$ relative to][]{Angr} suggests that the gas outside the cold front is representative of the pristine outskirts of typical clusters. This means that we are currently observing the very mechanism which spreads the metals into these pristine regions, broadening the metal distribution to that usually seen in relaxed clusters. Moreover, we are able to determine the chemical composition of this pristine gas, and find that probably around 20\% of all supernovae contributing to its enrichment are SN~Ia, that the metal abundance ratios of this gas are very similar to those near the centre of the BCG, and that an enrichment by SN$_{\mathrm{CC}}$ products only can be excluded with a very high significance.
\section*{Acknowledgments}
The authors would like to thank J. de Plaa, A. Finoguenov, J. S. Kaastra, E. Roediger, M. Br\"uggen, C. Jones, A. Baldi and H. Matsumoto for helpful discussion.
Support for this work was provided by NASA through Einstein Postdoctoral Fellowship grants number PF9-00070 and PF8-90056 awarded by the Chandra X-ray Center, which is operated by the Smithsonian Astrophysical Observatory for NASA under contract NAS8-03060. We acknowledge NASA grant NNX08AZ88G.
AS is grateful for the hospitality of the Harvard-Smithsonian CfA. This work is based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA member states and the USA (NASA). The authors thank the Suzaku operation team and Guest Observer Facility, supported by JAXA and NASA. HB acknowledges support by the DfG Schwerpunkt Programm SPP 1177 and the Transregio Program TR33, the Dark Universe.
\bibliographystyle{mn2e}
|
\section{Introduction}
\label{sec:introduction}
Many biochemical processes can successfully be described by dynamical
systems allowing some form of switching, where, depending on, for
example, initial conditions, solutions of the dynamical system
end up in different regions of state space (associated with different
biochemical functions). Often dynamical systems admitting
bistability (i.e.\ the existence of two stable steady states) are used
for this purpose. There is a long tradition of establishing
bistability, both experimentally and computationally, in areas ranging
from signal transduction (see e.g.\ \cite{sig-023}) to cell cycle (see
e.g.\ \cite{cyc-005}).
From our point of view, however, bistability is too strong a
requirement, as already a saddle type steady state with just one
algebraically simple positive eigenvalue and all other eigenvalues
having negative real part gives rise to the desired switching
behaviour (with the global stable manifold of the saddle as a
switching surface, see \cite[Remark~3.2]{cc-flo-003}). The approach
presented here tries to directly establish such points and is hence
capable of establishing switching that is not necessarily associated
to bistability. Therefore we expect this approach to be of particular
interest for researchers working in Systems Biology and other areas of
Quantitative Biology.
In many applications, bistability of a dynamical
system has been established numerically using bifurcation analysis or
simulations that can become arduous tasks even for relatively small
systems. Moreover parameter uncertainty is a predominant issue in
Systems Biology: the dynamical systems consist of a large number of
states and parameters, while measurement data are often very noisy and
data points and repetitions are usually few. Hence techniques allowing
the direct analytic computation of parameter vectors where a given
system exhibits switching are desirable.
In developing these techniques we identified two promising approaches:
(i) establishing multiple steady states as a mechanism for possible
switching and bistability
\cite{cc-flo-multi-002,fein-024,fein-025,cc-flo-003} and (ii)
establishing points where the dynamical system undergoes a saddle-node
bifurcation so that the global stable manifold of the saddle is acting
as a switching surface. The first approach is motivated by the so-called
Chemical Reaction Network Theory developed by Feinberg and co-workers
(see \cite{fein-013,fein-016,fein-017} and
\cite{fein-025,cc-flo-003,alg-006}). The second approach is based on
the structure of the Jacobian of a mass-action network
\cite{fein-032}. This approach was successful for a
double-phosphorylation mechanism where the nullspace of the Jacobian
admits a very special representation (cf.\, \cite{fein-032}).
Here we extend these ideas to mass action networks in general
(involving at least one conservation relation) by making use of a
property that is frequently observed in dynamical systems
originating in Systems Biology: one often faces dynamical systems
that involve so-called \emph{conservation relations} confining
trajectories to affine linear subspaces of state space.
Thus, the Jacobian of such a system evaluated at an arbitrary
point in state space has at least as many zero eigenvalues as there
are conservation relations. Consequently, for a saddle-node to occur
at a particular point in state space, the Jacobian has to have an
\textit{additional} zero eigenvalue at that point. Generically,
mass-action systems undergo a bifurcation at that point -- one can
state conditions guaranteeing a saddle-node bifurcation
(cf. Section~\ref{sec:saddle-node-mass-action} or
\cite{fein-032}). One can expect that such sufficient conditions for a
saddle-node bifurcation can be established for mass action networks
originating in Systems Biology since there are many parameters which
can be chosen as continuation parameters. Hence, such an additional
zero eigenvalue frequently entails a saddle-node bifurcation and thus
switching in a mass action network.
The main result of our paper are sufficient conditions
guaranteeing such an additional zero eigenvalue that take the form
of linear inequality systems and are thus easy to check. Moreover, our
result is constructive in the sense that the
solutions to one of the inequality systems determine a state and
parameter vector where the Jacobian has an additional zero
eigenvalue and thus fulfills the necessary degeneracy condition for a
saddle-node. Infeasibility of all inequality systems does not exclude
additional zero eigenvalues, hence feasibility of at least one
inequality system is a sufficient condition for an additional
eigenvalue. In case the remaining eigenvalues of the linearization
have negative real parts such feasibility is generically sufficient for
a saddle-node and the associated bifurcation into a saddle and a
node. We verify this splitting in our case studies by computing
bifurcation diagrams.
Finally we'd like to point out that our results are in a certain sense
complementary to those obtained in \cite{fein-008},
\cite{fein-009,fein-019,fein-020}, \cite{fein-021},
\cite{inj-003,inj-002,inj-001}: all these references present sufficient
conditions for the global injectivity of a dynamical system defined by
a biochemical reaction network (not necessarily restricted to mass
action systems). In particular, these conditions exclude
switching. More along the line of our work is the approach of Mincheva
and coworkers \cite{fein-044,fein-028}. There the tight connection
between the characteristic polynomial of the Jacobian and the cycles
of certain graphs associated to the Jacobian are exploited to derive
conditions for certain instabilities (e.g.\ saddle-node or Hopf
bifurcations). The major difference to our work is that we are not
working with the characteristic polynomial but rather exploit the
fact (reported in \cite{fein-032}) that Jordan blocks of size
$\geq 2$ imply additional zero eigenvalues (and thus candidates for,
for example, saddle-node bifurcations).
In the following we briefly describe the organization of the paper and
at the same time offer the conclusions that can be drawn. In
Section~\ref{sec:dyn-sys-mass-action} we describe dynamical systems
defined by mass action networks, recall some results from
\cite{fein-032} and \textit{characterize} positive state vectors where
the Jacobian has such an additional and thus defective zero eigenvalue
(Lemma~\ref{lem:zero_eigen_defective}, Theorem~\ref{theo:nasc}).
Those state vectors arise from elements of a semialgebraic set
that contains only polynomials of degree two or less -- regardless
of the exponents in the polynomial ODE system defined by a mass
action network. In Section~\ref{sec:suff-lin-ineq}, based on a
result from Qualitative Matrix Theory ensuring the existence of
positive null vectors, we present a sufficient condition allowing
the computation of elements of that semialgebraic set that takes the
form of linear inequality systems. The solvability of these inequality
systems is then sufficient for the existence of an additional zero
eigenvalue (Theorem~\ref{lem:main}). In Section~\ref{sec:application},
finally, we demonstrate the applicability of the results presented
here by analyzing as a proof of principle two competing mass action
networks describing the G1/S transition in the cell cycle of budding
yeast. These networks were originally presented in \cite{fein-025} and
\cite{cc-flo-003} where their investigation was based on subnetwork
analysis. Both networks are not accessible by the results of
\cite{fein-032}.
For the convenience of the reader we provide some additional
information in four appendices. In
Appendix~\ref{sec:saddle-node-mass-action} we recall some remarks
concerning saddle-node bifurcations in mass-action networks that
were made earlier in \cite{fein-032}, in Appendix~\ref{sec:data-tern}
and Appendix~\ref{sec:data-bin} we collect the relevant
structural information of the G1/S transition networks discussed in
Section~\ref{sec:application}.
The final Appendix \ref{sec:imequal}, using basic linear algebra,
discusses some of the assumptions and results in the present work.
\section{Dynamical systems defined by mass action systems}
\label{sec:dyn-sys-mass-action}
\mbox{}
To introduce the notation, we use the network depicted in
equation~(\ref{eq:exa_net}) below. This network is analysed in
\cite{fein-013}, where multiple steady states are established.
\begin{equation}
\label{eq:exa_net}
\begin{split}
&\xymatrix{
A+2\, S \ar @<.4ex> @{-^>}[r] ^{\bf k_1} & A\, S_2 \ar @{-^>}[l]^{\bf k_2}
} \\
&\xymatrix{
B+S \ar @<.4ex> @{-^>}[r] ^{\bf k_3} & B\, S \ar @{-^>}[l]^{\bf k_4}
} \\
&\xymatrix{
A\, S_2 + B\, S \ar [r] ^{\bf k_5} & C+3\, S
} \\
&\xymatrix{
A \ar @<.4ex> @{-^>}[r] ^{\bf k_6} & 0 \ar @{-^>}[l]^{\bf k_7}
\ar @<.4ex> @{-^>}[r] ^{\bf k_8} & B \ar @{-^>}[l]^{\bf k_9} \\
& C \ar [u] ^{\bf k_{10}}
}
\end{split}
\end{equation}
Network (\ref{eq:exa_net}) consists of $n$ \emph{species} ($n=6$)
and with each species we associate a variable $x_i$ representing its
concentration and the corresponding unit vector $e_i$ of $\ensuremath{I\!\!R}^6$: $x_1$ and
$e_1$ with $A$, $x_2$ and $e_2$ with $B$, $x_3$ and $e_3$ with
$S$, $x_4$ and $e_4$ with $A\, S_2$, $x_5$ and $e_5$ with $B\, S$
and $x_6$ and $e_6$ with $C$.
The nodes of the network graph are called \emph{complexes} and with
each complex we associate the sum of its constituent species. The
above network contains $m$
complexes ($m=10$): The complex $0$ will be denoted by the zero vector
$0\in \ensuremath{I\!\!R}^6$ and is used to encode that the system is open with respect
to $A$, $B$ and $C$: $A$ and $B$ can enter and leave the system
while $C$ can only leave the system. As a complex $A$
is associated with $e_1\in \ensuremath{I\!\!R}^6$, $B$ with $e_2$, $C$ with
$e_6$, $A+2\, S$ with $e_1+2\, e_3$, $A\, S_2$ with $e_4$, $B+S$
with $e_2+e_3$, $B\, S$ with $e_5$, $A\, S_2 + B\, S$ with
$e_4+e_5$ and $C + 3\, S$ with $e_6 + 3\, e_3$.
The network consists of $r$ reactions ($r=10$), e.g. $A+2\, S \to
A\, S_2$,
where the complex at the tail of the arrow is called \emph{educt
complex} and the complex at the tip of the arrow is called
\emph{product complex}. To each reaction is associated a \emph{reaction
rate} $v_i(k,x)$. For mass action systems $v_i(k,x)$ is proportional
to the product of (powers of) concentrations of the species forming the
educt complexes: let $y_i$ be an educt complex vector, then one has
$v_i(k,x) = k_i\, x^{y_i}$ (where
$x^{p} = \prod_{j}x_j^{p_j}$ for $n$-vectors $x$ and $p$). For
the above network one obtains
\begin{displaymath}
v(k,x) =
(\, k_{1} x_{1} x_{3}^2,\,
k_{2} x_{4},\,
k_{3} x_{2} x_{3},\,
k_{4} x_{5},\,
k_{5} x_{4} x_{5},\,
k_{6} x_{1},\,
k_{7},\,
k_{8},\,
k_{9} x_{2},\,
k_{10} x_{6}\, )^T\ .
\end{displaymath}
We collect the exponents $y_i$ of the monomials contained in
$v_i(k,x)$ in the \emph{rate-exponent matrix} $\ensuremath{\mathcal{Y}}$. For the above
network one obtains the $(n\times r)$-matrix
\begin{align*}
\ensuremath{\mathcal{Y}} &=\left[y_1,\, ...\, ,y_{10}\right] \\
&= \bigl[
e_{1} + 2\, e_3,\,
e_{4},\,
e_{2} + e_{3},\,
e_{5},\,
e_{4} + e_5 ,\,
e_{1},\,
0,\,
0,\,
e_{2},\,
e_{6}
\bigr]\ .
\end{align*}
The reactions are encoded in the \emph{stoichiometric matrix} $S$,
where each column corresponds to one reaction and is defined as the
difference between product and educt complex. For example for the
reaction $ A+2\, S \to A\, S_2 $ one obtains $r_1 = -(e_1 + 2\,e_3) +
e_4$. The stoichiometric matrix for the above network is
\begin{displaymath}
S = \left[
\begin{array}{rrrrrrrrrr}
-1 & 1 & 0 & 0 & 0 & -1 & \phantom{-}1 & 0 & 0 & 0 \\
0 & 0 & -1 & 1 & 0 & 0 & 0 & \phantom{-}1 & -1 & 0 \\
-2 & 2 & -1 & 1 & 3 & 0 & 0 & 0 & 0 & 0 \\
1 & -1 & 0 & 0 & -1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & -1 & -1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & -1
\end{array}
\right]\, .
\end{displaymath}
A reaction network then defines a dynamical system
\begin{equation}
\label{eq:system_xdot}
\dot x = S\, v(k,x),
\end{equation}
which in case of the above network translates to
\begin{align*}
\dot x_1 &= \phantom{-}k_{7}-k_{6} x_{1}-k_{1} x_{1} x_{3}^2+k_{2} x_{4} \\
\dot x_2 &= \phantom{-}k_{8}-k_{9} x_{2}-k_{3} x_{2} x_{3}+k_{4} x_{5} \\
\dot x_3 &= -k_{3} x_{2} x_{3}-2 k_{1} x_{1} x_{3}^2+2 k_{2}
x_{4}+k_{4} x_{5}+3 k_{5} x_{4} x_{5} \\
\dot x_4 &= \phantom{-}k_{1} x_{1} x_{3}^2-k_{2} x_{4}-k_{5} x_{4} x_{5} \\
\dot x_5 &= \phantom{-}k_{3} x_{2} x_{3}-k_{4} x_{5}-k_{5} x_{4} x_{5} \\
\dot x_6 &= \phantom{-}k_{5} x_{4} x_{5}-k_{10} x_{6}\, .
\end{align*}
In general we consider a mass action network with $n$ species, $m$
complexes and $r$ reactions. Any such system defines a dynamical
system in the form given in \eqref{eq:system_xdot}. Note that $v(k,x)\in \ensuremath{I\!\!R}^r$ is
a \textbf{monomial}, \textbf{vector-valued} function of the form
\begin{displaymath}
v(k,x) = \diag(k)\, \phi(x),
\end{displaymath}
where $\diag(k)$ is a $(r\times r)$ diagonal matrix with the $k_i$ on the
diagonal and $\phi(x) = \left(x^{y_i}\right)_{i=1,\ldots, r}\in
\ensuremath{I\!\!R}^r$ is a vector of monomials in $x$. Note that the rate-exponent
matrix $\ensuremath{\mathcal{Y}}$ defined above contains the exponent vectors of the
monomials contained in $\phi(x)$. In the sequel, we speak of steady
states $(k,x)$ of (\ref{eq:system_xdot}) when $S\, v(k,x)$ vanishes
for positive $(k,x)$.
For many realistic systems in Systems Biology the matrix $S\in
\ensuremath{I\!\!R}^{n\times r}$ does not have full row rank $s := \rank(S)$ (i.e. $s <
n$). This gives rise to $n-s$ conservation relations: let $Z$ be any matrix
whose columns form a basis of $\ker(S^T)$, the left kernel of $S$.
Solutions $x(t)$ to (\ref{eq:system_xdot}) then satisfy
\begin{subequations}\begin{equation}
\label{eq:system_con_rel}
Z^T\, x(t) = Z^T\, x(0) =:c\, ,
\end{equation}
that is, these solutions lie in invariant domains $x(0)+\im(S)$
that are parallel translates of $\im(S)$. For the above example
(\ref{eq:exa_net}) one obtains
\begin{equation}\label{eq:system_con_rel2}
x_3+2 x_4+x_5 = c\ .
\end{equation}
\end{subequations}
\subsection{The Jacobian associated to a mass action network}
\label{sec:jacobian}
At positive $(k,x)$ the Jacobian of a mass action network (by this we
mean the Jacobian of a dynamical system defined by a mass action
network) is given by:
\begin{equation}
\label{eq:def_Jac}
Jac(k,x) = S\, \diag\left(v(k,x)\right)\, \ensuremath{\mathcal{Y}}^T\,
\diag\left(x^{-1}\right),
\end{equation}
with stoichiometric matrix $S$ and rate-exponent matrix $\ensuremath{\mathcal{Y}}$.
Observe that a positive pair $(k,x)$ is a steady state of
(\ref{eq:system_xdot}) if and only if $v(k,x) \in
\text{int}\left(\ker\left(S\right) \cap \ensuremath{I\!\!R_{\geq 0}}^r\right)$ (where
$\text{int}(\cdot)$ denotes the relative interior). The pointed
polyhedral cone $\ker\left(S\right)\cap\ensuremath{I\!\!R_{\geq 0}}^r$ is generated by a
finite set of unique (up to scalar multiplication) extreme rays
\cite{lin-004}.
The calculation of these rays is in general
computationally hard, however, there exists a variety of algorithms
and software tools, for example \cite{lin-050,lin-051}.
Let $p$ be the
number of extreme rays and let $E$ be
a matrix whose columns are generators of
$\ker\left(S\right)\cap\ensuremath{I\!\!R_{\geq 0}}^r$. Then $(k,x)$ is a positive steady
state if and only if there exists a $\nu$ with
\begin{subequations}
\begin{equation}\label{eq:def_LAM-1}
v(k,x) = E\, \nu \, >0\, ,\quad \nu \in \ensuremath{I\!\!R_{\geq 0}}^p \, .
\end{equation}
So we ask for all components of $E\, \nu$ to be (strictly) positive.
We collect all such $\nu$ in the set
\begin{equation}
\label{eq:def_LAM}
{\cal V} := \left\{\nu\in\ensuremath{I\!\!R_{\geq 0}}^p | E\, \nu >0 \right\}.
\end{equation}
Since $E$ is a nonnegative matrix, ${\cal V}$ consists of
the positive orthant $\ensuremath{\R_{> 0}}^p$, i.e.\ the interior of $\ensuremath{I\!\!R_{\geq 0}}^p $, and
potentially certain faces of $\ensuremath{I\!\!R_{\geq 0}}^p$ (i.e.\ elements $\nu \in {\cal
V}$ are either positive or nonnegative with predefined sign
pattern).
As we are interested in the Jacobian $Jac(k,x)$ evaluated at a
positive steady state we use (\ref{eq:def_LAM-1}) in
(\ref{eq:def_Jac}) to obtain
\begin{equation}
\label{eq:def_Jac_ss}
Jac(k,x)\ \equiv \ J(\nu,x) \ = \ N(\nu)\,
\diag\left(x^{-1}\right),\quad (\nu,x) \in
{\cal V}\times\ensuremath{\R_{> 0}}^n\, ,
\end{equation}
with the $\nu$-linear
\begin{equation}
\label{eq:def_Jac_ssh}
N(\nu)\ := \ S\, \diag\left(E\, \nu\right)\, \ensuremath{\mathcal{Y}}^T\ \
\in \, \ensuremath{I\!\!R}^{n\times n}\, , \quad (\nu,x) \in
{\cal V}\times\ensuremath{\R_{> 0}}^n\, .
\end{equation}
\end{subequations}
We'd like to emphasize that points $(\nu,x) \in {\cal V}
\times \ensuremath{\R_{> 0}}^n$ define points $(k,x)\in\ensuremath{\R_{> 0}}^r\times \ensuremath{\R_{> 0}}^n$ via
\begin{equation}
\label{eq:def_k_lambda}
k = \diag\left(\phi\left(x^{-1}\right)\right)\, E\, \nu\ .
\end{equation}
Hence finding points $(\nu,x)$ where $J(\nu,x)$ is singular
is equivalent to finding points $(k,x)$ where the Jacobian
$Jac(k,x)$ is singular.
\begin{subequations}
Null vectors of $Jac(k,x)=J(\nu,x)$ of the form
$\diag\left(x\right)z$ will be obtained from the identity
\begin{equation}
\label{eq:def_Jac_sshh}
J(\nu,x)\diag(x)\, z\, = \, N(\nu)\, z \, = \,
H(z)\, \nu = 0\, ,\ \nu \in {\cal V}\, ,
\end{equation}
with the $z$-linear
\begin{equation}
\label{eq:def_Jac_sshg}
H(z)\ :=\ S\diag\left(\ensuremath{\mathcal{Y}}^T\, z\right)E \ \in \, \ensuremath{I\!\!R}^{n\times p}\, .
\end{equation}
Our goal in (\ref{eq:def_Jac_sshh}) is to
use a condition from Qualitative Matrix Theory that entails
the existence of a positive null vector $\nu$ for the matrix
$H(z)$ (cf. Theorem~\ref{lem:main}).
\end{subequations}
\begin{example}[$J(\nu,x)$ derived from network~(\ref{eq:exa_net})]
The generator matrix of $\ker\brac{S}\cap\ensuremath{I\!\!R_{\geq 0}}^r$ is given by
\begin{subequations}
\begin{equation}\label{eq:exa_net0}
E =
\left[
\begin{array}{ccccc}
1&0&0&0&1 \\
1&0&0&0&0 \\
0&1&0&0&1 \\
0&1&0&0&0 \\
0&0&0&0&1 \\
0&0&1&0&0 \\
0&0&1&0&1 \\
0&0&0&1&1 \\
0&0&0&1&0 \\
0&0&0&0&1
\end{array}
\right]
\end{equation}
and hence satisfies
\begin{equation}\label{eq:exa_net1}
E\, \nu >0 \Leftrightarrow \nu>0\quad \text{and thus ${\cal V}
\equiv \ensuremath{\R_{> 0}}^5$.}
\end{equation}
The matrix $J(\nu,x)$ is given by
\begin{displaymath}
J(\nu,x) =
\left[
\begin{array}{cccccc}
-\frac{\nu_{1}+\nu_{3}+\nu_{5}}{x_{1}} & 0 & -\frac{2
(\nu_{1}+\nu_{5})}{x_{3}} & \frac{\nu_{1}}{x_{4}} & 0 & 0 \\
0 & -\frac{\nu_{2}+\nu_{4}+\nu_{5}}{x_{2}} &
-\frac{\nu_{2}+\nu_{5}}{x_{3}} & 0 & \frac{\nu_{2}}{x_{5}}
& 0 \\
-\frac{2 (\nu_{1}+\nu_{5})}{x_{1}} &
-\frac{\nu_{2}+\nu_{5}}{x_{2}} & -\frac{4
\nu_{1}+\nu_{2}+5 \nu_{5}}{x_{3}} & \frac{2 \nu_{1}+3
\nu_{5}}{x_{4}} & \frac{\nu_{2}+3 \nu_{5}}{x_{5}} & 0 \\
\frac{\nu_{1}+\nu_{5}}{x_{1}} & 0 & \frac{2
(\nu_{1}+\nu_{5})}{x_{3}} &
-\frac{\nu_{1}+\nu_{5}}{x_{4}} & -\frac{\nu_{5}}{x_{5}} & 0 \\
0 & \frac{\nu_{2}+\nu_{5}}{x_{2}} &
\frac{\nu_{2}+\nu_{5}}{x_{3}} & -\frac{\nu_{5}}{x_{4}} &
-\frac{\nu_{2}+\nu_{5}}{x_{5}} & 0 \\
0 & 0 & 0 & \frac{\nu_{5}}{x_{4}} & \frac{\nu_{5}}{x_{5}}
& -\frac{\nu_{5}}{x_{6}}
\end{array}
\right]\ .
\end{displaymath}
\end{subequations}
\end{example}
\subsection{Zero eigenvalues of the Jacobian of a mass action system}
\label{sec:zero-eigenvalues}
We assume $s=\rank(S)<n$ so that the Jacobian always has $n-s$ zero
eigenvalues. In addition we assume
\begin{equation}
\label{eq:imequal}
\im (S)\, = \, \im (J(\nu ,x))\, .
\end{equation}
In other terms, we assume the columns of the matrix $Z$
from (\ref{eq:system_con_rel}) to form a basis for $\ker{(J^T(\nu ,x))}$
so that $J(\nu ,x)$ does not possess more conservation laws than $S$.
In the end, we will have to validate this condition (\ref{eq:imequal})
(cf.~Appendix \ref{sec:imequal1} and \ref{sec:imequal3}).
In looking for
bifurcations, we reduce the system to the affine subspaces
$x(0)+\im(S)$. To this end let $U$, $W$ be orthonormal bases of
$\im(S)$, $\im(S)^\perp$, respectively and introduce
\begin{subequations}\begin{align}\label{eq:reduced-system0}
\xi &= U^T\, x,\; \eta = W^T\, x\text{ and $x(\xi,\eta) = U\, \xi +
W\, \eta$} \\
\intertext{to obtain the reduced system}
\label{eq:reduced-system}
\dot \xi &= U^T\, S \, v\left(k, x(\xi,\eta)\right) =:
g\left(\xi,\eta , k \right) \\
\dot \eta &= 0.
\end{align}
\end{subequations}
Then the upper left block of the Jacobian of this mass action network
is given by
\begin{align}
\notag
D_\xi \, g(\xi , \eta , k) &= U^T\, Jac(k,x(\xi , \eta ))\, U\\
\intertext{and at $(\nu,x)\in{\cal V} \times \ensuremath{\R_{> 0}}^n$ by}
\label{eq:reduced_Jac_ss}
G(\xi , \eta ,\nu ) &=
U^T\, J(\nu,x(\xi , \eta ))\, U\ \in \, \ensuremath{I\!\!R}^{s\times s}\, ,
\end{align}
where we recall the relation (\ref{eq:def_k_lambda}) between $k$,
$\nu$ and $x$. In \cite{fein-032} we presented a method that
links zero eigenvalues of $G(\xi , \eta ,\nu )$ to zero
eigenvalues of
$J(\nu,x(\xi,\eta))$.
Lemma~\ref{lem:zero_eigen_defective} below is
required for Theorem~\ref{theo:nasc}, the main result of this
section. We state it here without proof, for a proof see
\cite{fein-032}.
\vspace{2mm}
We start with some notation and, as in \cite{lin-012}, call an
eigenvalue $\lambda$ of a matrix $A\in\ensuremath{I\!\!R}^{n\times n}$ \emph{defective}
if its \emph{algebraic multiplicity} $m_{alg}(\lambda)$ is greater
than its \emph{geometric multiplicity} $m_{geo}(\lambda)$, that is,
if the multiplicity of $\lambda$ as a root of the characteristic
polynomial is greater than the number of linear independent
eigenvectors corresponding to $\lambda$.
Hence, $\lambda_0=0$ is a defective eigenvalue of
$A$ if and only if $\dim(\ker(A) + \im(A))$ is less $n$.
This can be stated in the following way:
\vspace{2mm}\begin{fact}
\label{fact:x_in_im_ker}
$\lambda_0 = 0$ is a defective eigenvalue of a matrix $A\in\ensuremath{I\!\!R}^{n\times n}$
iff there exists an $x\neq 0$ with $x\in \im\left(A\right)\cap
\ker\left( A \right)$.
\end{fact}
\begin{remark}
An alternative argument for Fact~\ref{fact:x_in_im_ker}
is based on the Jordan Canonical Form of a matrix $A$ (cf.,
for example, \cite{lin-012}). Assume an $n \times n $ matrix $A$
with eigenvalue $\lambda_0=0$ and $m_{alg}(\lambda_0) >
m_{geo}(\lambda_0)$ in Jordan Canonical Form. Then the $m_{alg}
\times m_{alg}$ block matrix corresponding to $\lambda_0$ is
not the zero-matrix, implying the existence of nontrivial $u_1\neq u_2$ with
$A\, u_1 = 0$ and $A\, u_2=u_1$ and hence $u_1 \in \ker(A) \cap
\im(A)$.
\end{remark}
\vspace{2mm}
\noindent We recall another fact from Lemma~1 in \cite{fein-032}:
\begin{lemma}
\label{lem:zero_eigen_defective}
Let $A\in\ensuremath{I\!\!R}^{n\times n}$ be a matrix of rank $s<n$
and let $U$ be orthonormal basis for $\im \left(A\right)$.
Then $\lambda_0=0$ is a defective eigenvalue of $A$ if and only if
$\lambda_0 = 0$ is an eigenvalue of $B_1:= U^T\, A\, U \in \ensuremath{I\!\!R}^{s\times s}$.
\end{lemma}
\vspace{2mm}\noindent Based on Fact~\ref{fact:x_in_im_ker} and
Lemma~\ref{lem:zero_eigen_defective} one is led to following
observation:
\vspace{2mm}\begin{lemma}
\label{fact:zero-eigenvalues-reduced}Let $Z_0$ be a basis
of $\im\brac{S}^\perp$. Then
the Jacobian $G(\xi , \eta ,\nu )$ of the reduced system,
evaluated at $\nu \in {\cal V}$ and $x=U\, \xi + W\, \eta \in \ensuremath{\R_{> 0}}^n$ (cf.\
(\ref{eq:reduced_Jac_ss}) and (\ref{eq:reduced-system0})),
has a zero eigenvalue if and only if there
exist a nontrivial vector $z\in\ensuremath{I\!\!R}^n$, a vector $x\in\ensuremath{\R_{> 0}}^n$
and a vector $\nu\in{\cal V}$ with
\begin{subequations}
\begin{align}
\label{eq:bif_1}
H(z)\, \nu &= 0 \\
\label{eq:bif_2}
Z_0^T\, \diag(x)\, z &= 0.
\end{align}
\end{subequations}
\end{lemma}
In the sequel, we take for $Z_0$ the matrix $Z$ describing the conservation laws
(cf. (\ref{eq:system_con_rel})).
\begin{proof}
From Lemma~\ref{lem:zero_eigen_defective}
follows that $G(\xi , \eta, \nu )$ has $\lambda_0=0$ as an
eigenvalue, if and only if $J(\nu,x)$ has $\lambda_0=0$ as a
defective eigenvalue. From Fact~\ref{fact:x_in_im_ker} follows
that $J(\nu,x)$ has a defective eigenvalue, if and only if
there is a nontrivial vector $\tilde z \in\ker\brac{S}\cap\im\brac{S}$. That
is, $\tilde z$ must satisfy $N(\nu)\, \diag\brac{\frac{1}{x}}\,
\tilde z = 0$ and $Z^T\, \tilde z = 0$ (cf.\ (\ref{eq:def_Jac_ss})
and (\ref{eq:imequal})). Let $\tilde
z=\diag\brac{x}\, z$, then \eqref{eq:bif_1} and \eqref{eq:bif_2}
follow.
\end{proof}
\vspace{2mm}First we consider condition~(\ref{eq:bif_2}) and establish necessary
and sufficient conditions for the existence of solutions ($x$, $z$)
$\in \ensuremath{\R_{> 0}}^n\times\ensuremath{I\!\!R}^n$, where we assume that $z$ is given.
\vspace{2mm}\begin{lemma}
\label{sec:lem_pos_x_sign_z}
Let $M\in R^{q_1\times q_2}$ be any matrix and let $z\in\ensuremath{I\!\!R}^{q_1}$ be given. Then
there exists a positive vector $x\in\ensuremath{\R_{> 0}}^{q_1}$ such that
\begin{displaymath}
M^T\, \diag(x)\, z = 0,
\end{displaymath}
if and only if
\begin{equation}
\label{eq:omega_condi}
\exists \omega\in\ker\left(M^T\right)\; \text{with
$\sign(\omega)=\sign(z)$.}
\end{equation}
In this case $x=\left(x\right)_{i=1,\, \ldots, q_1}$ is given by
\begin{equation}
\label{eq:x_def}
x_i =
\begin{cases}
\frac{\omega_i}{z_i}\text{, if $z_i\neq 0$} \\
\bar x_i > 0\text{, arbitrary, if $z_i=0$.}
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
Assume $M^T\, \diag(x)\, z=0$ holds for positive $x$ and some
$z$. Then $\omega:= \diag(x)\, z \in\ker\left(M^T\right)$ and
$\sign(\omega)=\sign(z)$ follows from positivity of $x$. Vice versa,
let $z\in\ensuremath{I\!\!R}$ and $\omega\in\ker\left(M^T\right)$ with
$\sign(\omega)=\sign(z)$ be given. Let $x$ be as in
(\ref{eq:x_def}). Then $\sign(\omega)=\sign(z)$ implies positivity
of $x$ and one has $\diag(x)\, z = \omega \in \ker\left(M^T\right)$.
\end{proof}
\begin{remark}
Observe that, given a vector $z$, the condition (\ref{eq:omega_condi})
takes the form of linear inequalities: one has to establish
feasibility of the system
\begin{displaymath}
M^T\, \omega = 0\text{, $\sign\left(z_i\right)\, \omega_i > 0$, if
$z_i\neq 0$ and $\omega_i=0$, if $z_i = 0$.}
\end{displaymath}
\end{remark}
\begin{remark}[Connection to \cite{fein-032}]
The condition (\ref{eq:bif_1}) requires the \textbf{symbolic
computation} of $\ker\left(H(\nu)\right)$. This can be of
forbidding complexity, especially for large
networks, even though it is in principle possible. \\
So far, the only application of the simple fact in
Lemma~\ref{lem:zero_eigen_defective}
we are aware of was in \cite{fein-032}. There we analysed a mass
action network describing the double phosphorylation of a
protein. For this network we obtained a symbolic representation
of $\ker\left(N(\nu)\right)$ that could be brought into a
$\nu$-independent form. In general, the previous approach requires
positive solutions to some well-defined polynomial equations in
$\nu$ and is thus limited to certain classes of systems (cf.\
\cite{cc-flo-multi-002}).
\end{remark}
\vspace*{2mm}
In the sequel, we employ the structure of $H(z)$, given by
(\ref{eq:def_Jac_sshg}), when discussing $H(z)\, \nu = 0$.
Let the columns of $S_0\in\ensuremath{I\!\!R}^{r\times (r-s)}$ be a basis of $\ker(S)$
and let $S_\#\in\ensuremath{I\!\!R}^{r\times r}$ be a matrix such that
$S_\#\, S_0 = \left[
\begin{smallmatrix}
I_{r-s} \\ {\bf 0}_{s\times (r-s)}
\end{smallmatrix}
\right]$.
If we let the columns of $S_c$ be a basis for $\im(S^T)$ and if we denote the Moore-Penrose
inverse $(S_0^TS_0)^{-1}S_0^T$ by $S_0^\#$
we will consider a particular such $S_\#$ by setting $S_\#^{part}=\left[
\begin{smallmatrix}
S_0^\#\\ S_c^T
\end{smallmatrix}
\right]$.
Equation~(\ref{eq:bif_1}) is now equivalent to
\begin{equation}\label{eq:Ssharp0}
S_0\, \alpha \, =\, \diag\left(\ensuremath{\mathcal{Y}}^T\, z\right)\, E\, \nu
\end{equation}
for some vector $\alpha \in \ensuremath{I\!\!R}^{r-s}$ and, by left multiplication with
$S_\#^{part}$, to
\begin{equation}
\label{eq:Teq}
\left[
\begin{array}{c}
\alpha\\0
\end{array}
\right] = \left[\begin{array}{c}
P(z) \\ Q(z)
\end{array}
\right]\, \nu\, \quad \mbox{with} \
\left[
\begin{array}{c}
P(z) \\ Q(z)
\end{array}
\right] :=S_\#^{part} \diag\left(\ensuremath{\mathcal{Y}}^T\, z\right)E\, .
\end{equation}
Observe that $z$ and $\nu$ satisfy $H(z)\, \nu = 0$
(cf.(\ref{eq:def_Jac_sshh})\&(\ref{eq:def_Jac_sshg}))
if and only if one
has
\begin{equation}
\label{eq:Qz_condi}
Q(z)\, \nu = 0,\quad \nu \in {\cal V}\, .
\end{equation}
The corresponding $\alpha$ will be given by $P(z)\nu$.
We note that the elements of the matrices $P(z)$ and $Q(z)$
are linear forms in $z$. Appendix~\ref{sec:imequal2} shows that the
condition (\ref{eq:Qz_condi}) is independent from the chosen bases for
$\ker(S)$ and $\im(S^T)$.
\vspace{2mm}\begin{theorem}
\label{theo:nasc}
The Jacobian $G(\xi , \eta ,\nu)$ of the reduced system,
evaluated at $\xi$ and $\eta$ with
$x=x(\xi,\eta)\in \ensuremath{\R_{> 0}}^n$ as in (\ref{eq:reduced-system0}) and $\nu \in {\cal V}$,
has zero
as an eigenvalue with algebraic multiplicity $\geq 1$ if and only if there
exist $z\in \ensuremath{I\!\!R}^n$, $\omega\in \ensuremath{I\!\!R}^n$ and $\mu \in \ensuremath{I\!\!R_{\geq 0}}^p$ with
\begin{equation}\label{eq:nasc}
Q(z)\mu =0,\ \ E\mu >0,\ \ Z^T\omega=0,\ \ \sign(\omega) = \sign(z)\, .\end{equation}
\end{theorem}
\begin{proof}
With the settings $\omega=\diag(x)z$ as in (\ref{eq:x_def}),
$x=U\xi+W\eta$ as in (\ref{eq:reduced-system0}) and $\nu=\mu$,
Theorem~\ref{theo:nasc} follows immediately from the
Lemma~\ref{fact:zero-eigenvalues-reduced}, Lemma~\ref{sec:lem_pos_x_sign_z} and
the equivalence of $H(z)\nu=0$ with (\ref{eq:Qz_condi}).
\end{proof}
\vspace{2mm}\begin{remark}[Open condition (\ref{eq:Teq})]\label{openeq:Teq}
Observe that $Q(z)\, \mu = 0$ in (\ref{eq:nasc}) can also be written in the form
$\tilde{Q}(\mu)\, z=0$
since $N(\mu)z=H(z)\mu$ (cf.(\ref{eq:def_Jac_sshh}) with (\ref{eq:def_Jac_ssh})
\& (\ref{eq:def_Jac_sshg})) implies the equivalence of
(\ref{eq:Teq}) and
\begin{equation}
\label{eq:Teq1}
\left[
\begin{array}{c}
\alpha\\0
\end{array}
\right] = \left[\begin{array}{c}
\tilde{P}(\mu) \\ \tilde{Q}(\mu)
\end{array}
\right]\, z\, \quad \mbox{with} \
\left[
\begin{array}{c}\tilde{P}(\mu) \\ \tilde{Q}(\mu)\end{array}
\right] :=S_\#^{part} \diag\left(E\, \mu\right)\ensuremath{\mathcal{Y}}^T\ .
\end{equation}
This reformulation reveals that (\ref{eq:Teq}) is
an \textit{open} condition: Given a particular solution $(\tilde{z},\tilde{\omega},\tilde{\mu})$,
there will exist a solution $(z,\omega,\mu)$ for all $\mu$'s that are sufficiently close to
$\tilde{\mu}$. So, there is some freedom in the choice of $\mu$,
cf.~Appendix \ref{sec:imequal3}.
\end{remark}
\vspace{2mm}Note that the semialgebraic set given by (\ref{eq:nasc}) is always
defined by polynomials of degree two or less,
independent of the exponents in the polynomial ODEs. Any
element gives rise to a defective eigenvalue $0$ of the Jacobian
$Jac(k,x)$. For the computation of elements of that set we will
later on employ the following observation:
in case the vector $\mu$ in (\ref{eq:nasc}) can be chosen as a
positive null vector of $Q(z)$, the condition $E\mu>0$ is
automatically satisfied. Thus we arrive at a sufficient condition for
a defective eigenvalue $0$ of the Jacobian $Jac(k,x)$ by
imposing conditions on the matrix $Q(z)$ that imply the existence of
a positive null vector $\mu$ and conditions on the vector $z$ ensuring
the sign-compatibility of $z$ with $\ker{(Z^T)}$.
Since the elements of
$Q(z)$ derived from a mass action network are always linear forms in
$z$, one can determine all sign patterns that
$\sign\brac{Q(z)}$ can admit by analyzing the corresponding
inequality systems. The idea is to look for sign patterns
guaranteeing that \emph{every matrix} with that sign pattern has a
\emph{positive} kernel vector. To this end we resort in subsection \ref{sec:new_condi} to
\emph{Qualitative Matrix Theory} \cite{sign-009} and to
\emph{$L^+$-matrices} in particular \cite{sign-004}.
We first exemplify our approach by examining (\ref{eq:nasc}) for
network~(\ref{eq:exa_net}) and turn to the general case in
Section~\ref{sec:new_condi}.
\vspace{3mm}\section{Conditions for a singular reduced Jacobian $G$}
\label{sec:suff-lin-ineq}\mbox{}
\subsection{System (\ref{eq:nasc}) for network~(\ref{eq:exa_net})}
Note that for network~(\ref{eq:exa_net}) the matrix $E$ of (\ref{eq:exa_net0})
is also a
basis for $\ker\brac{S}$ (in general this need not be the
case). Using this $E$ we obtain for equation (\ref{eq:Teq})
(where gray indicates rows belonging to $Q(z)$):
\begin{displaymath}
\left[
\begin{array}{ccccc}
z_{4} & 0 & 0 & 0 & 0 \\
0 & z_{5} & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & -z_{6} \\
0 & 0 & 0 & z_{2} & 0 \\
0 & 0 & 0 & 0 & z_{6} \\ \hline
\rowcolor{shade80} z_{1}+2 z_{3}-z_{4} & 0 & 0 & 0 & z_{1}+2 z_{3}-z_{6} \\
\rowcolor{shade80} 0 & z_{2}+z_{3}-z_{5} & 0 & 0 & z_{2}+z_{3}-z_{6} \\
\rowcolor{shade80} 0 & 0 & 0 & 0 & z_{4}+z_{5}-z_{6} \\
\rowcolor{shade80} 0 & 0 & z_{1} & 0 & z_{6} \\
\rowcolor{shade80} 0 & 0 & 0 & -z_{2} & -z_{6}
\end{array}
\right]\, \nu = \left[
\begin{array}{c}
\alpha_1 \\
\\
\vdots \\
\\
\alpha_5 \\ \hline
\rowcolor{shade80} 0 \\
\rowcolor{shade80} \\
\rowcolor{shade80} \vdots \\
\rowcolor{shade80}\\
\rowcolor{shade80} 0
\end{array}
\right]\ .
\end{displaymath}
One has $s=5$ and $r=10$, hence the matrix $Q(z)$ is defined by rows
6--10. However, it is easy to see that $v\in {\cal V}$ (and hence
positive $\nu$ by (\ref{eq:exa_net1})) exist only if
$z_6=z_4+z_5$. Hence $Q(z)$ consists
only of the rows 6, 7, 9 and 10 as row 8 evaluated at $z_6=z_4+z_5$ is
identically zero. One obtains
\begin{displaymath}
Q(z) =
\left[
\begin{array}{ccccc}
z_{1}+2 z_{3}-z_{4} & 0 & 0 & 0 & z_{1}+2 z_{3}-z_{4}-z_{5} \\
0 & z_{2}+z_{3}-z_{5} & 0 & 0 & z_{2}+z_{3}-z_{4}-z_{5} \\
0 & 0 & z_{1} & 0 & z_{4}+z_{5} \\
0 & 0 & 0 & -z_{2} & -z_{4}-z_{5}
\end{array}
\right]
\end{displaymath}
For this $Q(z)$ one has positive $\nu$, iff the following pairs of
linear forms are either of opposite sign or both equal to zero:
\begin{equation}
\label{eq:linear_forms_exa1}
\begin{split}
\ell_1(z) := z_1+2 z_3-z_4\quad &\text{and}\quad \ell_2(z) :=
z_1+2 z_3-z_4-z_5,\\
\ell_3(z) := z_2+z_3-z_5\quad &\text{and}\quad \ell_4(z) :=
z_2+z_3-z_4-z_5, \\
\ell_5(z) := z_1\quad &\text{and}\quad \ell_6(z) := z_4+z_5, \\
\ell_7(z) := -z_2\quad &\text{and}\quad \ell_8(z):=
-z_4-z_5\ .
\end{split}
\end{equation}
These conditions can be expressed as \emph{linear inequality systems},
for example
\begin{subequations}
\begin{equation}
\label{eq:z_ineq_exa1_1}
\begin{split}
z_1+2 z_3-z_4>0,\, z_1+2 z_3-z_4-z_5<0,\\
z_2+z_3-z_5<0,\, z_2+z_3-z_4-z_5>0,\\
z_1>0,\, z_4+z_5<0,\\
-z_2<0,\, -z_4-z_5>0.
\end{split}
\end{equation}
This system is feasible; pick any $\tilde z\in\ensuremath{I\!\!R}^5$ satisfying
(\ref{eq:z_ineq_exa1_1}) and let $\tilde z_6=\tilde z_4 + \tilde
z_5$. Then $Q(z)$ evaluated at that $\tilde z$ has a positive
kernel vector $\nu$ (cf. Table~\ref{tab:z_w_exa1} and
(\ref{eq:nu14}).\\
We apply Lemma~\ref{sec:lem_pos_x_sign_z} with
$M^T=Z^T=\left(0,\, 0,\, 1,\, 2,\, 1,\, 0\right)$ from
(\ref{eq:system_con_rel2}) and need to find a vector
$\tilde\omega\in\ker\brac{Z^T}$ with $\sign\brac{\tilde\omega} =
\sign\brac{\tilde{z}}$. For the choice of
$\tilde\omega$ with $\tilde\omega_3<0$, $\tilde\omega_4<0$ and
$\tilde\omega_5>0$ we consequently add
\begin{equation}
\label{eq:z_ineq_exa1_2}
\tilde{z}_3<0,\, \tilde{z}_4<0,\, \tilde{z}_5>0
\end{equation}
\end{subequations}
to the inequality system. The overall system (\ref{eq:z_ineq_exa1_1})
\& (\ref{eq:z_ineq_exa1_2}) is feasible and one solution $\tilde z$ is
given in Table~\ref{tab:z_w_exa1}. Table~\ref{tab:z_w_exa1} also
contains a vector $\tilde\omega\in\ker\brac{Z^T}$ with
$\sign\brac{\tilde z} = \sign\brac{\tilde\omega}$ and the vector
$x=\frac{\tilde\omega}{\tilde z}$ (cf.\
Lemma~\ref{sec:lem_pos_x_sign_z}, equation (\ref{eq:x_def})).
\begin{table}[htb]
\centering
\begin{tabular}{|>{$}c<{$}|*{6}{>{$}r<{$}}|}\hline
& \tiny 1 & \tiny 2 & \tiny 3 & \tiny 4 & \tiny 5 & \tiny 6 \\ \hline
\tilde z& 4 & 1 & -5 & -8 & 3 & -5 \\
\tilde \omega & 1 & 1 & -3 & -1 & 5 & -1 \\
x & \frac{1}{4} & 1 & \frac{3}{5} & \frac{1}{8} & \frac{5}{3} &
\frac{1}{5} \\ \hline
\end{tabular}
\caption{Vectors $\tilde{z}$, $\tilde{\omega}$ and $x$}
\label{tab:z_w_exa1}
\end{table}
Evaluating $Q(z)$ at $\tilde z$ from Table~\ref{tab:z_w_exa1} one has the
matrix
\begin{displaymath}
Q(\tilde{z}) = \left[
\begin{array}{rrrrr}
2 & 0 & 0 & 0 & -1 \\
0 & -7 & 0 & 0 & 1 \\
0 & 0 & \phantom{-}4 & 0 & -5 \\
0 & 0 & 0 & -1 & 5
\end{array}
\right]
\end{displaymath}
that has the positive kernel vector
\begin{equation}\label{eq:nu14}
\nu =
\left(
14,\,
4,\,
35,\,
140,\,
28
\right)^T\ .
\end{equation}
Vector $x$ from Table~\ref{tab:z_w_exa1} and the above $\nu$
define a vector of rate constants:
\begin{displaymath}
k = \left(
\frac{1400}{3},\,
112,\,
\frac{160}{3},\,
\frac{12}{5},\,
\frac{672}{5},\,
140,\,
63,\,
168,\,
140,\,
140
\right)^T\ .
\end{displaymath}
Evaluation of $Jac\brac{k,x}$ at this $k$ and $x$ from
Table~\ref{tab:z_w_exa1} confirms $\lambda=0$ as a defective
eigenvalue.
All in all there are 81 different inequality systems where the
pairs from (\ref{eq:linear_forms_exa1}) are of different sign or
both zero. There are also 13 inequality systems
like~(\ref{eq:z_ineq_exa1_2}) that constrain $z$ such that there
is a $\omega\in\ker\brac{Z^T}$ with $\sign\brac{\omega} =
\sign\brac{z}$. Of these 13*81=1053 inequality systems only the
following four are feasible:
\begin{minipage}[t]{0.48\linewidth}
\begin{align*}
z_3<0&,\; z_4<0,\; z_5>0 \\
\ell_1(z) &>0,\; \ell_2(z) <0 \\
\tag{$P_1^+$}\label{eq:P1}
\ell_3(z) &<0,\; \ell_4(z) >0 \\
\ell_5(z) &>0,\; \ell_6(z) <0 \\
\ell_7(z) &<0,\; \ell_8(z) >0 \\
\intertext{and}
z_3>0&,\; z_4>0,\; z_5<0 \\
\ell_1(z) &<0,\; \ell_2(z) >0 \\
\tag{$P_1^-$}\label{eq:P3}
\ell_3(z) &>0,\; \ell_4(z) <0 \\
\ell_5(z) &<0,\; \ell_6(z) >0 \\
\ell_7(z) &>0,\; \ell_8(z) <0
\end{align*}
\end{minipage}
\begin{minipage}[t]{0.48\linewidth}
\begin{align*}
z_3<0&,\; z_4>0,\; z_5<0 \\
\ell_1(z)&<0,\; \ell_2(z) > 0 \\
\tag{$P_2^+$}\label{eq:P2}
\ell_3(z)&>0,\; \ell_4(z) < 0 \\
\ell_5(z)&>0,\; \ell_6(z) < 0 \\
\ell_7(z)&<0,\; \ell_8(z) > 0 \\
\intertext{\phantom{and}}
z_3>0&,\; z_4<0,\; z_5>0 \\
\ell_1(z)&>0,\; \ell_2(z) < 0 \\
\tag{$P_2^-$}\label{eq:P4}
\ell_3(z)&<0,\; \ell_4(z) > 0 \\
\ell_5(z)&<0,\; \ell_6(z) > 0 \\
\ell_7(z)&>0,\; \ell_8(z) < 0
\end{align*}
\end{minipage}
\vspace{2mm}\noindent Because of the definitions of $\ell_5,\ell_6$
and $\ell_7$ in (\ref{eq:linear_forms_exa1}), feasible $z$'s do not
have vanishing components. All in all we have established the
following necessary and sufficient condition for a defective
eigenvalue of $J(\nu,x)$ of (\ref{eq:exa_net}).
\vspace{2mm} \begin{fact}
The Jacobian $J(\nu,x)$ of (\ref{eq:exa_net}) evaluated at
$(\nu,x) \in {\cal V} \times \ensuremath{\R_{> 0}}^6$ (and hence $Jac(k,x)$
evaluated at positive $(k,x)$ via (\ref{eq:def_k_lambda})) has
$\lambda_0=0$ as a defective eigenvalue, if and only if $\nu$
and $x$ satisfy:
\begin{enumerate}
\item The vector $x$ can be written as $x = \frac{
\omega}{z}$ with
(i) $ z \in \ensuremath{I\!\!R}^6$ satisfies one of the inequality systems
($P_1^\pm$), ($P_2^\pm$) and $z_6=z_4 + z_5$ (implying ${z}_i\neq
0$ for $i=1,...,6$),\,
(ii) $\omega\in\ker\brac{Z^T}$ and (iii) $\sign\brac{z} =
\sign\brac{\omega}$.
\item The above $z$ and the vector $\nu>0$ are such that
$Q(z)\, \nu = 0$.
\end{enumerate}
\end{fact}
\vspace{2mm}\noindent Note that, if $z\in\ensuremath{I\!\!R}^6$ with $z_6=z_4+z_5$
satisfies one of the systems ($P_1^\pm$) and ($P_2^\pm$),
then the sign
pattern $\sign\brac{Q(z)}$ is one of the following:
\begin{itemize}
\item If $z$ satisfies ($P_1^\pm$) then
\begin{displaymath}
\sign\brac{Q(z)} = \pm \, \left[
\begin{array}{rrrrr}
\phantom{-} 1 & 0 & 0 & 0 & -1 \\
0 & -1 & 0 & 0 & 1 \\
0 & 0 & \phantom{-} 1 & 0 & -1 \\
0 & 0 & 0 & -1 & 1
\end{array}
\right]\ .
\end{displaymath}
\item If $z$ satisfies ($P_2^\pm$) then
\begin{displaymath}
\sign\brac{Q(z)} =\pm \, \left[
\begin{array}{rrrrr}
-1 & 0 & 0 & 0 & 1 \\
0 & \phantom{-}1 & 0 & 0 & -1 \\
0 & 0 & \phantom{-}1 & 0 & -1 \\
0 & 0 & 0 & -1 & 1
\end{array}
\right]\, .
\end{displaymath}
\end{itemize}
\subsection{A sufficient condition}
\label{sec:new_condi}
For the example of network~(\ref{eq:exa_net}) we obtained necessary and
sufficient conditions in form of linear inequalities in $z$
guaranteeing a \emph{positive} kernel vector $\nu $ of $Q(z)$.
The idea is to look for sign patterns $\sign\brac{Q(z)}$
guaranteeing that \emph{every matrix} with that sign pattern has a
\emph{positive} kernel vector (as it has been the case with the sign
patterns of the previous section).
By \emph{Qualitative Matrix Theory} \cite{sign-009} (see in particular \cite{sign-004})
one has the following Theorem`\ref{theo:L+-matrices}. In our application,
it can be stated in the following way: If a sign pattern
$\sign\brac{Q(z)}$ is an $L^+$-matrix, then every matrix with the same
sign pattern has a positive kernel vector.
\vspace{2mm}\begin{theorem}[cf. \cite{sign-004}, Theorem 2.4, p.6]
\label{theo:L+-matrices}
For a $(m\times n)$ sign pattern $A$, the following are equivalent:
\begin{enumerate}[{(}a{)}]
\item $A$ is an $L^+$-matrix.
\item Every matrix with the sign pattern $A$ has a \underline{positive null
vector} and $A$ has no zero row.
\item For each nonzero vector $\sigma\in\left\{-1,\, 0,\,
1\right\}^{m}$, some column of \,
$\diag\left(\sigma\right)A$\, is nonzero and nonnegative.
\item For each nonzero vector $\sigma\in\left\{-1,\, 0,\,
1\right\}^{m}$, some column of \,
$\diag\left(\sigma\right)A$\, is nonzero and nonpositive.
\end{enumerate}
\end{theorem}
Note that Theorem~\ref{theo:L+-matrices} already contains -- by the
parts $(c)$ or $(d)$ -- a primitive algorithm to determine whether or
not a given sign pattern is an $L^+$-matrix. So by
Theorem~\ref{theo:L+-matrices} one can decide whether or not a
particular sign pattern is an $L^+$-matrix.
\vspace{2mm} With respect to the $z$-linear matrix
$Q(z)=Q_{ij}(z)\in \ensuremath{I\!\!R}^{s\times p}$ from
(\ref{eq:Qz_condi}) we propose the following: We first stack the columns of $Q$
and consider the column vector
\begin{displaymath}
(Q_{11},\ldots,Q_{s1},\ldots\ldots ,Q_{1p},\ldots,Q_{sp})^T\, .
\end{displaymath}
Then we omit the components that are trivial linear forms to obtain a
bijective mapping of the form
\begin{equation}\label{map-vec}
\psi:\ Q(z)\in \ensuremath{I\!\!R}^{s\times p} \ \mapsto \ L\, z=(\ell_1z,\ldots \ldots ,\ell_\gamma z)^T\in \ensuremath{I\!\!R}^\gamma
\end{equation}
with nontrivial $n$-dimensional row-vectors $\ell_1$, \ldots,
$\ell_\gamma$, $\gamma \leq sp$. So, the $(\gamma
\times n)$-matrix
\begin{equation}
\label{eq:VD}
L=\left( \ell_i \right)_{i=1,\ldots,\gamma}\, .
\end{equation}
just corresponds to the nontrivial linear forms in $Q(z)$.
Since we look for a $\nu \in {\cal V}$ with $Q(z)\nu=0$,
we are interested in the sign patterns that $Lz$ can assume.
So we define the set $\mathcal{L}^+$ of all
\emph{sign pattern matrices} $\Sigma\in \{-1,0,1\}^{s\times p}$
that are \emph{$L^+$-matrices} and that are realized by $Q(z)$ for some $z$.
Since the mapping (\ref{map-vec}) associates a {\em signature vector}
$\sigma =\psi(\Sigma)\in \{-1,0,1\}^\gamma $ to $\Sigma\in \{-1,0,1\}^{s\times p}$
one arrives at
\begin{equation}\label{eq:LD}
\begin{split}
\mathcal{L}^+
&:= \Bigl\{ \Sigma \in \{-1,0,1\}^{s\times p} \, \Big| \ \text{$\Sigma$ is an $L^+$-matrix},\\
&\qquad \exists z\, \in \ensuremath{I\!\!R}^n\; \text{with $\sigma_i\, \left(Lz\right)_i >0$ if
$\sigma_i \neq 0$ and $\left(L\, z\right)_i =0$ if $\sigma_i = 0$}
\Bigr\}\, .
\end{split}
\end{equation}
\begin{fact}
\label{fact:f1}
Assume $\mathcal{L}^+$ is
nonempty and let $\Sigma\in\mathcal{L}^+$ and $\sigma=\psi(\Sigma)$.
Then there exists a vector
$z\in\ensuremath{I\!\!R}^n$ with
\begin{displaymath}
\sigma_i\, \left(L\, z\right)_i > 0\text{, if $\sigma_i\neq 0$},\ \
\left(L\, z\right)_i = 0\text{, if $\sigma_i = 0$}
\end{displaymath}
so that $\sigma=\sign\brac{L\, z}$. Moreover, for each such
$z\in\ensuremath{I\!\!R}^n$, there exists a positive $\nu=\nu(z)$ with $Q(z)\, \nu =
0$ by Theorem~\ref{theo:L+-matrices}. By the discussion of (\ref{eq:Qz_condi}),
this implies
that the pair ($z$, $\nu$) satisfies $H(\nu)\, z = 0$.
\end{fact}
\vspace{2mm}\begin{theorem}
\label{lem:main}
Consider a dynamical system defined by a mass action network as
described in Section~\ref{sec:dyn-sys-mass-action}. Recall the
matrix $Q(z)$ defined in (\ref{eq:Teq}), the matrix $L$ defined
in (\ref{map-vec})\&(\ref{eq:VD}) and the set $\mathcal{L}^+$ defined in
(\ref{eq:LD}). If there exist an element $\Sigma=\sign\brac{Q(z)}\in\mathcal{L}^+$
and an element $\omega\in\ker\brac{Z^T}$ with
\begin{equation}
\label{eq:li3}
\sign(\omega) = \sign(z),
\end{equation}
then there exists a solution $\nu\in\ensuremath{\R_{> 0}}^p$ to $Q(z)\, \nu =
0$ and the Jacobian $G(\xi , \eta ,\nu)$ of the reduced system has
zero as an eigenvalue with algebraic multiplicity $\geq 1$. The
corresponding steady state in original ($k$, $x$)-coordinates is
given by
\begin{subequations}
\begin{align}
x & = \left(x_i\right)_{i=1,\ldots,n} , \\
x_i &= \begin{cases}
\frac{\omega_i}{z_i}\text{, if $z_i\neq 0$ ,} \\
\bar x_i > 0\text{, arbitrary, if $z_i=0$ ,}
\end{cases} \\
k &= \diag\left(\phi\left(x^{-1}\right)\right)\, E\, \nu\, .
\end{align}
The corresponding $(\xi,\eta)$-coordinates in $G(\xi , \eta ,\nu)$
are then given by (\ref{eq:reduced-system0}).
\end{subequations}
\end{theorem}
\begin{proof} The statements follow directly from
Lemma~\ref{sec:lem_pos_x_sign_z} and Fact~\ref{fact:f1}.
\end{proof}
\vspace{2mm}The condition (\ref{eq:li3}) can be tested by examining
the following linear inequality systems defined by orthants of
$R^n$. To establish these we identify each orthant by its sign pattern
$\delta\in\{-1,0,1\}^n$: let $x\in\ensuremath{I\!\!R}^n$, then the sign pattern of $x$
is defined as $\delta:=\sign(x)$ and the orthant containing $x$ is
given by $\ensuremath{I\!\!R}_\delta^n := \{ x\in\ensuremath{I\!\!R}^n | \sign(x) = \delta\}$. To find
$z$ and $\omega$ satisfying (\ref{eq:li3}) for a given signature
$\sigma =\psi(\Sigma)$ for an $L^+$-matrix $\Sigma$
then amounts to finding an orthant $\ensuremath{I\!\!R}_\delta^n$ such that
\begin{subequations}
\begin{gather}
\label{eq:orthi_sys_1}
\sigma_i\, \left(Lz\right)_i > 0\, \text{ if
$\sigma_i\neq 0$},\;\ \ \left(Lz\right)_i = 0\, \text{ if $\sigma_i
= 0$}, \\
\label{eq:orthi_sys_2}
Z^T\, \omega = 0,\;\ \text{with }\ \;
\delta_i\, \omega_i >0,\, \delta_i\, z_i >0 \ \text{ if $\delta_i
\neq 0$}\ \text{and} \ \omega_i = 0,\, z_i=0\ \text{ if $\delta_i=0$}.
\end{gather}
\end{subequations}
\vspace{2mm}\begin{corollary}
If there exists a signature $\Sigma\in\mathcal{L}^+$ and an orthant
$\ensuremath{I\!\!R}_\delta^n$, such that the linear inequality system
(\ref{eq:orthi_sys_1}), (\ref{eq:orthi_sys_2}) is feasible, then the
reduced Jacobian $G(\xi , \eta ,\nu)$ has zero as an eigenvalue with
algebraic multiplicity $\geq 1$.
\end{corollary}
\vspace{2mm}\begin{remark}
In the previous discussion we have only considered positive kernel
vectors of $Q(z)$. However the set ${\cal V}$ can contain nonnegative
vectors $\nu$. Thus, suppose $\nu$ contains the facet of
$\ensuremath{I\!\!R_{\geq 0}}^p$ given by $\bigl\{\nu\in\ensuremath{I\!\!R_{\geq 0}}^p | \nu_i=0$,
$\nu_j>0$, $i\neq j=1$, \ldots, $p\bigr\}$. Then one may fix
$\nu_i=0$, replace $Q(z)$ in the discussion above by the
submatrix $\tilde Q(z)$ obtained by deleting the $i$-th column (and
eventually occurring zero rows) and obtain the remaining $\nu_i$
by asking for positive kernel vectors of $\tilde Q(z)$ (i.e.\ by
establishing the $L^+$-property for $\tilde Q(z)$).
\end{remark}
\vspace{2mm}\begin{remark}
The condition (\ref{eq:orthi_sys_1})
tests whether the given $L^+$-matrix $\Sigma$ belongs to $\mathcal{L}^+$.
By
the definition of the matrix $L$ this requires the labeling of the
hyperplane arrangement given by $Lz$, which is computationally
expensive (for an algorithm see \cite{arr-001},
\cite{arr-003}). We have shown that all $z$ satisfying (\ref{eq:orthi_sys_1})
lead to a positive null vector of $Q(z)$.
The condition (\ref{eq:orthi_sys_2}) then stands for the compatibility
with the kernel of $Z^T$: It tests whether there is a $z$ in the solution set of
(\ref{eq:orthi_sys_1}) that possesses a signature thats is compatible with $\ker\brac{Z^T}$.
Since one has to
decide whether or not one of the systems (\ref{eq:orthi_sys_1}),
(\ref{eq:orthi_sys_2}) is feasible the overall procedure can be
computationally demanding, even though the individual steps only
involve simple matrix computations.
\end{remark}
\vspace{3mm}\section{Saddle node bifurcations for the G1/S transition in budding
yeast}
\label{sec:application}\mbox{}
The networks displayed in (\ref{eq:net_ten}) and (\ref{eq:net_bin})
below are competing hypotheses describing the G1/S transition in
budding yeast. Both networks are biologically plausible and hard to
distinguish experimentally \cite{fein-025}.
\begin{equation}
\label{eq:net_ten}
\scalebox{0.8}{
\begin{minipage}{1.0\linewidth}
\begin{displaymath}
\begin{split}
&\xymatrix{
{\bf [Sic1P]} \ar [r] ^{\bf k_3} & {\bf [0]} \ar @<.4ex> @{-^>}
[r] ^{\bf k_1} & \ar @{-^>} [l] ^{\bf k_2} {\bf [Sic1]} } \\
&\xymatrix{
{\bf [Clb] + [Sic1]} \ar @<.4ex> @{-^>} [r] ^{\bf k_4} & \ar
@{-^>} [l] ^{\bf k_5} {\bf [Clb\cdot Sic1]} \ar [dr] ^{\bf k_6}
& \\
& & {\bf [Clb]} \\
{\bf [Clb] + [Sic1P]} \ar @<.4ex> @{-^>} [r] ^{\bf k_7} & \ar
@{-^>} [l] ^{\bf k_8} {\bf [Clb\cdot Sic1P]} \ar [ur] ^{\bf k_9}
& } \\
&\xymatrix{
{\bf [Clb\cdot Sic1] + [Clb]} \ar @<.4ex> @{-^>} [r] ^{\bf k_{10}}
& \ar @{-^>} [l] ^{\bf k_{11}} {\bf [Clb\cdot Sic1 \cdot Clb]} \ar
[r] ^{\bf k_{12}} & {\bf [Clb\cdot Sic1P] + [Clb]}
} \\
&\xymatrix{
{\bf [Sic1P] + [Cdc14]} \ar @<.4ex> @{-^>} [r] ^{\bf k_{13}} &
\ar @{-^>} [l] ^{\bf k_{14}} {\bf [Sic1P\cdot Cdc14]} \ar [r]
^{\bf k_{15}} & {\bf [Sic1] + [Cdc14]}
} \\
&\xymatrix{
{\bf [Clb\cdot Sic1P] + [Cdc14]}\ar @<.4ex> @{-^>} [r]
^{\bf k_{16}} & \ar @{-^>} [l] ^{\bf k_{17}} {\bf [Clb\cdot
Sic1P\cdot Cdc14]} \ar [r] ^{\bf k_{18}} & {\bf [Clb\cdot Sic1]
+ [Cdc14]}
}
\end{split}
\end{displaymath}
\end{minipage}
}
\end{equation}
\begin{equation}
\label{eq:net_bin}
\scalebox{0.8}{
\begin{minipage}{1.0\linewidth}
\begin{displaymath}
\begin{split}
&\xymatrix{
{\bf [Sic1P]}\ar [r] ^{\bf k_3} & {\bf [0]} \ar @<.4ex> @{-^>}[r]
^{\bf k_1} & {\bf [Sic1] } \ar @{-^>}[l]^{\bf k_2}& &
} \\
&\xymatrix{
{\bf [Sic1\cdot Clb]}\ar @<.4ex> @{-^>}[r] ^{\bf k_{4}} \ar
[ddr] ^{\bf k_9} & \ar @{-^>} ^{\bf k_5} [l] {\bf [Clb] +
[Sic1]} \ar @<.4ex> @{-^>}[r] ^{\bf k_6} & \ar @{-^>} ^{\bf
k_7} [l] {\bf [Clb\cdot Sic1]} \ar [dr] ^{\bf k_8} & \\
& & & {\bf [Clb]} \\
& {\bf [Clb] + [Sic1P]} \ar @<.4ex> @{-^>}
[r] ^{\bf k_{10}} & \ar
@{-^>}[l] ^{\bf k_{11}} {\bf [Clb\cdot Sic1P]} \ar [ur]
^{\bf k_{12}} &
}\\
&\xymatrix{
{\bf [Sic1P] + [Cdc14]} \ar @<.4ex> @{-^>} [r] ^{\bf k_{13}} &
\ar @{-^>} [l] ^{\bf k_{14}} {\bf [Sic1P\cdot Cdc14]} \ar [r]
^{\bf k_{15}} & {\bf [Sic1] +[Cdc14]} &
} \\
&\xymatrix{
{\bf [Clb\cdot Sic1P] + [Cdc14]} \ar @<.4ex> @{-^>} [r] ^{\bf
k_{16}} & \ar @{-^>} [l] ^{\bf k_{17}} {\bf [Clb \cdot
Sic1P\cdot Cdc14]} \ar [r] ^{\bf k_{18}} & {\bf [Clb\cdot
Sic1] + [Cdc14]}
}
\end{split}
\end{displaymath}
\end{minipage}
}
\end{equation}
Switching is a desired property of models describing the G1/S
transition: depending on its past a trajectory should move to
different regions of state space, associated with the G1 and the S
phase of cell cycle. Classically this has been realized by choosing
rate constants and total concentrations, such that the ODE system
shows bistability and hence hysteretic behaviour
\cite{cyc-005,sig-042}. For example in \cite{fein-025},
multistationarity has been established for both models, indicating
that both may be valid models. Here we demonstrate the applicability
of our results by confirming switching for both networks. We show that
both models satisfy the conditions of Theorem~\ref{lem:main} and
compute states and rate constants where the Jacobian has a defective
eigenvalue. We verify by numerical continuation that the system
undergoes a saddle-node bifurcation, as generically expected, so that
the codimension-$1$ stable manifold of the saddle-node and - after
bifurcation - the one of the saddle represents a switching surface.
For the network given in \eqref{eq:net_ten} one obtains using the
stoichiometric matrix $S$ given in Appendix~\ref{sec:data-tern}:
\begin{flushleft}
\scalebox{0.5}{
\begin{minipage}{1.0\linewidth}
\begin{displaymath}
Q(z) = \left[
\begin{array}{ccccccccccc}
-z_1 & 0 & 0 & 0 & 0 & 0 & -z_9 & -z_4 & -z_9 & z_7-z_9 &
z_8-z_9 \\
0 & 0 & 0 & 0 & 0 & 0 & z_2-z_9 & 0 & z_5-z_9 & z_7-z_9 &
z_8-z_9 \\
0 & z_1+z_3-z_4 & 0 & 0 & 0 & 0 & z_1+z_3-z_9 & z_1+z_3-z_4 &
z_1+z_3-z_9 & z_1+z_3-z_9 & z_8-z_9 \\
0 & 0 & z_2+z_3-z_5 & 0 & 0 & 0 & -z_5+z_9 & 0 & -z_5+z_9 &
-z_5+z_9 & -z_8+z_9 \\
0 & 0 & 0 & z_3+z_4-z_9 & 0 & 0 & z_3+z_4-z_9 & 0 & z_3+z_4-z_9
& z_3+z_4-z_9 & z_3+z_4-z_9 \\
0 & 0 & 0 & 0 & z_2+z_6-z_7 & 0 & 0 & 0 & 0 & z_2+z_6-z_7 & 0 \\
0 & 0 & 0 & 0 & 0 & z_5+z_6-z_8 & 0 & 0 & 0 & 0 & z_5+z_6-z_8
\end{array}
\right]
\end{displaymath}
\end{minipage}
}
\end{flushleft}
\vspace{2mm}\noindent From the last three rows of $Q(z)$ one has that
positive $\nu$ with $Q(z)\, \nu=0$ exist only if
\begin{equation}
\label{eq:eq_const_tern}
z_3+z_4-z_9 = 0,\; z_2+z_6-z_7 =0,\; z_5+z_6-z_8 =0
\end{equation}
and hence, for example,
\begin{displaymath}
z_2 = -z_6+z_7,\; z_3 = -z_4 + z_9,\; z_5 = -z_6+z_8\ .
\end{displaymath}
In this case colum 4, 5 and 6 will be the zero column, indicating that
$\nu_4$, $\nu_5$, $\nu_6>0$ are unconstrained. Thus we
need only consider the matrix
\begin{displaymath}
Q_s(z) =
\left[
\begin{array}{cccccccc}
-\pi_{8}\, z & 0 & 0 & -\pi_{10}\, z & -\pi_{9}\, z & -\pi_{10}\, z
& \pi_{6}\, z & \pi_{5}\, z \\
0 & 0 & 0 & \pi_{2}\, z & 0 & -\pi_{4}\, z & \pi_{6}\, z &
\pi_{5}\, z \\
0 & \pi_{3}\, z & 0 & \pi_{7}\, z & \pi_{3}\, z & \pi_{7}\, z &
\pi_{7}\, z & \pi_{5}\, z \\
0 & 0 & \pi_{1}\, z & \pi_{4}\, z & 0 & \pi_{4}\, z & \pi_{4}\, z &
-\pi_{5}\, z
\end{array}
\right]
\end{displaymath}
with
\begin{align*}
\pi_{1}\, z &:= -z_{4}+z_{7}-z_{8}+z_{9} &
\pi_{2}\, z &:= z_{5}+z_{7}-z_{8}-z_{9} \\
\pi_{3}\, z &:= z_{1}-2 z_{4}+z_{9} &
\pi_{4}\, z &:= -z_{5}+z_{9} \\
\pi_{5}\, z &:= z_{8}-z_{9} &
\pi_{6}\, z &:= z_{7}-z_{9} \\
\pi_{7}\, z &:= z_{1}-z_{4} &
\pi_{8}\, z &:= z_{1} \\
\pi_{9}\, z &:= z_{4} &
\pi_{10}\, z &:= z_{9}
\end{align*}
For example the system
\begin{equation}
\label{eq:ineq_const_tern}
\begin{split}
\pi_{1}\, z < 0,\; \pi_{2}\, z >0,\; \pi_{3}\, z <0,\; \pi_{4}\, z >0,\;
\pi_{5}\, z <0,\\ \pi_{6}\, z >0,\; \pi_{7}\, z > 0,\; \pi_{8}\, z >
0,\; \pi_{9}\, z > 0,\; \pi_{10}\, z < 0 \\
z_1>0,\; z_2>0,\; z_3<0,\; z_4>0,\; z_5<0,\; z_6>0,\; z_7>0,\;
z_8<0,\; z_9<0 \\
\omega_1>0,\; \omega_2>0,\; \omega_3<0,\; \omega_4>0,\;
\omega_5<0,\; \omega_6>0,\; \omega_7>0,\; \omega_8<0,\; \omega_9<0
\end{split}
\end{equation}
is feasible. Let $z\in\ensuremath{I\!\!R}^9$ such that \eqref{eq:eq_const_tern} and
\eqref{eq:ineq_const_tern} hold. Then
\begin{displaymath}
\sign\brac{Q_s(z)} = \left[
\begin{array}{rrrrrrrr}
-1&0&0&1&-1&1&1&-1 \\
0&0&0&1&0&-1&1&-1 \\
0&-1&0&1&-1&1&1&-1 \\
0&0&-1&1&0&1&1&1
\end{array}
\right]
\end{displaymath}
is an $L^+$-matrix (cf. (c,d) of Theorem~\ref{theo:L+-matrices}).
One obtains, for example, the feasible points
\begin{displaymath}
\tilde z=\left(13,\, 2,\, -10,\, 8,\, -6,\, 2,\, 4,\, -4,\, -2
\right)^T,\;
\tilde \omega = \left(1,\, 1,\, -1,\, 6,\, -1,\, 1,\, 1,\, -2,\, -1
\right)^T.
\end{displaymath}
Vectors $\tilde z$ and $\tilde \omega$ yield the state vector
\begin{equation}
\label{eq:x_tern}
\tilde x = \frac{\tilde\omega}{\tilde z} = \left(
\frac{1}{13},\,
\frac{1}{2},\,
\frac{1}{10},\,
\frac{3}{4},\,
\frac{1}{6},\,
\frac{1}{2},\,
\frac{1}{4},\,
\frac{1}{2},\,
\frac{1}{2}
\right)
\end{equation}
For the matrix $Q(z)$ evaluated at $\tilde z$ one has
\begin{displaymath}
Q(\tilde z) = \left[
\begin{array}{rrrrrrrrrrr}
-13&0&0&0&0&0&2&-8&2&6&-2 \\
0&0&0&0&0&0&4&0&-4&6&-2 \\
0&-5&0&0&0&0&5&-5&5&5&-2 \\
0&0&-2&0&0&0&4&0&4&4&2 \\
\end{array}
\right].
\end{displaymath}
The kernel of $Q(\tilde z)$ contains the following positive vector
\begin{displaymath}
\tilde\nu = \left(
18,\,
1,\,
2000,\,
1,\,
1,\,
1,\,
1,\,
\frac{246}{35},\,
\frac{5114}{105},\,
\frac{1996}{5},\,
\frac{23146}{21}
\right)^T.
\end{displaymath}
Vectors $\tilde x$ and $\tilde \nu$ yield the rate constants
\begin{equation}
\label{eq:k_tern}
\begin{split}
\tilde k &= \Biggl(
\frac{1121}{15},\,
234,\,
2,\,
\frac{178204}{3},\,
\frac{4}{3},\,
\frac{328}{35},\,
40000,\,
\frac{72006}{5},\,
\frac{10228}{35},\,
\frac{1303760}{63},\\
&\qquad
2,\,
\frac{65146}{21},\,
\frac{8004}{5},\,
4,\,
\frac{7984}{5},\,
\frac{92668}{7},\,
2,\,
\frac{46292}{21}
\Biggr)^T.
\end{split}
\end{equation}
A numerical continuation with this $\tilde k$ and initial
condition $\tilde x$ verifies that at the dynamical system undergoes a
saddle-node bifurcation at ($\tilde k$, $\tilde x$), cf.\
Fig.~\ref{fig:conti_tc}(a).
\vspace{2mm}
For the network given in \eqref{eq:net_bin} one obtains using the
stoichiometric matrix $S$ as given in Appendix~\ref{sec:data-bin}:
\begin{flushleft}
\scalebox{0.5}{
\begin{minipage}{1.0\linewidth}
\begin{displaymath}
Q(z) = \left[
\begin{array}{cccccccccccc}
-z_{1} & 0 & 0 & 0 & 0 & 0 & -z_{4} & -z_{9} & z_{7}-z_{9}
& -z_{9} & -z_{4}+z_{8}-z_{9} & z_{8}-z_{9} \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & z_{2}-z_{9} & z_{7}-z_{9} &
z_{5}-z_{9} & z_{8}-z_{9} & z_{8}-z_{9} \\
0 & -z_{1}-z_{3}+z_{9} & 0 & 0 & 0 & 0 & 0 &
-z_{1}-z_{3}+z_{9} & -z_{1}-z_{3}+z_{9} &
-z_{1}-z_{3}+z_{9} & -z_{1}-z_{3}+z_{9} &
-z_{1}-z_{3}+z_{9} \\
0 & 0 & z_{1}+z_{3}-z_{4} & 0 & 0 & 0 & z_{1}+z_{3}-z_{4}
& 0 & 0 & 0 & -z_{4}+z_{8} & -z_{4}+z_{8} \\ 0 & 0 & 0 &
z_{2}+z_{3}-z_{5} & 0 & 0 & 0 & 0 & 0 & z_{2}+z_{3}-z_{5}
& z_{2}+z_{3}-z_{8} & z_{2}+z_{3}-z_{8} \\
0 & 0 & 0 & 0 & z_{2}+z_{6}-z_{7} & 0 & 0 & 0 &
z_{2}+z_{6}-z_{7} & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & z_{5}+z_{6}-z_{8} & 0 & 0 & 0 & 0 &
z_{5}+z_{6}-z_{8} & z_{5}+z_{6}-z_{8}
\end{array}
\right]
\end{displaymath}
\end{minipage}
}
\end{flushleft}
\vspace{2mm}\noindent From rows 3, 6 and 7 of $Q(z)$ one has that
positive $\nu$ with $Q(z)\, \nu=0$ exist only if
\begin{equation}
\label{eq:eq_const_bin}
-z_{1}-z_{3}+z_{9}=0,\; z_{5}+z_{6}-z_{8}=0,\; z_{2}+z_{6}-z_{7}=0
\end{equation}
and hence, for example,
\begin{displaymath}
z_1 = -z_3+z_9,\; z_2 = z_5 -z_8 + z_7,\; z_6= -z_5 + z_8\ .
\end{displaymath}
In this case colum 2, 5 and 6 will be the zero column, indicating that
$\nu_2$, $\nu_5$, $\nu_6>0$ are unconstrained. Thus we need only
consider the matrix
\begin{displaymath}
Q_s(z) =
\left[
\begin{array}{ccccccccc}
\pi_{10}\, z & 0 & 0 & -\pi_{11}\, z & -\pi_{12}\, z & \pi_{7}\,
z & -\pi_{12}\, z & \pi_{4}\, z & \pi_{9}\, z \\
0 & 0 & 0 & 0 & \pi_{3}\, z & \pi_{7}\, z & \pi_{8}\, z &
\pi_{9}\, z & \pi_{9}\, z \\
0 & \pi_{5}\, z & 0 & \pi_{5}\, z & 0 & 0 & 0 & \pi_{6}\, z &
\pi_{6}\, z \\
0 & 0 & \pi_{2}\, z & 0 & 0 & 0 & \pi_{2}\, z & \pi_{1}\, z &
\pi_{1}\, z
\end{array}
\right]
\end{displaymath}
with
\begin{align*}
\pi_{1}\, z &= z_{3}+z_{5}+z_{7}-2 z_{8} &
\pi_{2}\, z &= z_{3}+z_{7}-z_{8} \\
\pi_{3}\, z &= z_{5}+z_{7}-z_{8}-z_{9} &
\pi_{4}\, z &= -z_{4}+z_{8}-z_{9} \\
\pi_{5}\, z &= -z_{4}+z_{9} &
\pi_{6}\, z &= -z_{4}+z_{8} \\
\pi_{7}\, z &= z_{7}-z_{9} &
\pi_{8}\, z &= z_{5}-z_{9} \\
\pi_{9}\, z &= z_{8}-z_{9} &
\pi_{10}\, z &= z_{3}-z_{9} \\
\pi_{11}\, z &= z_{4} &
\pi_{12}\, z &= z_{9}
\end{align*}
One obtains the feasible inequality system
\begin{equation}
\label{eq:ineq_const_bin}
\begin{split}
\pi_{1}\, z > 0,\;
\pi_{2}\, z < 0,\;
\pi_{3}\, z > 0,\;
\pi_{4}\, z = 0,\;
\pi_{5}\, z > 0,\;
\pi_{6}\, z < 0,\;
\pi_{7}\, z > 0,\ \\
\pi_{8}\, z > 0,\;
\pi_{9}\, z < 0,\;
\pi_{10}\, z < 0,\;
\pi_{11}\, z < 0,\;
\pi_{12}\, z < 0, \\
z_1 > 0,\;
z_2 > 0,\;
z_3 < 0,\;
z_4 < 0,\;
z_5 > 0,\;
z_6 < 0,\;
z_7 > 0,\;
z_8 < 0,\;
z_9 < 0, \\
\omega_1 > 0,\;
\omega_2 > 0,\;
\omega_3 < 0,\;
\omega_4 < 0,\;
\omega_5 > 0,\;
\omega_6 < 0,\;
\omega_7 > 0,\;
\omega_8 < 0,\;
\omega_9 < 0,
\end{split}
\end{equation}
where $Q(z)$ is an $L^+$-matrix. For example the feasible points
\begin{displaymath}
\tilde z = \left(9,\,
9,\,
-11,\,
-4,\,
2,\,
-8,\,
1,\,
-6,\,
-2 \right)^T,\;
\tilde \omega = \left(
1,\,
1,\,
-1,\,
-1,\,
4,\,
-1,\,
2,\,
-1,\,
-1
\right)^T
\end{displaymath}
define the state vector
\begin{equation}
\label{eq:x_bin}
\tilde x = \left(
\frac{1}{9},\,
\frac{1}{9},\,
\frac{1}{11},\,
\frac{1}{4},\,
2,\,
\frac{1}{8},\,
2,\,
\frac{1}{6},\,
\frac{1}{2}
\right)
\end{equation}
For the matrix $Q(z)$ evaluated at $\tilde z$ one obtains
\begin{displaymath}
Q\left(\tilde z\right) = \left[
\begin{array}{rrrrrrrrrrrr}
-9 & 0 & 0 & 0 & 0 & 0 & 4 & 2 & 3 & 2 & 0 & -4 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 11 & 3 & 4 & -4 & -4 \\
0 & 0 & 2 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & -2 & -2 \\
0 & 0 & 0 & -4 & 0 & 0 & 0 & 0 & 0 & -4 & 4 & 4
\end{array}
\right],
\end{displaymath}
with the positive kernel vector
\begin{displaymath}
\tilde \nu = \left(
297,\,
11,\,
22,\,
11,\,
11,\,
11,\,
440,\,
1,\,
11,\,
451,\,
456,\,
6
\right)^T.
\end{displaymath}
Finally one obtains for the rate constants
\begin{equation}
\label{eq:k_bin}
\begin{split}
\tilde k &= \bigl(
1645,\,
2673,\,
9,\,
22,\,
92664,\,
45738,\,
112,\,
3584,\,
1850,\,
91476,\\
&\qquad
\frac{11}{2},\,
\frac{451}{2},\,
1584,\,
\frac{11}{2},\,
\frac{11}{2},\,
1892,\,
66,\,
2772
\bigr)^T.
\end{split}
\end{equation}
Again, numerical continuation show a saddle node bifurcation at
($\tilde k$, $\tilde x$), cf.\ Fig.\ref{fig:conti_tc}(b).
\begin{figure}
\centering
\subfigure[Continuation for network \eqref{eq:net_ten}]{\includegraphics[width=.4\linewidth]{conti_tc}}
\subfigure[Continuation for network \eqref{eq:net_bin}]{\includegraphics[width=.4\linewidth]{conti_bc}}
\caption{\label{fig:conti_tc} Numerical continuation for the
networks \eqref{eq:net_ten} and \eqref{eq:net_bin} using rate
constants $\tilde k$ and initial condition $\tilde x$ given in
(\ref{eq:x_tern}), (\ref{eq:k_tern}) for network
(\ref{eq:net_ten}) and in (\ref{eq:x_bin}), (\ref{eq:k_bin}) for
network (\ref{eq:net_bin}). In both cases the upper and lower branches
correspond to exponentially stable steady states. The total
concentration $c_1$ is used as a bifurcation parameter.}
\end{figure}
\newpage
|
\section{Introduction}\label{intro}
Quantum spin chains play a fundamental role in the study of many-body systems
and quantum phase transitions. These phenomena take place at
zero temperature, that is in a purely quantum regime, and are induced by the variation of an internal parameter
causing a critical change in the ground state due to level crossing in the energy spectrum \cite{sachdev}.
In recent years, a renewed interest in quantum phase transitions has been developed to understand the behavior
of entanglement near the critical point \cite{osborne,fazio,vidal,vedral}.
On the contrary,
the study of interactions between spins and light is considered as the starting point to introduce
the framework of open systems, where the appearance of
dissipation and decoherence can be understood \cite{weiss}. A prototypical example is the so-called spin-boson model
\cite{leggett}, consisting of a single spin interacting with a multimode electromagnetic radiation,
modeled as distribution of quantum harmonic oscillators. Here, we discuss the case where a series of spins
(arranged in an isotropic $XY$ chain) interacts with one or more bosonic modes.
Despite the importance of spin systems, they are not always directly accessible
experimentally, and efficient experimental quantum simulation methods are needed to study their properties.
It has been shown that spin chains can be efficiently simulated using
internal and external degrees of freedom of trapped
particles \cite{porras04,deng,ciaramicoli,johanning,lewenstein}. In particular, arrays of
laser-cooled trapped ions seem to be very promising from an experimental point of view.
Ions can be trapped with high spatial accuracy,
and their internal states can be manipulated with high precision by means of
the interaction with electromagnetic fields.
The first experimental evidence of coupling between
two-level systems, consisting in the transition from
paramagnetic to ferromagnetic order in a two-spin quantum Ising model has recently been
reported \cite{Friedenauer}. Furthermore, ion traps offer the possibility of engineering
spin-boson coupling \cite{johanning,porras08}.
Since spin chains are also important for practical applications, like quantum communication
protocols \cite{bose}, their simulation would allow estimation of the decoherence
effects which derive from the interaction with the environment \cite{noi}.
As said before, we propose a model which allows to observe the phase transition
in an isotropic $XY$ spin chain coupled with external boson modes. The model we introduce in the following is exactly solvable. The phase transition manifests
itself in a nonanalytical variation of the amplitude of the bosonic field.
In correspondence of this change, the chain acquires a finite magnetization. The physical
implementation of this model can be done within the framework of trapped ions discussed before.
The paper is organized as follows. In Sec. \ref{II} we first introduce the general argument of simulation of spin chains with trapped ions.
Then, we show how it is possible to simulate a particular model, consisting of a chain of spins in contact with a bosonic environment.
The critical properties at zero temperature of this model are discussed in Sec. \ref{III},
while in Sec. \ref{IV} we study the phase diagram at finite temperature
in the case of a single slowly oscillating phonon. Finally, in Sec. \ref{V} we present our conclusions.
\section{Arrays of trapped ions and interaction between spins and bosons}\label{II}
As shown in Refs. \cite{porras04} and \cite{deng}, Coulomb chains in linear Paul traps can
simulate spin-spin interactions. An effective spin-spin Hamiltonian emerges as the result of the interaction of
the internal (electronic) states of ions with the
phononic degrees of freedom generated by Coulomb repulsion. The total Hamiltonian includes a
phonon bath ($H_\nu$), a state-dependent force produced by a set of lasers along the directions $\alpha=x,y,z$ ($H_f$), and an
effective magnetic field that can be generated by forcing transitions between
the internal ion states ($H_m$).
In units of $\hbar=1$, we have (see \cite{porras04})
\begin{eqnarray}
H_\nu&=&\sum_{\alpha,n}\omega_{\alpha,n}a^\dag_{\alpha,n}a_{\alpha,n},\nonumber\\
H_f&=&-2\sum_{\alpha,l}F_\alpha q_{\alpha,l}(1+\sigma^\alpha_l),\nonumber\\
H_m&=&\sum_{\alpha,l}B^\alpha\sigma^\alpha_l,
\end{eqnarray}
where $a_{\alpha,n}$ is the annihilation operator of the $n$ vibrational mode in the $\alpha$
direction of the phonon bath due to the Coulomb repulsion, $F_\alpha$ are the laser forces, $q_{\alpha,l}$ is the displacement, with respect to its equilibrium position, of ion $l$ in direction $\alpha$, and $B^\alpha$ is the effective
magnetic field. Expressing the coordinates in terms of collective modes through the matrices $M^\alpha$,
\begin{equation}
H_f=\sum_{\alpha,l,n}F_\alpha\frac{M^\alpha_{l,n}}{\sqrt{2m\omega_{\alpha,n}}}(a^\dag_{\alpha,n}+a_{\alpha,n})(1+\sigma^\alpha_l).
\end{equation}
A spin-spin Hamiltonian is obtained after the application of a suitable canonical
transformation $U=e^S$, introduced to eliminate $H_f$: if
\begin{equation}
S=\sum_{\alpha,l,n}F_\alpha\frac{M^\alpha_{l,n}}{\sqrt{2m\omega^3_{\alpha,n}}}(a^\dag_{\alpha,n}-a_{\alpha,n})(1+\sigma^\alpha_l),
\end{equation}
the total system Hamiltonian turns out to be
\begin{equation}\label{eqn:heff}
\tilde{H}=UHU^{\dag}=H_\nu+\frac{1}{2}\sum_{\alpha,l,j}J^\alpha_{l,j}\sigma_l^{\alpha}\sigma_j^{\alpha}+\sum_{\alpha,l}B^{\prime\alpha}\sigma_l^{\alpha}+H_E,
\end{equation}
where $B^{\prime\alpha}=B^{\alpha}-F_\alpha^2/(m\omega_\alpha^2)$, the coupling parameters $J$ are function of $F,M,\omega$, and $H_E$
is a residual term that
can be neglected at low temperatures or by using highly anisotropic traps \cite{porras04}.
A particular case of Eq. (\ref{eqn:heff}) is represented by the
isotropic $XY$ chain in the presence of a transverse field, numerically studied
in Ref. \cite{deng}.
Here, we show that a different interpretation of the unitary transformation allows us to obtain other kinds of phase
transitions, involving, for instance, the interaction between spins and bosons. A possible way to do that
consists in applying different transformations along the three axes. In fact, by limiting the
transformation given before to the directions $x$ and $y$, through
\begin{equation}
S_{x,y}=\sum_{\alpha=x,y}\sum_{l,n}F_\alpha\frac{M^\alpha_{l,n}}{\sqrt{2m\omega^3_{\alpha,n}}}(a^\dag_{\alpha,n}-a_{\alpha,n})(1+\sigma^\alpha_l),
\end{equation}
we obtain
\begin{eqnarray}
e^{-S_{x,y}}He^{S_{x,y}}&=&
\sum_{l,n}F_z\frac{M^z_{l,n}}{\sqrt{2m\omega_{z,n}}}(a^\dag_{z,n}+a_{z,n})(1+\sigma^z_l) \nonumber\\
&&+\frac{1}{2}\sum_{\alpha=x,y}\sum_{l,j}J^\alpha_{l,j}\sigma_l^{\alpha}\sigma_j^{\alpha}+\sum_{l}B^{z}\sigma_l^{z}\nonumber\\&&+
\sum_{\alpha=x,y}\sum_l B^{'\alpha}\sigma_l^{\alpha}+H_E^{\prime}+H_\nu,
\end{eqnarray}
Here, the correction $H_E^{\prime}$ can be neglected in the same limit of $H_E$. In the case of a spatially
homogeneous array, we can drop the dependence from the ion position in both $M$ and in $J$.
Moreover, if the Coulomb interaction can be
considered as a perturbation with respect to the trapping potential (stiff limit), the trapping frequencies can be tuned to obtain relevant
spin-spin coupling only between nearest neighbors \cite{porras04}.
Along the transverse direction, a boson displacement can be applied to eliminate
the term $\sum_n F_z M^z_{n}/(\sqrt{2m\omega_{z,n}})(a^\dag_{z,n}+a_{z,n})$.
As a result,
\begin{eqnarray}
\tilde{H}&=&\sum_{\alpha=x,y}\frac{J^\alpha}{2}\sum_{l}\sigma_l^{\alpha}\sigma_{l+1}^{\alpha} +\sum_{n}\frac{g_n^z}{\sqrt{N}}(a^\dag_{z,n}+a_{z,n})\sum_l\sigma^z_l\nonumber\\&&+ (B^{z}-2\sum_{n}\frac{g_n^{z2}}{\omega_n})\sum_l\sigma^z_l+\sum_{\alpha=x,y}\sum_l B^{'\alpha}\sigma_l^{\alpha}\nonumber\\&&+H_\nu-N\sum_{n}\frac{g_n^{z2}}{\omega_n},
\end{eqnarray}
where $g_n^z=F_z M^z_{n}\sqrt{(N/2m\omega_{z,n})}$. The last term is a constant that
translates all the energies, and the fields $B^\prime$ can be set to zero. By fixing the external
magnetic field $B^{z}=2\sum_{n}(g_n^{z2}/\omega_n)$, we arrive to
\begin{eqnarray}
\tilde{H}&=&\frac{1}{2}\sum_{l}(J^x\sigma_l^{x}\sigma_{l+1}^{x}+
J^y\sigma_l^{y}\sigma_{l+1}^{y}) \nonumber\\&&+\sum_{n}\frac{g_n^z}{\sqrt{N}}(a^\dag_{z,n}+a_{z,n})\sum_l\sigma^z_l+
H_\nu-\frac{NB^{z}}{2}\label{htilde}.
\end{eqnarray}
It is worth noting that, since the mutual distance between the ions makes any direct spin-spin interaction negligible, the spins we introduced are effective two-level systems that, because of the unitary
transformation, are not trivially related to the original internal degrees of freedom of the ions. Nevertheless, the bosonic
variables do refer to the real phonon modes along the $z$ direction, so their measurement appears to be the natural way
to get information about the overall system.
\section{Critical properties}\label{III}
To describe the critical properties of this model, let us first discuss the simpler single-mode version.
By assuming $J^x=J^y=J$, the Hamiltonian is
\begin{equation}
H_{sm}=J\sum_{l=1}^N[\sigma_{l}^{+}\sigma_{l+1}^{-}+H.c.]+\omega
a^{\dag}a+\frac{g}{\sqrt{N}}(a^{\dag}+a)\sum_{l=1}^N\sigma_{l}^{z},\label{acca}
\end{equation}
where $N$ is the total number of spin, and where periodic boundary conditions are imposed.
As it is useful to work with adimensional quantities, we use $J$ as energy unit and define the rescaled phonon
frequency $\gamma=\omega / J$ as well as the bare coupling constant $\lambda=g^2/(J \omega)$. Using this notation
the Hamiltonian reads
\begin{equation}
H_{sm}=\sum_{l=1}^N[\sigma_{l}^{+}\sigma_{l+1}^{-}+H.c.]+\gamma
a^{\dag}a+\sqrt{ \frac{\lambda \gamma }{N}}(a^{\dag}+a)\sum_{l=1}^N\sigma_{l}^{z}.\label{acca_ad}
\end{equation}
To diagonalize it, first we introduce the
Jordan-Wigner transformation \cite{lieb},
defined by $\sigma _{l}^{z}=1-2c_{l}^{\dagger }c_{l}$, $\sigma
_{l}^{+}=\prod_{j<l}\left( 1-2c_{j}^{\dagger }c_{j}\right) c_{l}$, and $\sigma _{l}^{-}=\prod_{j<l}\left( 1-2c_{j}^{\dagger }c_{j}\right)
c_{l}^{\dagger }$, mapping spins into spinless fermions. Then, the Fourier transform
\begin{equation}
c_k=\frac{1}{\sqrt{N}}\sum_{l=1}^N e^{i2\pi kl/N}c_l
\end{equation}
allows us to write, in the thermodynamic limit,
\begin{equation}
H_{sm}=\sum_{k}\epsilon_k c_k^\dag c_k+\gamma a^{\dag}a+\sqrt{ \frac{\lambda \gamma}{ N}}(a^{\dag}+a)\sum_{k}( 1-2c_{k}^{\dagger }c_{k}),
\end{equation}
with $\epsilon_k=-2\cos (2\pi k/N)$. Fermions can be decoupled from the bosonic degree of freedom by
applying the displacement operator $D=\exp{[\hat{\alpha}(a^\dag-a)]}$ such that $DaD^{\dag}=b$,
$\hat{\alpha}=-\sqrt{\lambda/(\gamma N)}\sum_{k}( 1-2c_{k}^{\dagger }c_{k})$ being an operator in the fermionic space.
Since $[\hat{\alpha},\sum_{k}\epsilon_k c_k^\dag c_k]=0$, fermions are not transformed into polarons because of
the application of $D$. Moreover, it is immediately seen that $[b,b^\dag]=1$. This displacement replaces the boson vacuum with a coherent state
whose amplitude is determined by the transverse component of the total spin.
The final Hamiltonian is
\begin{equation}
\tilde{H}_{sm}=\sum_{k}\epsilon_k c_k^\dag c_k+\gamma b^{\dag}b-\frac{\lambda}{N}[\sum_{k}(1-2c_{k}^{\dag}c_{k})]^{2}. \label{hd}
\end{equation}
Adding the boson displacement $D$ to $S_{z}$ and $S_{x,y}$ of Sec. \ref{II} amounts to building the total $S$ of Ref. \cite{porras04}
for a particular choice of parameters. If one disregards the term $\gamma b^{\dag}b$, the Hamiltonian \ref{hd}
can now be interpreted as a $XXZ$ spin chain with short-range coupling along the radial direction
and axial long-range coupling. These kinds of interaction ranges are realistic, since radial modes can mediate nearest-neighbor spin-spin interactions
(stiff limit), and transverse modes are known to generate
long-range interactions (soft limit). While the range of interactions along the transverse direction
depends only on the number of ions and cannot be manipulated by means of laser fields, the stiff
limit in the radial direction can be reached by increasing the trapping frequencies \cite{deng}.
Performing the unitary transformations in two separate steps helped us to interpret this model
in terms of the coupling with a real phonon. Without performing the two steps separately, we could not discuss the spin-boson phase transition.
At first sight, $\tilde{H}_{sm}$ is similar to that of the $XX$ chain in a transverse field \cite{katsura}, whose
Hamiltonian in the Jordan-Wigner space would be $\sum_{k}\epsilon_k c_k^\dag c_k-h\sum_{k}(1-2c_{k}^{\dag}c_{k})$. While the latter
model has the $U(1)$ symmetry, since it is invariant under rotations around the $z$ axis, the Hamiltonian (\ref{acca}) has an
additional symmetry, also being invariant under the action of the operator ${\cal S}=\prod_l\sigma_{l}^{x}\otimes\exp{[i\pi a^{\dag}a]}$.
The $U(1)$ symmetry implies, in the language of fermions, that eigenstates of $H$ have a fixed number of particles.
The symmetry ${\cal S}$ could be broken, for instance, by adding a displacing bosonic field, $a+a^\dag$, and then sending it to $0$.
The average
value of $a+a^\dag$ can be assumed as an order parameter to study the breakdown of ${\cal S}$.
Since we are interested in the implementation in an array of ion traps, finite-size effects should be taken into account.
A detailed analysis of $XYZ$ models in finite cycles has been made in several papers \cite{katsura,prb,depa}. In our model,
after the Jordan-Wigner transformation, the Hamiltonian becomes
\begin{eqnarray}
H_{sm}&=&\sum_{l=1}^{N-1}(c_{l+1}c_{l}^{\dag}+c_{l}c_{l+1}^{\dag})-
{\cal P}(c_{1}^{\dag}c_{N}+c_{N}^{\dag}c_{1})+\gamma a^{\dag}a\nonumber\\&&+\sqrt{ \frac{\lambda \gamma}{ N}}(a^{\dag}+a)\sum_{k}( 1-2c_{k}^{\dagger }c_{k}),
\end{eqnarray}
where $N$ is assumed to be an even number and where ${\cal P}=\prod_{l=1}^{N}(1-2c_{l}^{\dag}c_{l})$. Its possible
eigenvalues are $\pm1$. Note that ${\cal P}$ is a measure of the parity of the number of particles of each state.
Since $[H_{sm},{\cal P}]=0$, all eigenstates of $H_{sm}$ have definite parity, and we can proceed to
a separate diagonalization of $H_{sm}$ in the two subspaces corresponding
to ${\cal P}=\pm 1$. Then, if we introduce the two new Hamiltonians $H_{sm}^{\pm}=H_{sm}({\cal P}=\pm1)$, the complete set of
eigenvectors of $H_{sm}$
will be given by the odd eigenstates of $H_{sm}^{-}$ and the even eigenstates of $H_{sm}^{+}$. The choice of ${\cal P}=-1$ or ${\cal P}=+1$
amounts to having, respectively, periodic or antiperiodic boundary conditions. This implies that $H_{sm}^{+}$ and $H_{sm}^{-}$ are diagonalized
through two Fourier transforms which differ from each other because of the set of allowed values of $k$.
If half-integer values ($k=1/2,3/2,\ldots, N
-1/2$) are used to diagonalize $H_{sm}^{+}$, in the case of $H_{sm}^{-}$ we have $k=1,2,\ldots, N$. Now, the displacement operator $D$ can
be introduced, and a structure formally identical to that given in Eq. (\ref{hd}) can be done for both $H_{sm}^{+}$ and $H_{sm}^{-}$, with the
proper choice of $k$.
The ground state of the bosonic part can be identified as the vacuum state of the operator $b$, and it corresponds, in the
original representation, to a coherent state of amplitude equal to the average value of the operator $\hat{\alpha}$ calculated
on the fermionic ground state, whose structure is now enlightened. Since both $H_{sm}^{+}$ and $H_{sm}^{-}$ are exactly solvable, their
spectra can be calculated by measuring the energy of all possible configurations, obtained adding the desired number of fermions.
The energy of a state with $m$ particles is
\begin{equation}
E_{{\cal M}}=-\frac{\lambda(N-2m)^2}{N}-2\sum_{k\in {\cal M}}\cos \frac{2 \pi k}{N},\label{mm}
\end{equation}
where ${\cal M}$ is a string representing the occupation of different modes. The sum is extended only to the values of $k$ where a
particle is present.
For large values of $\lambda$, the first term will be larger than the sum, unless $N-2m$ is very small. The ground state
corresponds to the sequences which maximize $(N-2m)^2$, that is, to $m=0$ or $m=N$, and it is twofold degenerate. In fact, both the
state $|\Phi^{+}\rangle=|0\rangle\otimes|\alpha\rangle$ and the state $|\Phi^{-}\rangle=\prod_{k}c_{k}^{\dag}|0\rangle\otimes|-\alpha\rangle$
have energy equal to $E_0=-N\lambda$. Here the state $|\Phi^{+}\rangle$ is, in the original fermion-boson representation (before the application of $D$), the tensor
product of the fermionic vacuum and of a coherent bosonic state of amplitude $\alpha=-g\sqrt{N}/\omega$, while in $|\Phi^{-}\rangle$
the fermionic system is fully occupied, and the boson coherent state has amplitude $-\alpha$.
As $\lambda$ decreases, the two terms in Eq. (\ref{mm}) start to compete with each other.
Let us call $\lambda_m$ the value such that $E_0=\min\{E_{{\cal M}}\}$. Since, by symmetry, $\lambda_m=\lambda_{N-m}$,
we limit out consideration $m=1,2,\dots,N/2$. It can be shown that, if $m<m'$, then $\lambda_m<\lambda_{m'}$.
In other words, the first ground-state level crossing takes place between $|\Phi^{+}\rangle$ (or $|\Phi^{-}\rangle$)
and a state with $N/2$ particles. This transition occurs at the value of $\lambda_{N/2}=\lambda_c$ determined by
\begin{equation}
N \lambda_c=4 \sum_{k=1}^{N/4}\cos\frac{(2k-1)\pi}{N}.
\end{equation}
In the thermodynamic limit, performed replacing the sum with an integral, we obtain $\lambda_c=2/\pi$.
As a consequence of the half-filling, $\sum_{k}( 1-2c_{k}^{\dagger }c_{k})=0$, and the boson part is left in its vacuum.
This state can be written as $|\Phi^{HF}\rangle=\prod_{|k|<N/4}c_{k}^{\dag}|0\rangle\otimes|0\rangle$.
By further decreasing $\lambda$, no other transitions are observed.
Then the state $|\Phi^{HF}\rangle$ is the Hamiltonian ground state for every $\lambda<\lambda_c$.
It is worth noting that both $|\Phi^{HF}\rangle$ and $ |\Phi^{+}\rangle$ (or $ |\Phi^{-}\rangle$) are eigenstates of $H_{sm}^+$.
In any case, eigenstates of $H_{sm}^-$ are excitations. In Fig. \ref{energie}, we plot the lowest energy levels for a chain of eight spins.
The transition $ |\Phi^{\pm}\rangle\rightarrow |\Phi^{HF}\rangle$ is observed.
While $|\Phi^{HF}\rangle$ is an eigenstate of the symmetry operator ${\cal S}$, this is not true in the case of
$|\Phi^{+}\rangle$ and $|\Phi^{-}\rangle$. In fact, ${\cal S}|\Phi^{\pm}\rangle=|\Phi^{\mp}\rangle$.
The possibility of obtaining a coherent emission of light on the bosonic mode is inherently linked to
the breakdown of the Hamiltonian symmetry, since the only coherent state which is also an
eigenstate of $\exp{[i\pi a^{\dag}a]}$ is the vacuum. While symmetry breaking takes place
independently of the system size (the number of spins), only in the thermodynamic limit does a true phase
transition take place, since the Hilbert spaces that can be built up starting from $|\Phi^{+}\rangle$
and $|\Phi^{-}\rangle$ become unitarily nonequivalent.
\begin{figure}
\includegraphics[height=5cm]{FFL1}\\
\caption{Ground-state energy of $H$ as a function of $\lambda$ for a chain of 8 spins.
Both the energy and $\lambda$ are measured in units of $J$ or $\omega$ (we assume $J=\omega=1$).
For any given number of particles, only the lower level is plotted. The solid green line represents the energy of $|\Phi^{+}\rangle$
and $|\Phi^{-}\rangle$, while the solid black line corresponds to $|\Phi^{HF}\rangle$. As we can see, there is only one transition,
occurring at $\lambda_c\simeq 0.65$. All energies derived from numbers of fermions different from $0,N/2,N$ are excitations.
Specifically, the dotted green line represents the energy of the one-particle state, the dotted red line is for the two-particle state, and the dotted black
line corresponds to the three-particle state. Due to the particle-hole symmetry, states with $m$ and $N-m$ particles have
the same energy.}
\label{energie}
\end{figure}
The exact results we propose have been obtained in the presence of periodic boundary conditions.
By releasing this hypothesis, they are correct only in the thermodynamic limit,
while in the case of finite size, as in the example in Fig. \ref{energie}, corrections to the energy levels are expected.
Actually, the structure of eigenfrequencies giving rise to the phase transition from $|\Phi^{\pm}\rangle$ to $|\Phi^{HF}\rangle$ can be observed also
in the case of a very short open chain ($N=4$).
In the presence of multimode radiation,
Hamiltonian (\ref{htilde}) can be recovered modifying Hamiltonian (\ref{acca}) in the following way
\begin{eqnarray}
H_{mm}&=& \sum_{l=1}^N[\sigma_{l}^{+}\sigma_{l+1}^{-}+H.c.]+
\sum_n\gamma_n a^{\dag}_n a_n\nonumber\\&&+\sum_n\sqrt{\frac{\lambda_n \gamma_n}{N}}(a_n^{\dag}+a_n)\sum_{l=1}^N\sigma_{l}^{z},\label{accamm}
\end{eqnarray}
and it is subject to the symmetry ${\cal S}_{mm}=\prod_l\sigma_{l}^{x}\otimes\exp{[i\pi\sum_n a_n^{\dag}a_n]}$.
The diagonalization is performed through the application of the Jordan-Wigner transformation and of the operator
$\prod_n D_n$, with $D_n=\exp{[\hat{\alpha_n}(a_n^\dag-a_n)]}$, and
$\hat{\alpha_n}=-\sqrt{\lambda_n/(\gamma_n N)}\sum_{k}( 1-2c_{k}^{\dagger }c_{k})$.
As a result, we have
\begin{equation}
\tilde{H}_{mm}=\sum_{k}\epsilon_k c_k^\dag c_k+
\sum_n\omega_n b^{\dag}_n b_n-\frac{\Lambda}{N}[\sum_{k}(1-2c_{k}^{\dag}c_{k})]^{2},\label{hdm}
\end{equation}
where $\Lambda=\sum_ng_n^2/(J \omega_n)$ is related to the spectral density of the bath.
In Ref. \cite{porras08}, the authors showed how baths characterized by different spectral densities can be simulated.
Being $\tilde{H}_{mm}$ formally identical to $\tilde{H}_{sm}$, we expect a phase transition for $\Lambda_c=2/\pi$. For $\Lambda>\Lambda_c$,
the two degenerate ground states are $|\Phi^{+}_{mm}\rangle=|0\rangle\otimes\prod_n|\alpha_n\rangle$ and
$|\Phi^{-}_{mm}\rangle=\prod_{k}c_{k}^{\dag}|0\rangle\otimes\prod_{n}|-\alpha_n\rangle$, while for
$\Lambda<\Lambda_c$ the non degenerate half filled ground state
is $|\Phi^{HF}_{mm}\rangle=\prod_{|k|<N/4}c_{k}^{\dag}|0\rangle\otimes\prod_{n}|0\rangle$.
To clarify the proposal, we comment on the preparation scheme needed to observe phase transitions in this system. As mentioned in \cite{porras04}, the procedure consists
in initialization to the state $\left|\downarrow...\downarrow\right\rangle$, and then adiabatic switching of the spin-spin couplings. This process is repeated
several times for different ratios of $J/g$, while the phase transition is detected through fluorescence of the ions internal levels, as experimentally realized in \cite{Friedenauer}.
Here, we also need to perform motional ground-state cooling as far as axial degrees of freedom are concerned.
The possible errors in simulating this model with trapped ions comes from the canonical transformation and $H_E'$, where the error is quantified by the boson displacements
and the mean phonon number (given by temperature) of the neglected transverse modes. Therefore they have to be cooled down too \cite{porras04}.
\section{Temperature effects in the adiabatic limit}\label{IV}
In this section we analyze the phase transition at finite temperature in the so-called adiabatic
limit ($\gamma\ll 1$) with a single coupled mode.
This limit corresponds to the case of a very slowly oscillating phonon and it is asymptotically
exact at high temperatures.
The adiabatic Hamiltonian is obtained by neglecting the kinetic energy of the phonon and treating
the coordinate as a parameter
\begin{equation}
H_{AD}=\sum_{l=1}^N [\sigma_{l}^{+}\sigma_{l+1}^{-}+H.c.]+\frac{\nu^2}{4 \lambda}+
\frac{\nu}{\sqrt{N}}\sum_{l=1}^N\sigma_{l}^{z},
\end{equation}
where we introduced the adimensional coordinate $\nu=(g/J)\sqrt{2 \omega m} x$.
The Hamiltonian is a isotropic $XY$ model in an external field, parametrically dependent
on $\nu$, it can be diagonalized in the same
way as described previously and reduced in two blocks corresponding to
odd and even pseudofermion occupation number
\begin{equation}
H_\pm=\sum_k \epsilon^{AD}_k(\nu) c^\dag_k c_k +\frac{\nu^2}{4 \lambda}+\frac{\nu}{\sqrt{N}},
\end{equation}
with $\epsilon^{AD}_k(\nu)=-2\left[\cos(2 \pi k/N)+\nu/\sqrt{N}\right]$.
The energy of a configuration with a filling $m$ is
\begin{equation}
E_{{\cal M}}=-2\left(\frac{\nu m}{\sqrt{N}}+\sum_{k\in{\cal M}}\cos\frac{2 \pi}{N}k\right).
\end{equation}
The partition function, defined by
\begin{equation}
Z\propto \int d\nu Z_0(\nu,\beta) \exp{- \beta (\frac{\nu^2}{4 \lambda}+\frac{\nu}{\sqrt{N}})},
\end{equation}
with the adimensional parameter $\beta=J/k_B T$ and where
\begin{equation}
Z_0(\nu,\beta)=\tr{e^{-\beta \sum_k \epsilon^{AD}_k(\nu) c^\dag_k c_k}},
\end{equation}
can be written as $Z \propto \int d\nu \exp{- \beta V_{AD}(\nu,\beta)} $
so defining an adiabatic potential
\begin{equation}
V_{AD}(\nu,\beta)=\frac{\nu^2}{4 \lambda}+\frac{\nu}{\sqrt{N}}+
-\frac{1}{\beta} \ln Z_0(\nu,\beta).
\end{equation}
Above the critical temperature, $V_{AD}(\nu,\beta)$ is minimized by $\nu=0$, while,
at the critical point,
the potential admits two minima, allowing for a displacement of the phonon. The phase diagram
obtained is plotted in Fig. \ref{fig:ph_diag}.
The experimental procedure to measure the phase transition is exactly as in the zero-temperature case.
The radial vibrations should be cooled, while the axial degrees of freedom can be prepared in a ``thermal state''.
The temperature of this state can be controlled by reducing the power of the laser
(resolved sideband) cooling mechanism. To raise it further, the power of the
Doppler laser cooling has to be tuned.
We still expect to observe an abrupt change in fluorescence properties, as in the zero-temperature case, since for moderate temperatures
the coherent state $|\alpha\rangle$ becomes thermal, but the mean number of phonons does not disappear abruptly.
According to Fig. \ref{fig:ph_diag}, the order parameter value should change as a function of $T$.
\begin{figure}
\includegraphics[height=6cm,angle=-90]{phase_diagram_adiab2}\\
\caption{Phase diagram in the adiabatic approximation. Data were obtained with 10 spins and
numerical evaluation of the adiabatic potential.
The left region corresponds to a single-well adiabatic potential,
the right region to a double-well potential.}
\label{fig:ph_diag}
\end{figure}
\section{Conclusions}\label{V}
To resume, we have shown that a system of trapped ions can be mapped into an isotropic $XY$ chain interacting with phonons, following
the scheme in Ref. \cite{porras04}. Modifying the original canonical transformations needed to write a
spin-spin Hamiltonian, the vibrational
degrees of freedom along a fixed direction are coupled with the effective spins the other degrees of freedom are mapped onto.
The resulting model is exactly solvable and exhibits a quantum phase transition, due to the breaking of a symmetry which takes
into account both the spin and the boson degrees of freedom. In this phase, phonons are in a coherent state with
finite amplitude. Since the phonons refer to the real ionic vibration and not to
an effective quantity, phase transition detection should be possible using state-of-the-art techniques.
The experimental realization of the building block of such
simulations (Ref. \cite{Friedenauer}), encourages this kind of investigation.
\acknowledgments
GLG acknowledges discussion with F. de Pasquale.
SP acknowledges discussion with A. Sanpera and G. de Chiara.
This work was partially funded by CoQuSys project (200450E566). GLG and SP are supported by the Spanish Ministry of Science and Innovation
through the program Juan de la Cierva.
|
\section{Introduction}\label{intro}
\noindent Lattice reduction algorithms are essential tools in computational number theory and cryptography. A lattice is a discrete subset of $\mathbb{R}^n$ that is also a $\mathbb{Z}$-module. The goal of lattice reduction is to find a `nice' basis for a lattice, one which is near orthogonal and composed of short vectors. Since the publication of the 1982 Lenstra, Lenstra, Lov\'asz~\cite{LLL} lattice reduction algorithm many applications have been discovered, such as polynomial factorization~\cite{LLL,Hoeij} and attacking several important public-key cryptosystems including knapsack cryptosystems~\cite{knapsack}, RSA under certain settings~\cite{rsa}, and DSA and some signature schemes in particular settings~\cite{dsa}. One of the important features of the LLL algorithm was that it could approximate the shortest vector of a lattice in polynomial time. This is valuable because finding the exact shortest vector in a lattice is provably NP-hard~\cite{ajtai,svp}. Given a basis ${\bf b}_ 1, \ldots, {\bf b}_ d \in \mathbb{R}^n$ which satisfies $\parallel {\bf b}_{i} \parallel \leq X ~~\forall i$, the LLL algorithm has a running time of $\mathcal{O}(d^5 n \log^3{X})$ using classical arithmetic. Recently there has been a resurgence of lattice reduction work thanks to Nguyen and Stehl\'e's ${\rm L}^2$ algorithm~\cite{fpLLL,fpLLL2} which performs lattice reduction in $\mathcal{O}(d^4 n \log{X} [d+\log{X}])$ CPU operations. The primary result of ${\rm L}^2$ was that the dependence on $\log{X}$ is only quadratic allowing for improvement on applications using large input vectors.
\noindent \textbf{The main result:} Many applications of LLL (see the applications section below) involve finding a vector in a lattice whose norm is known to be small in advance. In such cases it can be more efficient to reduce a basis of a sub-lattice which contains all targeted vectors than reducing a basis of the entire lattice. In this paper we target short vectors in specific types of input lattice bases which we call knapsack-type bases. The new algorithm introduces a search parameter $B$ which the user provides. This parameter is used to bound the norms of targeted short vectors. To be precise:
\vspace{.1in}
The {\em rows} of the following matrices represent a knapsack-type basis
\begin{center}
$ \left( \begin{array}{ccc|ccc}
0 & \cdots & 0 & 0 & \cdots & P_N\\
0 & \cdots & 0 & 0 & \Ddots & 0\\
0 & \cdots & 0 & P_1 & \cdots & 0\\
\hline
1 & \cdots & 0 & x_{1,1} &\cdots & x_{1,N}\\
\vdots & \ddots & \vdots & \vdots & & \vdots \\
0 & \cdots & 1 & x_{r,1} &\cdots & x_{r,N}\\
\end{array} \right)$
\textrm{ or }
$ \left( \begin{array}{ccc|ccc}
1 & \cdots & 0 & x_{1,1} &\cdots & x_{1,N}\\
\vdots & \ddots & \vdots & \vdots & & \vdots \\
0 & \cdots & 1 & x_{r,1} &\cdots & x_{r,N}\\
\end{array} \right)$.
\end{center}
\noindent The specifications of our algorithm are as follows. It takes as input a knapsack-type basis ${\bf b}_ 1, \ldots, {\bf b}_ d \in \mathbb{Z}^{n}$ of a lattice $L$ with $\parallel {\bf b}_{i} \parallel \leq X$ $\forall i$ and a search parameter $B$; it returns a \em{reduced} basis generating a \em{sub-lattice} $L' \subseteq L$ such that if ${\bf v} \in L$ and $\parallel {\bf v} \parallel \leq B$ then $ {\bf v} \in L'$.
\noindent Our algorithm has the following complexity bounds for various input:
\vspace{.1in}
\begin{center}
\begin{tabular}{|c|c|}
\hline
No $P_i$ &\raisebox{11pt}{~}$\mathcal{O}(d^2(n+d^2)(d+\log{B})[\log{X}+n(d+\log{B})])$ \\
\hline
No restriction on $P_i$ &\raisebox{11pt}{~}$\mathcal{O}(d^4(d+\log{B})[\log{X}+d(d+\log{B})])$ \\
\hline
Many $P_i$ large w.r.t. $B$ &\raisebox{11pt}{~} $\mathcal{O}(dr^3(r+\log{B})[\log{X}+d(r+\log{B})])$\\
\hline
\end{tabular}
\end{center}
\vspace{.1in}
\noindent These complexity bounds have several distinct parameters, so a comparison with other algorithms is a bit subtle. The most significant parameter to explore is $B$, the search parameter. If one selects $B = X$ then our algorithm will return a reduced basis of $L'=L$ in $\mathcal{O}(d^2n(n+d^2)[d^2+ \log^2{X}])$. This is an interesting result because our algorithm, like the original LLL and the ${\rm L}^2$ algorithms, uses switches and size-reductions of the vectors to arrive at a reduced basis. The fact that we return a reduced basis with a complexity so similar to ${\rm L}^2$ implies that there are alternative orderings on the switches which lead to similar performance.
\noindent When using a smaller value of $B$ than $X$ the algorithm will return either:
\begin{itemize}
\item A reduced basis of a sub-lattice $L'$ which contains all vectors of norm $\leq B$. This sub-lattice may be different than the sub-lattice, $L''$, generated by all vectors of norm $\leq B$, and we do have $L'' \subseteq L' \subseteq L$. Also, because the basis of $L'$ is reduced, we have an approximation of the shortest non-zero vector of $L$.
\item The empty set, in which case the algorithm has proved that no non-zero vector of norm $\leq B$ exists in $L$.
\end{itemize}
\noindent We offer the following complexity comparison with ${\rm L}^2$~\cite{fpLLL} for some values of $B$ on square input lattices (with $P_j$'s). When a column has a non-zero $P_j$ we can reduce the $x_{i,j}$ modulo $P_j$. Thus, without loss of generality, we may assume that $P_j$ is the largest element in its column. Note that $r=d-N$.
\begin{center}
\begin{tabular}{|c|c|}
\hline
${\rm L}^2$ & \raisebox{11pt}{~}$\mathcal{O}(d^6\log{X}+d^5\log^2{X})$\\
\hline
\hline
$B=\mathcal{O}(X)$ & \raisebox{11pt}{~}$\mathcal{O}(d^7+d^5\log^2{X})$\\
\hline
$B=\mathcal{O}(X^{1/d})$ &\raisebox{11pt}{~} $\mathcal{O}(d^2r^5+r^3\log^2{X})$\\
\hline
$B=2^{\mathcal{O}(d)}$ &\raisebox{11pt}{~} $\mathcal{O}(d^4r^3+d^2r^3\log{X})$ \\
\hline
\end{tabular}
\end{center}
\noindent It should be noted that~\cite{fpLLL} explores running times of ${\rm L}^2$ on knapsack lattices with $N=1$ (such lattice bases are used in~\cite{goldstein}). In this case, ${\rm L}^2$ will have complexity $\mathcal{O}(d^5\log{X}+d^4\log^2{X})$.
\noindent \textbf{Our approach:} We reduce the basis gradually, using many separate calls to another lattice reduction algorithm. To get the above complexity results we chose H-LLL~\cite{HLLL} but there are many suitable lattice reduction algorithms we could use instead such as~\cite{kaltofen,LLL,fpLLL,schnorr2}. For more details on why we made this decision see the discussion in section~\ref{compsect}.
There are three important features to our approach. First, we approach the problem column by column. Beginning with the $r\times r$ identity and with each iteration of the algorithm we expand our scope to include one more column of the $x_{i,j}$. Next, within each column iteration, we reduce the new entries bit by bit, starting with a reduction using only the most significant bits, then gradual including more and more bits of data. Third, we allow for the removal of vectors which have become too large. This allows us to always work on small entries, but restricts us to a sub-lattice.
The proof of the algorithm's complexity is essentially a study of two quantities, the product of the Gram-Schmidt lengths of the current vectors which we call the active determinant and an energy function which we call progress. We amortize all of the lattice reduction costs using progress, and we bound the number of iterations and number of vectors using the active determinant. Neither of these quantities is impacted by the choice of lattice reduction algorithm.
\noindent \textbf{Applications of the algorithm:} As evidence for the usefulness of this new approach we show two new complexity results based on applications of the main algorithm. The first result is a new complexity for the classical problem of factoring polynomials in $\mathbb{Z}[x]$. If the polynomial has degree $N$, coefficients smaller than $\log(A)$, and when reduced modulo a prime $p$ has $r$ irreducible factors then we prove a complexity of $\mathcal{O}(N^3r^4+N^2r^4\log{A})$ for the lattice reduction costs using classical arithmetic. One must also add the cost of multi-factor Hensel lifting which is $\mathcal{O}(N^6+N^4 \log^2{A})$ ignoring the small terms $\log(r)$ and $\log^2{p}$ (see~\cite{mca} for details). This is the first improvement over the Sch\"onhage bound given in 1984~\cite{Schonhage} of $\mathcal{O}(N^8+N^5\log^3{A})$.
The second new complexity result comes in the problem of reconstructing a minimal polynomial from a complex approximation of the algebraic number. In this application we know $\mathcal{O}(d^2+d\log{H})$ bits of an approximation of some complex root of an unknown polynomial $h(x)$ with degree $d$ and with maximal coefficient of absolute value $\leq H$. Then our algorithm can be used to find the coefficients of $h(x)$ in $\mathcal{O}(d^7 + d^5 \log^2{H})$ CPU operations.
Other problems of common interest which might be impacted by our algorithm include integer relation finding (where $N = 1$) and simultaneous Diophantine approximation of several real numbers \cite{hanrot,bright} (where $r = 1$).
\noindent \textbf{Notations:} All costs are given for the bit-complexity model. A standard row vector will be denoted ${\bf v}$, ${\bf v}[i]$ represents the $i^{\textrm{th}}$ entry of ${\bf v}$, ${\bf v}[i, \ldots, j]$ a vector consisting of all entries of ${\bf v}$ from the $i^{\textrm{th}}$ entry to the $j^{\textrm{th}}$ entry, and ${\bf v}[-1]$ the final entry of ${\bf v}$. Also we will use $\| {\bf w} \|_{\infty}$ as the max-norm or the largest absolute value of an entry in the vector ${\bf w}$, $\parallel {\bf w} \parallel :=\sqrt{ \sum ({\bf w}[i])^2}$ which we call the norm of ${\bf w}$, and ${\bf w}^{T}$ as the transpose of ${\bf w}$. The scalar product will be denoted ${\bf v} \cdot {\bf w}:= \sum {\bf v}[i] \cdot {\bf w}[i]$. For a matrix $M$ we will use $M[1,\ldots, k]$ to denote the first $k$ columns of $M$. The $n$ by $n$ identity matrix will be denoted $I_{n \times n}$. For a real number $x$ we use $\lceil x \rceil$ and $\lfloor x \rfloor$ to denote the closest integer $\geq x$ and $\leq x$ respectively.
\noindent \textbf{Road map:} In section~\ref{background} we give a brief introduction to lattice reduction algorithms. In section~\ref{mainalgsec} we present the central algorithm of the paper and prove its correctness. In section~\ref{adsect} we prove several important features by studying quasi-invariants we call the active determinant and progress. In this section we treat lattice reduction as a black-box algorithm. In section~\ref{compsect} we prove the overall complexity and other important claims about the new algorithm by fixing a choice for a standard lattice reduction algorithm. In section~\ref{newcomplexities} we offer new complexity results for factoring polynomials in $\mathbb{Z}[x]$ and algebraic number reconstruction.
\section{Background on lattice reduction}\label{background}
\noindent The purpose of this section is to present some facts from~\cite{LLL} that
will be needed throughout the paper. For a more general treatment of lattice reduction see~\cite{lovasz}.
A lattice, $L$, is a discrete subset of $\mathbb{R}^n$ that is also a $\mathbb{Z}$-module.
Let ${\bf b}_1,\ldots,{\bf b}_d \in L$ be a basis of $L$ and denote ${{\bf b}_1^*},\ldots,{{\bf b}_d^*} \in \mathbb{R}^n$ as the
Gram-Schmidt orthogonalization over $\mathbb{R}$ of ${{\bf b}_ 1},\ldots,{{\bf b}_ d}$. Let $\delta \in (1/4, 1]$ and $\eta \in [1/2, \sqrt{\delta})$.
Let $l_i = \log_{1/\delta}{\parallel {\bf b}_{i}^* \parallel}^2$,
and denote $\mu_{i,j}=\frac{{\bf b}_{i} \cdot {\bf b}_ j^*}{{\bf b}_ j^*
\cdot {\bf b}_ j^*}$. Note that ${\bf b}_{i}, {\bf b}_{i}^*, l_i, \mu_{i,j}$ will change throughout the algorithm sketched below.
\begin{definition}\label{reducedbasis}
${{\bf b}_ 1}, \ldots, {{\bf b}_ d}$ is \emph{LLL-reduced}
if ${\parallel {\bf b}_{i}^* \parallel}^2 \leq \frac{1}{\delta-\mu_{i+1,i}^2}{\parallel {{\bf b}_{i+1}^*} \parallel}^2$ for $1 \leq i < d$ and $|\mu_{i,j}| \leq \eta$ for $1 \leq j < i \leq d$.
\end{definition}
In the original paper the values for $(\delta, \eta)$ were chosen as $(3/4, 1/2)$ so that $\frac{1}{\delta-\eta^2}$ would simply be 2.
\begin{alg}[Rough sketch of LLL-type algorithms]\label{LLL} \mbox{} \\
\noindent \emph{Input:} A basis ${{\bf b}_ 1},\ldots,{{\bf b}_ d}$ of a lattice $L$. \\
\noindent \emph{Output:} An LLL-reduced basis of $L$.
\begin{enumerate}[A -]
\item $\kappa:= 2$
\item \textbf{while} $\kappa \leq d$ \textbf{do:}
\begin{enumerate}[1 -]
\item{\em{(Gram-Schmidt over $\mathbb{Z}$)}}.\label{subtract} By subtracting suitable $\mathbb{Z}$-linear
combinations of ${{\bf b}_ 1},\ldots,{{\bf b}_ {\kappa-1}}$ from ${{\bf b}_{\kappa}}$ make sure that $| \mu_{i,\kappa} | \leq \eta$ for $i<\kappa$.
\item{\em{(LLL Switch)}}.\label{swap} If interchanging ${{\bf b}_{\kappa-1}}$ and ${{\bf b}_{\kappa}}$ will
decrease $l_{\kappa-1}$ by at least 1 then do so.
\item{\em{(Repeat)}}. If not switched $\kappa:=\kappa+1$, if switched $\kappa = \textrm{max}(\kappa-1,2)$.
\end{enumerate}
\end{enumerate}
\end{alg}
That the above algorithm terminates, and that the output is
LLL-reduced was shown in~\cite{LLL}. Step \ref{subtract} has no effect on the $l_i$. In step~\ref{swap} the only $l_i$ that change
are $l_{\kappa-1}$ and $l_{\kappa}$. The following lemmas present some standard facts which we will need.
\begin{lemma}\label{maxlll} An LLL switch can not increase $\max(l_1, \ldots, l_d)$,
nor can it decrease $\min(l_1, \ldots, l_d)$.
\end{lemma}
\begin{lemma}\label{Bbound} If ${\parallel {\bf b}_d^* \parallel} > B$ then any vector in $L$ with norm
$\leq B$ is
a $\mathbb{Z}$-linear combination of ${{\bf b}_ 1}, \ldots, {{\bf b}_ {d-1}}$.
\end{lemma}
In other words, if the current basis of the lattice is ${{\bf b}_ 1},\ldots,{{\bf b}_ d}$ and if the last vector
has sufficiently large G-S length then, provided the user is only interested in elements of $L$ with norm $\leq B$, the last basis element can be removed.
Lemma~\ref{Bbound} follows from the proof of~\cite[Eq.\ (1.11)]{LLL}, and is true
regardless of whether ${{\bf b}_ 1},\ldots,{{\bf b}_ d}$ is LLL-reduced or not.
However, if one chooses an arbitrary basis ${{\bf b}_ 1},\ldots,{{\bf b}_ d}$ of some lattice $L$, then it is unlikely that the last vector has large G-S length (after all, $\parallel {{\bf b}_ d}^* \parallel$
is the norm of ${{\bf b}_ d}$ reduced modulo ${{\bf b}_ 1},\ldots,{{\bf b}_ {d-1}}$ over $\mathbb{R}$).
The effect of LLL reduction is to move G-S length towards later
vectors.
\section{Main algorithm}\label{mainalgsec}
In this section we present the central algorithm of the paper and a proof of its correctness. Our algorithm is a kind of wrapper for other standard lattice reduction algorithms. We try to present it as independently as possible of the choice of lattice reduction algorithm. In order to be general we must first outline the features that we require of the chosen lattice reduction algorithm. Our first requirement is that the output satisfy the following slightly weakened version of LLL-reduction.
\begin{definition}\label{alphareduced}
Let $L \subseteq \mathbb{R}^n$ be a lattice and ${{\bf b}_ 1},\ldots, {{\bf b}_ s} \in L$ be $\mathbb{R}$-linearly independent. We call ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ an $\alpha$-reduced basis of $L$ if \ref{one},\ref{two}, and \ref{threea} hold, and an $(\alpha,B)$-reduced sequence (basis of a sub-lattice) if \ref{one},\ref{two}, and \ref{threeb} hold:
\begin{enumerate}
\item \label{one} $\parallel {\bf b}_{i}^* \parallel \leq \alpha \parallel {\bf b}_{i+1}^* \parallel$ for $i=1 \ldots s-1$.
\item \label{two} $\parallel {\bf b}_{i}^* \parallel \leq \parallel {\bf b}_{i} \parallel \leq \alpha^{i-1} \parallel {\bf b}_{i}^* \parallel$ for $i=1 \ldots s$.
\item \begin{enumerate}
\item \label{threea} $L=\mathbb{Z}{{\bf b}_ 1} +\cdots + \mathbb{Z}{{\bf b}_ s}$.
\item \label{threeb} $\parallel {\bf b}_s^* \parallel \leq B$ and for every ${\bf v} \in L$ with $\parallel {\bf v} \parallel \leq B$ we have ${\bf v} \in \mathbb{Z}{{\bf b}_ 1} + \cdots + \mathbb{Z}{{\bf b}_ s}$.
\end{enumerate}
\end{enumerate}
\end{definition}
The original LLL algorithm from~\cite{LLL} returns output with $\alpha=\sqrt{2}$, ${\rm L}^2$ from~\cite{fpLLL} with $\alpha=\sqrt{ \frac{1}{\delta-\eta^2}}$ for appropriate choices of $(\delta, \eta)$, and H-LLL from~\cite{HLLL} reduced with $\alpha= \frac{ \theta \eta + \sqrt{(1+\theta^2)\delta-\eta^2}}{\delta - \eta^2}$ for appropriate $(\delta, \eta, \theta)$. We may now also make a useful observation about an $(\alpha,B)$-reduced sequence.
\begin{lemma}\label{smallbreduced}
If the vectors ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ form an $(\alpha,B)$-reduced sequence and we let ${\bf b}_1^*, \ldots, {\bf b}_s^*$ represent the GSO, then the following properties are true:
\begin{itemize}
\item $\parallel {\bf b}_{i}^* \parallel \leq \alpha^{s-i} B$ for all $i$.
\item $\parallel {\bf b}_{i} \parallel \leq \alpha^{s-1} B$ for all $i$.
\end{itemize}
\end{lemma}
We use the concept of $\alpha$-reduction as a means of making proofs which are largely independent of which lattice reduction algorithm a user might choose. For a basis which is $\alpha$-reduced, a small value of $\alpha$ implies a strong reduction. In our algorithm we use the variable $\alpha$ as the worst-case guarantee of reduction quality. We make our proofs (specifically Lemma~\ref{ADincrease} and Theorem~\ref{numloops}) assuming an $\alpha \geq \sqrt{4/3}$. This value is chosen because~\cite{LLL,fpLLL,HLLL} cannot guarantee a stronger reduction. An $(\alpha,B)$-reduced bases is typically made from an $\alpha$-reduced basis by removing trailing vectors with large G-S length. The introduction of $(\alpha,B)$-reduction does not require creating new lattice reduction algorithms, just the minor adjustment of detecting and removing vectors above a given G-S length.
\begin{alg}\label{lllwithremovals}
\textbf{LLL\_with\_removals}
\textbf{Input:} ${{\bf b}_ 1}, \ldots, {{\bf b}_ s} \in \mathbb{R}^n $ and $B \in \mathbb{R}$.
\textbf{Output:} ${{\bf b'}_1}, \ldots, {{\bf b'}_{s'}} \in \mathbb{R}^n$ $(\alpha,B)$-reduced, $s' \leq s$.
\textbf{Procedure:} Use any lattice reduction procedure which returns an $\alpha$-reduced basis and follows Assumption~\ref{switchreqs}. However, when it is discovered that the final vector has G-S length provably $>B$ remove that final vector (deal with it no further).
\end{alg}
\begin{ass}\label{switchreqs}
The lattice reduction algorithm chosen for LLL\_with\_removals must use switches of consecutive vectors during its reduction process. These switches must have the following properties:
\begin{enumerate}
\item There exists a number $\gamma > 1$ such that every switch of vectors ${\bf b}_{i}$ and ${{\bf b}_ {i+1}}$ increases $\parallel {\bf b}_{i+1}^* \parallel^2$ by a factor provably $\geq \gamma$.
\item The quantity $\textrm{max}\{ \parallel {\bf b}_{i}^* \parallel, \parallel {\bf b}_{i+1}^* \parallel \}$ cannot be increased by switching ${\bf b}_{i}$ and ${{\bf b}_ {i+1}}$.
\item No steps other than switches can affect G-S norms $\parallel {\bf b}_1^* \parallel, \ldots, \parallel {\bf b}_s^* \parallel $.
\end{enumerate}
\end{ass}
Assumption~\ref{switchreqs} is not very strong as~\cite{LLL,fpLLL,HLLL,schnorr2,arne} and the sketch in Algorithm~\ref{LLL} all conform to these assumptions. We do not allow for the extreme case where $\gamma=1$, although running times have been studied in~\cite{akhavi,flags}. It should also be noted that in the floating point lattice reduction algorithms $\parallel {\bf b}_s ^* \parallel$ is only known approximately. In this case one must only remove vectors whose approximate G-S length is sufficiently large to ensure that the exact G-S length is $\geq B$.
The format of the input matrices was given in section~\ref{intro}. A search parameter $B$ is given to bound the norm of the target vectors. The algorithm performs its best when $B$ is small compared to the bit-length of the entries in the input matrix, although $B$ need not be small for the algorithm to work.
\begin{definition}\label{pj_def} We say the $P_j$ are {\em large enough} if:
\begin{equation}\label{pj}
|P_j| \geq 2\alpha^{4r+4k+2}B^2 \textrm{~for all but $k=\mathcal{O}(r)$ values of $j$}.
\end{equation}
\end{definition}
\noindent Note that if $N=\mathcal{O}(r)$ then the $P_j$ are trivially large enough. However, for applications where $N$ is potentially much larger than $r$ this becomes a non-trivial condition. In this case having $B$ close to $X$ means that the $P_j$'s are not large enough.
\medskip
In the following algorithm we will gradually reduce the input basis. This will be done one column at a time, similar to the experiments in~\cite{Belabas,bright}. The current basis vectors are denoted ${\bf b}_{i}$ and we will use $M$ to represent the matrix whose rows are the ${\bf b}_{i}$. We will use the notation ${{\bf x}_j}$ to represent the column vector $(x_{1,j}, \ldots, x_{r,j})^{T}$.
The matrix $M$ will begin as $I_{r \times r}$, and we will adjoin ${\bf x}_1$ and a new row $({\bf 0}, P_1)$ if appropriate. Each time we add a column ${\bf x}_j$ we will need to calculate the effects of prior lattice reductions on the new ${\bf x}_j$. We use ${\bf y}_j$ to represent a new column of entries which will be adjoined to $M$. In fact ${\bf y}_j = M[1,\ldots, r] \cdot {\bf x}_j$. Before adjoining the entries we also scale them by a power of 2, to have smaller absolute values. This keeps the entries in $M$ at a uniform absolute value. The central loop of the algorithm is the process of gradually using more and more bits of ${\bf y}_j$ until every entry in $M$ is again an integer. No rounding is performed: we use rational arithmetic on the last column of each row. Throughout the algorithm the number of rows of $M$ will be changing. We let $s$ be the current number of rows of $M$. If~\eqref{pj} is satisfied for some $k=\mathcal{O}(r)$ then we can actually bound $s$ by $2r+2k+1$. We use $c$ as an apriori upper bound on $s$, either $c:=2r+2k+1$ or $c:=r+N$. The algorithm has better performance when $c$ is small. We let $L$ represent the lattice generated by the rows of $A$.
\begin{alg}\label{mainalg}
{\rm \textbf{Gradual\_LLL}}
\textbf{Input:} A search parameter, $B \geq \sqrt{5} \in \mathbb{Q}$, an integer knapsack-type matrix, $A$, and an $\alpha \geq \sqrt{4/3}$.
\textbf{Output:} An $(\alpha,B)$-reduced basis ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ of a sub-lattice $L'$ in $L$ with the property that if ${\bf v} \in L$ and ${\parallel {\bf v} \parallel} \leq B$ then ${\bf v} \in L'$.
\end{alg}
\textbf{The Main Algorithm:}
\begin{enumerate}[1 -]
\item\label{setc} \textbf{if}~\eqref{pj} holds set $c:=\textrm{min}(2r+2k+1,r+N)$
\item\label{init} $s:=r; M := I_{r \times r}$
\item\label{clmn} \textbf{for } $j = 1 \ldots N$ \textbf{ do}:
\begin{enumerate}[a -]
\item\label{newent} ${\bf y}_j:= M[1,\ldots,r] \cdot {\bf x}_j$; $\ell :=\lfloor \log_2{(\textrm{max}\{|P_j|, \|{\bf y}_j\|_{\infty}, 2 \})} \rfloor$
\item\label{adjoin} $M := \left[ \begin{array}{c|c} 0 & P_j/2^{\ell} \\ \hline M & {\bf y}_j/{2^{\ell}} \\ \end{array} \right]$; \textbf{if }$P_j \neq 0$\textbf{ then }$s := s+1$ \textbf{else} remove zero row
\item\label{mainloop}\textbf{while } $(\ell \neq 0)$ \textbf{ do:}
\begin{enumerate}[i -]
\item\label{newprec} ${\bf y}_j := 2^{\ell} \cdot M \cdot [ 0, \cdots, 0,1]^{T}$; $\ell := \textrm{max} \{ 0, \lceil \log_2 {( \frac{\|{\bf y}_j\|_{\infty}}{{\alpha}^{2c}B^2} )} \rceil \}$
\item\label{scale} $M:= \left[ \begin{array}{c|c} M[1, \ldots, r+j-1] & {{\bf y}_j}/{2^{\ell}} \end{array} \right]$
\item\label{singlelll} Call $\textrm{LLL\_with\_removals}$ on $M$ and set $M$ to output; adjust $s$
\end{enumerate}
\end{enumerate}
\item \textbf{return} $M$
\end{enumerate}
First we will prove the correctness of the algorithm. We need to show that the Gram-Schmidt lengths are never decreased by scaling the final entry or adding a new entry.
\begin{lemma}\label{scalenondecrease}
Let ${\bf b}_1, \ldots, {\bf b}_s \in \mathbb{R}^n$ be the basis of a lattice and ${\bf b}_1^*, \ldots, {\bf b}_s^*$ its GSO. Let $\sigma:\mathbb{R}^{n} \to \mathbb{R}^{n}$ scale up the last entry by some
factor $\beta >1$, then we have $\parallel {\bf b}_i^* \parallel \leq \parallel {\sigma({\bf b}_i)}^* \parallel$. In other words, scaling the final entry of each vector by the same scalar $\beta > 1$ cannot decrease $\|{\bf b}_i^*\|$ for any $i$.
\end{lemma}
\begin{lemma}\label{extraentry}
Let ${\bf b}_1, \ldots, {\bf b}_s \in \mathbb{R}^{n}$ and let ${\bf b}_1^*, \ldots, {\bf b}_s^* \in \mathbb{R}^{n}$ be their GSO. The act of adjoining an ${(n+1)}^{\textrm{st}}$ entry to each vector and re-evaluating the GSO cannot decrease $\parallel {\bf b}_i^* \parallel$ for any $i$ (assuming that the new entry is in $\mathbb{R}$).
\end{lemma}
The proofs of these lemmas are quite similar and can be found in the appendix. Now we are ready to prove the first theorem, asserting the correctness of algorithm~\ref{mainalg}'s output.
\begin{theorem}\label{taccuracy}
Algorithm~\ref{mainalg} correctly returns an $\alpha$-reduced basis of a sub-lattice, $L'$, in $L$ such that if ${\bf v} \in L$ and $\parallel {\bf v} \parallel \leq B$ then ${\bf v} \in L'$.
\end{theorem}
\begin{proof}
When the algorithm terminates all entries are unscaled and each vector in the output is inside of $L$ as it is a linear combination of the original input vectors. Thus the output is a basis of a sub-lattice $L'$ inside $L$. Further, the algorithm terminates after a final call to step~\ref{singlelll} so returns an $(\alpha,B)$-reduced sequence.
Now we show that if ${\bf v} \in L$ and $\parallel {\bf v} \parallel \leq B$ then ${\bf v} \in L'$. The removed vectors correspond to vectors $\tilde{ {\bf b}_i } \in L$ that, by lemmas~\ref{scalenondecrease} and~\ref{extraentry}, have G-S length at least as large as those of ${\bf b}_i$. The claim then follows from lemmas~\ref{maxlll} and~\ref{Bbound}.
\end{proof}
\section{Two invariants of the algorithm}
Here we present the important proofs about the set-up of our algorithm. All proofs in this section and the next allow for a black-box lattice reduction algorithm up to satisfying assumption~\ref{switchreqs}. Each proof in this section involves the study of an invariant. The two invariants which we use are:
\begin{itemize}
\item The Active Determinant, $\textrm{AD}(M)$, which is the product of the G-S lengths of the active vectors. This remains constant under standard lattice reduction algorithms, and allows us to bound many features of the proofs.
\item The Progress, $PF= \sum_{i=1}^s (i-1) \log\parallel {\bf b}_{i}^* \parallel^2 + n_{\textrm{rm}} r \log(4\alpha^{4c}B^4)$, where $n_{\textrm{rm}}$ is the total number of vectors which have been removed so far. This function is an energy function which never decreases, and is increased by $\geq 1$ for each switch made in the lattice reduction algorithm.
\end{itemize}
\subsubsection*{A study of the active determinant}\label{adsect}
\begin{definition}
We call the active determinant of the vectors ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ the product of their Gram-Schmidt lengths. For notation we use, AD or $\textrm{AD}(\{ {\bf b}_{i} \} ):= \prod_{i=1}^s {\parallel {\bf b}_{i}^* \parallel}$. For a matrix $M$ with the $i^{\textrm{th}}$ row denoted by $M[i]$, we use $\textrm{AD}$ or $\textrm{AD}(M)=\textrm{AD}(\{M[1], \ldots, M[s]\})$.
\end{definition}
For an $(\alpha,B)$-reduced sequence we can nicely bound the AD. We have such a sequence after each execution of step~\ref{singlelll}.
\begin{lemma}\label{ADreduced}
If ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ are an $(\alpha,B)$-reduced sequence then $\textrm{AD} \leq {({\alpha}^{s-1}B^2)}^{s/2}$.
\end{lemma}
We now want to attack two problems, bounding the norm of each vector just before lattice reduction, and bounding the number of vectors throughout the algorithm.
\begin{lemma}\label{smallvectors}
If $s \leq c$ then just before step~\ref{singlelll} we have $\parallel {\bf b}_i \parallel^2 \leq 2\alpha^{4c}B^4$ for $i=1 \ldots s$.
\end{lemma}
The full details of this proof can be found in the appendix. The following theorem holds trivially when there is no condition on the $P_j$ or if $N=0$. When $N > r$ and $B$ is at least a bit smaller than $X$ we can show that not all of the extra vectors stay in the lattice. In other words, if there is enough of a difference between $B$ and $X$ then the sub-lattice aspect of the algorithm begins to allow for some slight additional savings. Here the primary result of this theorem is allowing $\mathcal{O}(r)$ vectors with a relatively weak condition on the $P_j$.
\begin{theorem}\label{activevectors}
Throughout the algorithm we have $s \leq c$.
\end{theorem}
\begin{proof}
If $c=r+N$ then $s \leq c$ is vacuously true. So assume $c= 2(r+k)+1$ and all but $k=\mathcal{O}(r)$ of the $P_j$ satisfy $|P_j| \geq 2\alpha^{4r+4k+2}B^2$. When the algorithm begins, $\textrm{AD}=1$ and $s=r$. For $s$ to increase step~\ref{clmn} must finish without removing a vector. If this happens during iteration $j$ then the $\textrm{AD}$ has increased by a factor $|P_j|$. The LLL-switches inside of step~\ref{singlelll} do not alter the AD by Assumption~\ref{switchreqs}. Each vector which is removed during step~\ref{singlelll} has G-S length $\leq 2\alpha^{4r+4k+2}B^2$ by Lemmas~\ref{smallvectors} and~\ref{maxlll}. After iteration $j$ we have $n_{\textrm{rm}}=r+j-s$ as the total number of removed vectors. All but $k$ of the $P_i$ have larger norm than any removed vector. Therefore the smallest $AD$ can be after iteration $j$ is $\geq {(2\alpha^{(4r+4k+2)}B^2)}^{j-k-n_{\textrm{rm}}}$. Rearranging we get $\textrm{AD} \geq {(2\alpha^{4r+4k+2}B^2)}^{s-r-k}$. This contradicts Lemma~\ref{ADreduced} when $s$ reaches $2r+2k$ for the first time because $(2\alpha^{4r+4k+2}B^2)^{r+k} \geq {(\alpha^{2r+2k-1}B^2)}^{r+k}$. \end{proof}
\begin{corollary}\label{smallvectors2}
Throughout the algorithm we have $\parallel {\bf b}_{i}^* \parallel \leq 2\alpha^{2c}B^2$.
\end{corollary}
We also use the active determinant to bound the number of iterations of the main loop, i.e. step~\ref{mainloop}. First we show in the appendix that AD is increased by every scaling which does not end the main loop.
\begin{lemma}\label{ADincrease}
Every execution of step~\ref{scale} either increases the $\textrm{AD}$ by a factor $\geq \frac{{\alpha}^{c}B}{2}$ or sets $\ell=0$.
\end{lemma}
Now we are ready to prove that the number of iterations of the main loop is $\mathcal{O}(r+N)$. This is important because it means that, although we look at all of the information in the lattice, the number of times we have to call lattice reduction is unrelated to $\log{X}$.
\begin{theorem}\label{numloops}
The number of iterations of step~\ref{mainloop} is $\mathcal{O}(r+N)$.
\end{theorem}
The strategy of this proof is to show that each succesful scaling increases the active determinant and to bound the number of iterations using Lemma~\ref{ADreduced} and Corollary~\ref{smallvectors2}. For space constraints this proof is provided in the appendix.
\subsubsection*{A study of the progress function}\label{psect}
We will now amortize the costs of lattice reduction over each of the $\mathcal{O}(r+N)$ calls to step~\ref{singlelll}. We do this by counting switches, using Progress $PF$ (defined below). In order to mimic the proof from~\cite{LLL} for our algorithm we introduce a type of Energy function which we can use over many calls to LLL (not only a single call).
\begin{definition}
Let ${{\bf b}_ 1}, \ldots, {{\bf b}_ s}$ be the current basis at any point in our algorithm, let ${\bf b}_1^*, \ldots, {\bf b}_ s^*$ be their GSO, and $l_i := \log_{\gamma}{\parallel {\bf b}_{i}^* \parallel^2}$ for all $i=1\ldots s$. We let $n_{\textrm{rm}}$ be the number of vectors which have been removed so far in the algorithm. Then we define the progress function $PF$ to be:
$$PF:= 0 \cdot l_0 + \cdots + (s-1)\cdot l_s + n_{\textrm{rm}}\cdot c \cdot \log_{\gamma}{(4\alpha^{4c}B^4)}.$$
\end{definition}
This function is designed to effectively bound the largest number of switches which can have occurred so far. To prove that it serves this purpose we must prove the following lemma:
\begin{lemma}\label{Pincreases}
After step~\ref{init} Progress $PF$ has value 0. No step in our algorithm can cause the progress $PF$ to decrease. Further, every switch which takes place in step~\ref{singlelll} must increase $PF$ by at least 1.
\end{lemma}
\begin{theorem}\label{switchcomplexity}
Throughout our algorithm the total number of switches used by all calls to step~\ref{singlelll} is $\mathcal{O}((r+N)c(c+\log{B}))$ with $P_j$ and $\mathcal{O}(c^2(c+\log{B}))$ with no $P_j$.
\end{theorem}
\begin{proof}
Since Lemma~\ref{Pincreases} shows us that $PF$ never decreases and every switch increases $PF$ by at least 1, then the number of switches is bounded by $PF$. However $PF$ is bounded by Lemma~\ref{smallvectors} which bounds $l_i \leq \log_{\gamma}{(\alpha^{4r}B^4)}$, Theorem~\ref{activevectors} which bounds $s \leq c$, and the fact that we cannot remove more vectors than are given which implies $n_{\textrm{rm}} \leq r+N$. Further we can see that $(s-1)l_s \leq (c-1)\log_{\gamma}{(4\alpha^{4c}B^4)}$ so $PF$ is maximized by making $n_{\textrm{rm}} = (r+N)$ (or $c$ if no vectors added) and $s=0$. In which case we have $\textrm{number of switches} \leq PF \leq (r+N)(c-1)(\log_{\gamma}{(4\alpha^{4c}B^4)} = \mathcal{O}((r+N)c(c+\log{B}))$. Also if there are no $P_j$, we can replace $r+N$ by $c$.
\end{proof}
\section{Complexity bound of main algorithm}\label{compsect}
In this section we wish to prove a bound for the overall bit-complexity of algorithm~\ref{mainalg}. The complexity bound must rely on the complexity bound of the lattice reduction algorithm we choose for step~\ref{singlelll}. The results in the previous sections have not relied on this choice. We will present our complexity bound using the H-LLL algorithm from~\cite{HLLL}. We choose H-LLL for this result because of its favorable complexity bound and because the analysis of our necessary adaptations is relatively simple. See~\cite{HLLL} for more details on H-LLL.
We make some minor adjustments to the H-LLL algorithm and its analysis. The changes to the algorithm are the following: \begin{itemize}
\item We have a single non-integer entry in each vector of bit-length $\mathcal{O}(c+\log{X})$.
\item Whenever the final vector has G-S length sufficiently larger than $B$, it is removed. This has no impact on the complexity analysis.
\end{itemize}
We use $\tau$ as the number of switches used in a single call to H-LLL. This allows the analysis of progress $PF$ to be applied directly. The following theorem is an adaptation of the main theorem in~\cite{HLLL} adapted to reflect our adjustments.
\begin{theorem}\label{taucost}
If a single call to step~\ref{singlelll}, with H-LLL~\cite{HLLL} as the chosen variation of LLL, uses $\tau$ switches then the CPU cost is bounded by $\mathcal{O}((\tau+c+\log{B})c^2[(r+N)(c+\log{B})+\log{X}])$ bit-operations.
\end{theorem}
Now we are ready to complete the complexity analysis of the our algorithm.
\begin{theorem}\label{overallcomplexity}
The cost of executing algorithm~\ref{mainalg} with H-LLL~\cite{HLLL} as the variant of LLL in step~\ref{singlelll} is $$ \mathcal{O}((r+N)c^3(c+\log{B})[\log{X}+(r+N)(c+\log{B})])$$ CPU operations, where $B$ is a search parameter chosen by the user, $|A[i,j]| \leq X$ for all $i,j$, and $c=r+N$ or $c=\mathcal{O}(r)$ (see definition~\ref{pj_def} for details). If there are no $P_j$'s then the cost is $$\mathcal{O}((r+N+c^2)(c+\log{B})c^2[\log{X}+(r+N)(c+\log{B})]).$$
\end{theorem}
\begin{proof}
Steps~\ref{init},~\ref{adjoin},~\ref{newprec}, and~\ref{scale} have negligible costs in comparison to the rest of the algorithm. Step~\ref{newent} is called $N$ times, each call performs $s$ inner products. While each inner product performs $r$ multiplications each of the form ${\bf b}_{i}[m] \cdot x_{m,j}$ appealing to Corollary~\ref{smallvectors2} we bound the cost of each multiplication by $\mathcal{O}( (c+\log{B}) \log{X})$. Since Theorem~\ref{activevectors} gives $s \leq c$ we know that the total cost of all calls to step~\ref{newent} is $\mathcal{O}(Ncr(c+\log{B})\log{X})$. Let $k =\mathcal{O}(r+N)$ be the number of iterations of the main loop. Let $\tau_i$ be the number of LLL switches used in the $i^{\textrm{th}}$ iteration. Theorem~\ref{taucost} gives the cost of the $i^{\textrm{th}}$ call to step~\ref{singlelll} as $= \mathcal{O}((\tau_i+c+\log{B})c^2 [(r+N)(c+\log{B})+\log{X}])$. Theorem~\ref{switchcomplexity} implies that $\tau_1 + \cdots + \tau_k =\mathcal{O}((r+N)c(c+\log{B}))$ (or $\mathcal{O}(c^2(c+\log{B}))$ when there are no $P_j$'s). The total cost of all calls to step~\ref{singlelll} is then $\mathcal{O} ([k(c+\log{B})+\tau_1 + \cdots + \tau_k ]c^2 [(r+N)(c+\log{B})+\log{X}])$. The term $[k(c+\log{B})+\tau_1 + \cdots + \tau_k]$ can be replaced by $\mathcal{O}((r+N)c(c+\log{B}))$ (if no $P_j$ then $\mathcal{O}((r+N+c^2)(c+\log{B}))$). The complete cost of is now $\mathcal{O}(Nrc(c+\log{B})\log{X} + (r+N)c^3[c+\log{B})(\log{X}+(r+N)(c+\log{B})])$. The first term is absorbed by the cost of the second term, proving the theorem. If there are no $P_j$ then we get $\mathcal{O}((r+N+c^2)(c+\log{B})c^2[\log{X}+(r+N)(c+\log{B})])$.
\end{proof}
\section{New complexities for applications of main algorithm}\label{newcomplexities}
Our algorithm has been designed for some applications of lattice reduction. In this section we justify the importance of this algorithm by directly applying it to two classical applications of lattice reduction.
\subsubsection*{New complexity bound for factoring in $\mathbb{Z}[x]$}\label{polynomials}
In~\cite{BHKS} it is shown that the problem of factoring a polynomial, $f \in \mathbb{Z}[x]$, can be accomplished by the reduction of a large knapsack-type lattice. In this subsection we merely apply our algorithm to the lattice suggested in ~\cite{BHKS}.
\vspace{.1in}
\noindent \textbf{Reminders from~\cite{BHKS}.} Let $f \in \mathbb{Z}[x]$ be a polynomial of degree $N$. Let $A$ be a bound on the absolute value of the coefficients of $f$. Let $p$ be a prime such that $f \equiv
l_f f_1 \cdots f_r \/~\mathrm{mod}\/~p^a$ a separable irreducible factorization of $f$ in the $p$-adics lifted to precision $a$, the $f_i$ are monic, and $l_f$ is the leading coefficient of $f$. For our purposes we choose $B:=\sqrt{r+1}$.
We will make some minor changes to the All-Coefficients matrix defined in~\cite{BHKS}
to produce a matrix that looks like:
$$\left( \begin{array}{cccccc}
& & & & & p^{a-b_N} \\
& & & & \Ddots & \\
& & & p^{a-b_1} & & \\
1 & & & x_{1,1} & \cdots & x_{1,N} \\
& \ddots & & \vdots & \ddots & \vdots \\
& & 1 & x_{r,1} & \cdots & x_{r,N}
\end{array} \right).$$
Here $x_{i,j}$ is the $j^{\textrm{th}}$ coefficient of $f'_i \cdot f / f_i \textrm{ mods }p^a$ divided by $p^{b_j}$ and $p^{b_j}$ represents $\sqrt{N}$ times a bound on the $j^{\mathrm{th}}$ coefficient of $g' \cdot f/g$ for any true factor $g \in \mathbb{Z}[x]$ of $f$. In this way the target vectors will be quite small. An empty spot in this matrix represents a zero entry. This matrix has $p^{a-b_j} > 2^{N^2+N\log(A)} > 2{\alpha}^{4r+2}B^2$ for all $j$. An $(\alpha,B)$-reduction of this matrix will solve the recombination problem by a similar argument to the one presented in~\cite{BHKS} and refined in~\cite{phd}. Now we look at the computational complexity of making and reducing this matrix which gives the new result for factoring inside $\mathbb{Z}[x]$.
\begin{theorem}\label{polytheorem}
Using algorithm~\ref{mainalg} on the All-Coefficients matrix above provides a complete irreducible factorization of a polynomial $f$ of degree $N$, coefficients of bit-length $\leq \log{A}$, and $r$ irreducible factors when reduced modulo a prime $p$ in $$\mathcal{O}(N^2r^4[N+\log{A}])$$ CPU operations. The cost of creating the All-Coefficients matrix adds $\mathcal{O}(N^4[N^2+\log^2{A}])$ CPU operations using classical arithmetic (suppressing small factors $\log{r}$ and $\log^2{p}$) to the complexity bound.
\end{theorem}
The following chart gives a complexity bound comparison of our algorithm with the factorization algorithm presented by Sch\"onhage in~\cite{Schonhage} we estimate both bounds using classical arithmetic and fast FFT-based arithmetic~\cite{fft}. We also suppress all $\log{N}$, $\log{r}$, $\log{p}$, and $\log \log {A}$ terms.
\begin{center}
\begin{tabular}{|c|c|}
\hline
Classical Gradual\_LLL & \raisebox{11pt}{~}$\mathcal{O}(N^3r^4 + N^2r^4\log{A} + N^6 + N^4\log^2{A})$\\
\hline
Classical Sch\"onhage & \raisebox{11pt}{~}$\mathcal{O}(N^8+N^5\log^3{A})$\\
\hline
Fast Gradual\_LLL & \raisebox{11pt}{~}$\mathcal{O}(N^3r^3+N^2r^3\log{A})$\\
\hline
Fast Sch\"onhage & \raisebox{11pt}{~}$\mathcal{O}(N^6+N^4\log^2{A})$\\
\hline
\end{tabular}
\end{center}
The Sch\"onhage algorithm is not widely implemented because of its impracticality. For most polynomials, $r$ is much smaller than $N$. Our main algorithm will reduce the All-Coefficients matrix with a competitive practical running time, but constructing the matrix itself will require more Hensel lifting than seems necessary in practice. In~\cite{phd} a similar switch-complexity bound to section~\ref{psect} is given on a more practical factoring algorithm.
\subsubsection*{Algebraic number reconstruction}\label{lenstra}
The problem of finding a minimal polynomial from an approximation of a complex root was attacked in~\cite{lenstra} using lattice reduction techniques using knapsack-type bases. For an extensive treatment see~\cite{lovasz}.
\begin{theorem}\label{algnumrecon}
Suppose we know $\mathcal{O}(d^2+d\log{H})$ bits of precision of a complex root $\alpha$ of an unknown irreducible polynomial, $h(x)$, where the degree of $h$ is $d$ and its maximal coefficient has absolute value $\leq H$. Algorithm~\ref{mainalg} can be used to find $h(x)$ in $\mathcal{O}(d^7 + d^5\log^2{H})$ CPU operations.
\end{theorem}
This new complexity is an improvement over the ${\rm L}^2$ algorithm which would use $\mathcal{O}(d^9 + d^7\log^2{H})$ CPU operations to reduce the same lattice. Although, one can prove a better switch-complexity with a two-column knapsack matrix by using~\cite[Lem.\ 2]{hanrot} to bound the determinant of the lattice as $\mathcal{O}(X^2)$ and thus the potential function from~\cite{LLL} is $\mathcal{O}(X^{2d})$, leading to a switch complexity of $\mathcal{O}(d\log{X})$ (posed as an open question in~\cite[sec.\ 5.3]{stehle}). Using this argument the complexity for ${\rm L}^2$ is reduced to $\mathcal{O}(d^8 + d^6\log^2{H})$.
\textbf{Acknowledgements.} We thank Damien Stehl\'e, Nicolas Brisebarre, and Val\'erie Berth\'e for many helpful discussions. Also Ivan Morel for introducing us to H-LLL. This work was partially funded by the LaRedA project of the Agence Nationale de la Recherche, it was also supported in part by a grant from the National Science Foundation. It was initiated while the second author was hosted by the Laboratoire d'Informatique de Robotique et de Micro\'electronique de Montpellier (LIRMM).
|
\section{Introduction}
The scattering theory of polarised neutrons including magnetic interactions
started with the early work of
by Halpern and Johnson \cite{halpern} and has been essentially completed by Blume\cite{blume} and Maleyev\cite{maleyev} more than fourty years ago.
The major milestones in development of experimental means and applications have been set by technique of
longitudinal polarisation analysis (Moon {\em et al.} \cite{Moon})
and by spherical neutron polarimetry (Tasset and Brown)
\cite{tasset,Brown2001},
which nowadays provides a high precision tool to analyse the full polarisation tensor for single crystal magnetic scattering.
The development for diffuse scattering utilizing a wide angle detector has been essentially driven by Sch\"arpf, who established with {\em his} D7 instrument at the ILL and the so-called XYZ method\cite{xyzTB,xyz} efficient means for measuring and analysing spin-correlations
in powder samples. With recent instrumental progress, the D7 at ILL \cite{Stewart} and the DNS at JCNS \cite{ws_dns2000}
have gained the required efficiency to explore more routinely the diffuse magnetic scattering from single crystals.
To overcome the pending problem that the XYZ-method is invalid for
the separation of the more complex scattering from single crystals,
here an appropriate extension of the XYZ-method
is derived using the symmetry properties of the polarised scattering.
The present study revisits the polarised scattering from single crystals with the useful achievement of a proper separation based on the XYZ-method.
that will certainly be useful for exploring the complexity of disorder involving magnetic properties.
\section{Polarised neutron scattering and symmetry}
According to Blume \cite{blume} and Maleyev \cite{maleyev}
the neutron scattering process including magnetic interactions can be completely described by two master equations, Eq.(1) for the
scattering cross-section, here for brevity denoted by the intensity $I$, and Eq.(2) for $\mathbf P' I$,
where $\mathbf P'$ denotes the final polarization:
\begin{eqnarray}
I &=& N^{\dag} N +I_{si} + \mathbf M_{\perp} ^{\dag} \mathbf M _{\perp}
+ \mathbf P \cdot \mathbf M_{\perp} ^{\dag} N
+ \mathbf P \cdot \mathbf M_{\perp} N ^{\dag}
+ \mathrm i \mathbf P (\mathbf M_{\perp} ^{\dag} \times \mathbf M _{\perp} )
\\
\mathbf P' I &=& \mathbf P ( N^{\dag} N -\frac{1}{3} I_{si})
+(\mathbf P \cdot \mathbf M_{\perp} ^{\dag})\mathbf M_{\perp}
+(\mathbf P \cdot \mathbf M_{\perp} )\mathbf M_{\perp} ^{\dag}
- \mathbf P (\mathbf M_{\perp} ^{\dag} \mathbf M_{\perp} ) \\
&&+\mathrm i N(\mathbf P \times \mathbf M_{\perp} ^{\dag} )
-\mathrm i N ^{\dag} (\mathbf P \times \mathbf M_{\perp} )
+N \mathbf M_{\perp} ^{\dag} +N ^{\dag} \mathbf M_{\perp}
- \mathrm i (\mathbf M_{\perp} ^{\dag}\times \mathbf M_{\perp} ) \;,
\nonumber
\end{eqnarray}
where
$
N(\mathbf Q) = \sum_n b_n {\exp}(\mathrm{i} {\bf Q}\cdot {\bf R_n})
$
is the nuclear structure factor and
$
\mathbf M_\perp = \mathbf {e_Q} \times \mathbf M (\mathbf Q) \times \mathbf {e_Q}
$,
is the vector of dipolar magnetic interaction of magnetic moments with the neutron spin, and $ \mathbf {e_Q}$ is
the unit vector along the direction of the scattering vector $\mathbf Q$.
Accordingly, the parallel components to the scattering vector $\mathbf Q$ do not contribute to magnetic scattering.
Here,
$
\mathbf M (\mathbf Q) =
\sum_n \mathbf M_n {\exp}(\mathrm{i} {\bf Q}\cdot {\bf R_n})
$
denotes the Fourier transform of the magnetic moments.
$I_{si}$ denotes the nuclear spin-incoherent scattering, assuming that the nuclear spins are randomly oriented.
This diffuse background, which is usually small compared to magnetic and nuclear Bragg peaks,
can be relatively large when considering diffuse magnetic and diffuse nuclear scattering.
The full information about the scattering terms in Eqs.~(1,2) can be retrieved by spherical neutron polarimetry \cite{Brown2001}.
Therefore, the most convenient and standard choice of a $x$,$y$,$z$-coordinate system for the setting of the polarisation $\mathbf P$
is to have one axis $x$ parallel to $\mathbf Q$, with the axes $y$ and $z$ perpendicular to $\mathbf Q$ pointing in- and out-of the scattering plane
respectively.
However, when using multi-detectors $\mathbf P$ can be set ideally parallel to $\mathbf Q$ only for
a single detector. Therefore, the XYZ-polarisation analysis for multi-detectors
the axes $x$ and $y$ are chosen to be in the (horizontal) scattering plane and $z$ to point (vertically) out of the scattering plane.
In an excellent contribution to the present subject, the polarised scattering has been revisited and treated
within the density matrix formalism by Sch\"arpf \cite{Schaerpf_school}.
Following Ref.\cite{Schaerpf_school} we consider first the general case of an arbitrarily rotated $x$,$y$,$z$-system
and the polarised intensities that can be measured in spin-flip and non-spin flip modes and their relation to the above equations.
With the definition of polarised intensities and polarisation
\begin{eqnarray}
I=I_{\nu \nu} +I_{ \nu \bar \nu} \;\;\;\; ,\;\;\;\; P'_{\nu} I = I_{\nu \nu} - I_{\nu \bar \nu}
\end{eqnarray}
for any cartesian coordinate $\nu=x,y,z$, the non-spin flip (\textit{nsf}) and spin-flip intensities (\textit{sf}), $I _{\nu \nu }$ and $I _{\nu \bar \nu }$ respectively,
have been derived \cite{Schaerpf_school} from Eqs.~(1) to (3):
\begin{eqnarray}
I _{\nu \nu }&=& N^{\dag} N + N M_{\perp \nu }^{\dag} +N^{\dag} M_{\perp \nu } + M_{\perp \nu } ^{\dag} M_{\perp \nu } +\frac{1}{3}I_{si} \\
I _{\nu \bar \nu}&=& \mathbf M_{\perp } ^{\dag} \mathbf M_{\perp } - M_{\perp \nu} ^{\dag} M_{\perp \nu} +\mathrm i (\mathbf M_{\perp } ^{\dag} \times \mathbf M_{\perp })_{\nu} +\frac{2}{3}I_{si} \;.
\nonumber
\end{eqnarray}
Eq.~4 shows a well known result namely that only components of $\mathbf M_{\perp}$ parallel to the neutron spin appear in the non-spin flip scattering and only components of $\mathbf M_{\perp}$ perpendicular
to the neutron spin can contribute to the spin-flip scattering
and is exemplified for $\nu=x$
\begin{eqnarray*}
I _{x \bar x}&=& M_{\perp y} ^{\dag} M_{\perp y} + M_{\perp z} ^{\dag} M_{\perp z} +\mathrm i ( M_{\perp y } ^{\dag} M_{\perp z} - M_{\perp z } ^{\dag} M_{\perp y}) +\frac{2}{3}I_{si} \;.
\end{eqnarray*}
It is further worthwhile to note that with any specific choice of the coordinate system the average of non-spin flip and spin-flip intensities
differ for the different directions
$$I_{x} = \frac{1}{2} ( I_{xx} +I_{x \bar x}) \neq I_y \neq I_z .$$
So far we followed the analysis of Sch\"arpf. For a further analysis and separation of terms in the polarised neutron scattering, here,
we continue with separating intensities in symmetric and antisymmetric parts.
Therefore, we consider scattering for reversed polarisation by the intensities $ I _{\bar \nu \bar \nu } $ and $ I _{\bar \nu \nu } $
which are experimentally likewise accessible, and we obtain the analogue
to Eq.~(4) from Eqs.~(1-3):
\begin{eqnarray}
I _{\bar \nu \bar \nu }&=& N^{\dag} N - N M_{\perp \nu }^{\dag} -N^{\dag} M_{\perp \nu } + M_{\perp \nu } ^{\dag} M_{\perp \nu } +\frac{1}{3}I_{si} \\
I _{\bar \nu \nu}&=& \mathbf M_{\perp } ^{\dag} \mathbf M_{\perp } - M_{\perp \nu} ^{\dag} M_{\perp \nu} -\mathrm i (\mathbf M_{\perp } ^{\dag} \times \mathbf M_{\perp })_{\nu} +\frac{2}{3}I_{si} \;.
\nonumber
\end{eqnarray}
In comparison to Eq.~4 and 5, a further simplification is achieved by considering the average non-spin-flip and spin flip intensities, $\Sigma^{nsf}_{\nu}$ and $\Sigma^{sf}_{\nu}$ respectively:
\begin{eqnarray}
\Sigma^{nsf}_{\nu} = \frac{1}{2} ( I _{\nu \nu} + I _{\bar \nu \bar \nu}) &=& N^{\dag} N + M_{\perp \nu} ^{\dag} M_{\perp \nu} +\frac{1}{3}I_{si} \\
\Sigma^{sf}_{\nu} = \frac{1}{2} ( I _{\nu \bar\nu } +I _{\bar\nu \nu })&=& \mathbf M_{\perp } ^{\dag} \mathbf M_{\perp } - M_{\perp \nu} ^{\dag} M_{\perp \nu} + \frac{2}{3}I_{si}
\nonumber
\end{eqnarray}
showing the expected property of the unpolarised intensity $I$
\begin{eqnarray}
I=\Sigma_{x} =\Sigma_{y} =\Sigma_{z} = \Sigma^{nsf}_{\nu} +\Sigma^{sf}_{\nu}
&=& N^{\dag} N + \mathbf M_{\perp } ^{\dag} \mathbf M_{\perp } +I_{si} \;. \nonumber
\end{eqnarray}
By this we have eliminated the interference terms related to nuclear-magnetic
correlations (spin-orbit coupling) and vector chirality terms (cross products) that now appear in
the corresponding averaged deviations $\Delta^{nsf}_{\nu}$ and $\Delta^{sf}_{\nu}$ respectively, :
\begin{eqnarray}
\Delta^{nsf}_{\nu}=\frac{1}{2} ( I _{\nu \nu} - I _{\bar \nu \bar\nu })
&=&N M_{\perp \nu }^{\dag} +N^{\dag} M_{\perp \nu } =2 \Re( N M_{\perp \nu }^{\dag})\\
\Delta^{sf}_{\nu } =
\frac{1}{2} ( I _{\nu \bar \nu } - I _{\bar \nu \nu })
&=& 2 \mathrm i (\mathbf M_{\perp } ^{\dag} \times \mathbf M_{\perp })_\nu \;. \nonumber
\end{eqnarray}
Eq.~(7) reveals a well known result that is the basis of experiments with polarized neutrons using polarisation reversal
without polarisation analysis.
For the purpose of separating the interference terms due to chirality or due to nuclear-magnetic
correlations, polarisation analysis is actually
not required and can be achieved by setting $\nu$, with polarisation reversal, either parallel (for chirality) or perpendicular (for nuclear-magnetic interference)
to the scattering vector $\mathbf Q$.
However, as can be seen, Eq.~(7) also provides a complete separation of these terms for any choice of the cartesian coordinate system
if polarisation analysis is applied.
Finally, one may note that it is also straightforward
to derive this symmetry decomposition from Brown's tensor representation of polarised scattering \cite{Brown2001}
for the case of the standard coordinate system ($x || \mathbf Q$) and also to include spin-incoherent scattering.
\section{XYZ-polarisation analysis, powders and single crystals}
The standard XYZ-method has been successfully applied in many studies of magnetic correlations using powder samples.
From the directional dependence of the polarised scattering, the purely magnetic scattering contribution can be separated
by experimental means of polarisation analysis. For further interest the reader is referred to a recent and detailed review \cite{Stewart}
and a recent own study \cite{kagome}.
Here we shall briefly recall the fundamental equations (Eq.~8), and discuss the additional potential of instruments like DNS and D7 for
studies of powder samples with polarisation analysis.
An interesting consequence is that the interference of nuclear and
induced magnetic scattering could be observed even for powders if chiral domains relate to satellites with different $Q$ moduli or in presence of a symmetry breaking external field, whereas
usually, chiral scattering is found in single crystals with a preferred domain orientation.
In the following (3.2)
the central issue will be to separate and analyse the different contributions to the polarised scattering from single crystals
based on the conventional tools of the standard XYZ method.
Hence, the separation will be achieved without approximations only from diagonal intensity elements.
\subsection{ The application to powders}
In case that orientational averaging applies, e. g. for powder samples,
we do not distinguish the components $M_{\perp \nu}$,
and the magnetic scattering can be separated
via the XYZ polarisation analysis \cite{xyzTB,xyz}
from the diagonal intensities.
In case of only weak guide fields
one does not need to distinguish the intensities with respect
to the sign of polarisation
\noindent
\raisebox{0mm}{\hspace{0mm}\begin{minipage}[b]{33pc}
as proposed in the previous section. It is a particular virtue of the XYZ-method that this
method applies even if the Cartesian coordination system
is arbitrarily rotated in the scattering plane, so that the new coordinates are $x',y',z$
(see figure).
In case of multi-detectors covering a larger Q-range it is trivially impossible to have the polarisation simultaneously parallel to all different Q vectors;
the difference between the polarisation axis $x'$ and the actual $\mathbf Q$ vector
is the \textit{Sch\"arpf angle} $\alpha$.
\end{minipage}}\hspace{1pc}\includegraphics[width=5.8pc,angle=90]{alpha.pdf}
The information about the in-plane magnetic intensities, the square of magnetic amplitudes
$\mathbf M _{\perp y}$,
is recovered from
$ \mathbf M _{\perp y}^2= \mathbf M _{\perp x'}^2+ \mathbf M _{\perp y'}^2 = \mathbf M _{\perp y}^2 (\sin^2 \alpha + \cos^2 \alpha) $.
Hence, the magnetic, nuclear coherent and spin-incoherent scattering can be separated by combinations of polarised intensities \cite{xyzTB,xyz,Stewart}
\begin{eqnarray}
|\mathbf M_\perp|^2
&=& 2(I_{x'\bar x'}+I_{y'\bar y'}-2I_{z\bar z})^{sf}
= - 2(I_{x' x'}+I_{y' y'}-2I_{z z})^{nsf}
\\
I_{si} & =& {3\over 2} \left(
-I_{x'\bar x'} -I_{y'\bar y'}+ 3 I_{z\bar z}
\right) \nonumber
\\ N^2 &=& I_{z z} -\frac{1}{2} |\mathbf M_\perp|^2
- \frac{1}{3} I_{si} \nonumber
\end{eqnarray}
Here, we have assumed that the intensities do not change for a common polarisation reversal of initial and final polarisation.
More carefully, for the application of this analysis we have to consider whether the antisymmetric part actually vanishes.
This is fulfilled only if there is no distinguished direction by any external field. Guide fields applied in the XZY method are
typically in the order of magnitude of ten Gauss and they are too weak to induce any significant sample magnetization in paramagnetic
or antiferromagnetic samples.
Strong fields invalidate the standard XYZ separation method, however, this allows for the study of
the correlations and interference terms between nuclear and magnetic scattering amplitudes by determining the antisymmetric
part of the polarised intensity, \textit{id est} measuring intensity differences for polarisation reversal. A unique response of chiral terms to
an applied horizontal external field is less likely, although not impossible.
A well-known strategy for single crystals used in a ``half-polarised" setup
\cite{schweitzer}
can also be applied to powder diffraction.
Such a determination of correlations between nuclear and magnetic scattering amplitudes from powders looks promising
particularly for detectors covering a large solid angle without polarisation analysis.
With respect to polarisation analysis of scattering it
may be advantageous that only the difference of the \textit{nsf}-intensities contributes to the signal,
and the background from \textit{sf}-scattering could be eliminated.
\subsection{The application to single crystals}
For a single crystal, even without applied external field, its intrinsic anisotropy and possible polarity
may give rise to antisymmetric scattering contributions.
Therefore, the symmetry decomposition resulting in Eqs.~(6,7)
is important and helpful to analyse the polarised scattering and its magnetic contributions.
For a Cartesian coordinate system ($x',y',z$) rotated
by the angle $\alpha$ around the vertical $z$-axis with respect to
$x$ parallel to $\mathbf Q$,
the following relations can be derived from Eqs.~(6) and (7) providing a complete separation
\begin{eqnarray}
N^{\dag} N&=& \frac{1}{2} \left( \Sigma^{nsf}_{x}+\Sigma^{nsf}_{y} -\Sigma^{sf}_{z} \right)
= \frac{1}{2} ( \Sigma^{nsf}_{x'}+\Sigma^{nsf}_{y'} -\Sigma^{sf}_{z} )
\\
I_{si} &= &\frac{3}{2}
( \Sigma^{nsf}_{x}-\Sigma^{nsf}_{y} +\Sigma^{sf}_{z} ) =
\frac{3}{2}
\frac{ \Sigma^{nsf}_{x'}-\Sigma^{nsf}_{y'}}
{\cos ^2 \alpha - \sin^2 \alpha }+
\frac{3}{2}
\Sigma^{sf}_{z}
\\
M_{\perp y} ^{\dag} M_{\perp y}
&=&\Sigma^{sf}_{z} - \frac{2}{3}I_{si}
\\
M_{\perp z} ^{\dag} M_{\perp z}
&=&\Sigma^{nsf}_{z} - \frac{1}{3}I_{si} -
N^{\dag} N
\\
I_{chiral,x'}&=&
2 \mathrm i (\mathbf M_{\perp } ^{\dag} \times \mathbf M_{\perp })_{x'}
=\Delta^{sf}_{x'} \;\;=\cos \alpha \, \Delta^{sf}_{x}
\\
I_{NM,y'}&=&2 \Re( N Q_{y'}^{\dag}) \, \; \;\;\;\;\;\;\; \; = \Delta^{nsf}_{y'} =\cos \alpha \, \Delta^{nsf}_{y}
\\
I_{NM,z'}&=&2 \Re( N Q_{z'}^{\dag}) \, \;\;\;\;\;\;\;\; \;= \Delta^{nsf}_{z'} = \;\; \;\;\;\;\;\;\, \Delta^{nsf}_{z} \;,
\end{eqnarray}
where the geometrical correction factors in Eqs.~(10,13,14) stem from the projections $M_{\perp x'}=\sin \alpha \,M_{\perp y}$ and $M_{\perp y'}= \cos \alpha \,M_{\perp y} $
to the scattering vector, see comment \cite{footnote}.
Compared to the information content of powder experiments, the study of diffuse polarised scattering from single crystals
yields additionally the anisotropy of magnetic correlation functions, possible vector chirality and all nuclear- magnetic correlations,
essential information about complex magnetic materials of high current interest.
It is worthwhile to note that a separation of the chiral term has already been given earlier by Eq.~50 in Ref.\cite{xyz}.
For diffuse scattering
the typical experimental strategy is to map out scattering planes by stepwise rotations of the crystal around the vertical axis $z$.
Note, the separation by Eqs.~(9) to (15) is independent of the sample rotation $\omega$ and
the convenient choice is a fixed coordinate system $x',y',z$.
In order to interpret the magnetic scattering amplitudes in terms of the crystal lattice,
the in-plane component $M_{\perp y} $ has to be decomposed further into independent lattice
components in the plane (ab) and is given by the projections of $M_{a} $ and $M_{b} $, while the
vertical component is simply identified by $M_{c} $. For orthogonal components, only the squares of $M_{a} $ and $M_{b} $
determine $M_{\perp y} $, see Fig.~1. Note this example demonstrates that the considerations here also apply to inelastic scattering.
\begin{figure}[h]
\includegraphics[width=12.9pc,
angle=90]{ub.pdf}\hspace{2pc}%
\caption{\label{label} (color online) Polarisation $\mathbf P$ and magnetic (inelastic) scattering $M_{\perp \nu}^2$ (white squares) for the case of a
\textit{Sch\"arpf angle} $\alpha$ between scattering vector $\mathbf Q$ and polarisation $\mathbf P$;
here $\omega$ denotes the angle between $\mathbf Q$ and reciprocal lattice vector $\mathbf a^*$ with $\mathbf M_a || \mathbf a^*$. Shaded squares represent \textit{sf}- and \textit{nsf}-intensities related to the projections of $\mathbf M_{\perp a}$ and $\mathbf M_{\perp b}$ $\perp$ and $|| \mathbf P$ respectively. The vertical component ($||z$ and $\perp$ to $\mathbf P$, \textit{sf}-intensity) is not shown.}
\end{figure}
\vspace*{-0.3mm}
The disentanglement of $M_{a} $ and $M_{b} $ is, analog to structure determination or to disorder in displacements variables,
relying on sufficient information in extended Brillouin zones of reciprocal space.
Diffuse scattering related to disorder phenomena typically implies an exponential decay of correlations in real space.
Hence a Fourier analysis will efficiently reveal finite real space pair correlation functions, an approach that is
a linear least squares problem having a unique solution. Therefore, the determination of three dimensional pair-correlations functions
requires accordingly measurements of more than a single scattering plane.
\section{Final remarks}
With recent instrumental developments, particularly on the DNS and D7, diffuse magnetic scattering of single crystals can be efficiently measured
with polarisation analysis,
which has already provided exiting results on frustrated magnetism in spin-ice compounds \cite{Fennell}; a further example can be found in the proceedings of the PNSXM 2009 \cite{Chang}.
For a systematic and thorough understanding the separation of the different components is essential, a task that to date
has been believed to be incapable by the conventional means used in XYZ polarisation analysis.
The solution here, utilizing the symmetry properties, also places the emphasis on additional measurements with polarisation reversal.
It is worthwhile to note that this request does not demand for more beamtime to achieve equal statistical weights.
So far we have not considered the off-diagonal terms.
These terms also represent the interference terms, Eqs.~(13-15), due to chiral correlations and due to coupling of spin-space
and real-space variables.
Spherical neutron polarimetry of off-diagonal terms distinguishes
these effects irrespectively of a possible depolarisation from different magnetic domains in the crystal.
Obviously, the diagonal antisymmetric polarised intensities Eq.~(7) also reveal the mentioned interference terms.
It should be noted that the same XYZ-setup for multi-detectors can be used to access the complete polarisation tensor
and off-diagonal terms can be determined with the same accuracy by spherical neutron polarimetry
with precessing incident polarization\cite{ws_vector,ws_nn}.
However, the here proposed XYZ-polarisation analysis, a method based on only diagonal terms and applicable for multi-detectors,
will certainly provide a much more convenient and efficient separation for the diffuse magnetic neutron scattering from single crystals.
\section*{Acknowledgments}
Otto Sch\"arpf is gratefully acknowlegded for discussions and valuable comments.
\section*{References}
|
\section{Introduction}
Planar two-dimensional graphene has been considered to be a very promising
new material since its preparation by
Novoselov {\em et al.}\cite{Novoselov} and Berger {\em et al.}\cite{Berger}
in 2004. Recent interest is focused on the appearance of magnetism
around point defects in graphene\cite{magnetism} and
the possibility of hydrogen storage\cite{HonGraphene,stress,HonGraphene2,
graphane}. A global understanding of the origin of magnetism in finite
graphene systems is provided by Lieb's theorem on
bipartite lattices\cite{lieb}.
Any unbalance between the numbers of sites belonging to each of the two
sublattices gives rise to a magnetic groundstate\cite{innombrables}.
This result rests on the validity of a simple Hubbard Hamiltonian which
certainly works for the semi-quantitative description of the spin states
of some polycyclic aromatic hydrocarbon molecules but not for the
corresponding charged states\cite{PAH}.
The saturation of a carbon $\pi$-electron by hydrogen is one of the simplest
ways to change the balance between sublattice sites and, consequently,
it produces spin polarization in the neighborhood. This provides an
interesting link between hydrogenation and spin production that
motivates this work.
We shall demonstrate that the 1/2 spin originated by the
presence of an isolated hydrogen atom on top of a carbon atom
belonging to planar graphene can be
quenched if an extra electron bounds to the defect. Numerical
results obtained by {\em ab initio} open-shell
Density Functional Methods (DFT) suggest that the complex defect
is energetically favorable. A similar spin quenching phenomenon
was discussed some years ago by
Duplock, Scheffler, and Lindan for hydrogen near
a Stone-Wales defect\cite{SW}. These authors have shown that the
spin polarized groundstate around chemisorbed hydrogen disappears
in the presence of a Stone-Wales defect. In their interpretation,
this result is probably due to the strong destruction of alternation
near the defect that eliminates the tendency to antiferromagnetic order.
In our case, however, we argue that the mere flow of charge is enough
to heal the local unbalance between sublattices due to the
existence of one chemisorbed hydrogen on an otherwise ideal
graphene material.
The rest of the paper is organized as follows: The {\it ab initio} methods
used in this work are presented in section II
followed by the discussion of our main numerical results.
Final section III just remarks our main message.
\begin{table*}
\caption{
Total energies (in hartree) for atomic hydrogen and its anion,
for coronene $({\rm C}_{24}{\rm H}_{12})$ and coronene anion
$({\rm C}_{24}{\rm H}_{12})^{-}$
both ideal and with a chemisorbed H atom and for a larger cluster
$({\rm C}_{54}{\rm H}_{18})$ sometimes called supercoronene
and its anion both planar and deformed by the presence of a chemisorbed
H atom. Since results are given for two or more gaussian basis sets
some rough estimation of error is possible.}
\begin{ruledtabular}
\begin{tabular}{lccccc}
Cluster & Total Spin & Energy(MIDI) & Energy(CCT) & Energy(PC2) & Energy(PC3) \\
\hline
${\rm H}$ & $\frac{1}{2}$
& -0.4953 & -0.4988 & -0.4990 & -0.4991 \\
$\rm H^{-}$ & 0
& -0.4602 & -0.5035 & -0.5177 & -0.5254 \\
${{\rm C}_{24}{\rm H}_{12}}$ & 0
& -915.9342 & -921.6070 & -921.6516 & -921.6950 \\
$({\rm C}_{24}{\rm H}_{12})^{-}$ & $\frac{1}{2}$
& -915.9342 & -921.6186 & -921.6660 & -921.7097\\
${{\rm C}_{24}{\rm H}_{13}}$ & $\frac{1}{2}$
& -916.4438 & -922.1269 & -922.1723 & - \\
$({\rm C}_{24}{\rm H}_{13})^{-}$ & 0
& -916.4853 & -922.1812 & -922.2300 & - \\
${{\rm C}_{54}{\rm H}_{18}}$ & 0
& -2055.5931 & - & -2068.3971 & - \\
$({\rm C}_{54}{\rm H}_{18})^{-}$ & $\frac{1}{2}$
& -2055.6323 & - & -2068.4469 & - \\
${{\rm C}_{54}{\rm H}_{19}}$ & $\frac{1}{2}$
& -2056.1072 & - & -2068.9095 & - \\
$({\rm C}_{54}{\rm H}_{19})^{-}$ & 0
& -2056.1868 & - & -2068.9919 & - \\
\end{tabular}
\end{ruledtabular}
\label{TotalEnergies}
\end{table*}
\section{{\it Ab initio} calculations: Methods and Results}
Two planar carbon clusters with the structure of graphene have been
chosen to calculate the energetics of hydrogen absorption both at the
neutral state and when the system is electronically charged
with an extra electron (anions).
The first system is the well known coronene polycyclic aromatic
hydrocarbon (PAH) represented in Fig.~\ref{Coronene}
while the second one is a larger PAH obtained from coronene adding an
extra shell of benzene rings.
This system is referred as supercoronene in the
literature. We notice that
this larger PAH molecule has not yet been synthesized
but nonetheless it provides a good theoretical benchmark for our purposes
(Fig.~\ref{Htop-SC} schematically shows one hydrogen
chemisorbed on supercoronene).
The groundstate for both PAH's does not show spin polarization
(total spin is zero).
This is an important difference with our previous study of hydrogen
chemisorption on graphene where spin one-half clusters were used to
preserve the point symmetry and facilitate the
computational effort\cite{stress}.
In the present work, however, we focus on the description
of spin polarization and we must start with an unpolarized
cluster to accurately simulate the graphene layer. Point symmetry is lost and
computational load is larger. Nevertheless, we have checked that
both structural and energetic results for chemisorbed H coincide
with the values given in our previous work\cite{stress}.
Quantum chemistry calculations have been done using the GAMESS
program \cite{GAMESS}. Several sets of gaussian basis functions
have been employed. Depending on the computational effort
we have been able to assess the convergence of
numerical results for some cases by comparing results obtained with bases of
different sizes. For the largest systems, however,
we have been forced to choose a minimal basis and convergence
could only be assessed by reference to the smaller clusters.
Specifically,
we have always started by trying the so-called MIDI basis\cite{MIDI},
and when possible we have moved to a correlation consistent basis
referred in the literature as cc-pVTZ\cite{CCT}
(CCT within GAMESS and this paper) and a DFT adapted hierarchy of basis
called PCn where n indicates the level of polarization\cite{PCn}.
Our best results correspond to the larger PCn basis that we have
been able to use in each case.
Unrestricted Hartree Fock (UHF) calculations
have been intentionally avoided because total spin of the wavefunction
is undefined in those cases. Therefore, HF calculations for an odd
total number of electrons have been performed using the Restricted-Open-Shell
variant. All results presented in this work for clusters have been obtained
using the Becke-Lee-Yang-Parr hybrid density functional RB3LYP \cite{B3LYP}.
Our choice of finite clusters of various sizes poses the question of
to what extent results are affected by the particular boundary conditions
imposed to solve the quantum problem for electrons. Therefore, we check
by comparing with calculations performed on extended
periodic models using periodic boundary conditions
and a plane-waves basis. This model is set up
so a single H atom is adsorbed on a $4 \times 4$ graphene supercell
including 32 carbon atoms on a honeycomb lattice
(see Fig.~\ref{fig4x4},
$a=b=9.84$ {\AA}, $c=23.4$ {\AA}, $\alpha=\beta=90^{o}$,
$\gamma=60^{o}$).
We use ultrasoft pseudopotentials\cite{Vanderbilt},
an energy cutoff of 310 eV,
and a Monkhrost-Pack mesh of $3 \times 3 \times 1$\cite{monk}.
Actual calculations are performed with the
CASTEP program allowing for spin polarization of
the different electronic bands\cite{Payne,Accelrys}.
To describe the exchange and correlation potential
we use the local density
approximation\cite{Kohn}.
Forces and stresses on the system are converged to the
usual thresholds ($9\times 10^{-3}$ eV/{\AA} and $0.01$ GPa)
and the total energy is minimized for the different
systems. This procedure cannot provide an accurate
description for the electron affinity energy since the extra
electron in the supercell is intentionally neutralized with
a uniform positive background to subtract infinite contributions
in the periodic system. This uniform background, however, has
little effect on the spatial distribution of the electronic
and spin densities, that
can be analyzed with reasonable confidence.
Table I compiles the bulk of our quantum-chemistry results. Notice
that together with the total energies needed to get chemisorption
energies for hydrogen on graphene, we have computed the energies
corresponding to systems charged with an extra electron (anions in the
molecular case). Total spin of the groundstate is given by the second
column of the table. It can be seen that the S=0 value of the neutral
molecules is recovered by the anions of the hydrogenated cases.
Table II gives the electron affinities that are obtained from
the results shown in Table I\cite{EA}. Nice convergence to the experimental
electron affinity of hydrogen is observed in the first entry of the table.
Results for larger clusters are limited by our computational means
but nonetheless our results at the PC2 level
are good enough to support our conclusions.
A word of caution is in order here: although these results for the spin seem
to fit nicely within a simple electron-counting scheme (i.e., zero spin for
even number of electrons and net spin for odd number of electrons) this is not
always true. In particular, we recall the case where two hydrogen atoms are
adsorbed on the graphene layer: while adsorption of the two hydrogens in
next-neighbor positions results in a ground state with no net spin,
adsorption in next to next-neighbor sites results in a ground state with
a net $S=2$ $\mu_{B}$.
This is related to the fact that graphene is a bipartite lattice,
and it is in accordance to Lieb's theorem\cite{lieb}, showing the limitations
of simple electron-counting rules.
\begin{table}
\caption{
Electron affinities of hydrogen, coronene, monohydrogenated coronene,
supercoronene and monohydrogenated supercoronene
obtained from the results compiled in Table I.
Energies are now given in eV.}
\begin{ruledtabular}
\begin{tabular}{lcccc}
Cluster & Energy(CCT) & Energy(PC2) & Energy(PC3) & \mbox{Experimental} \\
\hline
$\rm H$ & 0.13 & 0.51 & 0.72 & 0.75\footnotemark[1] \\
${{\rm C}_{24}{\rm H}_{12}}$ & 0.32 & 0.39 & 0.40 & 0.47\footnotemark[2] \\
${{\rm C}_{24}{\rm H}_{13}}$ & 1.48 & 1.57 & - & - \\
${{\rm C}_{54}{\rm H}_{18}}$ & 1.07\footnotemark[3] & 1.35 & - & - \\
${{\rm C}_{54}{\rm H}_{19}}$ & 2.17\footnotemark[3] & 2.24 & - & - \\
\end{tabular}
\end{ruledtabular}
\footnotetext[1]{See, for example,
http://www.chemicool.com/elements/hydrogen.html.}
\footnotetext[2]{The electron affinities of coronene and coronene dimer
have been measured by Duncan {\em et al.} as reported in
Ref.(\onlinecite{Duncan}).}
\footnotetext[3]{This result corresponds to the small MIDI basis
type of calculation that is described in Ref.(\onlinecite{MIDI}).}
\label{ElectronAffinities}
\end{table}
Let us briefly discuss the results given in Table II.
The electron affinity of ${{\rm C}_{24}{\rm H}_{13}}$ cluster, i.e.,
the cluster with one hydrogen chemisorbed on top of one of the six
carbon atoms on the inner ring of coronene (see, Fig. 1) is 1.18 eV
larger than the electron affinity of coronene. Also, the value
for ${{\rm C}_{54}{\rm H}_{19}}$, that is, H on top of supercoronene
is 0.89 eV larger than the supercoronene electron affinity. Although,
only two cluster sizes have been studied, it seems that the difference is
approaching a limiting value close to the electron affinity of
free hydrogen, that is, close to 0.75 eV.
If this were the case, the extra charge would be attracted by hydrogen
with a similar strength as in free space and the screening by the rest
of $\pi$-electrons of graphene would remain unnoticed. Nevertheless,
we show later that only part of the extra charge remains close to
the defect. Therefore, we assume that for an infinite system only a fraction
of 0.75 eV proportional to the localized charge would remain.
There is an alternative elaboration of the results given in Table I
focusing on the variation of the binding of a hydrogen atom on top
of a charged surface compared to the binding by the neutral one.
From this point of view, H binding energy increases from 0.59 eV to
1.77 eV on coronene and from 0.36 eV to 1.25 eV in supercoronene using
PC2 values in Table I\cite{binding}. This means that the charged systems
bind hydrogen about one eV stronger than the neutral ones.
From this perspective, the larger values of the electron
affinity obtained for hydrogenated clusters can be asigned
to a stronger hydrogen binding to graphene.
The relative facility to move charge across the overall system implied
by the semi-metallic character of graphene and our
results in Table II point towards the formation of a complex point defect
with an extra electron in the neighborhood of the chemisorbed hydrogen atom.
This picture is further supported by the spatial distribution of this
extra electron in the studied clusters as it is shown in
Figs.~\ref{resta} and ~\ref{restaSC}.
Charge densities obtained with PC2 gaussian basis for coronene anion
and neutral coronene have been subtracted in Fig.~\ref{resta} using the
MOLDEN package\cite{MOLDEN}. The same difference for supercoronene
is given in Fig.~\ref{restaSC}. In both cases, the charge around "on top" H is
similar to the extra charge of the H anion. On the other hand, the spreading
positive and negative densities are similar but not equal in
coronene and supercoronene. A closer inspection reveals that
the extra electron is occupying the partially occupied HOMO level of the
corresponding neutral clusters. Fig.~\ref{homoSC} gives a picture of the HOMO
orbital of H on supercoronene that nicely explains the charge
difference previously shown in Fig.~\ref{restaSC}.
It is interesting to notice, however, that although the extra charge clusters
around the adsorbate, it is partly delocalized as it is expected from quantum
mechanics principles. In fact, from independent tight-binding calculations in
periodic systems we don't find a true {\it exponential} localization around
the defect (see Appendix).
On the contrary, a percentage of the charge is extended all over
the system (e.g., see Fig.~\ref{restaSC}). However, as our detailed
quantum-mechanical calculations show, the net effect of the localized part
is enough to quench the spin.
Our {\em ab initio} results on periodic extended systems fully support
the interpretation given in the previous paragraph.
We observe in Fig.~\ref{fig4x4} how the
extra charge is accumulated around the adsorbed H.
In this case, bonding charges for the neutral and charged supercells
are depicted in the upper and lower panels of the figure, respectively.
Bonding charges are defined as density charge differences between the
whole system and conveniently defined fragments. In our case, hydrogen
atom is one fragment while the graphene $4 \times 4$ supercell is the
second one. The upper difference integrates to one electronic charge
since the whole system is charged while the fragments are neutral.
We notice that a small part of the extra electron is on carbon atoms while
an important part attaches to hydrogen (the violet negative density
in the lower panel does not appears in the upper panel meaning a
positive contribution to the extra charge in the upper panel).
The same picture is extracted from a Mulliken analysis of populations
around different atoms. In the neutral system, charge flows upon
adsorption from hydrogen to graphene, so approximately $-0.63$ e is
located around hydrogen while the transferred charge resides
mostly around the closest carbon behind
hydrogen ($-0.33$ e).
On the other hand, in the charged system we find $-1.42$ e around
hydrogen, while the carbon behind keeps nearly the same occupation
($-0.34$ e) and the rest of the charge is distributed over nearest
neighbors and next-nearest neighbors. Therefore, about 80\% of
the extra electron is located near the chemisorbed hydrogen.
Along the same line, the H-C bond population analysis is about
three times larger for the anion, although the bond length is
nearly not affected.
Finally, integration of the spin density and the absolute value
of the spin density over the simulation cell give further
support for this picture. In the neutral system these values
amount to $0.4$ and $0.5$ $\mu_{B}$ respectively, while in
the charged one decrease to values of
amount to $4\times 10^{-6}$ and
$5\times 10^{-4}$ $\mu_{B}$ respectively.
Therefore, the accumulation of charge around the adsorbed H and
the C nearest and next nearest neighbors plays the role to cancel
the extra spin polarization brought by the adsorption of H on the
clean graphene layer in accordance with the results obtained
on finite clusters.
There is a subtle chemical argument that helps the understanding of
our numerical results. In Ref.(\onlinecite{tryphenilmethyl}),
trivalent carbon atoms with an unpaired electron
were unraveled in the studied geometry of carbon tetrapod.
This carbon radicals were stabilized by steric protection giving
rise to unpaired localized electrons that
polarize the carbon neighborhood and explain the appearance of
magnetism in pure organic systems. The original paradigm
is tryphenilmethyl, synthesized by Gomberg in 1900\cite{Gomberg},
where a trivalent carbon atom is stabilized by three bonds to benzene
rings impeding the reaction with a similar molecule. Nevertheless,
the anion of tryphenilmethyl reacts with a proton
to form a strong C-H bond
(heat of reaction amounts to 15.65 eV per molecule):
$$
{{\rm C}_{19}{\rm H}_{15}}^{-} + {\rm H}^{+} \rightarrow
{\rm C}_{19}{\rm H}_{16},
$$
producing a neutral non-magnetic molecule resembling the clusters that
we have studied here\cite{NIST}. Both the number of H atoms and electrons are
even allowing an easy shell closing and stabilization of the
resulting molecule. We can adapt the underlying chemistry of these
phenomena to the binding of H on graphene using the following argument:
it can be thought that
when H forms a covalent bond with a C atom of graphene the two
binding electrons are paired but additionally and due to the particular
topology of graphene lattice one $\pi$-electron becomes unpaired and,
therefore, spin polarized. Since there is no steric protection
around the defect (a kind of radical) any more or less free electron
of the system will flow into the defect to restore the
equilibrium between sublattices. Consequently, spin polarization
around chemisorbed hydrogen disappears. This qualitative
argument is fully supported by our total energy calculations
showing a gain in potential energy following the spin
neutralization (remember that the binding energy of H increases about
one eV for the charged system).
\section{Concluding Remarks}
From a detailed analysis based in first-principles DFT calculations
we find that in the presence of an extra electron chemisorbed H plays
to keep most of the extra charge in its vicinity.
The electron affinities computed on finite cluster models seem to
converge to a value that is somewhat smaller than the free atomic hydrogen
value of $0.75$ eV. Our calculations suggest that being
graphene a semimetal with zero density of states at the Fermi
energy, screening of Coulomb interactions by the $\pi$-electrons
liquid is weak and allows the electron flow to
sites where H is chemisorbed, forming a complex point defect.
Accumulation of extra charge around the defect, otherwise
giving rise to spin polarization, works to quench it.
We suggest that this physical effect is behind the difficulties to
observe magnetism in graphene derived systems.
\begin{acknowledgments}
Financial support by the Spanish MICINN (MAT2006-03741,
MAT2008-1497, FIS2009-08744 and CSD2007-41) is gratefully acknowledged.
\end{acknowledgments}
|
\section{Introduction}
One of the most serious problems in modern physics is the quantization of strongly interacting quantum fields. This includes the confinement problem in quantum chromodynamics, quantization of gravitation, and probably also high temperature superconductivity with its strong interaction between Cooper electrons. The problem is that the algebra of quantum operators describing strongly interacting fields is unknown. Known commutation relationships of type
\begin{equation}
\label{1-10}
\left[ \hat \phi(x), \hat \phi(y) \right] = i \delta(x-y)
\end{equation}
describe \textcolor{blue}{\emph{free, noninteracting fields}} (here $ \hat \phi (x) $ is the operator of a free field $ \phi (x) $).
The need for nonperturbative techniques in strongly interacting, nonlinear quantum field theories is an old problem that has been around since the beginning of the study of quantum fields. Much effort has gone into trying to resolve this puzzle. The different approaches that have been tried include: (i) lattice QCD \cite{Teper:1998kw}, (ii) the dual Meissner effect in the QCD-vacuum \cite{Pandey}-\cite{Agasian}, (iii) instantons \cite{Suganuma} \cite{Shifman}, (iv) path integration \cite{Kondo5}, (v) analytic calculations \cite{Simonov}, (vi) Dyson-Schwinger equations \cite{Gentles}. Despite this, the problem is not yet fully resolved. All these approaches are approximations ones.
In the 1950's, Heisenberg \cite{Heisenberg1} \cite{Heisenberg2} studied a nonlinear spinor field and developed nonperturbative techniques for quantizing the nonlinear spinor field. His method assumes an infinite system of equations which relate all the n-point Green's functions of the theory (this can be compared to the infinite number of Feynman diagrams which must, in principle, be calculated for a given process in pertubative quantum field theory). In order to solve this system of equations, one must find some physically reasonable cut-off approximations, so that one reduces the infinite system of equations into a finite system. Nevertheless, the algebra of field operators in Heisenberg's approach remains unknown.
In Ref. \cite{coleman2} it was shown that radiative corrections could introduce a symmetry breaking ({\it i.e.} negative) mass term into a scalar Lagrangian. This effect is called dimensional transmutation. Similar reasonings in Ref. \cite{Zloshchastiev:2009aw} give rise to that the logarithmic nonlinearity in the quantum wave equation can cause the spontaneous symmetry breaking
and mass generation phenomena. One can presuppose that a \textcolor{blue}{\emph{nonperturbative quantization}} of any strongly interacting fields would yield similar terms. In Ref. \cite{Dzhunushaliev:2009na-dec} it is offered to quantize strongly interacting fields using nonassociative (n/a) decomposition of the field into products of n/a factors. In such approach the rearrangement of brackets gives rise to additional terms in the same way as the permutation of field operators (in standard quantum field theory) gives rise to the Planck constant.
\section{Nonassociative decomposition of quantum field operators}
In this section we follow to Ref. \cite{Dzhunushaliev:2009na-dec}. We assume that operators of strongly interacting fields $\Phi_m \left (x^\mu \right)$ can be decomposed into n/a constituents $f^i_\alpha$ and $b_{i \beta}$:
\begin{equation}
\label{2-10}
\Phi_m \left( x^\mu \right) = f^i_\alpha \left( x^\mu \right)
b_{i \beta} \left( x^\mu \right)
\end{equation}
here $m$ is an index where internal and Lorentzian indices are collected, $i$ is the summation index, and the $\alpha$, $\beta$ are contained in $m$ as: $m = \left \{\alpha, \beta \right \}$. Although the constituent operators $f^i_\alpha, b _ {i \beta}$ are not associative, the basic idea requires their product to model an associative operator. In more mathematical terms, the operators $f^i_\alpha$ and $b _ {i \beta}$ are elements in a n/a algebra
$\mathbb A $, i.e., $f^i_\alpha , b _ {i \beta} \in \mathbb A \setminus \mathbb G$, which contains an associative subalgebra $\mathbb G \subset \mathbb A $, such that $\Phi_m = f^i_\alpha b_{i \beta} \in \mathbb G$. It is necessary to note that the operator $\Phi_m$ models observable quantities, whereas the n/a $f^i_\alpha, b _ {i \beta}$ are unobservable.
Let us note that the number of n/a constituents does not need to be two, but may be more. For example Eq. \eqref{2-10} can be rewritten in the form
\begin{equation}
\label{2-20}
\Phi_m \left( x^\mu \right) = \left(
q^{\phantom{1}}_{1\alpha} \left( x^\mu \right)
\right)^i_j
\left(
q^{\phantom{2}}_{2\alpha} \left( x^\mu \right)
\right)^j_k
\left(
q^{\phantom{3}}_{3\gamma} \left( x^\mu \right)
\right)^k_i
\end{equation}
here $i,j,k$ are the summation indices, and the $\alpha$, $\beta$ and $\gamma$ are contained in $m$ as: $m = \left \{\alpha, \beta, \gamma \right \}$. One can say that the decompositions \eqref{2-10} and \eqref{2-20} somewhat correspond to slave-boson \cite{slaveboson} and spin-charge (or quark-like) \cite{spincharge} decompositions.
\section{Applications}
In this section we would like to consider a few examples: scalar field theory with polynomial $\phi^4$, pure gravity, and gravity interacting with electromagnetic field.
\subsection{Scalar field theory with strong self-interaction}
\label{scalar}
Let us consider scalar field theory with Lagrangian
\begin{equation}
\label{3-20}
\mathcal L = \frac{1}{2} \nabla^\mu \phi \nabla_\mu \phi - V(\phi)
\end{equation}
where the nonlinear potential term is
\begin{equation}
\label{3-25}
V(\phi) = \frac{\lambda}{4} \phi^4 \left( x^\mu \right) .
\end{equation}
We assume a n/a decomposition
$\phi \left( x^\mu \right) = f^{i} \left( x^\mu \right) b_{i} \left( x^\mu \right)$. Here, the operator $\phi$ is an observable associative quantity, but $f^{i}, b_{i}$ are unobservable n/a quantities. Using n/a factors one can rewrite the potential term from Lagrangian \eqref{3-20} as follows
\begin{equation}
\label{3-30}
\biggl(f^{i_1} \left( x^\mu \right) b_{i_1} \left( x^\mu \right) \biggr)
\biggl(f^{i_2} \left( x^\mu \right) b_{i_2} \left( x^\mu \right) \biggr)
\biggl(f^{i_3} \left( x^\mu \right) b_{i_3} \left( x^\mu \right) \biggr)
\biggl(f^{i_4} \left( x^\mu \right) b_{i_4} \left( x^\mu \right) \biggr).
\end{equation}
In order to calculate an expectation value we have to define the action of operators on a quantum state. We will require:
\begin{equation}
\label{3-40}
\phi \left| \psi \right\rangle = \left( f^{i} b_{i} \right) \left| \psi \right\rangle
\stackrel{def}{=}
f^{i} \left( b_{i} \left| \psi \right\rangle \right) .
\end{equation}
This is the same rule as for the associative case. But when having two
or more nonassociative operators, there is:
\begin{equation}
\begin{split}
\label{3-50}
\phi^2 \left| \psi \right\rangle = & \Bigl(
\left( f^{i} b_{i} \right) \left( f^{i} b_{i} \right)
\Bigr) \left| \psi \right\rangle = \biggl( \Bigl(
\left( f^{i} b_{i} \right) f^{i}
\Bigr) b_{i} \biggr) \left| \psi \right\rangle + \mathrm{assoc} \left| \psi \right\rangle =
\Bigl(
\left( f^{i} b_{i} \right) f^{i}
\Bigr) \left( b_{i} \left| \psi \right\rangle \right)
+ \mathrm{assoc} \left| \psi \right\rangle =
\\
&
\left(f^{i} b_{i} \right) \Bigl( \bigl( f^{i}
\left( b_{i} \left| \psi \right\rangle \right) \bigr) \Bigr)
+ \mathrm{assoc} \left| \psi \right\rangle =
f^{i} \Bigl( b_{i} \bigl( f^{i}
\left( b_{i} \left| \psi \right\rangle \right) \bigr) \Bigr)
+ \mathrm{assoc} \left| \psi \right\rangle
\end{split}
\end{equation}
where the associator Ass is defined as follows
\begin{equation}
\label{3-60}
\left( f^{i} b_{i} \right) \left( f^{i} b_{i} \right) = \biggl( \Bigl(
\left( f^{i} b_{i} \right) f^{i}
\Bigr) b_{i} \biggr) + \mathrm{assoc} .
\end{equation}
In order to define the associator, $assoc$, we recall that in the standard commutations relationship \eqref{1-10} on the LHS we have two operators (observables) and on the RHS we have $2-2=0$ operators. In a similar way, we assume that on the RHS of \eqref{3-60} should be $2-2=0$ \textcolor{blue}{\emph{associative}} operators. It means that $\mathrm{assoc} = m^2$ where $m$ can be a complex number. One can say that it is the second Planck constant but with different dimension.
The same can be done for $\phi^3 \left| \psi \right\rangle$
\begin{equation}
\label{3-70}
\phi^3 \left| \psi \right\rangle =
f^{i} \Biggl( b_{i} \biggl( f^{i} \Bigl( b_{i} \bigl( f^{i}
\left( b_{i} \left| \psi \right\rangle \right) \bigr) \Bigr) \biggr) \Biggr)
+ m^2 \phi \left| \psi \right\rangle
\end{equation}
and for $\phi^4 \left| \psi \right\rangle$
\begin{equation}
\label{3-80}
\phi^4 \left| \psi \right\rangle =
f^{i} \left( b_{i} \left(
f^{i} \Biggl( b_{i} \biggl( f^{i} \Bigl( b_{i} \bigl( f^{i}
\left( b_{i} \left| \psi \right\rangle \right) \bigr) \Bigr) \biggr) \Biggr)
\right) \right)
+ m^2 \phi^2 \left| \psi \right\rangle .
\end{equation}
Then the expectation value will be
\begin{eqnarray}
\label{3-90}
\left\langle \psi \left| \phi^3 \right| \psi \right\rangle &=& \left\langle \psi \left|
f^{i} \left( b_{i} \left( f^{i} \left( b_{i} \left( f^{i}
\left( b_{i} \right| \psi \right\rangle \right) \right) \right) \right) \right)
+ m^2 \left\langle \psi \left| \phi \right| \psi \right\rangle ,
\\
\label{3-100}
\left\langle \psi \left| \phi^4 \right| \psi \right\rangle &=& \left\langle \psi \left|
f^{i} \left( b_{i} \left(
f^{i} \left( b_{i} \left( f^{i} \left( b_{i} \left( f^{i}
\left( b_{i} \right| \psi \right\rangle \right) \right) \right) \right) \right)
\right) \right)
+ m^2 \left\langle \psi \left| \phi^2 \right| \psi \right\rangle .
\end{eqnarray}
Ordinarily, the vacuum expectation value
$\left\langle \psi \left| \phi \right| \psi \right\rangle = \left\langle \phi \right\rangle$ of a field is zero. Consequently, in the case where $\left| \psi \right\rangle$ is a vacuum state, there is
\begin{equation}
\label{3-110}
\left\langle \psi \left| \phi \right| \psi \right\rangle = 0.
\end{equation}
It means that the n/a terms appear with the potential $V(\phi) \propto \phi^n, n\geq 4$ only.
Thus the conclusion for this section is that \textcolor{blue}{\emph{the nonassociative corrections appear by calculation of an expectation value in nonperturbative quantum field theory for strongly enough nonlinear terms in the Lagrangian.}}
\subsection{Gravity}
\label{gravity}
General relativity is the most nonlinear theory we know. Consequently nonperturbative effects must be very important in quantum gravity. In general relativity the first nonlinearity is connected with a connection $\left\{^{\alpha}_{\beta \gamma}\right\}$ compatible with a metric
\begin{equation}
\label{4-10}
\left\{^{\alpha}_{\beta \gamma}\right\} = \frac{1}{2} g^{\alpha \delta} \left(
g_{\beta \delta ,\gamma} + g_{\gamma \delta ,\beta} -
g_{\beta \gamma ,\delta}
\right)
\end{equation}
which is called as Christoffel symbols. The nonlinearity is created by the quantity $g^{\alpha \delta}$ which is the matrix reciprocal to the matrix $g_{\alpha \delta}$. Such nonlinearity is much stronger than the potential term \eqref{3-20} in usual quantum field theory. How $\hat{g}^{\alpha \delta}$ can be described as an operator in nonperturbative quantization is unclear. In this section we denote operators as $\widehat{\ldots}$. Nevertheless, we continue to carry out the idea about appearance of additional terms from calculation of an expectational value. For the Christoffel symbols in this case we will have
\begin{equation}
\label{4-20}
\left\langle \psi \left|
\widehat{\left\{^{\alpha}_{\beta \gamma}\right\}}
\right| \psi \right\rangle = \left\langle
\left\{^{\alpha}_{\beta \gamma}\right\}
\right\rangle = \widetilde{
\left\{^{\alpha}_{\beta \gamma}\right\}} +
K_{\beta \gamma}^{\phantom{\beta \gamma} \alpha} =
\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}
\end{equation}
here $\left\langle \ldots \right\rangle$ is quantum averaging;
$\widetilde{ \left\{^{\alpha}_{\beta \gamma}\right\}} =
\Gamma_{(\beta \gamma)}^{\phantom{\beta \gamma} \alpha}$ is the symmetric part of the affine connection
$\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}$;
$K_{\beta \gamma}^{\phantom{\beta \gamma} \alpha} =
\Gamma_{[\beta \gamma]}^{\phantom{\beta \gamma} \alpha}$ is the antisymmetric part of
$\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}$ and describes the deviation quantum Christoffel symbols
$\left\langle \left\{^{\alpha}_{\beta \gamma}\right\} \right\rangle$ from the classical
$\widetilde{ \left\{^{\alpha}_{\beta \gamma}\right\}}$;
$\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}$ is the designation for
$\left\langle \left\{^{\alpha}_{\beta \gamma}\right\} \right\rangle$. The interesting question here is: why quantum averaging of the symmetric affine connection $\widetilde{ \left\{^{\alpha}_{\beta \gamma}\right\}}$ leads to a non-symmetric affine connection
$\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}$ ? In our opinion it happens because the wave functional $| \psi \rangle$ is non-symmetric one.
For the first approximation we assume that
\begin{equation}
\label{4-30}
\left\langle
\left\{^{\alpha}_{\beta \gamma}\right\} \left\{^{\mu}_{\nu \rho}\right\}
\right\rangle
\approx
\left\langle
\left\{^{\alpha}_{\beta \gamma}\right\}
\right\rangle
\left\langle
\left\{^{\mu}_{\nu \rho}\right\}
\right\rangle .
\end{equation}
In this case the quantum averaged Ricci tensor is
\begin{equation}
\label{4-40}
\left\langle R_{\mu \nu} \right\rangle \approx
\frac{\partial \Gamma_{\mu \nu}^{\phantom{\mu \nu} \rho}}
{x^{\rho}} -
\frac{\partial \Gamma_{\mu \rho}^{\phantom{\mu \rho} \rho}}
{x^{\nu}} +
\Gamma_{\mu \nu}^{\phantom{\mu \nu} \rho} \Gamma_{\rho \tau}^{\phantom{\mu \tau} \tau} -
\Gamma_{\mu \rho}^{\phantom{\mu \rho} \tau} \Gamma_{\nu \tau}^{\phantom{\mu \tau} \rho}.
\end{equation}
It means that we can think about the quantum averaged Ricci tensor
$\left\langle R_{\mu \nu}(\left\{ \right\}) \right\rangle$ as about the Ricci tensor $\mathcal R_{\mu \nu}(\Gamma)$ with a connection $\Gamma_{\beta \gamma}^{\phantom{\beta \gamma} \alpha}$ which is not-compatible with the metric $g_{\mu \nu}$.
Let us remember some notions from the differential geometry. In general the affine connection $\Gamma_{\beta \gamma}^{\alpha}$ can be written as
\begin{equation}
\label{4-50}
\Gamma_{\mu \nu}^{\phantom{\mu \nu}\rho} =
\left\{^{\rho}_{\mu \nu}\right\} +
K_{\mu \nu}^{\phantom{\mu \nu}\rho}
\end{equation}
where $\left\{^{\alpha}_{\beta \gamma}\right\}$ are the usual Christoffel symbols of the symmetric connection and the contortion tensor $K$ is called the contorsion tensor and is given in terms of the torsion tensor by
\begin{eqnarray}
\label{4-60}
K_{\mu \nu}^{\phantom{\mu \nu}\rho} &=& \frac{1}{2} g^{\rho \sigma}
\left(
T_{\mu \sigma \nu} + T_{\nu \sigma \mu} - T_{\mu \nu \sigma}
\right) ,
\\
\label{4-70}
K_{[\mu \nu ]}^{\phantom{[ \mu \nu ]}\rho} &=& - \frac{1}{2}
T_{\mu \nu}^{\phantom{\mu \nu}\rho},
\\
\label{4-80}
K_{\mu \nu \rho} &=& - K_{\mu \rho \nu}
\end{eqnarray}
here the torsion tensor $T_{\mu \nu}^{\phantom{\mu \nu}\rho}$ is the antisymmetric part of the affine connection coefficients $\Gamma_{\mu \nu}^{\phantom{\mu \nu}\rho}$.
\begin{equation}
\label{4-90}
T_{\mu \nu}^{\phantom{\mu \nu}\rho} = -2
\Gamma_{[\mu \nu ]}^{\phantom{[ \mu \nu ]}\rho}
\end{equation}
According to \eqref{4-40} and definitions \eqref{4-60} -- \eqref{4-80} we can say that the connection in classical general relativity (Christoffel symbols) after nonperturbative quantization acquires an additional contribution in the form of torsion. As well one can say that the torsion appears in quantum gravity as the result of nonperturbative quantization.
\subsection{Gravity coupled with electrodynamics}
Above we have shown that by nonperturbative quantization of gravity the torsion appears as a quantum correction to Christoffel symbols. It means that the Einstein -- Cartan gravity is the first approximation for quantum gravity. In classical and perturbative quantum electrodynamics there is a big problem with an infinite energy of static electric field created by a point charge. One hope is that quantum gravity resolves this problem.
\subsubsection{Quantum corrections from gravitational nonlinearities}
\label{ed_gravity_1}
In this subsection we would like to show that above mentioned quantum corrections smooth this problem. These corrections should be taking into account on a small distance only. For simplicity, we consider a metric with torsion only (but otherwise flat). We can then consider Maxwell electrodynamics in Minkowski spacetime (following to Ref. \cite{Ponomarev:1978zy} we use $(-,+,+,+)$ signature here and in the next subsection). The Lagrangian in this case is
\begin{equation}
\label{4-1-10}
\mathcal L = \frac{1}{16 \pi c} F_{\mu \nu}\left( \Gamma \right)
F^{\mu \nu}\left( \Gamma \right)
\end{equation}
where
\begin{equation}
\label{4-1-20}
F_{\mu \nu}\left( \Gamma \right) = \nabla_\mu A_\nu - \nabla_\nu A_\mu =
\partial_\mu A_\nu - \partial_\nu A_\mu - T_{\mu \nu}^{\phantom{\mu \nu} \rho} A_\rho
\end{equation}
here $A_\rho$ is 4-potential of the electromagnetic field and $F_{\mu \nu}\left( \Gamma \right)$ is corresponding tensor of electromagnetic field for the affine connection $\Gamma$. Let us note that the electromagnetic tensor \eqref{4-1-20} is not gauge invariant in the consequence of the torsion term $T_{\mu \nu}^{\phantom{\mu \nu} \rho}$. Maxwell equations can be written in the form \cite{Ponomarev:1978zy}
\begin{eqnarray}
\label{4-1-30}
\frac{1}{\sqrt{-g}} \frac{\partial }{\partial x^\mu}
\left(
\sqrt{-g} F^{\mu \nu}\left( \Gamma \right)
\right) &=& \frac{4 \pi}{c} J^\nu ,
\\
\label{4-1-35}
F_{\mu \nu}\left( \Gamma \right) &=& \mathcal F_{\mu \nu} +
\frac{2 G}{c^4} \frac{A_{[ \mu} F_{\nu ] \rho} A^\rho}
{1 + \frac{G}{c^4} A_\mu A^\mu}
\\
\label{4-1-40}
J^\nu &=& - \frac{G}{4 \pi c^3} F^{\mu \nu}\left( \Gamma \right)
F_{\mu \rho}\left( \Gamma \right) A^\rho
\end{eqnarray}
here $\mathcal F_{\mu \nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$ and $G$ is Newton constant.
Allowing quantum corrections from torsion, we now consider an electrostatic spherically symmetric
solution with 4-potential
\begin{equation}
\label{4-1-50}
A_\mu = \left(
\phi (r), 0,0,0
\right).
\end{equation}
Then Eq. \eqref{4-1-30} has the form
\begin{eqnarray}
\label{4-1-60}
\mathrm{div} \vec E &=& 4 \pi \rho,
\\
\label{4-1-70}
\rho &=& \frac{G}{4 \pi c^4} \vec E^2 \phi ,
\\
\label{4-1-80}
E_r &=& F_{0r}\left( \Gamma \right) =
\frac{\mathcal F_{0r}}{1 + \frac{G}{c^4} \phi^2}.
\end{eqnarray}
The spherical solution is
\begin{eqnarray}
\label{4-1-90}
\phi &=& \frac{c^2}{\sqrt{G}} \sinh \left(
\frac{q \sqrt{G}}{c^2} \frac{1}{r}
\right),
\\
\label{4-1-100}
E_r &=& \frac{q}{\cosh \left(
\frac{q \sqrt{G}}{c^2} \frac{1}{r}
\right)} \frac{1}{r^2} ,
\\
\label{4-1-110}
\rho &=& \frac{\sqrt{G}}{4 \pi c^2}
\frac{\tanh \left( \frac{q \sqrt{G}}{c^2} \frac{1}{r} \right)}
{\cosh\left( \frac{q \sqrt{G}}{c^2} \frac{1}{r} \right)}
\frac{q^2}{r^4}
\end{eqnarray}
where $q = \int \limits_V \rho d^3 x$ is the electric charge. Interestringly, the electric field $E_r$ and charge density $\rho$ are nonsingular at the origin
\begin{equation}
\label{4-1-120}
E_r(0) = \rho(0) = 0.
\end{equation}
However, the total energy of the field becomes infinite:
\begin{equation}
\label{4-1-130}
\int \limits_V \left(
\frac{1}{8 \pi} \vec E^2 - \frac{1}{2} \rho \phi
\right) d^3 x = \infty .
\end{equation}
While the integral $\int \limits_V \vec E^2 d^3 x < \infty$ is finite, the total energy becomes
infinite due to the self-interaction term $\rho \phi$. Evidently, the coupling between electrodynamics and gravity is nonlinear. Consequently, one can hope that including nonperturbative corrections for the electromagnetic field connected with gravity will give rise to a finite energy of the electric field of a point charge.
\subsubsection{Quantum corrections from nonlinear gravitoelectric coupling}
\label{ed_gravity_2}
Quantum corrections coming from the interaction between electromagnetic and gravitational fields are unknown in the case of very strong nonlinear gravitoelectric coupling. We assume that the corrections can be written as $- \frac{d V(A_\mu)}{d A_\nu}$ in following form in Maxwell equations
\begin{equation}
\label{4-2-10}
\frac{1}{\sqrt{-g}} \frac{\partial }{\partial x^\mu}
\left(
\sqrt{-g} F^{\mu \nu}\left( \Gamma \right)
\right) = \frac{4 \pi}{c} J^\nu - \frac{d V(A_\mu)}{d A_\nu}
\end{equation}
where $V(A_\mu)$ are quantum nonperturbative corrections for the electromagnetic potential in the consequence of nonlinear system: gravity + electromagnetism. These corrections cannot be calculated in general, but for some $V(A_\mu)$ there exist \textcolor{blue}{\emph{regular}} solutions.
We now assume quantum corrections of the form
$V(A_\mu) = \frac{\lambda}{4} \left( A_\mu A^\mu + A_0^2 \right)^2$. For the spherically symmetric 4-potential \eqref{4-1-50} Maxwell equation is
\begin{equation}
\label{4-2-20}
\frac{1}{r^2} \left( r^2 \eta^\prime \right)^\prime =
-\tilde \lambda \sinh \eta \left[
\sinh^2 \left( \frac{\eta}{2} \right) - m^2
\right]
\end{equation}
where $\phi(r) = \frac{c^2}{\sqrt{G}} \sinh\left[ \frac{\eta(r)}{2} \right]$;
$\tilde \lambda = \frac{c^4}{G} \lambda$ and $m^2 = \frac{G}{c^4} A_0^2$. We are searching for the regular solution at the origin. Consequently the boundary conditions are
\begin{eqnarray}
\label{4-2-25}
\eta(0) &=& \eta_0, \\
\label{4-2-27}
\eta^\prime(0) &=& 0.
\end{eqnarray}
We doubt that there exists an analytical solution of Eq.\eqref{4-2-20}. Therefore we are searching for a numerical solution. The numerical investigation shows that a special regular solution $\eta^*(r)$ does exist for some special choice of $\eta(0) = \eta^*_0$ only. The results of numerical solution of Eq.\eqref{4-2-20} in Fig's \ref{eta} - \ref{rho} are presented. Since the solution is special, this makes mass, charge, and other physical properties of such an electric charge unique.
\begin{figure}[h]
\begin{minipage}[t]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.8\linewidth]{eta_phi}}
\caption{The profiles of $\eta(r)$ and $\frac{\sqrt{G}}{c^2} \phi(r)$.
$\tilde \lambda = 1, m = 0.1, \eta^*_0 = .9083$.}
\label{eta}
\end{center}
\end{minipage}\hfill
\begin{minipage}[t]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.8\linewidth]{efield}}
\caption{The profile of the electric field $\frac{\sqrt{G}}{c^2} E_r(r)$.
$\tilde \lambda = 1, m = 0.1, \eta^*_0 = .9083$.}
\label{electric}
\end{center}
\end{minipage}\hfill
\end{figure}
\begin{figure}[ht]
\begin{minipage}[t]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.8\linewidth]{rho}}
\caption{The profile of the charge density $16 \pi \frac{\sqrt{G}}{c^2} \rho(r)$.
$\tilde \lambda = 1, m = 0.1, \eta^*_0 = .9083$.}
\label{rho}
\end{center}
\end{minipage}\hfill
\end{figure}
\section{How one can measure nonassociativity in physics}
The question stated in the title of this section is analogous to standard quantum theory: How can one measure \textcolor{blue}{\emph{the noncommutativity}} of operators? For the standard quantum theory, the answer is: The Planck constant measures the noncommutativity of conjugated operators.
In this paper, we proposed an approach to the nonperturbative quantization through nonassociative decomposition of quantum field operators. During a lunch with Geoffrey Dixon, Tevian Dray, John Huerta, Jens K\"oplinger and Shahn Majid (on 2$^{nd}$ Mile High Conference on Nonassociative Mathematics, Denver, Colorado, USA) the question came up, on how one could measure nonassociativity in physics ?
We have considered two examples: (1) polynomial potential in Section \ref{scalar}; (2) gravity in \ref{gravity}. The calculations presented in \ref{scalar} for the scalar fields can be extended to any field theory having a polynomial potential term (for example, for a gauge theory). In gauge theory the Lagrangian is
\begin{equation}
\label{5-10}
\mathcal L = \frac{1}{4} F^a_{\mu \nu} F^{a \mu \nu}.
\end{equation}
where $F^a_{\mu \nu} = \partial_\mu A^a_\nu - \partial_\nu A^a_\mu + g f^{abc} A^b_\mu A^c_\nu$ is the field strength; $A^a_\mu$ is the gauge potential; $a,b,c = 1, \ldots ,N$ are the SU(N) color indices; $g$ is the coupling constant; $f^{abc}$ are the structure constants for the SU(N) gauge group.
The Lagrangian \eqref{5-10} has the term $f^{abc} f^{ade} A^b_\mu A^c_\nu A^{d \mu} A^{e \nu}$ which is similar to the scalar potential $\phi^4$. Reasoning similar to section \ref{scalar} leads to appearance of an additional mass term $m^2 A^a_\mu A^{a \mu}$. Such mass term controls the radius of the interaction $A_\mu (r) \sim e^{- m r}/r$. It means that the nonassociativity manifests itself as the radius of the interaction $r_{int} \sim 1/m$. The appearance of the mass term $m^2 A^a_\mu A^{a \mu}$ signifies breaking of gauge invariance. Quantum chromodynamics is a
field theory with strong interaction, for which above arguments are applicable. The radius $r_{int} \sim 1/m$ can be considered as a radius of a flux tube filled with a chromoelectric field and stretched between quark - antiquark. Thus n/a parameter $m^{-1}$ can be measured in principle.
In Section \ref{gravity} we have considered quantum corrections for gravity. Exact calculations cannot be done in this case because we do not know the exact form of the nonperturbative operators $\hat g^{\mu \nu}$, $\widehat{\sqrt{-g}}$, and so on. We have proposed the torsion as a quantum correction from the Christoffel symbols. In subsection \ref{ed_gravity_1}1 we have shown that such corrections lead to smoothing of singularities, together with infinite total field energy of a point charge. In subsection \ref{ed_gravity_2} we have considered the nonlinear system of gravity + electrodynamics. In such system we also can not calculate quantum corrections. But we have shown that if the quantum correction has some definite form then all infinities in point charge disappear.
In summarizing, nonassociative quantum corrections can be measured:
\begin{itemize}
\item In nonperturbative quantum field theory, nonperturbative (nonassociative) quantum corrections can be measured as a radius of corresponding forces.
\item In gauge theories with big enough coupling constant, the nonassociativity gives rise to breaking of gauge invariance and formation of nonlocal objects (flux tubes) with characteristic length reciprocal to n/a parameter $m$.
\item In pure gravity, n/a quantum corrections appear in affine connections as torsion. It can in principle be measured, but probably only on very small (Planck) distances.
\item In gravity + electromagnetism, system n/a quantum corrections probably gives rise to smoothing of all singularities connected with a point charge. It leads to the possibility to modeling of a zero-spin charged particles.
\end{itemize}
\section{Discussion and conclusions}
Here we have considered nonperturbative quantization procedures based on nonassociative decomposition of quantum field operators, into nonassociative constituents. We have seen that such approach gives rise to quantum corrections, by calculation of expectation values of field
operators on nonlinear functions \footnote{potential term $V(\phi) = \phi^n, n \geq 4$ in a scalar field theory; $f^{abc} f^{ade} A^b_\mu A^c_\nu A^{d \mu} A^{e \nu}$ in a gauge theory, Christoffel symbols in gravity}. We have shown that these corrections can in principle be measured as a radius of a force, characteristic length of nonlocal objects, and the failure of connection compatibility with metric. Also in such way one can regularize singularities of a point-like charge. Let us note that in Ref. \cite{Zaslavskii:2010yi} similar idea was considered. It has been shown that in non-linear electrodynamics in the framework of general relativity there exist ``weakly singular'' configurations such that (i) the proper mass is finite in spite of divergences of the energy density, (ii) all field and energy distributions are concentrated in the core region.
In Section \ref{ed_gravity_2} we have shown that quantum corrections having a false vacuum and two true vacuums gives rise to regular solution describing a regular charge distribution. The corrections considered allow us to assume that such peculiarity will be retained for nonperturbative quantum corrections similar to a Mexican hat potential.
According to \eqref{4-1-80} the torsion is controlled by the factor
$\frac{G}{c^4} \phi^2 = \sinh(\eta/2)$. If $ \sinh(\eta/2) \ll 1$ then quantum corrections can be neglected. Otherwise, the corrections join the solutions for electric Coulomb fields. Such construction may model an isolated electric charge in Minkowski spacetime.
Quantum mechanics manifests itself through appearance of the Planck constant $\hbar$. In this paper, we showed that nonassociativity may exhibit a constant in quantum field theory, $m^2$;
or otherwise become evident as a geometric property in quantum gravity, torsion. Further investigation is be needed to clarify these relations.
\section{Acknowledgments}
I am grateful to the Research Group Linkage Program of the Alexander von Humboldt Foundation for financial support, and to Geoffrey Dixon, Tevian Dray, John Huerta, Jens K\"oplinger and Shahn Majid for fruitful discussions. Special thanks to F.-W. Hehl for the fruitful discussion and criticism.
|
\section{Introduction}
Einstein-Podolsky-Rosen (EPR) paradox \cite{EPR} and Bell's theorem \cite{Bell} --- cornerstones of quantum information and cryptography --- deal with relativistic issues of locality, but methodology of proofs is non-relativistic. In the mid-1980s some authors, including myself, tried to amend the inconsistency and reformulated Bell's theorem in relativistic formalisms (relativistic quantum mechanics \cite{MC84}, algebraic field theory \cite{SW0}). First relativistic results were not entirely trivial (violation of Bell's inequality by vacuum fluctuations, non-invariance of EPR correlations for spins) but did not attract much attention. The turning point for the emerging field of relativistic quantum information was the paper by Peres, Scudo and Terno \cite{PST} on non-invariance of entropy defined in terms of `reduced spin density matrices'. The line of research following from \cite{PST} culminated in the review paper \cite{PT} and is still continued by various authors.
The goals of the present paper are, in a sense, complementary to those of \cite{PT}. I want to concentrate on manifestly covariant 2-spinor approach to storage and communication qubits, technically and philosophically inspired by the works of Penrose \cite{PR,PR2}, but not widely known and not included in \cite{PT}. One of the unusual features of qubits formulated in a 2-spinor way, and going beyond the standard helicity formalism, is the fundamental role of projections of spin on {\it null\/} directions in space-time, even if massive particles are concerned. In particular, it is always possible to project the Pauli-Lubanski vector on principal null directions of SL(2,C) transformations. Decoherence of the Peres-Scudo-Terno type is then eliminated, at least for massive particles \cite{MCMW}.
The second class of results comes from application of null directions to electromagnetic fields. It can be shown that a class of twistor-type spin-frames has covariance properties leading to 4-potentials that do not change gauge after Lorentz transformations (i.e. $A_a(x)$ is a world-vector field). Simultaneously, the same twistor-like transformation implies that momentum-space amplitudes of $A_a(x)$ split into classes belonging to different representations of the Poincar\'e group: Spin-1 zero-mass representation which we regard as photons, and two additional scalar massless fields. Electromagnetic qubits are associated only with the spin-1 part.
The construction can be performed for both first- and second-quantized fields. We concentrate on quantized Maxwell fields, but quantization procedure is not the usual one. I believe the procedure I employ, based on reducible representations of harmonic-oscillator Lie algebras, is in many respects superior to the standard one, where irreducibility is implicitly assumed. Readers interested in more details of the relativistic formalism based on reducible representations are referred to \cite{MCKW}.
In order to analyze relativistic EPR correlations of photons we have to make the notion of an EPR state more precise. The EPR paradox is based, technically speaking, on maximal entanglement in at least two complementary bases --- I term this property the EPR condition. In non-relativistic quantum mechanics EPR states are simultaneously maximally entangled in all bases, and maximally symmetric. In the relativistic case we have a similar object, the two-index antisymmetric spinor $\varepsilon_{AB}$. The problem is that $\varepsilon_{AB}$ is invariant with respect to spinor representations, while these are the unitary representations we have to work with. Scalar states were used in the EPR context by Caban \cite{Caban}, but it turns out that typical scalar states cannot be identified with EPR states, unless the two photons have the same momenta. The EPR condition is not SL(2,C) invariant if the two photons have different momenta, a fact responsible for relativistic non-invariance of EPR correlations for linear polarizations (and certain interferometric qubits).
Another conceptual difficulty with relativistic qubits is the choice of yes-no observables. For massive particles a natural definition is given by the Pauli-Lubanski vector (cf. \cite{MC-PRA}) but the resulting observable is not linear in number of particles. I show how to define an analogue of the Pauli-Lubanski vector which is linear in number of particles, but this is done in representations of harmonic oscillator Lie algebras that lead to a well behaved vacuum part of 4-momentum (this is one of the reasons why the reducible representations are here useful). Still, components of the Pauli-Lubanski vector commute in the massless case, so are useless for EPR correlations.
Since the field is quantized in a nonstandard way, one might ask to what extent the EPR averages for linear polarizations and interferometric observables are sensitive to modifications of field quantization paradigm. The answer is negative: In both reducible and irreducible representations of oscillator algebras the EPR averages possess the well known $\cos 2(\alpha-\beta)$ term, multiplied by $p$ parameterized by sets of wave vectors analyzed by photon detectors.
The paper is organized as follows. In Sec.~2 I recall basic links between 2-spinors and null vectors, and stress certain similarity between tetrads in Minkowski space and 2-qubit states. This section simultaneously introduces abstract-index notation needed later in the paper. Sec.~3 deals with first-quantized Dirac fields. I analyze properties of the Pauli-Lubanski vector and show why it pays to work at a level more general than spin defined via helicity. Sec.~3.5 discusses an unorthodox approach to unitary representations of the Poincar\'e group, regarded as passive versions of active spinor representations. In Sec.~3.7 I concentrate on the Peres-Scudo-Terno problem and show that a formalism based on principal null directions plays in this context a role of error correction. Sec.~4 introduces basic 2-spinor concepts needed for massless fields. I emphasize the special role played by twistor-like spin-frames and, in Sec.~5, apply them to first-quantized electromagnetic qubits. In Sec.~6 I perform field quantization in terms of reducible representations of harmonic oscillator Lie algebras, and define polarization operators that are linear in numbers of particles. In Sec.~7 I concentrate on relativistic aspects of EPR states. EPR correlations are computed in Sec.~8. The main result shows that the reducible representations I advertise predict essentially the same formula for EPR averages as the usual formalism. This is perhaps the most important new result of the paper.
Sec.~9 discusses the issues of `entanglement with vacuum' and `vacuum violations' of Bell's inequality. I illustrate on simple examples the ideas behind these approaches, and show that they are related to different meanings of the notion of `vacuum'.
I end the paper with technical Appendices related to the choice of representations employed in my approach to field quantization.
\section{Two-spinors: Geometric and algebraic preliminaries}
Let us consider three components $(x,y,z)$ of a unit vector,
\begin{eqnarray}
x^2+y^2+z^2 &=& 1,\label{S}
\end{eqnarray}
taken in some basis in $\bm R^3$. (\ref{S}) describes a unit sphere whose points can be equivalently parametrized by means of stereographic projection on the plane $z=0$. A particularly useful form of this map
is obtained if the plane is parametrized by complex coordinates $\zeta$, linked to the point on the sphere by $\zeta =(x+iy)/(1-z)$. Let us note that the `north pole' $(0,0,1)$ corresponds to $\zeta=\infty$, whereas the `south pole' $(0,0,-1)$ implies $\zeta=0$. Yet another possible parametrization of the sphere is in terms of projective coordinates $(\xi,\eta)$ satisfying $\zeta=\xi/\eta$. Now the poles correspond to $(\xi,0)$ and $(0,\eta)$. The coordinates are projective since $(\xi,\eta)$ and $(\lambda\xi,\lambda\eta)$, for any $\lambda\neq 0$, define the same point on the sphere.
Let us now treat our sphere as a light cone in Minkowski space,
\begin{eqnarray}
1-x^2-y^2-z^2=0.
\end{eqnarray}
The four components $(1,x,y,z)$ of a null future-pointing world-vector are linked to projective coordinates by
\begin{eqnarray}
(1,x,y,z) &=& \Big(1,\frac{\xi\bar\eta+\eta\bar\xi}{\xi\bar\xi+\eta\bar\eta},-i\frac{\xi\bar\eta-\eta\bar\xi}{\xi\bar\xi+\eta\bar\eta},
\frac{\xi\bar\xi-\eta\bar\eta}{\xi\bar\xi+\eta\bar\eta}\Big).
\end{eqnarray}
Multiplying both sides by $(\xi\bar\xi+\eta\bar\eta)/\sqrt{2}$ we obtain coordinates
\begin{eqnarray}
(T,X,Y,Z) &=&
\frac{1}{\sqrt{2}}\Big(\xi\bar\xi+\eta\bar\eta,\xi\bar\eta+\eta\bar\xi,-i(\xi\bar\eta-\eta\bar\xi),\xi\bar\xi-\eta\bar\eta\Big)\label{TXYZ}
\end{eqnarray}
of a point belonging to the same null line in Minkowski space as the point corresponding to $(1,x,y,z)$. (\ref{TXYZ}) is future-pointing since $T> 0$ for a non-vanishing $(\xi,\eta)$. There is, of course, nothing fundamentally special about $1/\sqrt{2}$, but this choice is convenient, as we shall see later. Given a null future-pointing world-vector with components $(T,X,Y,Z)$, and solving the four equations (\ref{TXYZ}), one can determine $\xi$ and $\eta$ up to a common phase factor.
$(T,X,Y,Z)$ is termed the flagpole of $(\xi,\eta)$. A flag plane can be defined by a vector tangent at $(x,y,x)$ to the sphere (\ref{S}). The flagpole and the flag plane determine $(\xi,\eta)$ up to a sign. This sign ambiguity is related to topological properties of the rotation roup.
I have purposefully stressed that $(T,X,Y,Z)$ are {\it components\/} of a Minkowski-space world-vector evaluated with respect to {\it some\/} basis. The pair $(\xi,\eta)$ can be associated with any basis in a 2-dimensional complex space, but there is no obvious link between the Minkowski space and this 2-dimensional linear space.
Putting it differently, if $\bm t$, $\bm x$, $\bm y$, $\bm z$, define a Minkowski tetrad the vector $X \bm x+Y \bm y+Z \bm z+T \bm t$ can be linked with a vector (a 2-spinor) $\xi \bm o+ \eta \bm \iota$, where $\bm o$ and $\bm \iota$ are basis vectors in $\bm C^2$, but there is no {\it a priori\/} link between $\bm t$, $\bm x$, $\bm y$, $\bm z$ and $\bm o$, $\bm \iota$. However, let us consider the tensor product of this spinor with its complex conjugate,
\begin{eqnarray}
(\xi \bm o+ \eta \bm \iota)\otimes(\bar\xi \bar{\bm o}+ \bar\eta \bar{\bm \iota})
&=&
\xi\bar\xi \bm o\otimes\bar{\bm o}
+
\xi\bar\eta \bm o \otimes\bar{\bm \iota}
+
\eta\bar\xi \bm \iota\otimes\bar{\bm o}
+
\eta \bar\eta \bm \iota\otimes\bar{\bm \iota}
\\
&=&
\underbrace{\frac{1}{\sqrt{2}}\Big(\xi\bar\eta+\eta\bar\xi\Big)}_X
\tilde{\bm x}
\underbrace{-
i\frac{1}{\sqrt{2}}\Big(\xi\bar\eta-\eta\bar\xi\Big)}_Y
\tilde{\bm y}
+
\underbrace{\frac{1}{\sqrt{2}}\Big(\xi\bar\xi-\eta\bar\eta\Big)}_Z
\tilde{\bm z}
\nonumber\\
&\pp=&
+
\underbrace{\frac{1}{\sqrt{2}}\Big(\xi\bar\xi+\eta\bar\eta\Big)}_T
\tilde{\bm t}
,
\end{eqnarray}
where
\begin{eqnarray}
\tilde{\bm t}
&=&
\frac{1}{\sqrt{2}}\Big(\bm o\otimes\bar{\bm o}+\bm \iota\otimes\bar{\bm \iota}\Big),\label{tilde t}\\
\tilde{\bm x}
&=&
\frac{1}{\sqrt{2}}\Big(\bm o\otimes\bar{\bm \iota}+\bm \iota\otimes\bar{\bm o}\Big),\label{tilde x}\\
\tilde{\bm y}
&=&
\frac{i}{\sqrt{2}}\Big(\bm o\otimes\bar{\bm \iota}-\bm \iota\otimes\bar{\bm o}\Big),\label{tilde y}\\
\tilde{\bm z}
&=&
\frac{1}{\sqrt{2}}\Big(\bm o\otimes\bar{\bm o}-\bm \iota\otimes\bar{\bm \iota}\Big),\label{tilde z}
\end{eqnarray}
and
\begin{eqnarray}
\bm o\otimes\bar{\bm o}
&=&
\frac{1}{\sqrt{2}}\Big(\tilde{\bm t}+\tilde{\bm z}\Big),\label{oo}\\
\bm o \otimes\bar{\bm \iota}
&=&
\frac{1}{\sqrt{2}}\Big(\tilde{\bm x}-i\tilde{\bm y}\Big)\label{oi},\\
\bm \iota\otimes\bar{\bm o}
&=&
\frac{1}{\sqrt{2}}\Big(\tilde{\bm x}+i\tilde{\bm y}\Big)\label{io},\\
\bm \iota\otimes\bar{\bm \iota}
&=&
\frac{1}{\sqrt{2}}\Big(\tilde{\bm t}-\tilde{\bm z}\Big).\label{ii}
\end{eqnarray}
A change of basis $\{\bm o,\bm\iota\}\to \{\bm o',\bm\iota'\}$ implies an associated change of components $(\xi,\eta)\to (\xi',\eta')$. If $(\xi,\eta)\to (\xi',\eta')$ is a linear transformation with unit determinant (an element of the group SL(2,C)) then the associated transformation
$(X,Y,Z,T)\to (X',Y',Z',T')$ turns out to be a proper isochronous Lorentz transformation. Since, $X\tilde{\bm x}+Y\tilde{\bm y}+Z\tilde{\bm z}+T\tilde{\bm t}=X'\tilde{\bm x}{'}+Y'\tilde{\bm y}{'}+Z'\tilde{\bm z}{'}
+T'\tilde{\bm t}{'}$, the map $\{\tilde{\bm x},\tilde{\bm y},\tilde{\bm z},\tilde{\bm t}\}\to
\{\tilde{\bm x}{'},\tilde{\bm y}{'},\tilde{\bm z}{'},\tilde{\bm t}{'}\}$ is a proper isochronous Lorentz transformation as well. Tensor products of 2-spinors with their complex conjugates transform under SL(2,C) transformations in the same way as world-vectors in Minkowski space under boosts and rotations. This leads us to the abstract-index formalism.
\subsection{Abstract-index formalism of Penrose}
Second-rank spinors $\{\tilde{\bm x},\tilde{\bm y},\tilde{\bm z},\tilde{\bm t}\}$ have transformation properties of a Minkowski tetrad.
Spinor approach to space-time geometry is based on the assumption that one can {\it identify\/} the two bases, $\{\tilde{\bm x},\tilde{\bm y},\tilde{\bm z},\tilde{\bm t}\}$ in $\bm C^2\otimes \bm C^2$ and
$\{{\bm x},{\bm y},{\bm z},{\bm t}\}$ in Minkowski space, and simply skip the tildas in (\ref{tilde t})--(\ref{tilde z}).
From this perspective 2-spinors are more fundamental than world-vectors in Minkowski space. The Minkowski space can be regarded as a structure derived from a more fundamental spinorial level.
This is the starting point of the abstract-index formalism of Penrose \cite{PR}, where (\ref{tilde t})--(\ref{tilde z}) would be written as
\begin{eqnarray}
t^a&=&
\frac{1}{\sqrt{2}}\Big(o^A \bar o^{A'}+\iota^A\bar \iota^{A'}\Big)=t^{AA'},\\
x^a
&=&
\frac{1}{\sqrt{2}}\Big(o^A \bar \iota^{A'}+\iota^A \bar o^{A'}\Big)=x^{AA'},\\
y^a
&=&
\frac{i}{\sqrt{2}}\Big(o^A \bar \iota^{A'}-\iota^A \bar o^{A'}\Big)=y^{AA'},\\
z^a&=&
\frac{1}{\sqrt{2}}\Big(o^A \bar o^{A'}-\iota^A\bar \iota^{A'}\Big)=z^{AA'}.
\end{eqnarray}
The indices are just labels (analogous to `Alice' and `Bob'; think of $t^{\rm alice}=t^{\rm Alice,Alice'}$) and do not take numerical values.
The null tetrad corresponding to (\ref{oo})--(\ref{ii}) is expressed in the abstract-index form as
\begin{eqnarray}
l^a &=& o^A \bar o^{A'},\\
m^a &=& o^A \bar \iota^{A'},\\
\bar m^a &=& \iota^A \bar o^{A'},\\
n^a &=& \iota^A \bar \iota^{A'}.
\end{eqnarray}
In this approach there are no operations that would involve (anti-)symmetrizations of primed and unprimed indices. In consequence, we do not loose generality by assuming that the primed labels occur to the right of the unprimed ones. In particular, we identify $\alpha_A\beta_{A'}$ with $\beta_{A'}\alpha_A$ for all 2-spinors $\alpha_A$ and $\beta_{A'}$, and one does not need to distinguish between hermiticity and reality. As an illustration, let us consider reality properties of the null tetrad:
\begin{eqnarray}
\overline{l^a} &=& \overline{o^A} \overline{\bar o^{A'}}=\bar o^{A'}o^A=o^A\bar o^{A'}=l^a ,\\
\overline{m^a} &=& \overline{o^A} \overline{\bar \iota^{A'}}=\bar o^{A'}\iota^A=\iota^A\bar o^{A'}=\bar m^a,
\end{eqnarray}
and so on. This formalism is very convenient and does not lead to ambiguities in practical computations but may seem somewhat counterintuitive, especially to those who are accustomed to the more traditional spinor notation employed in relativistic quantum mechanics. This is why in the next subsection we shall describe a variant of the Penrose formalism that can be directly translated into formulas we know from quantum mechanics textbooks.
But before we do that let us discuss relations between abstract and numerical indices. Numerical indices, 0 and 1, are sometimes needed but then we denote them by upright boldface fonts. Accordingly, the symbol $\phi^A$ denotes a 2-spinor (`a spinor of Alice'), but $\phi^{\bf A}$ may equal $\phi^{0}$ or $\phi^{1}$. $\phi^A$ is basis independent, but $\phi^{\bf A}$ implicitly depends on a basis.
Now consider two 2-spinors, $\alpha^A=\alpha^0 o^A+\alpha^1 \iota^A=\tilde\alpha^0 \tilde o^A+\tilde\alpha^1 \tilde\iota^A$, $\beta^A=\beta^0 o^A+\beta^1 \iota^A=\tilde\beta^0 \tilde o^A+\tilde\beta^1 \tilde\iota^A$, where the two bases are related by $\tilde o^A=(So)^A$, $\tilde\iota^A=(S\iota)^A$, $\det S=1$. The determinant
\begin{eqnarray}
\left|
\begin{array}{cc}
\alpha^0 & \alpha^1\\
\beta^0 & \beta^1\\
\end{array}
\right|
=
\left|
\begin{array}{cc}
\tilde\alpha^0 & \tilde\alpha^1\\
\tilde\beta^0 & \tilde\beta^1\\
\end{array}
\right|=\alpha^0\beta^1-\alpha^1\beta^0=\varepsilon_{\bf AB}\alpha^{\bf A}\beta^{\bf B}=\alpha_{\bf B}\beta^{\bf B}=-
\alpha^{\bf A}\beta_{\bf A}\label{contr}
\end{eqnarray}
is independent of $S$ and defines
\begin{eqnarray}
\varepsilon_{\bf AB}
=
\left(
\begin{array}{cc}
0 & 1\\
-1 & 0\\
\end{array}
\right).
\end{eqnarray}
The last two equalities in (\ref{contr}) show how to lower spinor indices in SL(2,C)-invariant manner: $\alpha_{\bf B}=\alpha^{\bf A}\varepsilon_{\bf AB}$. The inverse rule is $\varepsilon^{\bf AB}\alpha_{\bf B}=\alpha^{\bf A}$, with $\varepsilon^{\bf AB}=\varepsilon_{\bf AB}$.
The lower-index $\alpha_{\bf B}$ may be regarded as a component of a spinor $\alpha_{A}$ dual to $\alpha^{A}$, the duality being given by
$\alpha_{A}\beta^{A}=\alpha_{\bf A}\beta^{\bf A}$. The formulas $\alpha_{B}=\alpha^{A}\varepsilon_{AB}$ and $\varepsilon^{AB}\alpha_{B}=\alpha^{A}$ define at the abstract-index level the isomorphisms between modules of upper-index 2-spinor fields and their duals. Note that $\varepsilon^{CB}\varepsilon_{CA}=\varepsilon{_A}{^B}$ acts by
$\varepsilon{_A}{^B}\phi_B=\phi_A$, and
\begin{eqnarray}
\varepsilon{_{\bf A}}{^{\bf B}}
=
\left(
\begin{array}{cc}
1 & 0\\
0 & 1\\
\end{array}
\right).
\end{eqnarray}
$\varepsilon{_A}{^B}$ thus plays a role of spinorial Kronecker delta; at the abstract-index level $\varepsilon{_A}{^B}$ is the isomorphism between `the spinor of Bob' and `the spinor of Alice', a map quite similar to the teleportation protocol \cite{MC-CQG}.
The basis $o^A$ and $\iota^A$, satisfying $o_A\iota^A=1$, is termed the spin-frame. Since $\phi_A\phi^A=0$ for any $\phi_A$, spin-frames possess a kind of gauge freedom: $(o_A+\lambda \iota_A)\iota^A=o_A(\iota^A+\mu o^A)=o_A\iota^A=1$. It is interesting that in electrodynamics this ambiguity indeed manifests itself in a form of gauge transformation associated with Lorentz transformations of four-potentials. Any spin-frame satisfies
\begin{eqnarray}
\varepsilon_{AB} &=& o_A\iota_B-\iota_A o_B,\label{ve_A_B}\\
\varepsilon{_A}{^B} &=& o_A\iota^B-\iota_A o^B.\label{ve_A^B}
\end{eqnarray}
If $g_{ab}$ is the Minkowski-space metric tensor then $g{_a}{^b}$ must be the abstract-index form of Kronecker's delta in Minkowski space. Accordingly, $g{_a}{^b}=\varepsilon{_A}{^B}\varepsilon{_{A'}}{^{B'}}$ ($\overline{\varepsilon{_A}{^B}}=\varepsilon{_{A'}}{^{B'}}$). Lowering the indices we obtain the following three useful abstract-index forms of the metric:
\begin{eqnarray}
g_{ab} &=& \varepsilon_{AB}\varepsilon_{A'B'}\label{g1}\\
&=&
t_at_b-x_ax_b-y_ay_b-z_az_b\label{g2}\\
&=&
n_al_b+l_an_b-\bar m_am_b-m_a\bar m_b.\label{g3}
\end{eqnarray}
The reader may have noticed that in the Penrose formalism the Minkowski tetrad has all the properties of the Bell basis of two-qubit entangled states. The null basis $l=\bm o\otimes\bar{\bm o}=l^a$, $m=\bm o\otimes\bar{\bm \iota}=m^a$, $\bar m=\bm \iota\otimes\bar{\bm o}=\bar m^a$, $n=\bm \iota\otimes\bar{\bm \iota}=n^a$ is analogous to the basis of product states. There exists an unexplored `duality' between space-time geometry and quantum information theory. Some preliminary considerations in this spirit can be found in \cite{MC-CQG}, where analogies between teleportation protocols and metric tensors of Lorentzian manifolds were discussed. Lorentzian techniques, with applications to multi-qubit entanglement, can be found also in \cite{Levay,Frumosu,Jaeger,Jaeger2}.
\subsection{Abstract-index analogues of Pauli matrices: Infeld-van der Waerden tensors}
Let us return to the formula (\ref{TXYZ}), but written as
\begin{eqnarray}
X
&=&(X^0,X^1,X^2,X^3) \nonumber\\
&=&
\frac{1}{\sqrt{2}}\Big(\phi^0\bar\phi^{0'}+\phi^1\bar\phi^{1'},\phi^0\bar\phi^{1'}+\phi^1\bar\phi^{0'},-i(\phi^0\bar\phi^{1'}-\phi^1\bar\phi^{0'}),
\phi^0\bar\phi^{0'}-\phi^1\bar\phi^{1'}\Big)\label{TXYZ1}\\
&=&
(g{^0}{_{\bf A}}{_{{\bf A}'}},g{^1}{_{\bf A}}{_{{\bf A}'}},g{^2}{_{\bf A}}{_{{\bf A}'}},g{^3}{_{\bf A}}{_{{\bf A}'}})\phi^{\bf A}\bar\phi^{{\bf A}'},\\
X^{\bf a}
&=&
g{^{\bf a}}{_{\bf A}}{_{{\bf A}'}}\phi^{\bf A}\bar\phi^{{\bf A}'}.
\end{eqnarray}
The coefficients $g{^{\bf a}}{_{\bf A}}{_{{\bf A}'}}$ are known as Infeld-van der Waerden symbols, and can be grouped into four matrices
\begin{eqnarray}
g{^0}{_{\bf A}}{_{{\bf A}'}}
&=&
\left(
\begin{array}{cc}
g{^0}{_{0}}{_{0'}} & g{^0}{_{0}}{_{1'}}\\
g{^0}{_{1}}{_{0'}} & g{^0}{_{1}}{_{1'}}
\end{array}
\right)
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
1 & 0\\
0 & 1
\end{array}
\right)=g{_0}{^{\bf A}}{^{{\bf A}'}},\\
g{^1}{_{\bf A}}{_{{\bf A}'}}
&=&
\left(
\begin{array}{cc}
g{^1}{_{0}}{_{0'}} & g{^1}{_{0}}{_{1'}}\\
g{^1}{_{1}}{_{0'}} & g{^1}{_{1}}{_{1'}}
\end{array}
\right)
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
0 & 1\\
1 & 0
\end{array}
\right)=g{_1}{^{\bf A}}{^{{\bf A}'}},\\
g{^2}{_{\bf A}}{_{{\bf A}'}}
&=&
\left(
\begin{array}{cc}
g{^2}{_{0}}{_{0'}} & g{^2}{_{0}}{_{1'}}\\
g{^2}{_{1}}{_{0'}} & g{^2}{_{1}}{_{1'}}
\end{array}
\right)
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
0 & -i\\
i & 0
\end{array}
\right)=-g{_2}{^{\bf A}}{^{{\bf A}'}},\\
g{^3}{_{\bf A}}{_{{\bf A}'}}
&=&
\left(
\begin{array}{cc}
g{^3}{_{0}}{_{0'}} & g{^3}{_{0}}{_{1'}}\\
g{^3}{_{1}}{_{0'}} & g{^3}{_{1}}{_{1'}}
\end{array}
\right)
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
1 & 0\\
0 & -1
\end{array}
\right)=g{_3}{^{\bf A}}{^{{\bf A}'}}.
\end{eqnarray}
The numerical indices are lowered or raised by means of the epsilons and $g^{\bf ab}=g_{\bf ab}={\rm diag}(1,-1,-1,-1)$.
In quantum mechanics textbooks one works with matrix 4-vectors $\sigma=(1,\sigma_x,\sigma_y,\sigma_z)$ and $\tilde\sigma=(1,-\sigma_x,-\sigma_y,-\sigma_z)=\varepsilon\bar\sigma\varepsilon^T$. Infeld-van der Waerden symbols, with appropriately raised or lowered indices, play precisely the same role, but make the formalism more flexible and prepared for more advanced applications.
An important relation between the four types of coefficients,
\begin{eqnarray}
g_{\bf ab}
&=&
g{_{\bf a}}{^{\bf A}}{^{{\bf A}'}}g{_{\bf b}}{^{\bf B}}{^{{\bf B}'}}\varepsilon_{\bf AB}\varepsilon_{{\bf A}'{\bf B}'},
\end{eqnarray}
can be rewritten in several {\it equivalent\/} ways, each of them revealing another aspect of their mutual relations:
\begin{eqnarray}
g_{\bf ab}g{^{\bf a}}{_{\bf A}}{_{{\bf A}'}}g{^{\bf b}}{_{\bf B}}{_{{\bf B}'}}
&=&
\varepsilon_{\bf AB}\varepsilon_{{\bf A}'{\bf B}'},\\
g{^{\bf a}}{_{\bf A}}{_{{\bf A}'}}g{_{\bf a}}{^{\bf B}}{^{{\bf B}'}}
&=&
\varepsilon{_{\bf A}}{^{\bf B}}\varepsilon{_{{\bf A}'}}{^{{\bf B}'}},
\\
g{_{\bf a}}{^{\bf A}}{^{{\bf A}'}}g{^{\bf b}}{_{\bf A}}{_{{\bf A}'}}
&=&
g{_{\bf a}}{^{\bf b}},\\
g{_{\bf a}}{_{\bf A}}{_{{\bf A}'}}g{_{\bf b}}{^{\bf B}}{^{{\bf A}'}}
+
g{_{\bf b}}{_{\bf A}}{_{{\bf A}'}}g{_{\bf a}}{^{\bf B}}{^{{\bf A}'}}
&=&
\varepsilon{_{\bf A}}{^{\bf B}}g_{\bf ab}.\label{Cl}
\end{eqnarray}
Eq.~(\ref{Cl}) is the 2-spinor form of the anticommutator known from the algebra of Dirac matrices.
The above formulas involve boldface (numerical) indices but can be regarded as components of appropriate abstract-index relations, say,
\begin{eqnarray}
g_{ab}
&=&
g{_{a}}{^{A}}{^{{A}'}}g{_{b}}{^{B}}{^{{B}'}}\varepsilon_{AB}\varepsilon_{{A}'{B}'},\label{43}
\end{eqnarray}
and the like. Infeld-van der Waerden {\it tensors\/} $g{_{a}}{^{A}}{^{{A}'}}$, of a mixed world-vector and 2-spinor type, play the roles of the isomorphisms allowing to identify world-vector indices with pairs of the spinor ones: $X_a=g{_{a}}{^{A}}{^{{A}'}}X_{AA'}$, and so forth with tensors of any rank.
Let us note that, in general, $g{_{a}}{^{A}}{^{{A}'}}$ may possess a nontrivial dependence on the point $x$ of an appropriate Lorentzian manifold. The identification $X_a=X_{AA'}$ implicitly assumes that in all differential equations one encounters in a given theory one can freely commute
$g{_{a}}{^{A}}{^{{A}'}}$ with derivatives. The latter means that one works with covariant derivatives satisfying $\nabla_a g{_{b}}{^{C}}{^{{C}'}}=0$ (plus analogous equations obtained by raising and lowering the indices). In most applications to Minkowski space we do not have to worry about such subtleties.
In the Penrose formalism no special role is given to Infeld-van der Waerden tensors --- they are just regarded as one of the possible instances of the abstract-index rule $g{_a}{^b}=g{_a}{^{BB'}}$. However, it is sometimes useful to work with Infeld-van der Waerden tensors occurring explicitly, if we need to translate some formulas into their more standard shapes. As an exercise, let us show several equivalent abstract-index versions of (\ref{43}):
\begin{eqnarray}
g_{ab}
&=&
g{_{a}}{^{C}}{^{{C}'}}g{_{b}}{^{D}}{^{{D}'}}\varepsilon_{CD}\varepsilon_{{C}'{D}'}
=
g{_{a}}{^{c}}g{_{b}}{^{d}}g_{cd}
=
g{_{a}}{^{c}}g{_{b}}{^{d}}g_{CC'DD'}
=
g{_{a}}{^{CC'}}g{_{b}}{^{DD'}}g_{cDD'}\nonumber\\
&=&
g{_{AA'}}{^{c}}g{_{BB'}}{^{d}}g_{cd}
=
g{_{AA'}}{^{c}}g{_{BB'}}{^{d}}g_{CC'd}
=
\varepsilon{_{A}}{^{C}}\varepsilon{_{A'}}{^{C'}}
\varepsilon{_{B}}{^{D}}\varepsilon{_{B'}}{^{D'}}
\varepsilon{_{C}}{_{D}}\varepsilon{_{C'}}{_{D'}}
\nonumber\\
&=&
g{_{AB'}}{^{CD'}}
g{_{BA'}}{^{DC'}}
\varepsilon{_{C}}{_{D}}\varepsilon{_{C'}}{_{D'}}=g{_{AB'}}{^{CD'}}
g{_{BA'}}{^{DC'}}
g_{cd}=\varepsilon{_{A}}{_{B}}\varepsilon{_{A'}}{_{B'}}.
\nonumber
\end{eqnarray}
Flexibility of switching between forms is indeed immense. We shall use this freedom in derivation of generators of certain important representations of SL(2,C).
\subsection{Tetrad and diad notation}
Another notational element we will need is based on the following observation. Let us note that, by definition of the relation between abstract and numerical indices, any world-vector satisfies
\begin{eqnarray}
X_a &=& g{_a}{^b}X_b=g{_a}{^{\bf b}}X_{\bf b}=g{_a}{^{0}}X_{0}+g{_a}{^{1}}X_{1}+g{_a}{^{2}}X_{2}+g{_a}{^{3}}X_{3},
\end{eqnarray}
so that the four {\it world-vectors\/} $g{_a}{^{\bf b}}$ play a role of a basis. Actually, this is a Minkowski tetrad, since
\begin{eqnarray}
g{_a}{^{\bf b}}g{^a}{^{\bf c}}
&=&
g{_{\bf a}}{^{\bf b}}g{^{\bf a}}{^{\bf c}}=g{^{\bf b}}{^{\bf c}}.
\end{eqnarray}
Analogously,
\begin{eqnarray}
g{^{a}}{^{b}}
&=&
g{_c}{^{a}}g{^c}{^{b}}
=
g{_{\bf c}}{^{a}}g{^{\bf c}}{^{b}}
=
g^{\bf cd}g{_{\bf c}}{^{a}}g{_{\bf d}}{^{b}}
\\
&=&
g{_{0}}{^{a}}g{_{0}}{^{b}}
-
g{_{1}}{^{a}}g{_{1}}{^{b}}
-
g{_{2}}{^{a}}g{_{2}}{^{b}}
-
g{_{3}}{^{a}}g{_{3}}{^{b}}
\end{eqnarray}
(compare Eq.~(\ref{g2})). A similar construction can be performed
with 2-spinors, starting with
\begin{eqnarray}
\phi_A
&=&
\varepsilon{_A}{^B}\phi_B=\varepsilon{_A}{^{\bf B}}\phi_{\bf B},
\end{eqnarray}
showing that $\varepsilon{_A}{^{\bf B}}$ is a diad of spinorial basis vectors. In particular, since $\varepsilon^{01}=1$ we find $\varepsilon{_A}{^{0}}\varepsilon{^A}{^{1}}=
\varepsilon{_{\bf A}}{^{0}}\varepsilon{^{\bf A}}{^{1}}=1$, so that $o_A=\varepsilon{_A}{^{0}}$, $\iota_A=\varepsilon{_A}{^{1}}$ form a spin-frame.
\subsection{Simplest spinor representations of SL(2,C) and their generators}
SL(2,C) transformations act by $\phi_A\to \tilde\phi_A=\Lambda{_A}{^B}\phi_B$, $\phi^A\to \tilde\phi^A=\phi^B\Lambda^{-1}{_B}{^A}$,
$\phi_{A'}\to \tilde\phi_{A'}=\Lambda{_{A'}}{^{B'}}\phi_{B'}$, $\phi^{A'}\to \tilde\phi^{A'}=\phi^{B'}\Lambda^{-1}{_{B'}}{^{A'}}$. Here $\Lambda{_{A}}{^{B}}$ and $\Lambda{_{A'}}{^{B'}}=\overline{\Lambda{_{A}}{^{B}}}$ are the two inequivalent 2-dimensional representations of $\Lambda\in$~SL(2,C). Let us note that transformation properties of lower- and upper-index spinors do not have to be separately postulated. Indeed, for SL(2,C) transformations we have
$\Lambda{_A}{^C}\Lambda{_B}{^D}\varepsilon{_C}{_D}=\varepsilon{_A}{_B}$. Raising appropriate indices and employing antisymmetry of the epsilon, we can transform this formula as follows
$-\Lambda{_A}{^C}\Lambda{^B}{_C}=\varepsilon{_A}{^B}$, which shows that $-\Lambda{^B}{_C}=\Lambda^{-1}{_C}{^B}$.
Representations $(\frac{1}{2},\frac{1}{2})$ (in Minkowski space), $(\frac{1}{2},0)$ (unprimed spinors), and $(0,\frac{1}{2})$ (primed spinors) of an element $\Lambda\in$~SL(2,C) are linked to the Lie-algebra of generators by:
$\Lambda{_r}{^s}=\exp\frac{-i}{2}{}y^{ab}\sigma_{ab}{_r}{^s}$, $\Lambda{_R}{^S}=\exp\frac{-i}{2}{}y^{ab}\sigma_{ab}{_R}{^S}$, and
$\Lambda{_{R'}}{^{S'}}=\exp\frac{-i}{2}{}y^{ab}\sigma_{ab}{_{R'}}{^{S'}}$. ${}y^{ab}$ is the antisymmetric tensor whose six independent components correspond to boosts ($y^{0\bf n}$) and rotations ($y^{\bf mn}$), ${\bf n,m}=1,2,3$. Accordingly, the generators can be obtained from
\begin{eqnarray}
\sigma_{ab}{_{\dots}}{^{\dots}}
&=&i\frac{\partial \Lambda{_{\dots}}{^{\dots}}}{\partial {}y^{ab}}\Big|_{{}y^{ab}=0}.
\end{eqnarray}
Differentiating at ${}y^{ab}=0$ both sides of $\Lambda{_A}{^C}\Lambda{_B}{^D}\varepsilon{_C}{_D}=\varepsilon{_A}{_B}$ we obtain
$\sigma_{ab}{_A}{_B}=\sigma_{ab}{_B}{_A}$. Similarly, differentiating $\Lambda{_{r}}{^{s}}=\Lambda{_{R}}{^{S}}\Lambda{_{R'}}{^{S'}}$, we find
\begin{eqnarray}
\sigma{_{ab}}{_r}{^s}
&=&
\sigma_{ab}{_R}{^S}\varepsilon{_{R'}}{^{S'}}
+
\varepsilon{_R}{^S}\sigma_{ab}{_{R'}}{^{S'}}
\label{gen rs}.
\end{eqnarray}
Components $\sigma{_{\bf ab}}{_{\bf r}}{^{\bf s}}$ of the Minkowski-space generators can be collected into six matrices, the ones corresponding to boosts being symmetric, as opposed to the antisymmetric generators of rotations. Their manifestly covariant abstract-index form,
\begin{eqnarray}
\sigma{_{ab}}{_r}{^s}
&=&
i\,(g_{ar}g_b{^s} - g_{br}g_a{^s})
\end{eqnarray}
can be rewritten in all the possible equivalent ways by means of the tricks I have described above. In particular, we find
\begin{eqnarray}
\sigma{_{ab}}{_r}{^s}
&=&
i\,(g_{aRR'}g_b{^{SS'}} - g_{bRR'}g_a{^{SS'}}).
\label{gen rs'}
\end{eqnarray}
Since any symmetric spinor $\phi_{AB}=\phi_{BA}$ satisfies $\phi{_{A}}{^{A}}=\varepsilon^{AB}\phi_{AB}=\varepsilon^{BA}\phi_{BA}=-\varepsilon^{AB}\phi_{AB}=0$,
comparing (\ref{gen rs}) with (\ref{gen rs'}) we get
\begin{eqnarray}
\sigma_{ab}{_R}{^S}
&=&
\frac{i}{2}(g_{aRX'}g_b{^{SX'}} - g_{bRX'}g_a{^{SX'}})
,
\\
\sigma_{ab}{_{R'}}{^{S'}}
&=&
\frac{i}{2}(g_{aXR'}g_b{^{XS'}} - g_{bXR'}g_a{^{XS'}})
.
\end{eqnarray}
Alternatively, after some simplifications,
\begin{eqnarray}
\sigma_{ab}{_R}{^S}
&=&
\frac{i}{2}\varepsilon_{A'B'}(\varepsilon_{AR}\varepsilon{_{B}}{^{S}} + \varepsilon{_{B}}{_{R}}\varepsilon{_{A}}{^{S}})
\label{sigma RS},\\
\sigma_{ab}{_{R'}}{^{S'}}
&=&
\frac{i}{2}\varepsilon_{AB}(\varepsilon_{A'R'}\varepsilon{_{B'}}{^{S'}} + \varepsilon_{B'R'}\varepsilon{_{A'}}{^{S'}})
\label{sigma R'S'}.
\end{eqnarray}
We will encounter several types of spinor fields, transforming {\it non-unitarily\/} (unitary representations will appear later) and denoted as follows
\begin{eqnarray}
\Lambda\phi{_{A_1\dots A_nA'_1\dots A'_m}}(x) &=& \Lambda{_{A_1}}{^{B_1}}\dots \Lambda{_{A'_m}}{^{B'_m}}
\phi{_{B_1\dots B_nB'_1\dots B'_m}}(\Lambda^{-1}x),\nonumber\\
&\pp=&
\textrm{(fields on Minkowski space)},\\
\Lambda\phi{_{A_1\dots A_nA'_1\dots A'_m}}(\bm p) &=& \Lambda{_{A_1}}{^{B_1}}\dots \Lambda{_{A'_m}}{^{B'_m}}\phi{_{B_1\dots B_nB'_1\dots B'_m}}(\bm{\Lambda^{-1}p}), \nonumber\\
&\pp=&\textrm{(fields on mass-$m$ hyperboloid)}.
\end{eqnarray}
$\bm{\Lambda^{-1}p}$ denotes the spacelike part
$\big(({\Lambda^{-1}p})^1,({\Lambda^{-1}p})^2,({\Lambda^{-1}p})^3\big)$.
Recall that $\Lambda p_a=\Lambda{_a}{^b}p_b$, $\Lambda p^a=p^b\Lambda^{-1}{_b}{^a}$, so that $\Lambda^{-1} p^a=p^b\Lambda{_b}{^a}$.
Generators of these representations,
\begin{eqnarray}
J{_{ab}}{_{A_1}}{^{B_1}}\dots {_{A'_m}}{^{B'_m}}
&=&
L_{ab}\varepsilon{_{A_1}}{^{B_1}}\dots \varepsilon{_{A'_m}}{^{B'_m}}
+
\sigma{_{ab}}{_{A_1}}{^{B_1}}\dots {_{A'_m}}{^{B'_m}},
\end{eqnarray}
split into orbital parts $L_{ab}$ and the spin parts
\begin{eqnarray}
\sigma{_{ab}}{_{A_1}}{^{B_1}}\dots {_{A'_m}}{^{B'_m}}
&=&
i
\frac{\partial}{\partial{}y^{ab}}
\Lambda{_{A_1}}{^{B_1}}\dots \Lambda{_{A'_m}}{^{B'_m}}\Big|_{{}y^{ab}=0}.
\end{eqnarray}
The explicit forms of the orbital parts,
\begin{eqnarray}
L_{ab}\phi{_{B_1\dots B_nB'_1\dots B'_m}}(x)
&=&
\sigma{_{ab}}{^r}{^s}
x_r\partial_s
\phi{_{B_1\dots B_nB'_1\dots B'_m}}(x)\\
L_{ab}\phi{_{B_1\dots B_nB'_1\dots B'_m}}(\bm p)
&=&
\sum_{{\bf j}=1}^3\sigma{_{ab}}{^{r}}{^{\bf j}}
p{_{r}}\frac{\partial}{\partial p^{\bf j}}
\phi{_{B_1\dots B_nB'_1\dots B'_m}}({\bm p}),
\end{eqnarray}
in general, will not be very important if we define spins by means of Pauli-Lubanski vectors.
\subsection{Bispinors and 2-spinors}
Simplest unitary representations of inhomogeneous SL(2,C) are most naturally introduced at the level of the Dirac equation, i.e. by means of bispinors. A bispinor is a direct sum of primed and unprimed 2-spinors. We will use the following abstract-index convention:
$
\psi^{\tt A}
=
\left(
\begin{array}{c}
\psi{^{A}}\\
\psi{^{A'}}
\end{array}
\right)
$ is a bispinor, Dirac's gamma matrices are represented by
\begin{eqnarray}
\gamma{_q}{_{\tt A}}{^{\tt B}}
=
\left(
\begin{array}{cc}
\gamma{_q}{_{A}}{^{B}} & \gamma{_q}{_{A}}{^{B'}}\\
\gamma{_q}{_{A'}}{^{B}} & \gamma{_q}{_{A'}}{^{B'}}
\end{array}
\right)
=
\sqrt{2}\left(
\begin{array}{cc}
0 & g{_{qA}}{^{B'}}\\
-g{_q}{^B}{_{A'}} & 0
\end{array}
\right),
\end{eqnarray}
and contractions of the bispinor indices are defined by $\alpha{^{\dots{\tt A}\dots}}\beta{_{\dots{\tt A}\dots}}=\alpha{^{\dots{A}\dots}}\beta{_{\dots{A}\dots}}+
\alpha{^{\dots{A'}\dots}}\beta{_{\dots{A'}\dots}}$. Generators of the bispinor $(\frac{1}{2},0)\oplus (0,\frac{1}{2})$ representation,
\begin{eqnarray}
\Lambda{_{\tt A}}{^{\tt B}}
&=&
\left(
\begin{array}{cc}
\Lambda{_{A}}{^{B}} & \Lambda{_{A}}{^{B'}}\\
\Lambda{_{A'}}{^{B}} & \Lambda{_{A'}}{^{B'}}
\end{array}
\right)
=
\left(
\begin{array}{cc}
\Lambda{_{A}}{^{B}} & 0\\
0 & \Lambda{_{A'}}{^{B'}}
\end{array}
\right),
\end{eqnarray}
satisfy
\begin{eqnarray}
\sigma{_{rs}}{_{\tt A}}{^{\tt B}}
=
\left(
\begin{array}{cc}
\sigma{_{rs}}{_{A}}{^{B}} & \sigma{_{rs}}{_{A}}{^{B'}}\\
\sigma{_{rs}}{_{A'}}{^{B}} & \sigma{_{rs}}{_{A'}}{^{B'}}
\end{array}
\right)
=
\left(
\begin{array}{cc}
\sigma{_{rs}}{_{A}}{^{B}} & 0\\
0 & \sigma{_{rs}}{_{A'}}{^{B'}}
\end{array}
\right).
\end{eqnarray}
Product of two gamma matrices,
\begin{eqnarray}
\gamma{_q}{_{\tt A}}{^{\tt B}}\gamma{_r}{_{\tt B}}{^{\tt C}}&=&
-2
\left(
\begin{array}{cc}
g{_{qA}}{^{B'}}g{_r}{^C}{_{B'}} & 0\\
0 & g{_q}{^B}{_{A'}} g{_{rB}}{^{C'}}
\end{array}
\right)\nonumber\\
&=&
\left(
\begin{array}{cc}
g{_{qr}}\varepsilon{_A}{^C}-2i\sigma{_{qr}} {_A}{^C} & 0\\
0 & g{_{qr}}\varepsilon{_{A'}}{^{C'}}-2i\sigma{_{qr}}
{_{A'}}{^{C'}}
\end{array}
\right)=
g{_{qr}}\varepsilon{_{\tt A}}{^{\tt C}}-2i\sigma{_{qr}} {_{\tt A}}{^{\tt C}},\nonumber
\end{eqnarray}
implies
\begin{eqnarray}
\gamma{_q}{_{\tt A}}{^{\tt B}}\gamma{_r}{_{\tt B}}{^{\tt C}}
+
\gamma{_r}{_{\tt A}}{^{\tt B}}\gamma{_q}{_{\tt B}}{^{\tt C}}
&=&
2g{_{qr}}\varepsilon{_{\tt A}}{^{\tt C}},\\
\gamma{_q}{_{\tt A}}{^{\tt B}}\gamma{_r}{_{\tt B}}{^{\tt C}}
-
\gamma{_r}{_{\tt A}}{^{\tt B}}\gamma{_q}{_{\tt B}}{^{\tt C}}
&=&
-4i\sigma{_{qr}} {_{\tt A}}{^{\tt C}}.
\end{eqnarray}
It is interesting that only at the abstract-index level one can realize that, as opposed to what one usually reads in relativistic quantum mechanics textbooks, $\gamma{_0}{_{\tt A}}{^{\tt B}}$ {\it cannot\/} be identified with the matrix $\beta$ occurring in the standard-notation formulas $\bar\psi\psi=\psi^{\dag}\beta\psi$ and $j_a=\bar\psi\gamma_a\psi=\psi^{\dag}\beta\gamma_a\psi$. Indeed, in the abstract-index notation the Dirac current reads \cite{PR}
\begin{eqnarray}
j_a
&=&
\sqrt{2}g_a{^{AA'}}\bigl(\psi_A\bar \psi_{A'} +
\psi_{A'}\bar \psi_{A}\bigr)
=\sqrt{2}\bigl(\psi_A\bar \psi_{A'} +
\psi_{A'}\bar \psi_{A}\bigr)
\label{spin-cur}\\
&=&
\sqrt{2}(\bar \psi^{A'},\bar \psi^A)
\left(
\begin{array}{cc}
0 & \varepsilon{_{A'}}{^{B'}}\\
-\varepsilon{_A}{^{B}} & 0
\end{array}
\right)
\left(
\begin{array}{cc}
0 & g{_{aB}}{^{C'}}\\
-g{_a}{^C}{_{B'}} & 0
\end{array}
\right)
\left(
\begin{array}{c}
\psi_C\\
\psi_{C'}
\end{array}
\right).
\end{eqnarray}
In particular,
\begin{eqnarray}
j_0&=&
\underbrace{
(\bar \psi^{A'},\bar \psi^A)}_{\bar\psi{^{\tt A'}}}
\underbrace{
\left(
\begin{array}{cc}
0 & \varepsilon{_{A'}}{^{B'}}\\
-\varepsilon{_A}{^{B}} & 0
\end{array}
\right)}_{\beta{_{\tt A'}}{^{\tt B}}}
\underbrace{\sqrt{2}
\left(
\begin{array}{cc}
0 & g{_{0B}}{^{C'}}\\
-g{_0}{^C}{_{B'}} & 0
\end{array}
\right)}_{\gamma_0{_{\tt B}}{^{\tt C}}}
\underbrace{\left(
\begin{array}{c}
\psi_C\\
\psi_{C'}
\end{array}
\right)}_{\psi{_{\tt C}}}\nonumber\\
&=&
\sqrt{2}
(\overline{\psi_{A}},\overline{\psi_{A'}})
\left(
\begin{array}{cc}
g{_0}{^A}{^{A'}} & 0\\
0 & g{_0}{^A}{^{A'}}
\end{array}
\right)
\left(
\begin{array}{c}
\psi_A\\
\psi_{A'}
\end{array}
\right).
\end{eqnarray}
Replacing Infeld-van der Waerden {\it tensors\/} by matrices of Infeld-van der Waerden {\it symbols\/} we find that the {\it matrices\/} corresponding to $\beta$ and $\gamma_0$ are indeed the same. However, a glimpse at the indices shows that these are matrices of maps that act in different linear spaces and in addition transform under SL(2,C) according to inequivalent representations (roughly speaking, $\beta{_{\tt A'}}{^{\tt B}}$ is a scalar, and not a timelike component of a world-vector). This is an example of a relativistic formal subtlety that can be appreciated only at the abstract-index 2-spinor level. In the following sections I will discuss more such subtleties, clarifying certain controversies about relativistic properties of qubits.
\section{Qubits associated with first-quantized Dirac equation}
Although 2-spinors and qubits are vectors from ${\bm C}^2$, not all qubits are 2-spinors, and not all 2-spinors are qubits. The reason is simple: 2-spinors correspond to finite-dimensional non-unitary representations of SL(2,C), but qubits --- by definition --- have to transform unitarily. Qubit representations of inhomogeneous SL(2,C) are infinite dimensional. Physically, the infinity of dimension means that relativistic qubits are {\it spinor fields\/} of some sort. These new types of spinors are directly related to 2-spinors by a kind of duality. The duality will be discussed at the end of this section, but before we do that we have to analyze some important preliminaries.
\subsection{Relativistic generalization of spin}
One expects that relativistic qubits are related to spin.
A conceptual difficulty one immediately encounters in this context is that there is no generally accepted definition of relativistic spin operator. It is not {\it a priori\/} evident that the discussed splitting of generators into `orbital' and `spin' parts corresponds, physically, to orbital angular momentum and spin. A simple illustration of the difficulty is the following. Let us take the Dirac Hamiltonian (units with $c=1$, $\hbar=1$)
\begin{eqnarray}
H &=& \bm\alpha\cdot\bm P+m\beta
\end{eqnarray}
in the non-manifestly-covariant formulation. The generator of rotations reads
\begin{eqnarray}
\bm J=\bm x\times \bm P+\bm s,
\end{eqnarray}
where $\bm s=\frac{1}{2}\left(\begin{array}{cc}\bm \sigma & 0\\0 & \bm \sigma\end{array}\right)$. It seems natural to identify spin with $\bm s$. However, since $[H,\bm s]\neq 0$ the projection of $\bm s$ on a unit vector $\bm a$ does not lead to a well defined quantum number, unless $\bm a$ is an operator parallel to momentum ${\bm P}=-i\bm \nabla$. The projection of spin on momentum $\bm P\cdot\bm s$ is a differential operator of first order, commuting with $H$. Knowledge of its eigenvectors is sufficient for vast majority of applications, but there are exceptions. Formulation of a relativistic generalization of the Bell inequality for electrons requires projections of spin on four different directions, so the knowledge of helicity states is not enough.
A possible solution is suggested by analogous problems with velocity of a free Dirac particle: $d\bm x/dt=\bm v=-i[\bm x,H]=\bm\alpha$, $[H,\bm v]\neq 0$. A way out is to start with the position operator
\begin{eqnarray}
\bm Q &=&\Pi_+ \bm x \Pi_+ +\Pi_- \bm x \Pi_-
\end{eqnarray}
where $\Pi_\pm$ project on states of positive or negative energy. Then, $d\bm Q/dt=\bm V=-i[\bm Q,H]=\bm P/H$, $[H,\bm V]=0$. $\bm Q$ has properties of relativistic center of mass, and $\bm V$ is the center-of-mass velocity. Applying this reasoning to spin, we arrive at
\begin{eqnarray}
\bm S &=&\Pi_+ \bm s \Pi_+ +\Pi_- \bm s \Pi_-=\bm J-\bm Q\times \bm P,\\
\bm J &=& \bm x\times \bm P+\bm s = \bm Q\times \bm P+\bm S.
\end{eqnarray}
Now $[\bm S,H]=0$, and $\bm a\cdot \bm S$ is a natural projection of spin with no restriction on the direction of $\bm a$. Apparently, the first application of the above idea to relativistic qubits (relativistic Einstein-Podolsky-Rosen-Bohm experiment) can be found in the unpublished preprint of mine from 1984 \cite{MC84}. The resulting spin operator depends on momentum in a very complicated way and does not satisfy the rotation Lie algebra typical of angular momentum. If $\bm a=\bm P/|\bm P|$ then $\bm a\cdot \bm S=\bm a\cdot \bm s=\bm a\cdot \bm J$ is the helicity operator.
If $\bm a$ is a unit vector perpendicular to $\bm P$ then the eigenvalues of $\bm a\cdot \bm S$ tend to 0 with $\bm p\to\infty$ (in momentum space), or with $m\to 0$. The latter property can be understood as a combined effect of two classical relativistic phenomena: the Lorentz flattening of the particle, and the M{\o}ller shift of the center of mass \cite{Moller,Fleming1}.
Another, manifestly covariant approach is to start with first-order differential operator known as the Pauli-Lubanski vector
\begin{eqnarray}
W^a{_{\tt X}}{^{\tt Y}} &=& P_b {}^*M^{ab}{_{\tt X}}{^{\tt Y}}=P_b {}^*\sigma^{ab}{_{\tt X}}{^{\tt Y}}
=\frac{1}{2}e^{abcd}P_b\sigma_{cd}{_{\tt X}}{^{\tt Y}}.
\end{eqnarray}
Its projection $W^0=\bm P\cdot \bm s$ on the `time' direction turns out to be the helicity times $|\bm P|$. Moreover, $W^a$ commutes with 4-momentum and $W_aW^a$ is a Casimir operator of the Poincar\'e group. In irreducible representations $W_aW^a=-m^2 j(j+1)$ where $j$ is the dimension of the associated representation of su(2). This is why another popular definition of spin is $w^a=W^a/m$. Its spacelike component $\bm w$ equals $\bm s$ for $\bm p=0$ (again, in momentum space). If $t^a$ is a constant world-vector then eigenvalues of $t^aW_a$ tend to infinity with $\bm p\to \infty$ (this is clear since in momentum space $W^0=\bm p\cdot\bm s$). It is interesting that $\bm S$ and $\bm W$ are proportional to each other, $\bm S=\bm W/H$, and thus possess the same eigenvectors. From the point of view of the Bell inequality, say, they are equivalent (cf. the analysis of this point given in \cite{MC-PRA,SPIE}). The components of $W^a$ do not satisfy the angular momentum Lie algebra.
The simplest way of deriving an explicit form of $W^a{_{\tt X}}{^{\tt Y}}$ is based on the following abstract-index identity:
If
$F_{ab}=-F_{ba}$
then (cf. Eq.~(3.4.21) in \cite{PR})
\begin{eqnarray}
F_{ab}
&=&
\phi_{AB}\varepsilon_{A'B'}+\varepsilon_{AB}\psi_{A'B'},\\
\phi_{AB}
&=&
\frac{1}{2}F{_{AX'}}{_{B}}{^{X'}},\\
\psi_{A'B'}
&=&
\frac{1}{2}F{_{XA'}}{^{X}}{_{B'}},\\
^*F_{ab}
&=&
-i\phi_{AB}\varepsilon_{A'B'}+i\varepsilon_{AB}\psi_{A'B'}.
\end{eqnarray}
Comparison with (\ref{sigma RS}), (\ref{sigma R'S'}) implies
\begin{eqnarray}
^*\sigma_{ab}{_R}{^S}
&=&
\frac{1}{2}\varepsilon_{A'B'}(\varepsilon_{AR}\varepsilon{_{B}}{^{S}} + \varepsilon{_{B}}{_{R}}\varepsilon{_{A}}{^{S}})
\\
&=&
\frac{1}{2}(g_{aRX'}g_b{^{SX'}} - g_{bRX'}g_a{^{SX'}})
\label{*sigma RS},\\
^*\sigma_{ab}{_{R'}}{^{S'}}
&=&
-\frac{1}{2}\varepsilon_{AB}(\varepsilon_{A'R'}\varepsilon{_{B'}}{^{S'}} + \varepsilon_{B'R'}\varepsilon{_{A'}}{^{S'}})\\
&=&
-\frac{1}{2}(g_{aXR'}g_b{^{XS'}} - g_{bXR'}g_a{^{XS'}})
\label{*sigma R'S'}.
\end{eqnarray}
It follows that
\begin{eqnarray}
W_{a}{_R}{^S}
&=&
\frac{1}{2}(g_{aRX'}P{^{SX'}} - P_{RX'}g_a{^{SX'}}),\label{W_a_R^S}\\
W_{a}{_{R'}}{^{S'}}
&=&
-\frac{1}{2}(g_{aXR'}P{^{XS'}} - P_{XR'}g_a{^{XS'}}),\\
W_{a}{_{\tt R}}{^{\tt S}}
&=&
\left(
\begin{array}{cc}
W_{a}{_R}{^S} & 0\\
0 & W_{a}{_{R'}}{^{S'}}
\end{array}
\right).
\end{eqnarray}
Projections of the Pauli-Lubanski vector on various directions will be discussed in a separate section.
Yet another possibility is to begin with the Newton-Wigner position operator $\bm Q_{NW}$, which differs from $\bm Q$ by a term commuting with $H$ \cite{NW}. The components of $\bm Q_{NW}$ commute, as opposed to those of $\bm Q$. One obtains $\bm S_{NW}=\bm J-\bm Q_{NW}\times \bm P$ whose components satisfy the angular momentum Lie algebra. It can be shown \cite{BLT} that this is the only axial vector operator linear in the Pauli-Lubanski vector, satisfying su(2), and transforming under rotations as a 3-dimensional vector. The above properties look natural. Spin based on $\bm S_{NW}$ has been extensively studied in the context of relativistic qubits and Bell inequalities by Rembieli\'nski and his group \cite{CRW1,CRW2,CR,RS}.
One can argue, however, that in classical mechanics of spinning particles one does {\it not\/} arrive at {\it commuting\/} center-of-mass coordinates. This happens whenever one imposes mass or spin constraints, which is the case here, and replaces Poisson brackets by Dirac brackets according to the rules of constrained dynamics. The position variable that commutes with respect to the unconstrained bracket becomes noncommutative with respect to the constrained one \cite{Mukunda,Zakrzewski}, and this is the bracket that should be employed if spin is 1/2 or mass is $m_{\rm electron}$. What is interesting, the resulting algebra is analogous to this of $\bm Q$ and not to $\bm Q_{NW}$. This is not so surprising if one realizes that position generates shifts of momentum. Mass hyperboloid in not a flat manifold so translations do not commute. If we relax the mass constraint, the momentum space becomes the Minkowski space, which is flat, so positions should commute, as it indeed happens in off-shell formalisms in quantum mechanics.
Classification of all spin operators linear in momentum can be found in \cite{Bagrov,IBB}, and different spins are used in practical applications to exact solutions of the Dirac equation coupled to electromagnetic fields.
From a formal point of view all the possible spin-like observables are possible candidates for quantum yes-no observables. Which of them are actually measured in experiments depends on experimental procedures.
\subsection{Beyond helicity: Two-spinor approach to polarization operators based on the Pauli-Lubanski vector}
Generators of four-translations are defined by
\begin{eqnarray}
e^{iy^a P_a}\psi{_{\tt A}}(x)=\psi{_{\tt A}}(x-y)=e^{-y^a \partial_a}\psi{_{\tt A}}(x),
\end{eqnarray}
so $P_a=i\partial_a$. The Pauli-Lubanski vector thus satisfies
\begin{eqnarray}
W_{a}{_{\tt R}}{^{\tt S}}e^{\mp i p\cdot x}
&=&
\underbrace{\pm\frac{1}{2}
\left(
\begin{array}{cc}
g_{aRX'}p{^{SX'}} - p_{RX'}g_a{^{SX'}} & 0\\
0 & -g_{aXR'}p{^{XS'}} + p_{XR'}g_a{^{XS'}}
\end{array}
\right)}_{W_{a}{_{\tt R}}{^{\tt S}}(\pm p)=\pm W_{a}{_{\tt R}}{^{\tt S}}(\bm p)}e^{\mp i p\cdot x}.\nonumber\\
\end{eqnarray}
In all the formulas we assume that $p^a$ is future-pointing, i.e. $p_0>0$.
Now, let us consider any symmetric spinor $\phi_{AB}$. Using $\phi{_A}{^A}=\phi{_0}{^0}+\phi{_1}{^1}=0$ we find
\begin{eqnarray}
\left|
\begin{array}{cc}
\phi{_0}{^0}-\lambda & \phi{_0}{^1}\\
\phi{_1}{^0} & \phi{_1}{^1}-\lambda
\end{array}
\right|
&=&
\lambda^2
+
\frac{1}{2}\phi{_A}{_B}\phi{^A}{^B},
\end{eqnarray}
so that the eigenvalues $\lambda^{(\pm)}$ of $\phi{_A}{^B}$ are given by
\begin{eqnarray}
\lambda^{(\pm)}
&=&
\pm\sqrt{-\phi{_X}{_Y}\phi{^X}{^Y}/2}.
\end{eqnarray}
Applying this formula to
$W{_{\tt R}}{^{\tt S}}(t,\bm p)=t^a(\bm p)W_{a}{_{\tt R}}{^{\tt S}}(\bm p)$, where $t^a(\bm p)$ is an arbitrary field of (real) world-vectors, we conclude that
the eigenvalues of such a general projection of the Pauli-Lubanski vector are given by
\begin{eqnarray}
\lambda^{(\pm)}(t,\bm p)
&=&
\pm \sqrt{-W{_{X}}{_{Y}}(t,\bm p)W{^{X}}{^{Y}}(t,\bm p)/2}
=
\pm \sqrt{-W{_{X'}}{_{Y'}}(t,\bm p)W{^{X'}}{^{Y'}}(t,\bm p)/2}
\nonumber\\
&=&
\pm \frac{1}{2}\sqrt{(t\cdot p)^2-t^2p^2}.\label{W(p,t)}
\end{eqnarray}
Here $t\cdot p=p_at^a(\bm p)$, $t^2=t_a(\bm p)t^a(\bm p)$, $p^2=p_ap^a=m^2$. Helicity, corresponding to
$t^{\bm a}(\bm p)=(1/|\bm p|,0,0,0)$, has eigenvalues
\begin{eqnarray}
\lambda^{(\pm)}(t,\bm p)
&=&
\pm \frac{1}{2}\sqrt{p_0^2/\bm p^2-m^2/\bm p^2}=\pm \frac{1}{2}.
\end{eqnarray}
If $t^a(\bm p)=p^a$ then $\lambda^{(\pm)}(t,\bm p)=0$, since $P^aW_a=0$ in any representation. The latter incidentally shows that projections of the Pauli-Lubanski vector possess a gauge freedom: $t^aW_a=(t^a+\theta P^a)W_a$, for any $\theta$. The same concerns the eigenvalues: $\lambda^{(\pm)}(t,\bm p)=\lambda^{(\pm)}(t+\theta p,\bm p)$.
For a particle at rest, $p^{\bf a}=(m,\bm 0)$, the Pauli-Lubanski vector reduces to non-relativistic spin multiplied by mass:
$W^{\bf a}=m(0,\bm s)$. The non-relativistic observable $\bm a\cdot\bm s=\frac{1}{m}\bm a\cdot\bm W$ can be covariantly written as $t^aW_a$, where
\begin{eqnarray}
t^{\bm a}=(t^0,\bm a/m)=(0,\bm a/m)+(t^0,\bm 0),\label{t gauge}
\end{eqnarray}
the last term being proportional to $p^{\bf a}$. This is the simplest explanation of the gauge freedom inherently present in the Pauli-Lubanski vector.
The fact that it is only the spacelike part of $t^a$ that counts in the definition of $t^aW_a$ does not mean that $t^a$ itself should be spacelike. A spacelike $t^a$ when shifted by a timelike $\theta p^a$ can become timelike or null (or remain spacelike).
Of particular interest turns out to be the null case: $t^2=0$, $t^ap_a=1$. Indeed, first of all the corresponding eigenvalues are identical to those of the helicity, $\lambda^{(\pm)}(t,\bm p)=\pm \frac{1}{2}$. Let $t^a=\omega^a(\bm p)=\omega^A(\bm p)\bar\omega^{A'}(\bm p)$, and define
\begin{eqnarray}
\pi{^{A}}(\bm p)
&=&
p{^{A}}{^{B'}}\bar\omega{_{B'}}(\bm p).\label{pi-omega}
\end{eqnarray}
The trace-reversal formula (Eq.~(3.4.13) in \cite{PR})
\begin{eqnarray}
p^{AB'}p^{BA'}
&=&
p^{AA'}p^{BB'}-\frac{m^2}{2}\varepsilon^{AB}\varepsilon^{A'B'}
=
p^ap^b-\frac{m^2}{2}g^{ab},
\end{eqnarray}
implies
\begin{eqnarray}
\pi^A(\bm p)\bar\pi^{A'}(\bm p)
&=&
p{^{A}}{^{B'}}p{^{B}}{^{A'}}\omega{_{B}}(\bm p)\bar\omega{_{B'}}(\bm p)
\nonumber\\
&=&
p{^{a}}
\underbrace{
p{^{b}}\omega{_{B}}(\bm p)\bar\omega{_{B'}}(\bm p)}_{p^bt_b=1}-\frac{m^2}{2}\omega{^{A}}(\bm p)\bar\omega{^{A'}}(\bm p),
\end{eqnarray}
and we arrive at the following important decomposition of $p^a$ into a combination of two null directions:
\begin{eqnarray}
p{^{a}}
&=&
\pi^A(\bm p)\bar\pi^{A'}(\bm p)
+
\frac{m^2}{2}\omega{^{A}}(\bm p)\bar\omega{^{A'}}(\bm p).\label{p}
\end{eqnarray}
Formula (\ref{p}) implies (\ref{pi-omega}). Indeed,
\begin{eqnarray}
p{^{AB'}}\bar\omega{_{B'}}(\bm p)
&=&
\Big(\pi^A(\bm p)\bar\pi^{B'}(\bm p)
+
\frac{m^2}{2}\omega{^{A}}(\bm p)\bar\omega{^{B'}}(\bm p)
\Big)\bar\omega{_{B'}}(\bm p)=\pi^A(\bm p).\nonumber
\end{eqnarray}
It follows that $\omega_A(\bm p)$ supplemented by (\ref{p}) determines its spin-partner $\pi^A(\bm p)$ uniquely. Moreover, since
\begin{eqnarray}
p_{AC'}\pi{^{A}}(\bm p)
&=&
p_{AC'}p{^{A}}{^{B'}}\bar\omega{_{B'}}(\bm p)=\frac{m^2}{2}\bar\omega{_{C'}}(\bm p),
\end{eqnarray}
the knowledge of $\pi{^{A}}(\bm p)$ and (\ref{p}) determines $\omega{_{A}}(\bm p)$ uniquely, unless $m=0$.
As opposed to the special case $m=0$, where $\pi^A(\bm p)$ is defined by its flagpole $p^a$ up to a phase (U(1) internal symmetry), the internal symmetry group is here U(2). The pair $(\pi^A(\bm p),\frac{m}{\sqrt{2}}\omega{^{A}}(\bm p))$ can be replaced by
$(\tilde\pi^A(\bm p),\frac{m}{\sqrt{2}}\tilde\omega{^{A}}(\bm p))$, where
\begin{eqnarray}
\left(
\begin{array}{c}
\tilde\pi^A(\bm p)\\
\frac{m}{\sqrt{2}}\tilde\omega{^{A}}(\bm p)
\end{array}
\right)
&=&
\left(
\begin{array}{cc}
\alpha(\bm p) & \beta(\bm p)\\
\gamma(\bm p) & \delta(\bm p)
\end{array}
\right)
\left(
\begin{array}{c}
\pi^A(\bm p)\\
\frac{m}{\sqrt{2}}\omega{^{A}}(\bm p)
\end{array}
\right),
\end{eqnarray}
and the matrix is unitary. Hughston showed \cite{H} that the existence of such splitting-of-$p^a$ ambiguities may be the actual geometric reason for internal symmetries occurring in gauge theories. From our point of view it is more important that a similar mechanism is responsible for the structure of unitary representations of the Poincar\'e group \cite{MC-BW}.
\subsection{The simplest case: Projection on the null direction $t^a=\omega^a(\bm p)$}
Since $\omega^ap_a=\omega^A\bar \omega^{A'}\pi_A\bar \pi_{A'}=1$ independently of the value of $m$, we can assume --- without loss of generality --- spin-frame normalization $\omega_A\pi^A=1$.
Projections on the null direction,
\begin{eqnarray}
W(\omega,\bm p){_{A}}{^{B}}&=&
\frac{1}{2}\Bigl(
\pi_{A}\omega^{B}+ \omega_{A}\pi^{B}\Bigr),\label{W W omega}\\
W(\omega,\bm p){_{A'}}{^{B'}}&=&-
\frac{1}{2}\Bigl(
\bar \pi_{A'}\bar \omega^{B'}+
\bar \omega_{A'}\bar \pi^{B'}\Bigr),
\end{eqnarray}
look the same for both $m=0$ and $m\neq 0$. Also spin eigenvalue problems become in the null formalism particularly simple:
\begin{eqnarray}
W(\omega,\bm p){_{A}}{^{B}}\omega_{B}&=&
\frac{1}{2}\omega_{A},\label{pl000}\\
W(\omega,\bm p){_{A'}}{^{B'}}\bar \pi_{B'}&=&
\frac{1}{2}\bar \pi_{A'},\label{pl00'}\\
W(\omega,\bm p){_{A}}{^{B}}\pi_{B}&=&
-\frac{1}{2}\pi_{A},\label{pl00}\\
W(\omega,\bm p){_{A'}}{^{B'}}\bar \omega_{B'}&=&
-\frac{1}{2}\bar \omega_{A'}.\label{pl000'}
\end{eqnarray}
Spectral projectors associated with $\lambda^{(\pm)}(\omega,\bm p)=\pm \frac{1}{2}$,
\begin{eqnarray}
\Pi^{(+)}(\omega,\bm p){_{A}}{^{B}} &=& \omega{_{A}}\pi{^{B}},
\\
\Pi^{(-)}(\omega,\bm p){_{A}}{^{B}} &=& -\pi_{A}\omega^{B},
\\
\Pi^{(+)}(\omega,\bm p){_{A'}}{^{B'}}
&=&
-\bar\pi{_{A'}}\bar\omega{^{B'}},
\\
\Pi^{(-)}(\omega,\bm p){_{A'}}{^{B'}}
&=&
\bar\omega{_{A'}}\bar\pi{^{B'}},
\end{eqnarray}
trivially satisfy 2-spinor resolutions of unity (\ref{ve_A^B}),
\begin{eqnarray}
\Pi^{(+)}(\omega,\bm p){_{A}}{^{B}}+\Pi^{(-)}(\omega,\bm p){_{A}}{^{B}}
&=&
\omega{_{A}}\pi{^{B}}-\pi_{A}\omega^{B}=\varepsilon{_{A}}{^{B}},\\
\Pi^{(-)}(\omega,\bm p){_{A'}}{^{B'}}
+
\Pi^{(+)}(\omega,\bm p){_{A'}}{^{B'}}
&=&
\bar\omega{_{A'}}\bar\pi{^{B'}}-\bar\pi{_{A'}}\bar\omega{^{B'}}=\varepsilon{_{A'}}{^{B'}}.
\end{eqnarray}
At the bispinor level the projectors read
\begin{eqnarray}
\Pi^{(+)}(\omega,\bm p){_{\tt A}}{^{\tt B}}
&=&
\left(
\begin{array}{cc}
\Pi^{(+)}(\omega,\bm p){_{A}}{^{B}} & 0\\
0 & \Pi^{(+)}(\omega,\bm p){_{A'}}{^{B'}}
\end{array}
\right)
=
\left(
\begin{array}{cc}
\omega{_{A}}\pi{^{B}} & 0\\
0 & -\bar\pi{_{A'}}\bar\omega{^{B'}}
\end{array}
\right),\label{Pi^+}\\
\Pi^{(-)}(\omega,\bm p){_{\tt A}}{^{\tt B}}
&=&
\left(
\begin{array}{cc}
\Pi^{(-)}(\omega,\bm p){_{A}}{^{B}} & 0\\
0 & \Pi^{(-)}(\omega,\bm p){_{A'}}{^{B'}}
\end{array}
\right)
=
\left(
\begin{array}{cc}
-\pi_{A}\omega^{B} & 0\\
0 & \bar\omega{_{A'}}\bar\pi{^{B'}}
\end{array}
\right).\label{Pi^-}
\end{eqnarray}
\subsection{Wave functions associated with $W(\omega,\bm p){_{\tt A}}{^{\tt B}}$}
In order to introduce wave functions corresponding to $W(\omega,\bm p){_{\tt A}}{^{\tt B}}$ we first have to expand solutions of Dirac equation in terms of appropriate eigen-bispinors. A systematic procedure of deriving the bispinors can be based on `spin-energy' projectors.
The free Dirac equation
\begin{eqnarray}
D{_{\tt A}}{^{\tt B}}\psi{_{\tt B}}=0, \quad
D{_{\tt A}}{^{\tt B}}=i\partial{^r}\gamma{_r}{_{\tt A}}{^{\tt B}}-m\varepsilon {_{\tt A}}{^{\tt B}}\label{Dirac}
\end{eqnarray}
is implicitly an orthogonality condition for `sign-of-energy' projectors $\Pi_\pm{_{\tt A}}{^{\tt B}}(\bm p)$ constructed as follows.
One begins with
\begin{eqnarray}
D{_{\tt A}}{^{\tt B}}e^{\mp i p\cdot x}
&=&
D_\pm{_{\tt A}}{^{\tt B}}(\bm p)e^{\mp i p\cdot x},
\end{eqnarray}
where
\begin{eqnarray}
D_\pm{_{\tt A}}{^{\tt B}}(\bm p)
&=&
\pm p{^r}\gamma{_r}{_{\tt A}}{^{\tt B}}-m\varepsilon {_{\tt A}}{^{\tt B}}
=
\left(
\begin{array}{ccc}
-m\varepsilon {_{A}}{^{B}} &, &\pm\sqrt{2} p{_{A}}{^{B'}}\\
\mp\sqrt{2}p{^B}{_{A'}} &, & -m\varepsilon {_{A'}}{^{B'}}
\end{array}
\right).
\end{eqnarray}
The orthogonality conditions
\begin{eqnarray}
D_\pm{_{\tt A}}{^{\tt B}}(\bm p)
D_\mp{_{\tt B}}{^{\tt C}}(\bm p)
&=&
0,\\
D_\pm{_{\tt A}}{^{\tt B}}(\bm p)
D_\pm{_{\tt B}}{^{\tt C}}(\bm p)
&=&
-2m D_\pm{_{\tt A}}{^{\tt C}}(\bm p),
\end{eqnarray}
imply that the possible eigenvalues of $D_\pm{_{\tt A}}{^{\tt B}}(\bm p)$ are 0 and $-2m$.
Accordingly, the projectors we are interested in read
\begin{eqnarray}
\Pi_\pm{_{\tt A}}{^{\tt B}}(\bm p)=D_\pm{_{\tt A}}{^{\tt B}}(\bm p)/(-2m)
=
\frac{1}{2}
\left(
\begin{array}{cc}
\varepsilon {_{A}}{^{B}} &\mp\frac{\sqrt{2}}{m} p{_{A}}{^{B'}}\\
\pm\frac{\sqrt{2}}{m}p{^B}{_{A'}} & \varepsilon {_{A'}}{^{B'}}
\end{array}
\right).\label{Pi pm}
\end{eqnarray}
Multiplying (\ref{Pi pm}) by (\ref{Pi^+}) or (\ref{Pi^-}) we obtain `spin-energy' projectors,
\begin{eqnarray}
\Pi_\pm^{(+)}{_{\tt A}}{^{\tt C}}(\omega,\bm p)
&=&
\Pi_\pm{_{\tt A}}{^{\tt B}}(\bm p)
\Pi^{(+)}{_{\tt B}}{^{\tt C}}(\omega,\bm p)
=
\frac{1}{2}
\left(
\begin{array}{cc}
\omega{_{A}}\pi{^{C}} &\mp\frac{m}{\sqrt{2}}\omega{_{A}}\bar\omega{^{C'}}\\
\pm\frac{\sqrt{2}}{m}\bar\pi{_{A'}}\pi{^{C}} & -\bar\pi{_{A'}}\bar\omega{^{C'}}
\end{array}
\right),\label{Pi pm +}\\
\Pi_\pm^{(-)}{_{\tt A}}{^{\tt C}}(\omega,\bm p)
&=&
\Pi_\pm{_{\tt A}}{^{\tt B}}(\bm p)
\Pi^{(-)}{_{\tt B}}{^{\tt C}}(\omega,\bm p)
=
\frac{1}{2}
\left(
\begin{array}{cc}
-\pi{_{A}}\omega^{C} &\mp\frac{\sqrt{2}}{m} \pi{_{A}}\bar\pi{^{C'}}\\
\pm\frac{m}{\sqrt{2}}\bar\omega{_{A'}}\omega^{C} & \bar\omega{_{A'}}\bar\pi{^{C'}}
\end{array}
\right).\label{Pi pm -}
\end{eqnarray}
The completeness relation is
\begin{eqnarray}
\sum_{s,s'=\pm}\Pi_s^{(s')}{_{\tt A}}{^{\tt B}}(\omega,\bm p)=\varepsilon{_{\tt A}}{^{\tt B}}
=\left(
\begin{array}{cc}
\varepsilon {_{A}}{^{B}} & 0\\
0 & \varepsilon {_{A'}}{^{B'}}
\end{array}
\right).
\end{eqnarray}
It is interesting and important for later applications that (\ref{Pi pm +}) and (\ref{Pi pm -}) allow us to define basis bispinors that can be used also in the massless case. The trick is the following. Let
$
\phi{_{\tt A}}(\bm p)
=
2\left(
\begin{array}{c}
\pi{_{A}}(\bm p)\\
\bar\pi{_{A'}}(\bm p)
\end{array}
\right).
$
The eigen-bispinors
\begin{eqnarray}
\phi_\pm^{(+)}{_{\tt A}}(\omega,\bm p)
&=&
\Pi_\pm^{(+)}{_{\tt A}}{^{\tt B}}(\omega,\bm p)
\phi{_{\tt B}}(\bm p)
=
\left(
\begin{array}{c}
\pm\frac{m}{\sqrt{2}}\omega{_{A}}(\bm p)\\
\bar\pi{_{A'}}(\bm p)
\end{array}
\right),\\
\phi_\pm^{(-)}{_{\tt A}}(\omega,\bm p)
&=&
\Pi_\pm^{(-)}{_{\tt A}}{^{\tt B}}(\omega,\bm p)
\phi{_{\tt B}}(\bm p)
=
\left(
\begin{array}{c}
\pi{_{A}}(\bm p) \\
\mp\frac{m}{\sqrt{2}}\bar\omega{_{A'}}(\bm p)
\end{array}
\right),
\end{eqnarray}
are well defined also for $m=0$, and satisfy
\begin{eqnarray}
W(\omega,\bm p){_{\tt A}}{^{\tt B}}\phi_s^{(\pm)}{_{\tt B}}(\omega,\bm p)
&=&
\pm \frac{1}{2}\phi_s^{(\pm)}{_{\tt A}}(\omega,\bm p),\\
D{_{\tt A}}{^{\tt B}}
\phi_\pm^{(s')}{_{\tt B}}(\omega,\bm p)
e^{\mp ip\cdot x}
&=&
-2m\,
\phi_\pm^{(s')}{_{\tt A}}(\omega,\bm p)
e^{\mp ip\cdot x},\\
D{_{\tt A}}{^{\tt B}}
\phi_\mp^{(s')}{_{\tt B}}(\omega,\bm p)
e^{\mp ip\cdot x}
&=&
0.
\end{eqnarray}
A general solution of free Dirac equation can be finally written as
\begin{eqnarray}
\psi_{\tt A}(x)
&=&
\int dp \Big(\psi_{-\tt A}(\bm p)e^{-i p\cdot x}
+
\psi_{+\tt A}(\bm p)e^{+i p\cdot x}\Big)\nonumber\\
&=&
\sum_{{s}=\pm}\int dp \Big(\phi_-^{({s})}{_{\tt A}}(\omega,\bm p)f(s,\bm p)e^{-i p\cdot x}
+
\phi_+^{({s})}{_{\tt A}}(\omega,\bm p)\overline{g(-s,\bm p)}e^{+i p\cdot x}\Big).
\end{eqnarray}
Here $dp=d^3p/[(2\pi)^3 2\sqrt{\bm p^2+m^2}]$ is the invariant measure on mass-$m$ hyperboloid, and complex conjugation and the minus sign in $\overline{g(-s,\bm p)}$ are convenient for later applications. By definition,
\begin{eqnarray}
\psi_{+\tt A}(\bm p)
&=&
\left(
\begin{array}{c}
\frac{m}{\sqrt{2}}\omega{_{A}}(\bm p)\\
\bar\pi{_{A'}}(\bm p)
\end{array}
\right)
f(+,\bm p)
+
\left(
\begin{array}{c}
\pi{_{A}}(\bm p) \\
-\frac{m}{\sqrt{2}}\bar\omega{_{A'}}(\bm p)
\end{array}
\right)
f(-,\bm p),\\
\psi_{-\tt A}(\bm p)
&=&
\left(
\begin{array}{c}
-\frac{m}{\sqrt{2}}\omega{_{A}}(\bm p)\\
\bar\pi{_{A'}}(\bm p)
\end{array}
\right)
\overline{g(-,\bm p)}
+
\left(
\begin{array}{c}
\pi{_{A}}(\bm p) \\
\frac{m}{\sqrt{2}}\bar\omega{_{A'}}(\bm p)
\end{array}
\right)
\overline{g(+,\bm p)}.
\end{eqnarray}
Wave functions $f(s,\bm p)$ and $g(s,\bm p)$ are simultaneously scalar fields and spinors of a new type, as we will see shortly. They can be extracted from $\psi_{\pm\tt A}(\bm p)$ in a simple way:
\begin{eqnarray}
f(+,\bm p) &=& \bar\omega{_{A'}}(\bm p)\psi{_{+}}{^{A'}}(\bm p),\\
f(-,\bm p) &=& \omega{_{A}}(\bm p)\psi{_{+}}{^{A}}(\bm p),\\
\overline{g(-,\bm p)} &=& \bar\omega{_{A'}}(\bm p)\psi{_{-}}{^{A'}}(\bm p),\\
\overline{g(+,\bm p)} &=& \omega{_{A}}(\bm p)\psi{_{-}}{^{A}}(\bm p).
\end{eqnarray}
For future reference let us explicitly write the two components of the $m=0$ case:
\begin{eqnarray}
\psi_{A}(x)
&=&
\int dk\,
\pi{_{A}}(\bm k)\Big(f(-,\bm k)e^{-ik\cdot x}+\overline{g(+,\bm k)}e^{ik\cdot x}\Big),\label{psi m=01}\\
\psi_{A'}(x)
&=&
\int dk\,
\bar\pi{_{A'}}(\bm k)\Big(f(+,\bm k)e^{-ik\cdot x}+\overline{g(-,\bm k)}e^{ik\cdot x}\Big).
\label{psi m=02}
\end{eqnarray}
Massless 4-momenta are denoted here by $k^a$ and $dk=d^3k/[(2\pi)^3 2 |\bm k|]$ is the invariant measure on the light cone.
This is the right moment to explain the issue of unitary representations of the (covering space of the) Poincar\'e group. I will not follow the usual Wigner-Mackey procedure \cite{Wigner,Mackey} of induction from little groups. Instead, I will show by means of 2-spinor techniques that unitary representations of the group are encoded in shapes of solutions of relativistic wave equations. The latter statement is in itself not very original since links between unitary representations and relativistic wave equations were discussed already in \cite{BW} (for a modern analysis cf. \cite{BR}). Nevertheless, the 2-spinor tricks I will use do not seem to be widely known and, apparently, were introduced for the first time in \cite{MC-BW}.
\subsection{Duality between active and passive SL(2,C) transformations: $\omega$-spinors}
Let us now consider the transformed solution
\begin{eqnarray}
\Lambda\psi_{\tt A}(x)
&=&
\Lambda{_{\tt A}}{^{\tt B}}\psi_{\tt B}(\Lambda^{-1}x),\nonumber\\
&=&
\sum_{{s}}\int dp \,\Lambda{_{\tt A}}{^{\tt B}}\Big(\phi_-^{({s})}{_{\tt B}}(\omega,\bm p)f(s,\bm p)e^{-i p\cdot \Lambda^{-1}x}
+
\phi_+^{({s})}{_{\tt B}}(\omega,\bm p)\overline{g(-s,\bm p)}e^{+i p\cdot \Lambda^{-1}x}\Big)\nonumber
\end{eqnarray}
Employing $p\cdot \Lambda^{-1}x=\Lambda p\cdot x$, changing variables in integrals, keeping in mind that $dp=d(\Lambda p)$, and finally expressing the transformed solution again in the basis $\phi_\pm^{({\pm})}{_{\tt A}}(\omega,\bm p)$, we get
\begin{eqnarray}
\Lambda\psi_{\tt A}(x)
&=&
\sum_{{s}}\int dp \Big(\Lambda{_{\tt A}}{^{\tt B}}\phi_-^{({s})}{_{\tt B}}(\omega,\bm{\Lambda^{-1}p})f(s,\bm{\Lambda^{-1}p})e^{-i p\cdot x}
\nonumber\\
&\pp=&
\pp{\sum_{{s}}\int dp \Big(}
+
\Lambda{_{\tt A}}{^{\tt B}}\phi_+^{({s})}{_{\tt B}}(\omega,\bm{\Lambda^{-1}p})\overline{g(-s,\bm{\Lambda^{-1}p})}e^{+i p\cdot x}\Big)\nonumber\\
&=&
\sum_{{s}}\int dp \Big(\phi_-^{({s})}{_{\tt A}}(\omega,\bm p)\Lambda f(s,\bm{p})e^{-i p\cdot x}
+
\phi_+^{({s})}{_{\tt A}}(\omega,\bm p)\overline{\Lambda g(-s,\bm{p})}e^{+i p\cdot x}\Big)\nonumber\\
\end{eqnarray}
Comparing appropriate terms we find
\begin{eqnarray}
&{}&
\left(
\begin{array}{c}
\frac{m}{\sqrt{2}}\Lambda \omega{^{A}}(\bm{p})\\
\overline{\Lambda\pi}{^{A'}}(\bm{p})
\end{array}
\right)
f(+,\bm{\Lambda^{-1}p})
+
\left(
\begin{array}{c}
\Lambda \pi{^{A}}(\bm{p}) \\
-\frac{m}{\sqrt{2}}\overline{\Lambda\omega}{^{A'}}(\bm{p})
\end{array}
\right)
f(-,\bm{\Lambda^{-1}p})
\nonumber\\
&{}&\pp=
=
\left(
\begin{array}{c}
\frac{m}{\sqrt{2}}\omega{^{A}}(\bm p)\\
\bar\pi{^{A'}}(\bm p)
\end{array}
\right)
\Lambda f(+,\bm p)
+
\left(
\begin{array}{c}
\pi{^{A}}(\bm p) \\
-\frac{m}{\sqrt{2}}\bar\omega{^{A'}}(\bm p)
\end{array}
\right)
\Lambda f(-,\bm p),
\end{eqnarray}
where $\Lambda\pi{_{A}}(\bm p)=\Lambda{_{A}}{^{B}}\pi{_{B}}(\bm{\Lambda^{-1}p})$, $\Lambda\omega{_{A}}(\bm p)=\Lambda{_{A}}{^{B}}\omega{_{B}}(\bm{\Lambda^{-1}p})$, and analogously with the primed spin-frames and complex-conjugated anti-particle wave functions
$\overline{\Lambda g(-s,\bm{p})}$. Denoting
\begin{eqnarray}
\left(
\begin{array}{c}
\psi_{-\it 0}(\bm p)\\
\psi_{-\it 1}(\bm p)\\
\psi_{+\it 0}(\bm p)\\
\psi_{+\it 1}(\bm p)
\end{array}
\right)
&=&
\left(
\begin{array}{c}
f(+,\bm p)\\
f(-,\bm p)\\
g(+,\bm p)\\
g(-,\bm p)\\
\end{array}
\right),\label{t-bispinor}
\end{eqnarray}
we obtain
\begin{eqnarray}
\Lambda \psi_{\pm\cal A}(\bm p) &=& U(\Lambda,\bm p){_{\cal A}}{^{\cal B}}\psi_{\pm\cal B}(\bm{\Lambda^{-1}p}),\label{unitary0}
\end{eqnarray}
\begin{eqnarray}
\left(
\begin{array}{c}
\Lambda \psi_{\pm\it 0}(\bm p)\\
\Lambda \psi_{\pm\it 1}(\bm p)
\end{array}
\right)
&=&
\underbrace{
\left(
\begin{array}{cc}
{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\pi}{^{A'}}(\bm{p}) & -\frac{m}{\sqrt{2}}{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\omega}{^{A'}}(\bm{p})
\\
\frac{m}{\sqrt{2}}\omega{_{A}}(\bm{p})\Lambda \omega{^{A}}(\bm{p}) &
{\omega}{_{A}}(\bm{p}){\Lambda\pi}{^{A}}(\bm{p})
\end{array}
\right)}_{U(\Lambda,\bm p){_{\cal A}}{^{\cal B}}}
\left(
\begin{array}{c}
\psi_{\pm\it 0}(\bm{\Lambda^{-1}p})\\
\psi_{\pm\it 1}(\bm{\Lambda^{-1}p})
\end{array}
\right).\nonumber\\
\label{unitary1}
\end{eqnarray}
Notice that new a type of binary indices has been introduced: ${\cal A},{\cal B}={\it 0}, {\it 1}$. They correspond to local SU(2) spinors. (Local since spinor fields taken at different points of the mass-$m$ hyperboloid transform by different SU(2) transformations.)
It is obvious from the construction that the latter {\it passive\/} \cite{PR} transformation of wave functions is equivalent
to the bispinor-field {\it acive\/} transformation $\Lambda\psi_{\tt A}(x)
=
\Lambda{_{\tt A}}{^{\tt B}}\psi_{\tt B}(\Lambda^{-1}x)$ of {\it solutions of the Dirac equation\/}. One can explicitly check \cite{MC-BW}
that the matrix in (\ref{unitary1}) is unitary and has unit determinant, and that the map $f(s,\bm p)\to \Lambda f(s,\bm p)$ is a representation of SL(2,C). The choice of $f(s,\bm p)$ is completely arbitrary and unrelated to the Dirac equation itself, but the equation is encoded in the form of the matrix occurring in (\ref{unitary1}) ({\it via\/} its implicit dependence on `sign-of-energy' projectors). Translations $x^a\to x^a+y^a$ are represented at the level of wave functions by $f(s,\bm p)\to e^{ip\cdot y}f(s,\bm p)$. The map $f(s,\bm p)\to e^{ip\cdot y}\Lambda f(s,\bm p)$ is thus nothing else but the {\it unitary\/} spin-1/2, mass-$m$ {\it representation\/} of inhomogeneous SL(2,C). The massless limit $m=0$ is trivially obtained from (\ref{unitary1}) and shows that the representation splits for $m=0$ into direct sum of two representations.
The above properties are well known in the context of induced representations of inhomogeneous SL(2,C) \cite{Wigner}: The case $m>0$ corresponds to the little group SU(2) of $p^{\bm a}=(m,\bm 0)$; for $m=0$ the little group is E(2) and its irreducible discrete-spin representations are 1-dimensional. The formulation I have presented in this section depends on two crucial technical elements, the decomposition (\ref{p}) and the choice of $t^a=\omega^a$, but does not use the idea of induction from little groups.
Local SU(2) spinors associated with projections of $W^a$ on $t^a$ are termed the $t$-spinors. A link between general $t$-spinors and the $\omega$-spinors is given by an SU(2) transformation whose explicit form can be found in \cite{MC-BW}. For a non-null $t^a$ the transformation is much more cumbersome than the case of null $\omega^a$, but choosing $t^{\bm a}=(1/|\bm p|,\bm 0)$ we reconstruct the standard helicity formalism \cite{Weinberg}.
The direct sum (\ref{t-bispinor}) of particle and anti-particle representations can be called an $\omega$-bispinor.
\subsection{Pauli-Lubanski vector associated with (\ref{unitary1})}
It pays to compute explicitly the Pauli-Lubanski vector corresponding to (\ref{unitary1}),
\begin{eqnarray}
W_a(\bm p){_{\cal A}}{^{\cal B}}
&=&
p^b {^*}i\frac{\partial}{\partial{}y^{ab}}
\left(
\begin{array}{cc}
{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\pi}{^{A'}}(\bm{p}) & -\frac{m}{\sqrt{2}}{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\omega}{^{A'}}(\bm{p})
\\
\frac{m}{\sqrt{2}}\omega{_{A}}(\bm{p})\Lambda \omega{^{A}}(\bm{p}) &
{\omega}{_{A}}(\bm{p}){\Lambda\pi}{^{A}}(\bm{p})
\end{array}
\right)\Bigg|_{{}y^{ab}=0}\nonumber\\
&=&
-\left(
\begin{array}{cc}
{\bar\omega}{^{R'}}(\bm{p})W_a(\bm p){_{R'}}{^{S'}}\overline{\pi}{_{S'}}(\bm{p})
&
-\frac{m}{\sqrt{2}}{\bar\omega}{^{R'}}(\bm{p})W_a(\bm p){_{R'}}{^{S'}}\overline{\omega}{_{S'}}(\bm{p})
\\
\frac{m}{\sqrt{2}}\omega{^{R}}(\bm{p})W_a(\bm p){_{R}}{^{S}}\omega{_{S}}(\bm{p})
&
{\omega}{^{R}}(\bm{p})W_a(\bm p){_{R}}{^{S}}{\pi}{_{S}}(\bm{p})
\end{array}
\right)\nonumber\\
&=&
\frac{1}{2}\left(
\begin{array}{cc}
\pi{_{A}}(\bm p)\bar\pi{_{A'}}(\bm p)-\frac{m^2}{2}\omega{_{A}}(\bm p)\bar\omega{_{A'}}(\bm p)
&
-m\sqrt{2}\pi_{A}(\bm p)\bar\omega{_{A'}}(\bm p)
\\
-m\sqrt{2}\omega_{A}(\bm p)\bar\pi{_{A'}}(\bm p)
&
\frac{m^2}{2}\omega{_{A}}(\bm p)\bar\omega{_{A'}}(\bm p)
-\pi{_{A}}(\bm p)\bar\pi{_{A'}}(\bm p)
\end{array}
\right)\nonumber\\
\label{W_ARS}
\end{eqnarray}
The projection
\begin{eqnarray}
\omega{^{A}}(\bm p)\bar\omega{^{A'}}(\bm p)W_a(\bm p){_{\cal A}}{^{\cal B}}
&=&
W(\omega,\bm p){_{\cal A}}{^{\cal B}}
=
\frac{1}{2}
\left(
\begin{array}{cc}
1 & 0
\\
0 & -1
\end{array}
\right),
\end{eqnarray}
is the same for both $m=0$ and $m>0$. For $m\neq 0$, one can employ
\begin{eqnarray}
\pi{^{A}}(\bm p)\bar\omega{^{A'}}(\bm p)W_a(\bm p){_{\cal A}}{^{\cal B}}
&=&
\frac{m}{\sqrt{2}}
\left(
\begin{array}{cc}
0 & 0
\\
1 & 0
\end{array}
\right),\\
\omega{^{A}}(\bm p)\bar\pi{^{A'}}(\bm p)W_a(\bm p){_{\cal A}}{^{\cal B}}
&=&
\frac{m}{\sqrt{2}}
\left(
\begin{array}{cc}
0 & 1
\\
0 & 0
\end{array}
\right),
\end{eqnarray}
to define relativistically invariant analogs of the remaining two Pauli matrices,
The massless case,
\begin{eqnarray}
W_a(\bm k){_{\cal A}}{^{\cal B}}
&=&
\frac{1}{2}\left(
\begin{array}{cc}
1
&
0
\\
0
&
-1
\end{array}
\right)k_a
\label{W_ARSm=0}
\end{eqnarray}
explains why projections of the Pauli-Lubanski vectors cannot violate the Bell inequality if $m=0$: All such observables commute with one another.
The ultrarelativistic limit of massive particles is more complicated \cite{MC-PRA,CRW1}.
\subsection{Peres-Scudo-Terno problem and an $\omega$-spinor way to circumvent it}
True relativistic qubits are, by definition, fields that transform unitarily under inhomogeneous SL(2,C), i.e. the $t$-spinors, or the $t$-bispinors. Let us concentrate on the $t$-spinor case. The corresponding Hilbert space is equipped with the scalar product
\begin{eqnarray}
\langle f_1|f_2\rangle
&=&
\sum_{s=\pm}\int dp\,
\overline{f_1(s,\bm p)}f_2(s,\bm p).
\end{eqnarray}
The expression
$
\rho(s,\bm p;s',\bm p')=f(s,\bm p)\overline{f(s',\bm p')}
$
is an example of a pure-state one-particle density matrix of a relativistic qubit. The qubit is characterized by two quantum numbers --- spin $s$ and momentum $\bm p$.
Peres, Scudo and Terno posed the following problem \cite{PST}: Assume that our detectors do not distinguish between different $\bm p$, but detect the sign $s$ in $f(s,\bm p)$. Can we use this $s$ as a quantum bit? The probability of finding a given $s$ is given by $\int dp\,\rho(s,\bm p;s,\bm p)$. It is therefore justified to introduce the reduced spin density matrix $\rho(s,s')=\int dp\,\rho(s,\bm p;s',\bm p)$. The problem is that the entropy of $\rho(s,s')$ is not relativistically invariant. Indeed, let $f(s,\bm p)=F(s)G(\bm p)$, where $\int dp\,|G(\bm p)|^2=1$, $\sum_s|F(s)|^2=1$, and $\rho(s,s')=F(s)\overline{F(s')}$. The wave function $F(s)$ is a pure state so the entropy of $\rho(s,s')$ is zero. Now, the matrix $U(\Lambda,\bm p){_{\cal A}}{^{\cal B}}$ depends on momenta in a highly nontrivial way, and its form is even more complicated if one works with helicity or other $t$-spinors. In consequence, $\Lambda f(s,\bm p)$ does not separate into a product $\Lambda F(s)\Lambda G(\bm p)$, and thus involves a nontrivial entanglement between $s$ and $\bm p$. But then it is well known that the reduced density matrix $\Lambda\rho(s,s')$ is mixed: Its entropy is nonzero. The result has attracted attention of many authors and various variations on the theme can be found in the literature. With the exception of \cite{MCMW,MC05} apparently all the authors work in the helicity formalism.
But we know that qubits can be defined in terms of all $t$-spinors --- there is nothing fundamental about helicity, a $t$-spinor associated with direction parallel to some $t^{\bm a}=(1,0,0,0)$. In particular, let us take an arbitrary null direction $t^a=\tau^A\bar\tau^{A'}$. If $p^2=m^2>0$ then $p\cdot t$ is nonzero for all $p^a$. The null vector $\omega^a=\tau^A\bar\tau^{A'}/(p\cdot t)$ satisfies $p\cdot \omega=1$, so $\omega^A(\bm p)
=\tau^A/\sqrt{p\cdot t}$ can be used to split 4-momentum $p^a$ according to (\ref{p}), and the whole construction of $\omega$-spinors can be repeated.
Now, let $\Lambda{_A}{^B}$ be an arbitrary SL(2,C) transformation. It is known (Sec.~3.6 in \cite{PR}) that any SL(2,C) transformation possess at least one eigenvector, $\Lambda{_A}{^B}\tau_B=\lambda \tau_A$, $\lambda=|\lambda|e^{i\varphi}$, so let us take this $\tau_A$ in our definition of $t^a$. Then,
\begin{eqnarray}
\Lambda{_A}{^B}\omega_B(\bm{\Lambda^{-1}p})
&=&
\frac{\Lambda{_A}{^B}\tau_B}{\sqrt{\Lambda^{-1}p\cdot t}}=
\frac{\lambda\tau_A}{\sqrt{p\cdot \Lambda t}}=
e^{i\varphi}\omega_A(\bm p),\\
\Lambda{_A}{^B}\pi_B(\bm{\Lambda^{-1}p})
&=&
\frac{\Lambda{_A}{^B}(\Lambda^{-1}p){_B}{^{B'}}\bar\tau{_{B'}}}{{\sqrt{\Lambda^{-1}p\cdot t}}}
=
\frac{\Lambda{_A}{^B}\Lambda^{-1}{_B}{^C}p{_C}{^{B'}}\Lambda{_{B'}}{^{C'}}\bar\tau{_{C'}}}{{\sqrt{p\cdot \Lambda t}}}\nonumber\\
&=&
\frac{p{_A}{^{B'}}\bar\lambda\bar\tau{_{B'}}}{{|\lambda|\sqrt{p\cdot t}}}=e^{-i\varphi}\pi_A(\bm{p}).
\end{eqnarray}
The transformation matrix becomes
\begin{eqnarray}
&{}&
\left(
\begin{array}{cc}
{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\pi}{^{A'}}(\bm{p}) & -\frac{m}{\sqrt{2}}{\bar\omega}{_{A'}}(\bm{p})\overline{\Lambda\omega}{^{A'}}(\bm{p})
\\
\frac{m}{\sqrt{2}}\omega{_{A}}(\bm{p})\Lambda \omega{^{A}}(\bm{p}) &
{\omega}{_{A}}(\bm{p}){\Lambda\pi}{^{A}}(\bm{p})
\end{array}
\right)
\nonumber\\
&{}&\pp==
\left(
\begin{array}{cc}
e^{i\varphi}{\bar\omega}{_{A'}}(\bm{p})\overline{\pi}{^{A'}}(\bm{p}) & -e^{-i\varphi}\frac{m}{\sqrt{2}}{\bar\omega}{_{A'}}(\bm{p})\overline{\omega}{^{A'}}(\bm{p})
\\
e^{i\varphi}\frac{m}{\sqrt{2}}\omega{_{A}}(\bm{p})\omega{^{A}}(\bm{p}) &
e^{-i\varphi}{\omega}{_{A}}(\bm{p}){\pi}{^{A}}(\bm{p})
\end{array}
\right)
=
\left(
\begin{array}{cc}
e^{i\varphi} & 0
\\
0 &
e^{-i\varphi}
\end{array}
\right),
\end{eqnarray}
which is independent of $\bm p$. Actually, the transformation matrix is simply an ordinary rotation by $2\varphi$ which, of course, cannot change any entanglement. The reduced density matrix of Peres, Scudo and Terno has not changed its entropy, even though the SL(2,C) transformation was arbitrary.
The construction I have just presented was introduced in \cite{MCMW}, and can be regarded as a relativistic error correction algorithm. What it means is that it is always possible to adjust $t^a$ in $t^a W_a$ in a way that ensures that qubits defined by this concrete observable do not change their Peres-Scudo-Terno entropy. This trick can be generalized to any number of particles that propagate in arbitrary ways, not necessarily inertially, and described by any states --- entangled or not. Another consequence is that entanglement between Peres-Scudo-Terno qubits will not change, no matter what kind of motion is considered, if one defines qubits by means of $t^a$ which are principal null directions \cite{PR,PR2} of Lorentz transformations that define the motion of observers.
\section{Transformation properties of fields of spin frames for massless particles}
Let us consider a massless particle whose 4-momentum $k^a$ is a future-pointing null but nonzero world-vector, $k^ak_a=0$, $k^a\neq 0$. We know that such a $k^a$ determines, up to a phase, a 2-spinor $\pi^A$ satisfying $k^a=\pi^A\bar\pi^{A'}$. The latter equation can be understood in several ways. The approach I prefer takes the future light cone as a given 3-dimensional manifold on which a spinor field $\pi^A=\pi^A(\bm k)$ is defined, but treats this spinor field as a fundamental physical object related to 4-momentum by $k^a=\pi^A(\bm k)\bar\pi^{A'}(\bm k)$. A consequence is apparently paradoxical. The spinor field transforms under SL(2,C) transformations as
$\pi_A(\bm k)\mapsto\Lambda\pi_A(\bm k)=\Lambda{_A}{^B}\pi_B(\bm{\Lambda}^{-1}\bm k)$. But then $\Lambda\pi_A(\bm k)\overline{\Lambda\pi}_{A'}(\bm k)=\Lambda{_a}{^b}\Lambda{^{-1}}{_b}{^c}k_c=k_a$. Accordingly, the 4-momentum does not change. This type of approach is consistent with the idea that the fundamental fields are those of 2-spinors and not the 4-vectors. An important consequence is that both $\Lambda\pi_A(\bm k)$ and $\pi_A(\bm k)$ possess the same flagpole and, thus, differ at most by a phase factor: $\Lambda\pi_A(\bm k)=e^{-i\Theta(\Lambda,\bm k)}\pi_A(\bm k)$. For any spin-frame partner $\omega_A(\bm k)$ of $\pi_A(\bm k)$, we find $\omega_A(\bm k)\Lambda\pi^A(\bm k)=e^{-i\Theta(\Lambda,\bm k)}$ and
$\bar\omega{_{A'}}(\bm k)\overline{\Lambda\pi}{^{A'}}(\bm k)=e^{i\Theta(\Lambda,\bm k)}$. The latter two expressions have been encountered before in the diagonal elements of the matrix in (\ref{unitary1}).
Now let us consider the `orbital' part of the 6-angular momentum antisymmetric tensor of the massless particle: $M_{ab}=x_a k_b-x_bk_a$. Any antisymmetric (real) second-rank tensor can be written as (Eq. (3.4.20) in \cite{PR})
\begin{eqnarray}
M_{ab} &=& \mu_{AB}\varepsilon_{A'B'}+\varepsilon_{AB}\bar\mu_{A'B'},\\
\mu_{AB} &=& \frac{1}{2}M_{AA'BB'} \varepsilon^{A'B'}=\mu_{BA},\\
\bar\mu_{A'B'} &=& \overline{\mu_{AB}},
\end{eqnarray}
the relation between $M_{ab}$ and $\mu_{AB}$ being expressible also by means of the generator of $\Lambda{_A}{^B}$:
\begin{eqnarray}
M{^{ab}}\sigma_{ab}{_R}{^S}
&=&
M{^{ab}}
\frac{i}{2}(g_{aRX'}g_b{^{SX'}} - g_{bRX'}g_a{^{SX'}})
=i\, M{^{ab}}g_{aRX'}g_b{^{SX'}}\nonumber\\
&=&
i\, M_{RX'}{^{SX'}}=2i\,\mu_{R}{^{S}}.
\end{eqnarray}
Let us apply this procedure to $M_{ab}$:
\begin{eqnarray}
\mu_{AB}
&=&
\frac{1}{2}(x_a k_b-x_bk_a)\varepsilon^{A'B'}\nonumber\\
&=&
\frac{1}{2}(x_{AA'} \pi_B\bar\pi_{B'}-x_{BB'}\pi_A\bar\pi_{A'})\varepsilon^{A'B'}
\nonumber\\
&=&
\frac{1}{2}(\underbrace{x_{AA'}\bar\pi^{A'}}_{\omega_A} \pi_B\underbrace{x_{BB'}\bar\pi^{B'}}_{\omega_B}\pi_A)
\nonumber\\
&=&
\frac{1}{2}(\omega_A\pi_B+\omega_B\pi_A).\label{W W W omega}
\end{eqnarray}
The orbital part thus turns out to be determined by two spinors: $\pi_A$ whose flagpole is the 4-momentum, and $\omega_A=\omega_A(x,\pi)=x_{AA'}\bar\pi^{A'}$.
The bispinor $(\omega_A,\bar\pi_{A'})$, with the characteristic bilinear dependence of $\omega_A$ on $x$ and $\bar\pi$, is known as a {\it twistor} \cite{PR2}.
Treating $\pi_A$ as a spinor field $\pi_A(\bm k)$ we obtain $\omega_A=\omega_A(x,\bm k)=x_{AA'}\bar\pi^{A'}(\bm k)$ which is a field defined on the Cartesian product of the Minkowski space with the momentum-space light cone. The field transforms under SL(2,C) transformations as follows
\begin{eqnarray}
\omega_{A}(x,\bm k)
&\to&
\Lambda{_A}{^B}\omega_B(\Lambda^{-1}x,{\bm \Lambda}^{-1}{\bm k})
\nonumber\\
&=&
\Lambda{_A}{^B}\Lambda^{-1}{_B}{^C}\Lambda^{-1}{_{B'}}{^{C'}}x_{CC'}\bar\pi^{B'}({\bm \Lambda}^{-1}{\bm k})
\nonumber\\
&=&
\varepsilon{_A}{^C}x_{CC'}\bar\pi^{B'}({\bm \Lambda}^{-1}{\bm k})\Lambda^{-1}{_{B'}}{^{C'}}
\nonumber\\
&=&
x_{AC'}\overline{\Lambda\pi}{^{C'}}({\bm k})
\nonumber\\
&=&
x_{AA'}e^{i\Theta(\Lambda,\bm k)}\bar\pi{^{A'}}({\bm k})
\nonumber\\
&=&
e^{i\Theta(\Lambda,\bm k)}\omega_A(x,\bm k).
\end{eqnarray}
The contraction
\begin{eqnarray}
\omega_A(x,\bm k)\pi^A(\bm k)
&=&
x_{AA'}\pi^A(\bm k)\bar\pi^{A'}(\bm k)=x_ak^a=x\cdot k
\end{eqnarray}
shows that for $x_ak^a\neq 0$ we can define a spin-frame field
\begin{eqnarray}
\iota_A(\bm k) &=& \pi_A(\bm k),\label{!1}\\
o_A(x,\bm k) &=& \frac{\omega_A(x,\bm k)}{x\cdot k}=\frac{x_{AA'}\bar\pi^{A'}(\bm k)}{x\cdot k},
\end{eqnarray}
satisfying simultaneously
\begin{eqnarray}
o_A(x,\bm k)\iota^A(\bm k) &=& 1
\\
\iota_A(\bm k)\bar\iota_{A'}(\bm k) &=& k_a,\\
\Lambda{_A}{^B}\iota_B(\bm{\Lambda}^{-1}\bm k)
&=&
e^{-i\Theta(\Lambda,\bm k)}\iota_A(\bm k),\\
\Lambda{_A}{^B}o_B(\Lambda^{-1}x,{\bm \Lambda}^{-1}{\bm k})
&=&
e^{i\Theta(\Lambda,\bm k)}o_A(x,\bm k).\label{!2}
\end{eqnarray}
It is an instructive exercise to check what would have been the consequences of assuming that
$x_{AA'}\bar\pi^{A'}(\bm k)/(x\cdot k)$ is a spinor field defined on the light cone only, and not on the `phase space' $(x,\bm k)$. Then we would have to assume the transformation rule
\begin{eqnarray}
\Lambda o_A({\bm k})
&=&
\Lambda{_A}{^B}o_B({\bm \Lambda}^{-1}{\bm k})
=
e^{i\Theta(\Lambda,\bm k)}\Lambda x_{AA'}\bar\pi^{A'}(\bm k)/(\Lambda x\cdot k).
\label{wrong}
\end{eqnarray}
The right side of (\ref{wrong}) is not proportional to $o_A({\bm k})$ but also contains a component parallel to $\iota_A({\bm k})$.
In effect, we find
\begin{eqnarray}
\Lambda o_A({\bm k})
&=&
\Lambda{_A}{^B}o_B({\bm \Lambda}^{-1}{\bm k})
=
e^{i\Theta(\Lambda,\bm k)}\tilde o_A({\bm k})\label{tilde o}
\end{eqnarray}
where, in spite of the fact that $\tilde o_A({\bm k})\iota^A({\bm k})=1$, $\tilde o_A({\bm k})=o_A({\bm k})+\phi(\Lambda,\bm k)\iota^A({\bm k})$.
The appearance of the function $\phi(\Lambda,\bm k)=o_A({\bm k})\tilde o^A({\bm k})$, which is different for different explicit forms of the spin frame, is responsible, as we shall see later, for a gauge transformation of the electromagnetic four-potential and has consequences for field quantization.
This change of gauge seems to be a generic property of spin frames defined on mass-$m$ hyperboloids.
However, we will see that there are reasons to believe that Eq.~(\ref{!2}) is of fundamental importance for relativistic transformations of electromagnetic qubits, but one should not identify the point $x^a$ with an {\it event\/} in Minkowski space. A more natural interpretation arises if one treats the twistor as defined with respect to some internal coordinate, say $R^a$, which is timelike (and thus $R\cdot k$ is nowhere vanishing if $k^a\neq 0$). At the level of quantized electromagnetic field the approach will lead to an intriguing splitting of the field in momentum space into three massless fields: one field transforming according to the unitary spin-1 representation of the Poincar\'e group, and two scalars. Photon degrees of freedom will be defined with respect to the spin-1 part.
\section{Qubits associated with electromagnetic field}
Qubits associated with electromagnetic field can be discussed at both first- and second-quantized levels. In this section we concentrate on the first-quantized description, which is important since transformation properties of polarization degrees of freedom will remain unchanged after field quantization.
Electromagnetic field tensor is antisymmetric, so can be decomposed in the usual way: $F_{ab}(x)=F_{AB}(x)\varepsilon_{A'B'}+\varepsilon_{AB}\bar F_{A'B'}(x)$, $F_{A'B'}(x)=\overline{F_{AB}(x)}$, $F_{AB}(x)=F_{BA}(x)$. Source-free Maxwell equations, $\partial^a F_{ab}(x)=\partial^a {^*}F_{ab}(x)=0$, are equivalent to the spinor equation
\begin{eqnarray}
\partial^{AA'}F_{AB}(x)=0=\partial^{AA'}F_{A'B'}(x).
\end{eqnarray}
In Fourier space
\begin{eqnarray}
F_{AB}(x)
=
\int dk\,\Big(F_{-AB}(\bm k)e^{-ik\cdot x}+F_{+AB}(\bm k)e^{ik\cdot x}\Big),\quad
F_{\pm AB}(\bm k)\pi{^{A}}(\bm k)\bar\pi{^{A'}}(\bm k)
=
0.\nonumber\\
\label{F-AB}
\end{eqnarray}
Expanding $F_{\pm AB}(\bm k)$ in terms of spin frames we find that the only symmetric solution of (\ref{F-AB}) is $F_{\pm AB}(\bm k)\sim
\pi{_{A}}(\bm k)\pi{_{B}}(\bm k)$. The field spinor can be thus written as
\begin{eqnarray}
F_{AB}(x)
&=&
-\int dk\,
\pi_A(\bm k)\pi_B(\bm k)
\Big(\alpha(-,\bm k)e^{-ik\cdot x}
+\overline{\alpha(+,\bm k)}e^{ik\cdot x}\Big),\label{F AB1}\\
F_{A'B'}(x)
&=&
-\int dk\,
\bar\pi_{A'}(\bm k)\bar\pi_{B'}(\bm k)
\Big(\alpha(+,\bm k)e^{-ik\cdot x}
+\overline{\alpha(-,\bm k)}e^{ik\cdot x}\Big).\label{F AB2}
\end{eqnarray}
The signs and conjugations are a matter of convention.
Similarity of (\ref{F AB1})-(\ref{F AB2}) to the zero-mass solution of Dirac's equation (\ref{psi m=01})-(\ref{psi m=02}) is not accidental and is typical of massless Bargmann-Wigner equations of any spin \cite{MC-BW,Woodhouse}. SL(2,C) transformations of (\ref{F AB1})-(\ref{F AB2}) can be deduced from analogous formulas found for the Dirac field,
\begin{eqnarray}
&{}&
\Lambda{_{A}}{^{C}}\Lambda{_{B}}{^{D}}F_{CD}(\Lambda^{-1}x)
\nonumber\\
&{}&\pp==
-\int dk\,
\pi_A(\bm k)\pi_B(\bm k)
e^{-2i\Theta(\Lambda,\bm k)}\Big(\alpha(-,\bm{\Lambda^{-1}k})e^{-ik\cdot x}
+\overline{\alpha(+,\bm{\Lambda^{-1}k})}e^{ik\cdot x}\Big),\nonumber\\
\label{F AB1'}\\
&{}&
\Lambda{_{A'}}{^{C'}}\Lambda{_{B'}}{^{D'}}F_{C'D'}(\Lambda^{-1}x)
\nonumber\\
&{}&\pp==
-\int dk\,
\bar\pi_{A'}(\bm k)\bar\pi_{B'}(\bm k)
e^{2i\Theta(\Lambda,\bm k)}\Big(\alpha(+,\bm{\Lambda^{-1}k})e^{-ik\cdot x}
+\overline{\alpha(-,\bm{\Lambda^{-1}k})}e^{ik\cdot x}\Big).\nonumber\\
\label{F AB2'}
\end{eqnarray}
\subsection{Four-potential $A_a(x)$ based on spin frames defined on the light cone}
Four-potential is defined by
\begin{eqnarray}
{F}_{ab}(x)
=
\partial_a A_b(x)-\partial_b A_a(x)
\label{F}.
\end{eqnarray}
We know that given $\pi^A(\bm k)$ we cannot find a unique spin-partner $\omega^A(\bm k)$, since
for any function $\phi(\bm k)$ the new field
\begin{eqnarray}
\tilde \omega^A(\bm k)=\omega^A(\bm k)+\phi(\bm k)\pi^A(\bm k)
\label{sim}
\end{eqnarray}
also satisfies
$\tilde \omega_A(\bm k)\pi^A(\bm k)=1$. This leads to the equivalence relation:
$\tilde \omega^A(\bm k)\sim \omega^A(\bm k)$ iff
$\tilde \omega^A(\bm k)-\omega^A(\bm k)$ is proportional to
$\pi^A(\bm k)$.
The 4-potential in a Lorenz gauge\footnote{Note: `Lorenz gauge' (L. Lorenz) and `Lorentz group' (H. A. Lorentz).}, can be taken in the form
(cf.~\cite{Ashtekar,MCJN})
\begin{eqnarray}
A_a(x) &=& i\int dk\,e^{-ik\cdot x}
\Big(
\omega_A(\bm k)\bar\pi_{A'}(\bm k)
\alpha(+,\bm k)
+
\pi_{A}(\bm k)\bar\omega_{A'}(\bm k)
\alpha(-,\bm k)
\Big)
+ {\,\rm c.c.}\label{A}
\end{eqnarray}
Now, if we replace $\omega^A(\bm k)$ by $\tilde \omega^A(\bm k)$
satisfying (\ref{sim}),
then
\begin{eqnarray}
{A}_a(x) &\mapsto&
\tilde {A}_a(x)
=
{A}_a(x)
-
\partial_a
\Phi(x)\label{tilde omega}
\end{eqnarray}
where
\begin{eqnarray}
\Phi(x)
&=&
\int dk\,
\phi(\bm k)
\Big(\alpha(+,\bm k)e^{-ik\cdot x}
+
\overline{\alpha(-,\bm k)}e^{ik\cdot x}
\Big)
+\,{\rm c.c.}
\end{eqnarray}
is a solution of $\partial^a\partial_a \Phi(x)=0$. The equivalence
class of spin-frames corresponds to an equivalence class of
Lorenz-gauge potentials.
(\ref{F AB1'})-(\ref{F AB2'}) show the duality between active spinor transformations and passive zero-mass, spin-1 unitary transformations of wave functions. An analogous construction does not exactly work for $A_a(x)$. Let us have a closer look at this important subtlety since it has nontrivial implications for field quantization.
The problem is with changes of gauge of the form (\ref{tilde omega}) and (\ref{tilde o}):
\begin{eqnarray}
\Lambda{_a}{^b}A_b(\Lambda^{-1}x)
&=& i\int dk\,e^{-ik\cdot x}
e^{2i\Theta(\Lambda,\bm k)}
\tilde\omega_A(\bm k)\bar\pi_{A'}(\bm k)
\alpha(+,\bm{\Lambda^{-1} k})
+\dots\nonumber
\\
&=&i\int dk\,e^{-ik\cdot x}
e^{2i\Theta(\Lambda,\bm k)}
\omega_A(\bm k)\bar\pi_{A'}(\bm k)
\alpha(+,\bm{\Lambda^{-1} k})
+
\partial_a\Phi(x)+\dots\nonumber
\\
\end{eqnarray}
In consequence, if wave functions transform according to massless spin-1 unitary representations of inhomogeneous SL(2,C), then
\begin{eqnarray}
A_a(x)\to \Lambda{_a}{^b}A_b(\Lambda^{-1}x)-\partial_a\Phi(x),
\end{eqnarray}
meaning that $A_a(x)$ is not a four-vector field. And the other way around, if we assume that the four-potential is a four-vector field, its corresponding wave functions do not carry massless spin-1 unitary representations.
\subsection{Null vs. Minkowski tetrads, or circular vs. linear polarizations}
The polarization world vectors occurring in (\ref{A}) can be regarded as elements of a null tetrad in Minkowski space (we skip the arguments $\bm k$):
\begin{eqnarray}
\left(
\begin{array}{c}
g^{a}_{~~00'}\\
g^{a}_{~~01'}\\
g^{a}_{~~10'}\\
g^{a}_{~~11'}
\end{array}
\right) =
\left(
\begin{array}{c}
\varepsilon^{A}_{~~0}\varepsilon^{A'}_{~~0'}\\
\varepsilon^{A}_{~~0}\varepsilon^{A'}_{~~1'}\\
\varepsilon^{A}_{~~1}\varepsilon^{A'}_{~~0'}\\
\varepsilon^{A}_{~~1}\varepsilon^{A'}_{~~1'}
\end{array}
\right)=
\left(
\begin{array}{c}
\omega^{A}\bar\omega^{A'}\\
\omega^{A}\bar\pi^{A'}\\
\pi^{A}\bar\omega^{A'}\\
\pi^{A}\bar\pi^{A'}
\end{array}
\right)
=
\left(
\begin{array}{c}
\omega^{a}\\
m^{a}\\
\bar{m}^{a}\\
k^{a}
\end{array}
\right)\label{4}
\end{eqnarray}
and dually
\begin{eqnarray}
\left(
\begin{array}{c}
g_{a}^{~~00'}\\
g_{a}^{~~01'}\\
g_{a}^{~~10'}\\
g_{a}^{~~11'}
\end{array}
\right) =
\left(
\begin{array}{c}
\varepsilon_{A}^{~~0}\varepsilon_{A'}^{~~0'}\\
\varepsilon_{A}^{~~0}\varepsilon_{A'}^{~~1'}\\
\varepsilon_{A}^{~~1}\varepsilon_{A'}^{~~0'}\\
\varepsilon_{A}^{~~1}\varepsilon_{A'}^{~~1'}
\end{array}
\right)
=
\left(
\begin{array}{c}
\pi_{A}\bar\pi_{A'}\\
-\pi_{A}\bar\omega_{A'}\\
-\omega_{A}\bar\pi_{A'}\\
\omega_{A}\bar\omega_{A'}
\end{array}
\right)
=
\left(
\begin{array}{c}
k_{a}\\
-\bar{m}_{a}\\
-m_{a}\\
\omega_{a}
\end{array}
\right)\label{5}
\end{eqnarray}
The associated Minkowski tetrad
can be expressed by means of Infeld-van der Waerden symbols,
$g^{a}{_{\bf{a}}}
=
g_{\bf{a}}{^{\bf{BB}'}}g^{a}{_{\bf{BB}'}}
$,
\begin{eqnarray}
\left(
\begin{array}{c}
t^{a}\\
x^{a}\\
y^{a}\\
z^{a}
\end{array}
\right) = \left(
\begin{array}{c}
g^{a}_{~~0}\\
g^{a}_{~~1}\\
g^{a}_{~~2}\\
g^{a}_{~~3}
\end{array}
\right) = \frac{1}{\sqrt{2}} \left(
\begin{array}{cccc}
1&0&0&1\\
0&1&1&0\\
0&i&-i&0\\
1&0&0&-1
\end{array}
\right) \left(
\begin{array}{c}
\omega^{A}\bar\omega^{A'}\\
\omega^{A}\bar\pi^{A'}\\
\pi^{A}\bar\omega^{A'}\\
\pi^{A}\bar\pi^{A'}
\end{array}
\right)
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{c}
\omega^{a}+k^{a}\\
m^{a}+\bar{m}^{a}\\
i m^{a} - i\bar{m}^{a}\\
\omega^{a}-k^{a}
\end{array}
\right)
\end{eqnarray}
and dually,
$g_{a}{^{\bf{a}}}
=
g^{\bf{a}}{_{\bf{BB}'}}g_{a}{^{\bf{BB}'}}
$,
\begin{eqnarray}
\left(
\begin{array}{c}
t_{a}\\
-x_{a}\\
-y_{a}\\
-z_{a}
\end{array}
\right) = \left(
\begin{array}{c}
g_{a}^{~~0}\\
g_{a}^{~~1}\\
g_{a}^{~~2}\\
g_{a}^{~~3}
\end{array}
\right) = \frac{1}{\sqrt{2}} \left(
\begin{array}{cccc}
1&0&0&1\\
0&1&1&0\\
0&-i&i&0\\
1&0&0&-1
\end{array}
\right) \left(
\begin{array}{c}
\pi_{A}\bar\pi_{A'}\\
-\pi_{A}\bar\omega_{A'}\\
-\omega_{A}\bar\pi_{A'}\\
\omega_{A}\bar\omega_{A'}
\end{array}
\right)=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{c}
k_{a}+\omega_{a}\\
-\bar{m}_{a}-m_{a}\\
i \bar{m}_{a} -i m_{a}\\
k_{a}-\omega_{a}
\end{array}
\right)\nonumber
\end{eqnarray}
The meaning of $x_a$ and $y_a$ can be deduced from
\begin{eqnarray}
x_a(\bm k)
\alpha_1(\bm k)
+
y_a(\bm k)
\alpha_2(\bm k)
&=&
m_a(\bm k)
\alpha(+,\bm k)
+
\bar m_a(\bm k)
\alpha(-,\bm k),\\
\alpha(\pm,\bm k)
&=&
\frac{1}{\sqrt{2}}\Big(\alpha_1(\bm k)\pm i\,\alpha_2(\bm k)\Big).
\end{eqnarray}
Elements $m_a(\bm k)$, $\bar m_a(\bm k)$ of the null tetrad correspond to two circular polarizations (since after SL(2,C) transformations they get multiplied by appropriate Wigner phase factors). The two elements of the Minkowski tetrad, $x_a(\bm k)$, $y_a(\bm k)$, define linear polarizations,
\begin{eqnarray}
A_a(x) &=& i\int dk\,e^{-ik\cdot x}
\Big(
g{_{a}}{^{1}}(\bm k)
\alpha_1(\bm k)
+
g{_{a}}{^{2}}(\bm k)
\alpha_2(\bm k)
\Big)
+ {\,\rm c.c.}\label{A L}
\end{eqnarray}
\subsection{$A_a(x)$ based on twistor-like spin-frames}
Let $R^2=1$, $R_0>0$, $dR=d^3R/[(2\pi)^3 2\sqrt{1+\bm R^2}]$, and consider spin-frames analogous to (\ref{!1})-(\ref{!2}):
\begin{eqnarray}
\omega_A(R,\bm k)\pi^A(\bm k) &=& 1,\label{!1'}
\\
\pi_A(\bm k)\bar\pi_{A'}(\bm k) &=& k_a,\\
\Lambda{_A}{^B}\pi_B(\bm{\Lambda}^{-1}\bm k)
&=&
e^{-i\Theta(\Lambda,\bm k)}\pi_A(\bm k),\\
\Lambda{_A}{^B}\omega_B(\bm{\Lambda^{-1}R},{\bm \Lambda}^{-1}{\bm k})
&=&
e^{i\Theta(\Lambda,\bm k)}\omega_A(\bm R,\bm k).\label{!2'}
\end{eqnarray}
The world-vector $R^a$ has properties of four-velocity. In this sense it is analogous to the four-velocity parameter occurring in theories with preferred reference frame or based on non-standard clock synchronization schemes \cite{RS,R,CR2}. Another interesting analogy is with fields defined by Kaiser on phase-space \cite{Kaiser,Kjmp} --- here $R^a$ would be an analogue of the temper vector.
The four-potential
\begin{eqnarray}
A_a(x) &=&
i\int dk\,dR\,e^{-ik\cdot x}
\nonumber\\
&{}&\times
\Big(
\omega_A(\bm R,\bm k)\bar\pi_{A'}(\bm k)
\alpha(+,\bm R,\bm k)
+
\pi_{A}(\bm k)\bar\omega_{A'}(\bm R,\bm k)
\alpha(-,\bm R,\bm k)
\Big)
+ {\,\rm c.c.}\nonumber\\
&=& i\int dk\,dR\,e^{-ik\cdot x}
\Big(
g{_{a}}{^{1}}(\bm R,\bm k)
\alpha_1(\bm R,\bm k)
+
g{_{a}}{^{2}}(\bm R,\bm k)
\alpha_2(\bm R,\bm k)
\Big)
+ {\,\rm c.c.}\label{A L R}
\end{eqnarray}
transforms as follows
\begin{eqnarray}
\Lambda{_a}{^b}A_b(\Lambda^{-1}x)
&=&
i\int dk\,dR\,e^{-ik\cdot \Lambda^{-1}x}
\Lambda{_a}{^b}
\omega_B(\bm R,\bm k)\bar\pi_{B'}(\bm k)
\alpha(+,\bm R,\bm k)
+\dots\nonumber\\
&=&
i\int dk\,dR\,e^{-ik\cdot x}
\Lambda{_A}{^B}
\omega_B(\bm R,\bm{\Lambda^{-1}k})\overline{\Lambda\pi}{_{B'}}(\bm k)
\alpha(+,\bm R,\bm{\Lambda^{-1}k})
+\dots\nonumber\\
&=&
i\int dk\,dR\,e^{-ik\cdot x}
e^{2i\Theta(\Lambda,\bm k)}
\omega_A(\bm{R},\bm{k})\bar\pi{_{A'}}(\bm k)
\alpha(+,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k})
+\dots\nonumber\\
\end{eqnarray}
Note that linear polarization vectors are elements of the Minkowski tetrad transforming as follows:
\begin{eqnarray}
\left(
\begin{array}{c}
\Lambda t^{a}(\bm R,\bm k)\\
\Lambda x^{a}(\bm R,\bm k)\\
\Lambda y^{a}(\bm R,\bm k)\\
\Lambda z^{a}(\bm R,\bm k)
\end{array}
\right)
&=&
\frac{1}{\sqrt{2}} \left(
\begin{array}{cccc}
1&0&0&1\\
0&1&1&0\\
0&i&-i&0\\
1&0&0&-1
\end{array}
\right) \left(
\begin{array}{c}
\Lambda \omega^{A}(\bm R,\bm k)\overline{\Lambda \omega}{^{A'}}(\bm R,\bm k)\\
\Lambda \omega^{A}(\bm R,\bm k)\overline{\Lambda \pi}{^{A'}}(\bm k)\\
\Lambda \pi^{A}(\bm k)\overline{\Lambda \omega}{^{A'}}(\bm R,\bm k)\\
\Lambda \pi^{A}(\bm k)\overline{\Lambda \pi}{^{A'}}(\bm k)
\end{array}
\right)\label{LLL-R}\\
&=&
\left(
\begin{array}{c}
t^a(\bm R,\bm k)
\\
\cos 2\Theta(\Lambda,\bm k) x^a(\bm R,\bm k)
+\sin 2\Theta(\Lambda,\bm k) y^a(\bm R,\bm k)
\\
-\sin 2\Theta(\Lambda,\bm k) x^a(\bm R,\bm k)
+\cos 2\Theta(\Lambda,\bm k) y^a(\bm R,\bm k)
\\
z^a(\bm R,\bm k)
\end{array}
\right)\label{LLLL-R}
\end{eqnarray}
$t^a(\bm R,\bm k)$ and $z^a(\bm R,\bm k)$ do not appear in (\ref{A L R}), but the space-like plane spanned by $x^a(\bm R,\bm k)$ and $y^a(\bm R,\bm k)$ is invariant under SL(2,C). So the twistor-like spin-frames lead to Minkowski tetrads that span two orthogonal subspaces, both invariant under SL(2,C). This does not hold for tetrads corresponding to spin-frames that depend only on $\bm k$.
The passive unitary representation of the amplitudes reads
\begin{eqnarray}
\Lambda\alpha(\pm,\bm{R},\bm{k})
&=&
e^{\pm 2i\Theta(\Lambda,\bm k)}
\alpha(\pm,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k}).
\end{eqnarray}
There is no change of gauge. (Change of gauge means that the plane spanned by $m_a$ and $\bar m_a$ is not invariant, since a contribution parallel to $k_a$ occurs. In quantum-field-theory terms this would mean that SL(2,C) transformations mix transverse photons with `longitudinal or timelike' spin-0 `photons'; see the discussion in the next Section.)
The field spinors
\begin{eqnarray}
F_{AB}(x)
&=&
-\int dk\,dR\,
\pi_A(\bm k)\pi_B(\bm k)
\Big(\alpha(-,\bm R,\bm k)e^{-ik\cdot x}
+\overline{\alpha(+,\bm R,\bm k)}e^{ik\cdot x}\Big)\label{F AB1 R}\\
F_{A'B'}(x)
&=&
-\int dk\,dR\,
\bar\pi_{A'}(\bm k)\bar\pi_{B'}(\bm k)
\Big(\alpha(+,\bm R,\bm k)e^{-ik\cdot x}
+\overline{\alpha(-,\bm R,\bm k)}e^{ik\cdot x}\Big),\label{F AB2 R}
\end{eqnarray}
are identical to (\ref{F AB1})-(\ref{F AB2}) if $\int dR\,\alpha(\pm,\bm R,\bm k)=\alpha(\pm,\bm k)$. Moreover, relativistic invariance of $dR$ implies that
\begin{eqnarray}
\Lambda\alpha(\pm,\bm k)
&=&
e^{\pm 2i\Theta(\Lambda,\bm k)}
\int dR\,\alpha(\pm,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k})
=
e^{\pm 2i\Theta(\Lambda,\bm k)}\alpha(\pm,\bm{\Lambda^{-1}k}).
\label{int dR a}
\end{eqnarray}
Let $\int dR\,\alpha_{j}(\bm R,\bm k)=\alpha_{j}(\bm k)$, $j=1,2$.
SL(2,C) simply rotates the amplitude 2-dimensional vector $(\alpha_{1}(\bm k), \alpha_{2}(\bm k))$ by spin-1 Wigner angle $2\Theta(\Lambda,\bm k)$.
The world-vector $R^a$ plays a role of an internal degree of freedom, invisible at the level of field tensors. Its role can be fully appreciated only at the level of quantized fields \cite{MCKW}. Let us stress that the choice of a time-like $R^a$ is perhaps not necessary --- in principle $R^a$ could be space-like or null, but only for a time-like $R^a$ the product $R\cdot k$ is, for $k^a\neq 0$, non-vanishing, which allows us to define spin-frames globally for all $k^a$ and $R^a$. The globality condition is convenient but not necessary.
\section{Qubits associated with quantized electromagnetic field}
An automatic identification of the Pauli-Lubanski vector with polarization of a quantum field leads to a problem: $W^a$ is {\it quadratic\/} in elements of Poincar\'e algebra, $W^a=P_b{^*}M^{ab}$, and thus --- in terms of number operators --- $W^a$ is quadratic as well. However, spin observables, similarly to 4-momentum and 6-angular momentum, are expected to be linear in numbers of particles. The subtlety is not so visible in first quantization since observables are there defined at 1-particle levels. A possible way out is to treat the {\it harmonic oscillator algebra\/} as the fundamental algebraic level, and treat inhomogeneous SL(2,C) as a structure derived from it. The choice of spin-fames --- possessing appropriate transformation properties --- becomes then essential for a correct relativistic formulation of the formalism.
\subsection{Manifestly covariant quantization of $A_a(x)$ based on twistor-like spin-frames}
One typically quantizes $A_a(x)$ in a way that does not guarantee the four-vector behavior of the field. An approach to relativistic EPR correlations of photons based on such a non-manifestly covariant formalism was recently discussed in \cite{Caban,Caban2}. In what follows I will outline an alternative but not widely known formulation based on tetrads built from twistor-like spin-frames \cite{MCKW}.
The potential is defined by
\begin{eqnarray}
A_a(x)
&=& i\int dk\,dR\,e^{-ik\cdot x}
\nonumber\\
&\pp=&
\times
\Big(
g{_{a}}{^{1}}(\bm R,\bm k)
a_1(\bm R,\bm k)
+
g{_{a}}{^{2}}(\bm R,\bm k)
a_2(\bm R,\bm k)
\nonumber\\
&\pp=&
\pp{\times\Big(}
+
g{_{a}}{^{3}}(\bm R,\bm k)
a_3(\bm R,\bm k)
+
g{_{a}}{^{0}}(\bm R,\bm k)
a_0(\bm R,\bm k)^{\dag}
\Big)
+ {\,\rm c.c.}
\end{eqnarray}
The field $g{_{a}}{^{\bf a}}(\bm R,\bm k)$, by definition of a Minkowski tetrad, satisfies $g{_{a}}{^{\bf a}}(\bm R,\bm k)g{^{a}}{^{\bf b}}(\bm R,\bm k)
=g^{\bf ab}$, $g{_{a}}{_{\bf a}}(\bm R,\bm k)g{_{b}}{^{\bf a}}(\bm R,\bm k)
=g_{ab}$.
The amplitude operators are assumed to satisfy the ordinary harmonic-oscillator Lie algebra,
\begin{eqnarray}
{[a_{\bf a}(\bm R,\bm k),a_{\bf b}(\bm R',\bm k')^{\dag}]}
&=&
\delta_{\bf ab}\delta(\bm R,\bm k,\bm R',\bm k')I(\bm R,\bm k),\label{ccr1}\\
{[a_{\bf a}(\bm R,\bm k),n_{\bf b}(\bm R',\bm k')]}
&=&
\delta_{\bf ab}\delta(\bm R,\bm k,\bm R',\bm k')a_{\bf a}(\bm R,\bm k),\label{nccr1}\\
{[a_{\bf a}(\bm R,\bm k)^{\dag},n_{\bf b}(\bm R',\bm k')]}
&=&
-\delta_{\bf ab}\delta(\bm R,\bm k,\bm R',\bm k')a_{\bf a}(\bm R,\bm k)^{\dag},\label{nccr2}\\
{[a_{\bf a}(\bm R,\bm k),I(\bm R',\bm k')]}
&=&
{[a_{\bf a}(\bm R,\bm k)^{\dag},I(\bm R',\bm k')]}
={[n_{\bf a}(\bm R,\bm k)^{\dag},I(\bm R',\bm k')]}=0,\label{ccr2}\\
{[a_{\bf a}(\bm R,\bm k),a_{\bf b}(\bm R',\bm k')]}
&=&
{[a_{\bf a}(\bm R,\bm k)^{\dag},a_{\bf b}(\bm R',\bm k')^{\dag}]}=0.
\end{eqnarray}
Here $\delta_{\bf ab}$, ${\bf a}, {\bf b}=0,1,2,3$, is the Kronecker delta; $\delta(\bm R,\bm k,\bm R',\bm k')=\delta_1(\bm R,\bm R')
\delta_0(\bm k,\bm k')$, where the latter two distributions are SL(2,C) invariant Dirac deltas on $R^2=1$ and $k^2=0$ hyperboloids, respectively. Central element
$I(\bm R,\bm k)$ is a multiple of identity in irreducible representations of the algebra. However, I believe that a class of reducible representations discussed below is more physical \cite{MCJN,MCKW,MWMC} so I keep $I(\bm R,\bm k)$ at this stage unspecified. How to choose the representation of the algebra becomes particularly important if one considers the so called {\it entanglement with vacuum\/}, which turns out to be a notion representation dependent \cite{MPMC}. Number operator $n_{\bf b}(\bm R',\bm k')$ is added explicitly as an element of the Lie algebra as there exist representations where $n_{\bf b}(\bm R,\bm k)\neq a_{\bf b}(\bm R,\bm k)^{\dag}a_{\bf b}(\bm R,\bm k)$.
Let us note that $e^{-ik\cdot x}g{_{a}}{^{0}}(\bm R,\bm k)$ is accompanied by the {\it creation\/} operator $a_0(\bm R,\bm k)^{\dag}$. Had we employed in this place an annihilation operator, as it is typically done in the literature, we would have to resign either from hermiticity of the 4-potential or positivity of the scalar product in the space of states. Both cases lead to nonunitary evolutions. Our formulation leads to a manifestly covariant quantization procedure with Hermitian 4-potential and does not involve unphysical indefinite-metric `Hilbert space' \cite{MCJN,MCKW}.
The field commutator
\begin{eqnarray}
{[A_a(x),A_b(y)]}
&=&
ig_{ab} D(x-y)\label{[A,A]}
\end{eqnarray}
involves a generalized Jordan-Pauli `function'
\begin{eqnarray}
D(x)
&=&
i\int d k\,dR\,I(\bm R,\bm k)
\big(
e^{-ik\cdot x}
-
e^{ik\cdot x}
\big).
\label{J-P0}
\end{eqnarray}
The fact that we work with twistor-like spin-frames guarantees that the four operators occurring in (\ref{ccr1})--(\ref{ccr2}) split into two different massless representations of inhomogeneous SL(2,C):
$a_{1}(\bm R,\bm k)$, $a_{2}(\bm R,\bm k)$ (spin 1), and $a_{0}(\bm R,\bm k)$, $a_{3}(\bm R,\bm k)$ (spin 0). Transformation
$x_a\mapsto \Lambda{_a}{^b}x_b+y_a$ has to be represented unitarily at the operator level, i.e. there must exist $U(\Lambda,y)$ such that
\begin{eqnarray}
U(\Lambda,y)^{\dag}a(\pm,\bm{R},\bm{k})U(\Lambda,y)
&=&
e^{ip\cdot y}e^{\pm 2i\Theta(\Lambda,\bm k)}
a(\pm,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k}),\label{spin1}\\
U(\Lambda,y)^{\dag}a_3(\bm{R},\bm{k})U(\Lambda,y)
&=&
e^{ip\cdot y}
a_3(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k}),\\
U(\Lambda,y)^{\dag}a_0(\bm{R},\bm{k})^{\dag}U(\Lambda,y)
&=&
e^{ip\cdot y}
a_0(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k})^{\dag}.
\end{eqnarray}
Alternatively,
\begin{eqnarray}
U(\Lambda,y)^{\dag}A_a(x)U(\Lambda,y)
&=&
\Lambda{_a}{^b}A_b\big(\Lambda^{-1}(x-y)\big).
\end{eqnarray}
For twistor-like spin-frames the above two sets of transformations are {\it not\/} inconsistent.
Electromagnetic qubits are associated only with the spin-1 part (\ref{spin1}).
How to explicitly construct $U(\Lambda,y)$ depends on the choice of representation of the oscillator algebra. In general, formulas involving irreducible representations with $I(\bm R,\bm k)=Z=$~const lead to mathematical inconsistencies typical of standard quantum field theory.
In order to define a mathematically precise representation one has to carefully define Dirac delta, and take a nontrivial $I(\bm R,\bm k)$.
This type of quantization is nonstandard but I believe it is exactly what should be done (to understand how it works for fields interacting with sources, see \cite{MCKW}).
\subsection{Reducible representations of harmonic-oscillator Lie algebra and transformation properties of field operators}
One begins with four operators, $a_0$, $a_1$, $a_2$, $a_3$,
satisfying commutation relations typical of an
{\it irreducible\/} representation of CCR:
$[a_{\bf a},a_{\bf b}^{\dag}]=\delta_{\bf ab}1$.
Let $|0\rangle$ denote their common vacuum,
i.e.~$a_{\bf a}|0\rangle=0$.
Now take kets $|\bm k\rangle$ and $|\bm R\rangle$, normalized with respect
to SL(2,C)-invariant M-shaped delta functions (see Appendix),
$
\langle\bm k|\bm k'\rangle
=
\delta_0(\bm k,\bm k')
$,
$\langle\bm R|\bm R'\rangle
=
\delta_1(\bm R,\bm R')$.
The reducible representation is parametrized by a natural number $N$. For $N=1$ the Hilbert space, denoted by ${\cal H}(1)$, is spanned by
kets of the form
\begin{eqnarray}
|\bm R,\bm k,n_0,n_1,n_2,n_3\rangle
=
|\bm R,\bm k\rangle\otimes
\frac{(a_0^{\dag})^{n_0}(a_1^{\dag})^{n_1}(a_2^{\dag})^{n_2}(a_3^{\dag})^{n_3}}
{\sqrt{n_0!n_1!n_2!n_3!}}|0\rangle,\nonumber
\end{eqnarray}
$|\bm R,\bm k\rangle=|\bm R\rangle\otimes|\bm k\rangle$.
The 1-oscillator representation is defined by
\begin{eqnarray}
a_{\bf a}(\bm R,\bm k,1) &=&|\bm R,\bm k\rangle\langle\bm R,\bm k|\otimes a_{\bf a},\\
n_{\bf a}(\bm R,\bm k,1) &=&|\bm R,\bm k\rangle\langle\bm R,\bm k|\otimes a_{\bf a}^{\dag}a_{\bf a},\label{n_a}\\
I(\bm R,\bm k,1) &=&|\bm R,\bm k\rangle\langle\bm R,\bm k|\otimes 1.
\end{eqnarray}
Operators
$I(\bm R,\bm k,1)$ form a resolution of unity
\begin{eqnarray}
\int d k{\,}d R{\,} I(\bm R,\bm k,1)
&=&
I\otimes I\otimes 1
=I(1),\label{I(1)-R}
\end{eqnarray}
Let us note that (\ref{n_a}) is constructed independently of $a_{\bf a}(\bm R,\bm k,1)$. The form $n_{\bf a}(\bm R,\bm k,1)=a_{\bf a}(\bm R,\bm k,1)^{\dag}a_{\bf a}(\bm R,\bm k,1)$ is possible if one works with M-shaped Dirac deltas `regular at zero', i.e. $\delta_0(\bm k,\bm k)=\delta_1(\bm R,\bm R)=1$ (see Appendix). Quantized analogs of the wave functions occurring in (\ref{int dR a}),
\begin{eqnarray}
a_{\bf a}(\bm k,1) &=&\int dR\,a_{\bf a}(\bm R,\bm k,1)=I\otimes|\bm k\rangle\langle\bm k|\otimes a_{\bf a},
\end{eqnarray}
are in this representation well defined.
Now, let
\begin{eqnarray}
P_a(1)
&=&
-I\otimes {\int} d k{\,}k_a|\bm k\rangle\langle \bm k|\otimes
a^{\dag}_{\bf a}a^{\bf a}
\label{P-R}\\
&=&
{\int} dk{\,}k_a
\big(
n_1(\bm k,1)
+
n_2(\bm k,1)
+
n_3(\bm k,1)
-
n_0(\bm k,1)
\big)
\label{Pn-R},\\
&=&
{\int} dk{\,}k_a
\big(
n_+(\bm k,1)
+
n_-(\bm k,1)
+
n_3(\bm k,1)
-
n_0(\bm k,1)
\big)
\label{Pn-R+-},\\
n_{\bf a}(\bm k,1) &=&
\int dR\,n_{\bf a}(\bm R,\bm k,1)=I\otimes|\bm k\rangle\langle\bm k|\otimes a_{\bf a}^{\dag}a_{\bf a},\quad {\bf a}=0,1,2,3,+,-,\\
a_\pm
&=&
\frac{1}{\sqrt{2}}(a_1\pm i a_2),\\
J_3
&=&
i(a_{1}^{\dagger}a_{2}-a_{2}^{\dagger}a_{1})=
a_{+}^{\dagger}a_{+}-a_{-}^{\dagger}a_{-}\label{L=J}.
\end{eqnarray}
The representation of inhomogeneous SL(2,C) then reads
\begin{eqnarray}
U(\Lambda,0,1) &=&\int dk{\,}d R{\,}|\bm R,\bm k\rangle\langle
\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k}|\otimes
e^{2i \Theta(\Lambda,\bm k) J_3},\label{U-R}\\
U(\bm 1,y,1) &=&e^{iy^aP_a(1)}.
\end{eqnarray}
One immediately checks that
\begin{eqnarray}
U(\Lambda,y,1)^{\dag}a(\pm,\bm{R},\bm{k},1)U(\Lambda,y,1)
&=&
e^{ip\cdot y}e^{\pm 2i\Theta(\Lambda,\bm k)}
a(\pm,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},1),\\
U(\Lambda,y,1)^{\dag}a_3(\bm{R},\bm{k},1)U(\Lambda,y,1)
&=&
e^{ip\cdot y}
a_3(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},1),\\
U(\Lambda,y,1)^{\dag}a_0(\bm{R},\bm{k},1)^{\dag}U(\Lambda,y,1)
&=&
e^{ip\cdot y}
a_0(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},1)^{\dag},\\
U(\Lambda,y,1)^{\dag}A_a(x,1)U(\Lambda,y,1)
&=&
\Lambda{_a}{^b}A_b\big(\Lambda^{-1}(x-y),1\big).
\end{eqnarray}
Moreover, splitting the 4-momentum into electromagnetic and scalar parts,
\begin{eqnarray}
P_a(1)
&=&
P^{\{12\}}_a(1)+P^{\{3\}}_a(1)-P^{\{0\}}_a(1),\\
P^{\{12\}}_a(1)
&=&
{\int} dk{\,}k_a
\big(
n_+(\bm k,1)
+
n_-(\bm k,1)
\big)
\label{P 12},\\
P^{\{3\}}_a(1)
&=&
{\int} dk{\,}k_a
n_3(\bm k,1)
\label{P 3},\\
P^{\{0\}}_a(1)
&=&
{\int} dk{\,}k_a
n_0(\bm k,1)
\label{P 0},
\end{eqnarray}
we note that the three types of 4-momenta are not mixed with one another by SL(2,C) transformations:
\begin{eqnarray}
U(\Lambda,y,1)^{\dag}P^{\{12\}}_a(1)U(\Lambda,y,1)
&=&
{\int} dk{\,}k_a
\big(
n_+(\bm{\Lambda^{-1}k},1)
+
n_-(\bm{\Lambda^{-1}k},1)
\big)\nonumber\\
&=&
\Lambda{_a}{^b}P^{\{12\}}_b(1),\\
U(\Lambda,y,1)^{\dag}P^{\{3\}}_a(1)U(\Lambda,y,1)
&=&
{\int} dk{\,}k_a
n_3(\bm{\Lambda^{-1}k},1)
=
\Lambda{_a}{^b}P^{\{3\}}_b(1),\\
U(\Lambda,y,1)^{\dag}P^{\{0\}}_a(1)U(\Lambda,y,1)
&=&
{\int} dk{\,}k_a
n_0(\bm{\Lambda^{-1}k},1)
=
\Lambda{_a}{^b}P^{\{0\}}_b(1).
\end{eqnarray}
For arbitrary $N$ the representation is constructed as follows.
Define
\begin{eqnarray}
{\cal H}(N)=\underbrace{{\cal H}(1)\otimes\dots\otimes {\cal H}(1)}_N={\cal H}(1)^{\otimes N}
\end{eqnarray}
and let $A$ be an arbitrary operator defined for $N=1$. Let
\begin{eqnarray}
A^{(n)}=\underbrace{I(1)\otimes\dots\otimes I(1)}_{n-1}\otimes A
\otimes \underbrace{I(1)\otimes\dots\otimes I(1)}_{N-n}.
\end{eqnarray}
The $N$ oscillator extension of $a_{\bf a}(\bm R,\bm k,1)$ is defined by
\begin{eqnarray}
a_{\bf a}(\bm R,\bm k,N)
&=&
\frac{1}{\sqrt{N}}\sum_{n=1}^N a_{\bf a}(\bm R,\bm k,1)^{(n)},\\
n_{\bf a}(\bm R,\bm k,N)
&=&
\sum_{n=1}^N n_{\bf a}(\bm R,\bm k,1)^{(n)},\\
I(\bm R,\bm k,N)
&=&
\frac{1}{N}\sum_{n=1}^N I(\bm R,\bm k,1)^{(n)},\\
U(\Lambda,y,N) &=& U(\Lambda,y,1)^{\otimes N}.
\end{eqnarray}
Note that for $N>1$ $n_{\bf a}(\bm R,\bm k,N)\neq a_{\bf a}(\bm R,\bm k,N)^{\dag}a_{\bf a}(\bm R,\bm k,N)$.
The 4-momentum describes $N$ noninteracting 4-dimensional oscillators of indefinite frequency
\begin{eqnarray}
P_a(N)
&=&
\sum_{n=1}^NP_a(1)^{(n)}\\
&=&{\int} dk{\,}k_a
\big(
n_+(\bm k,N)
+
n_-(\bm k,N)
+
n_3(\bm k,N)
-
n_0(\bm k,N)
\big).
\end{eqnarray}
The latter formula justifies the definition of $n_{\bf a}(\bm R,\bm k,N)$.
$I(\bm R,\bm k,N)$ again form a resolution of unity
\begin{eqnarray}
\int d k{\,}d R{\,} I(\bm R,\bm k,N)
&=&
I(N)=I(1)^{\otimes N}.\label{I(N)-R}
\end{eqnarray}
Of course,
\begin{eqnarray}
U(\Lambda,y,N)^{\dag}a(\pm,\bm{R},\bm{k},N)U(\Lambda,y,N)
&=&
e^{ip\cdot y}e^{\pm 2i\Theta(\Lambda,\bm k)}
a(\pm,\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},N),\\
U(\Lambda,y,N)^{\dag}a_3(\bm{R},\bm{k},N)U(\Lambda,y,N)
&=&
e^{ip\cdot y}
a_3(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},N),\\
U(\Lambda,y,N)^{\dag}a_0(\bm{R},\bm{k},N)^{\dag}U(\Lambda,y,N)
&=&
e^{ip\cdot y}
a_0(\bm{\Lambda^{-1}R},\bm{\Lambda^{-1}k},N)^{\dag},
\end{eqnarray}
and
\begin{eqnarray}
U(\Lambda,y,N)^{\dag}A_a(x,N)U(\Lambda,y,N)
&=&
\Lambda{_a}{^b}A_b\big(\Lambda^{-1}(x-y),N\big),
\end{eqnarray}
so the representation of inhomogeneous SL(2,C) again splits into spin-1 and spin-0 parts in the Fourier space, but at the level of Minkowski space
$A_a(x,N)$ is an ordinary world-vector field.
\subsection{Polarization operators for fields quantized in reducible representations of harmonic-oscillator Lie algebras}
Our goal is to define polarization operators that have transformation properties analogous to those based on Pauli-Lubanski vectors, but that are nevertheless linear in number-of-particles operators. In order to do so, let us note that for $N=1$ there exists an analog of first-quantized 4-momentum,
\begin{eqnarray}
K_a &=& I\otimes\int dk\,k_a |\bm k\rangle\langle \bm k|\otimes 1,\\
U(\Lambda,y,1)^{\dag}K_aU(\Lambda,y,1)
&=&
\Lambda{_a}{^b}K_b.
\end{eqnarray}
This operator naturally occurs if one starts with 4-momentum that includes vacuum energy (see Appendix). Indeed,
\begin{eqnarray}
P_a(1)+K_a
&=&
P_a(1)
+
I\otimes {\int} d k{\,}k_a|\bm k\rangle\langle \bm k|\otimes
\frac{1}{2}\Big(1+1+1-1\Big)\nonumber\\
&=&
-I\otimes {\int} d k{\,}k_a|\bm k\rangle\langle \bm k|\otimes
\frac{1}{2}\Big(a^{\dag}_{\bf a}a^{\bf a}+a^{\bf a}a^{\dag}_{\bf a}\Big)
.\nonumber
\end{eqnarray}
$K_a$ is in this representation well defined and commutes with elements of the oscillator Lie algebra. Generator
\begin{eqnarray}
M_{ab}(1)
&=&
i\frac{\partial U(\Lambda,0,1)}{\partial {}y^{ab}}\Big|_{{}y^{ab}=0}=L_{ab}(1)+S_{ab}(1)
\end{eqnarray}
splits into orbital and spin parts, where
\begin{eqnarray}
S_{ab}(1)
&=&
I\otimes
\int dk\,|\bm k\rangle\langle\bm{k}|\otimes
i\frac{\partial }{\partial {}y^{ab}}
e^{2i \Theta(\Lambda,\bm k) J_3}
\Big|_{{}y^{ab}=0}\nonumber\\
&=&
2I\otimes
\int dk\,|\bm k\rangle\langle\bm{k}|i\frac{\partial }{\partial {}y^{ab}}
{\bar\omega}{_{A'}}(\bm{k})\overline{\Lambda\pi}{^{A'}}(\bm{k})\Big|_{{}y^{ab}=0}\otimes J_3
\end{eqnarray}
and $\omega_A(\bm k)$ is any spin-partner of $\pi^A(\bm k)$.
Now consider
\begin{eqnarray}
W_a(1)
&=&
K^b(1){^*}S_{ab}(1)=
I\otimes
\int dk\,k_a|\bm k\rangle\langle\bm{k}|\otimes
J_3\\
&=&
\int dk\,k_a \Big(n_+(\bm k,1)-n_-(\bm k,1)\Big).
\end{eqnarray}
An extension to $N>1$, which is linear in $n_{\bf a}(\bm k,N)$, satisfies
\begin{eqnarray}
W_a(N)
&=&\sum_{n=1}^N W_a(1)^{(n)}=
\int dk\,k_a \Big(n_+(\bm k,N)-n_-(\bm k,N)\Big).
\end{eqnarray}
This type of extension is analogous to the relation between $P_a(1)$ and $P_a(N)$.
\subsection{Linear polarizations}
For $N=1$ let us define
\begin{eqnarray}
V_{\theta}(1)=I\otimes \int dk|\bm k\rangle\langle\bm k|\otimes e^{\theta(\bm k) (a_1^{\dag}a_2-a_2^{\dag}a_1)}
\end{eqnarray}
satisfying
\begin{eqnarray}
V_{\theta}(1)a_1(\bm k,1)V_{\theta}(1)^{\dag}
&=&
a_1(\bm k,1)\cos\theta(\bm k) - a_2(\bm k,1)\sin\theta(\bm k)=a_{\theta}(\bm k,1),\\
V_{\theta}(1)a_2(\bm k,1)V_{\theta}(1)^{\dag}
&=&
a_2(\bm k,1)\cos\theta(\bm k) + a_1(\bm k,1)\sin\theta(\bm k)=a_{\theta'}(\bm k,1).
\end{eqnarray}
For arbitrary $N\geq 1$ we define
\begin{eqnarray}
V_{\theta}(N)
&=&
V_{\theta}(1)^{\otimes N},\\
V_{\theta}(N)a_1(\bm k,N)V_{\theta}(N)^{\dag}
&=&
a_1(\bm k,N)\cos\theta(\bm k) - a_2(\bm k,N)\sin\theta(\bm k)=a_{\theta}(\bm k,N),\\
V_{\theta}(N)a_2(\bm k,N)V_{\theta}(N)^{\dag}
&=&
a_2(\bm k,N)\cos\theta(\bm k) + a_1(\bm k,N)\sin\theta(\bm k)=a_{\theta'}(\bm k,N),\\
V_{\theta}(N)a(\pm,\bm k,N)V_{\theta}(N)^{\dag}
&=&
e^{\pm i\theta(\bm k)}a(\pm,\bm k,N)=\frac{1}{\sqrt{2}}\Big(a_{\theta}(\bm k,N)\pm i\,a_{\theta'}(\bm k,N)\Big).
\end{eqnarray}
Number operators corresponding to linear polarizations are defined analogously
\begin{eqnarray}
V_{\theta}(N)n_1(\bm k,N)V_{\theta}(N)^{\dag}
&=&
n_{\theta}(\bm k,N),\\
V_{\theta}(N)n_2(\bm k,N)V_{\theta}(N)^{\dag}
&=&
n_{\theta'}(\bm k,N).
\end{eqnarray}
A yes-no observable associated with linear polarization can be defined as
\begin{eqnarray}
Y_{\theta}(\bm k,N)
&=&
n_{\theta}(\bm k,N)-n_{\theta'}(\bm k,N).\label{Y-N}
\end{eqnarray}
\section{Problem of relativistic analogues of EPR states of photons}
EPR states involve maximal entanglement in at least two different polarization bases (say, circular and linear). Such states can be easily invented also for photons, but the problem is with relativistic invariance of the EPR condition. At first glance the difficulty should not be a serious one since we know that the form (\ref{ve_A_B}) is SL(2,C) invariant: $\varepsilon_{AB}$ is a scalar representation of SL(2,C). However, for relativistic qubits the scalar-field condition has to be formulated at the level of unitary representation of SL(2,C), and it seems the problem is far from being fully systematized as yet. The basic question is if we can find scalar fields that involve maximal entanglement in two different polarization degrees of freedom, and in all reference frames.
\subsection{Vacuum and multiphoton states}
To begin with, let us note that vacuum is any pure state annihilated by all annihilation and number operators.
In reducible representations
\begin{eqnarray}
|0,N\rangle
&=&
|0,1\rangle^{\otimes N},\\
|0,1\rangle
&=&
\int dk\,dR\, O(\bm R,\bm k)|\bm R,\bm k\rangle \otimes|0\rangle,\quad \int dk\,dR\, |O(\bm R,\bm k)|^2=1.
\end{eqnarray}
The analysis given in \cite{MCKW} suggests that the most natural choice of vacuum wave function is the separable state
$O(\bm R,\bm k)=O_0(\bm k)O_1(\bm R)$. Vacuum states are in this representation Bose-Einstein condensates at zero temperature. As such they are not unique, but the whole subspace of vacuum states is a relativistically invariant subspace of the Hilbert space of states. In field quantization based on twistor-like spin-frames an inhomogeneous SL(2,C) transformation cannot produce particles, a fact which is not so obvious in general since Lorentz transformations in more standard approaches squeeze vacuum (a detailed discussion of this problem can be found in \cite{MCKW}).
It is important to understand that the parameter $N$ that characterizes a given reducible representation is unrelated to the number of photons. For example,
\begin{eqnarray}
n_{\bf a}(\bm R,\bm k,N)
\Big(a_{\bf a}(\bm R,\bm k,N)^{\dag}\Big)^j|0,N\rangle
&=&
j\Big(a_{\bf a}(\bm R,\bm k,N)^{\dag}\Big)^j|0,N\rangle.
\end{eqnarray}
Weak (i.e. performed at the level of averages) limits $N\to\infty$ reconstruct standard {\it regularized\/} formulas known from irreducible representations (for details see \cite{MCJN,MCKW,MWMC}).
\subsection{Scalar states}
The scalar-field condition means that
\begin{eqnarray}
\Psi(N)=\sum_{s,s'=\pm}\int dR\,dR'\,dk \,dk'\,
\Psi(s,\bm R,\bm k;s',\bm R',\bm k')
a(s,\bm R,\bm k,N)^{\dag}
a(s',\bm R',\bm k',N)^{\dag}\nonumber\\
\end{eqnarray}
satisfies
\begin{eqnarray}
U(\Lambda,0,N)^{\dag}\Psi(N)U(\Lambda,0,N)
=
\Psi(N).
\end{eqnarray}
Gauge independence implies that $\Psi(s,\bm R,\bm k;s',\bm R',\bm k')$ can be constructed only by means of spinor fields $\pi_A$ (and thus does not depend on $\bm R$ and $\bm R'$) so that
\begin{eqnarray}
\Psi(N) &=&
\sum_{s,s'=\pm}\int dR\,dR'\,dk \,dk'\,
\Psi(s,\bm k;s',\bm k')
a(s,\bm R,\bm k,N)^{\dag}
a(s',\bm R',\bm k',N)^{\dag}\nonumber\\
&=&
\sum_{s,s'=\pm}\int dk \,dk'\,
\Psi(s,\bm k;s',\bm k')
a(s,\bm k,N)^{\dag}
a(s',\bm k',N)^{\dag}\label{Psi bez R}.
\end{eqnarray}
Since
\begin{eqnarray}
&{}&
U(\Lambda,0,N)^{\dag}a(s,\bm k,N)^{\dag}a(s',\bm k',N)^{\dag}U(\Lambda,0,N)
\nonumber\\
&{}&\pp=
=
e^{-2is\Theta(\Lambda,\bm k)}e^{-2is'\Theta(\Lambda,\bm k')}
a(s,\bm{\Lambda^{-1}k},N)^{\dag}a(s',\bm{\Lambda^{-1}k'},N)^{\dag},
\nonumber
\end{eqnarray}
the problem reduces to finding $\Psi$ satisfying
\begin{eqnarray}
\Psi(s,\bm k;s',\bm k')
e^{-2is\Theta(\Lambda,\bm k)}e^{-2is'\Theta(\Lambda,\bm k')}
=
\Psi(s,\bm{\Lambda^{-1}k};s',\bm{\Lambda^{-1}k'}).
\label{Psi}
\end{eqnarray}
Recalling that
\begin{eqnarray}
e^{-i\Theta(\Lambda,\bm k)}\pi_A(\bm k)
&=&
\Lambda{_A}{^B}\pi_B(\bm{\Lambda}^{-1}\bm k)
,\\
e^{i\Theta(\Lambda,\bm k)}\bar\pi_{A'}(\bm k)
&=&
\Lambda{_{A'}}{^{B'}}\bar\pi_{B'}(\bm{\Lambda}^{-1}\bm k)
\end{eqnarray}
we observe that only two types of contractions have the required form (\ref{Psi}):
\begin{eqnarray}
\Big(\pi_A(\bm k)\pi^A(\bm k')\Big)^2e^{-2i\Theta(\Lambda,\bm k)}e^{-2i\Theta(\Lambda,\bm k')}
&=&
\Big(\pi_A(\bm{\Lambda}^{-1}\bm k)\pi^A(\bm{\Lambda}^{-1}\bm k')\Big)^2
,\\
\Big(\bar\pi_{A'}(\bm k)\bar\pi^{A'}(\bm k')\Big)^2e^{2i\Theta(\Lambda,\bm k)}e^{2i\Theta(\Lambda,\bm k')}
&=&
\Big(\bar\pi_{A'}(\bm{\Lambda}^{-1}\bm k)\bar\pi^{A'}(\bm{\Lambda}^{-1}\bm k')\Big)^2.
\end{eqnarray}
Since for two null vectors, $k_a=\pi_A(\bm k)\bar\pi_{A'}(\bm k)$,
$l_a=\pi_A(\bm l)\bar\pi_{A'}(\bm l)$, we find
\begin{eqnarray}
k\cdot l=k_al^a = \pi_A(\bm k)\pi^A(\bm l)\bar\pi_{A'}(\bm k)\bar\pi^{A'}(\bm l)=\big|\pi_A(\bm k)\pi^A(\bm l)\big|^2,
\end{eqnarray}
the expressions $\big(\bar\pi_{A'}(\bm k)\bar\pi^{A'}(\bm k')\big)^{-1}$ and $\pi_A(\bm k)\pi^A(\bm k')$ are proportional to each other, the proportionality factor $k\cdot k'$ being SL(2,C) invariant.
So the corresponding solution reads
\begin{eqnarray}
\Psi(N)
&=&
\int dk\,dk'\,
F_+\big(\pi_A(\bm k)\pi^A(\bm k')\big)
a(+,\bm k,N)^{\dag}a(+,\bm k',N)^{\dag}\nonumber\\
&+&
\int dk\, dk'\,
F_-\big(\bar\pi_{A'}(\bm k)\bar\pi^{A'}(\bm k')\big)
a(-,\bm k,N)^{\dag}a(-,\bm k',N)^{\dag},
\label{Psi1}
\end{eqnarray}
where $F_\pm$ are any functions satisfying $F_\pm(e^{i\phi} z)=e^{2i\phi} F_\pm(z)$.
In reducible representations the integrals in (\ref{Psi1}) are well defined.
Another solution of (\ref{Psi}) is $\Psi(\pm,\bm k;\pm,\bm k')=0$,
\begin{eqnarray}
\Psi(\pm,\bm k;\mp,\bm k')
&=&
c_\pm \delta_0(\bm k,\bm k'),\\
\Psi(N)
&=&
c\int dk \,a(+,\bm k,N)^{\dag}a(-,\bm k,N)^{\dag}.\label{Psi2}
\end{eqnarray}
States generated by both types of $\Psi(N)$ are maximally entangled in circular polarizations. Now take a closer look at linear-polarization entanglement of (\ref{Psi1}). Inserting
\begin{eqnarray}
a(\pm,\bm k,N)^{\dag}=\frac{1}{\sqrt{2}}e^{\pm i\theta(\bm k)}\Big(a_{\theta}(\bm k,N)^{\dag}\mp i\,a_{\theta'}(\bm k,N)^{\dag}\Big),
\end{eqnarray}
into (\ref{Psi1}),
\begin{eqnarray}
\Psi(N)
&=&
\int dk\,dk'\,
\frac{1}{2}
\Big(
F_+(\dots)
e^{i[\theta(\bm k)+\theta(\bm k')]}
+
F_-(\dots)
e^{-i[\theta(\bm k)+\theta(\bm k')]}
\Big)
\nonumber\\
&\pp=&\times
\Big(
a_{\theta}(\bm k,N)^{\dag}a_{\theta}(\bm k',N)^{\dag}
-
a_{\theta'}(\bm k,N)^{\dag}a_{\theta'}(\bm k',N)^{\dag}
\Big)
\nonumber\\
&+&
\int dk\,dk'\,
\frac{1}{2i}
\Big(
F_+(\dots)
e^{i[\theta(\bm k)+\theta(\bm k')]}
-
F_-(\dots)
e^{-i[\theta(\bm k)+\theta(\bm k')]}
\Big)
\nonumber\\
&\pp=&\times
\Big(
a_{\theta'}(\bm k,N)^{\dag}a_{\theta}(\bm k',N)^{\dag}
+
a_{\theta}(\bm k,N)^{\dag}a_{\theta'}(\bm k',N)^{\dag}
\Big),
\end{eqnarray}
we obtain a condition for maximally entangled linear polarizations,
\begin{eqnarray}
F_-\big(\bar\pi_{A'}(\bm k)\bar\pi^{A'}(\bm k')\big)
=
\pm
F_+\big(\pi_A(\bm k)\pi^A(\bm k')\big)e^{2i[\theta(\bm k)+\theta(\bm k')]}
.
\end{eqnarray}
Combining this with (\ref{Psi1}), we arrive at
\begin{eqnarray}
\Psi(N)
&=&
\int dk\,dk'\,
F_+\big(\pi_A(\bm k)\pi^A(\bm k')\big)
\nonumber\\
&\pp=&\times
\Big(a(+,\bm k,N)^{\dag}a(+,\bm k',N)^{\dag}\pm e^{2i[\theta(\bm k)+\theta(\bm k')]}a(-,\bm k,N)^{\dag}a(-,\bm k',N)^{\dag}\Big),
\nonumber\\
\label{Psi5}
\end{eqnarray}
which {\it does not\/} satisfy (\ref{Psi}). State $\Psi(N)|0,N\rangle$ is scalar but does not satisfy linear-polarization EPR condition. It seems there is no difficulty with extending the above argument to general elliptic polarizations.
The second solution (\ref{Psi2}) is more interesting. Indeed, (\ref{Psi2}) can be written as
\begin{eqnarray}
\Psi(N)
&=&
\frac{c}{2}
\int dk
\Big(
a_{\theta}(\bm k,N)^{\dag}a_{\theta}(\bm k,N)^{\dag}
+
a_{\theta'}(\bm k,N)^{\dag}a_{\theta'}(\bm k,N)^{\dag}
\Big),\label{Psi4}
\end{eqnarray}
which is simultaneously maximally entangled in all linear polarizations. Let us note that although (\ref{Psi4}) involves photons of the same momenta, this does not exclude applications to experiments with pairs of detectors localized in space in arbitrary locations.
\subsection{EPR states involving different momenta}
Let us again consider the operator (\ref{Psi bez R}) but with the kernel satisfying
\begin{eqnarray}
\Psi(s,\bm k;s',\bm k')=-\Psi(s',\bm k;s,\bm k')=-\Psi(s,\bm k';s',\bm k).\label{anti-s}
\end{eqnarray}
Denoting $\Psi(+,\bm k;-,\bm k')=\psi(\bm k,\bm k')/2=-\psi(\bm k',\bm k)/2$, we find
\begin{eqnarray}
\Psi(N) &=&
\int dk \,dk'\,
\psi(\bm k,\bm k')
a(+,\bm k,N)^{\dag}a(-,\bm k',N)^{\dag}
\label{Psi6}\\
&=&
(\alpha\delta-\beta\gamma) \int dk \,dk'\,
\psi(\bm k,\bm k')
b_1(\bm k,N)^{\dag}b_2(\bm k',N)^{\dag}
\nonumber
\end{eqnarray}
where
\begin{eqnarray}
\left(
\begin{array}{c}
b_1(\bm k,N)^{\dag}\\
b_2(\bm k,N)^{\dag}
\end{array}
\right)
&=&
\left(
\begin{array}{cc}
\alpha & \beta\\
\gamma & \delta
\end{array}
\right)^{-1}
\left(
\begin{array}{c}
a(+,\bm k,N)^{\dag}\\
a(-,\bm k,N)^{\dag}
\end{array}
\right)
,
\end{eqnarray}
and the matrix is unitary and independent of $\bm k$. In particular, if
\begin{eqnarray}
\left(
\begin{array}{cc}
\alpha & \beta\\
\gamma & \delta
\end{array}
\right)^{-1}
=
\frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
e^{-i\theta} & e^{i\theta}\\
ie^{-i\theta} & -ie^{i\theta}
\end{array}
\right)
\end{eqnarray}
the state $\Psi(N)|0,N\rangle$ is maximally entangled in both linear and circular polarizations.
One can replace the antisymmetric condition (\ref{anti-s}) by a symmetric one, but then we arrive at the state of the form (\ref{Psi5}).
Let us now consider the issue of SL(2,C) invariance of linear-polarization entanglement of (\ref{Psi6}). The transformed operator
\begin{eqnarray}
&{}&
U(\Lambda,0,N)^{\dag}\Psi(N)U(\Lambda,0,N)
\nonumber\\
&{}&\pp==
\int dk \,dk'\,
\psi(\bm{\Lambda k},\bm{\Lambda k'})
e^{-2i\Theta(\Lambda,\bm{\Lambda k})}a(+,\bm{k},N)^{\dag}
e^{2i\Theta(\Lambda,\bm{\Lambda k'})}a(-,\bm{k'},N)^{\dag}
\label{Psi8}
\end{eqnarray}
is maximally entangled in all linear polarizations defined by operators
\begin{eqnarray}
\frac{1}{\sqrt{2}}
\Big(e^{-i\theta(\bm k)}a(+,\bm{k},N)^{\dag}+e^{i\theta(\bm k)}a(-,\bm{k},N)^{\dag}\Big)
,\\
\frac{i}{\sqrt{2}}
\Big(e^{-i\theta(\bm k)}a(+,\bm{k},N)^{\dag}-e^{i\theta(\bm k)}a(-,\bm{k},N)^{\dag}\Big),
\end{eqnarray}
where
$\theta(\bm k)=\theta+2\Theta(\Lambda,\bm{\Lambda k})$. In general, components corresponding to different wave vectors get rotated by different angles.
Putting it differently, the antisymmetry of $\psi(\bm{k},\bm{k'})$ is not conserved by SL(2,C) transformations since
\begin{eqnarray}
\psi(\bm{\Lambda k},\bm{\Lambda k'})
e^{-2i\Theta(\Lambda,\bm{\Lambda k})}
e^{2i\Theta(\Lambda,\bm{\Lambda k'})}
\end{eqnarray}
is not antisymmetric in $\bm k$ and $\bm k'$, so that the state is no longer maximally entangled in the original linear polarizations.
The latter effect is essentially the massless version of the Peres-Scudo-Terno phenomenon. In the massless case the polarization observables are not defined with respect to projections of the Pauli-Lubanski vector, and thus the argument based on $\omega$-spinors cannot be directly employed. On the other hand, Wigner phases depend only on directions of $\bm k$, so all parallel wave vectors correspond to the same rotation angle. In order to maintain maximal entanglement in EPR-type experiments, one has to employ momentum-dependent polarization operators that compensate the presence of the Wigner angle $2\Theta(\Lambda,\bm{\Lambda k})$.
\subsection{Normalization of 2-photon states in reducible representation}
Let us pause here for a moment to discuss the issue of normalization of 2-photon states.
As an exercise consider the simple scalar case $F_+(z)=z^2$, $F_-(z)=0$:
\begin{eqnarray}
&{}&\int dk\,dk'\,dl\,dl' \Big(\bar\pi_{A'}(\bm l)\bar\pi^{A'}(\bm l')\Big)^2 \Big(\pi_A(\bm k)\pi^A(\bm k')\Big)^2
\nonumber\\
&{}&\pp{\int dk\,dk'\,dl\,dl'}\times
\langle 0,N|a(+,\bm l,N)a(+,\bm l',N)a(+,\bm k,N)^{\dag}a(+,\bm k',N)^{\dag}|0,N\rangle\nonumber\\
&{}&\pp=
=
\int dk\,dk'\,dl\,dl' \Big(\bar\pi_{A'}(\bm l)\bar\pi^{A'}(\bm l')\Big)^2 \Big(\pi_A(\bm k)\pi^A(\bm k')\Big)^2
\nonumber\\
&{}&\pp{\int dk\,dk'\,dl\,dl'}\times
\Big(\delta_0(\bm k,\bm l')\delta_0(\bm k',\bm l)+\delta_0(\bm k',\bm l')\delta_0(\bm k,\bm l)\Big)
\langle 0,N|I(\bm k,N)I(\bm k',N)|0,N\rangle\nonumber\\
&{}&\pp=
=
2\int dk\,dk'\,(k\cdot k')^2 \langle 0,N|I(\bm k,N)I(\bm k',N)|0,N\rangle\nonumber\\
&{}&\pp=
=
2\Big(1-\frac{1}{N}\Big)\int dk\,dk'\,(k\cdot k')^2 |O_0(\bm k)|^2|O_0(\bm k')|^2. \nonumber
\end{eqnarray}
Let us note that in irreducible representations, where $I(\bm k)=Z\bm 1$, $Z=$~const, an analogous calculation would involve the divergent integral
\begin{eqnarray}
2\int dk\,dk'\,(k\cdot k')^2 \langle 0|I(\bm k)I(\bm k')|0\rangle
=
2Z^2\int dk\,dk'\,(k\cdot k')^2.
\end{eqnarray}
Invariance of the norm under inhomogeneous SL(2,C) is a consequence of the fact that the vacuum wave function is a scalar field:
$O_0(\bm k)\to O_0(\bm{\Lambda}^{-1}\bm k)$ is equivalent to $|0,N\rangle\to U(\Lambda,y,N)|0,N\rangle$. (Inclusion of vacuum 4-momentum $K_a$ would imply $O_0(\bm k)\to e^{ik\cdot y}O_0(\bm{\Lambda}^{-1}\bm k)$.)
Performing an analogous calculation for (\ref{Psi6}), we obtain
\begin{eqnarray}
\langle 0,N|\Psi(N)^{\dag}\Psi(N)|0,N\rangle
=
\Big(1-\frac{1}{N}\Big)
\int dk\,dk'\,
|\psi(\bm k,\bm k')|^2
|O(\bm k)|^2 |O(\bm k')|^2.
\end{eqnarray}
In reducible representations the state $\Psi(N)|0,N\rangle$, generated by (\ref{Psi2}), is normalizable as well. Since this exercise is also instructive, let us explicitly compute the norm for $N=1$ in two alternative ways (put $c=1$ for simplicity). The first approach explicitly employs the form of the vacuum state and normalization of kets to M-shaped deltas,
\begin{eqnarray}
\Psi(1)|0,1\rangle
&=&
\int dk \,a(+,\bm k,1)^{\dag}a(-,\bm k,1)^{\dag}\int dk'\, dR'\, O_1(\bm R')O_0(\bm k')|\bm R',\bm k',0,0,0,0\rangle\nonumber\\
&=&
\int dk \,dR\,|\bm R,\bm k\rangle\langle\bm R,\bm k|\otimes a_+^{\dag}a_-^{\dag}\int dk'\, dR'\, O_1(\bm R')O_0(\bm k')|\bm R',\bm k',0,0,0,0\rangle\nonumber\\
&=&
\int dk\, dR\, O_1(\bm R)O_0(\bm k)|\bm R,\bm k\rangle\otimes a_+^{\dag}a_-^{\dag}|0\rangle\nonumber,
\end{eqnarray}
so that $\langle 0,1|\Psi(1)^{\dag}\Psi(1)|0,1\rangle=\int dR\, |O_1(\bm R)|^2\int dk\,|O_0(\bm k)|^2= 1$. The second way is based on properties of the M-shaped deltas (see Appendix \ref{subsec 75}), and the fact that annihilation operators annihilate vacuum:
\begin{eqnarray}
{}&{}&
\langle 0,1|\Psi(1)^{\dag}\Psi(1)|0,1\rangle
\nonumber\\
&{}&\pp==
\int dk \int dk'
\langle 0,1|a(+,\bm k,1)a(-,\bm k,1)a(+,\bm k',1)^{\dag}a(-,\bm k',1)^{\dag}|0,1\rangle\nonumber\\
&{}&\pp=
=
\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty} \lim_{n_4\to\infty}
\int dk \int dk'
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}}
\times
\langle 0,1|a(-,\bm k,1,\textstyle{\frac{1}{n_1}})a(+,\bm k,1,\textstyle{\frac{1}{n_2}})a(+,\bm k',1,\textstyle{\frac{1}{n_3}})^{\dag}a(-,\bm k',1,\textstyle{\frac{1}{n_4}})^{\dag}|0,1\rangle\nonumber\\
&{}&\pp=
=
\lim_{n_1\to\infty}\lim_{n_4\to\infty}
\int dk \int dk'
\langle 0,1|a(-,\bm k,1,\textstyle{\frac{1}{n_1}})\delta_0(\bm k,\bm k')I(\bm k,1)a(-,\bm k',1,\textstyle{\frac{1}{n_4}})^{\dag}|0,1\rangle\nonumber\\
&{}&\pp=
=
\int dk
\langle 0,1|I(\bm k,1)^2|0,1\rangle
=
\langle 0,1|\int dk\,I(\bm k,1)|0,1\rangle=\langle 0,1|0,1\rangle=1.
\nonumber
\end{eqnarray}
One analogously checks that, for arbitrary $N\geq 1$ and $c=1$, (\ref{Psi2}) implies
\begin{eqnarray}
\langle 0,N|\Psi(N)^{\dag}\Psi(N)|0,N\rangle
&=&
\frac{1}{N^2}+\Big(1-\frac{1}{N}\Big)\int dk\,|O_0(\bm k)|^4.
\end{eqnarray}
\section{EPR averages for linearly polarized photons --- the reducible representation approach}
Let us turn to the question of EPR averages for quantum electromagnetic fields quantized in reducible representations of harmonic oscillator Lie algebras. The issue is important since all the experiments performed so far were analyzed in terms of {\it irreducible\/} representations, so one might be tempted to conclude that the standard approach to quantization is supported by EPR-type predictions.
Let us consider the EPR state (\ref{Psi6}). Yes-no observables are defined by (\ref{Y-N}). The EPR average
\begin{eqnarray}
&{}&
\frac{\langle 0,N|\Psi(N)^{\dag}Y_{\beta}(\bm l',N)Y_{\alpha}(\bm l,N)\Psi(N)|0,N\rangle}
{\langle 0,N|\Psi(N)^{\dag}\Psi(N)|0,N\rangle}
\nonumber\\
&{}&\pp=
=
2\cos 2(\alpha-\beta)\frac{\delta_0(\bm l',\bm l)
|O_0(\bm l)|^2
\int dk\,
|\psi(\bm l,\bm k)|^2
|O_0(\bm k)|^2
-
|\psi(\bm l,\bm l')|^2
|O_0(\bm l)|^2 |O_0(\bm l')|^2}
{\int dk\,dk'\,
|\psi(\bm k,\bm k')|^2
|O_0(\bm k)|^2 |O_0(\bm k')|^2}
\nonumber\\
\end{eqnarray}
involves sharp wave vectors $\bm l$, $\bm l'$, and is independent of $N$ if $N>1$ ($\Psi(N)|0,N\rangle=0$ for $N=1$). Since localization of detectors leads to momentum spread, $\bm l\in \Omega$, $\bm l'\in \Omega'$, say, yes-no operators integrated over both sets
\begin{eqnarray}
Y_{\alpha}(N)
&=&
\int_\Omega dl\,Y_{\alpha}(\bm l,N),\\
Y'_{\beta}(N)
&=&
\int_{\Omega'} dl'\,Y_{\beta}(\bm l',N),
\end{eqnarray}
can be used to compute more realistic cases. For disjoint detectors, $\Omega \cap \Omega'=\phi$,
\begin{eqnarray}
\frac{\langle 0,N|\Psi(N)^{\dag}Y'_{\beta}(N)Y_{\alpha}(N)\Psi(N)|0,N\rangle}
{\langle 0,N|\Psi(N)^{\dag}\Psi(N)|0,N\rangle}
&=&
-
\cos 2(\alpha-\beta)p_{\Omega\times\Omega'},
\label{Y'Y}
\end{eqnarray}
where
\begin{eqnarray}
p_{\Omega\times\Omega'}
&=&
p_{\Omega'\times\Omega}\nonumber\\
&=&
2\frac{
\int_\Omega dl\,\int_{\Omega'} dl'\,|\psi(\bm l,\bm l')|^2
|O_0(\bm l)|^2 |O_0(\bm l')|^2
}
{\int_{\bm R^3} dk \int_{\bm R^3}dk'\,
|\psi(\bm k,\bm k')|^2
|O_0(\bm k)|^2 |O_0(\bm k')|^2}
\nonumber\\
&=&
2\frac{
\int_\Omega dl\,\int_{\Omega'} dl'\,|\psi(\bm l,\bm l')|^2
\chi(\bm l)\chi(\bm l')
}
{\int_{\bm R^3} dk \int_{\bm R^3}dk'\,
|\psi(\bm k,\bm k')|^2
\chi(\bm k)\chi(\bm k')}.\label{p red}
\end{eqnarray}
The form (\ref{p red}) employs the cutoff function $\chi(\bm k)=|O(\bm k)|^2/Z$, $Z=\max_{\bm k}|O(\bm k)|^2$, $0\leq \chi(\bm k)\leq 1$, whose appearance is typical of predictions based on the reducible representation.
For identical detectors, $\Omega =\Omega'$,
\begin{eqnarray}
\frac{\langle 0,N|\Psi(N)^{\dag}Y'_{\beta}(N)Y_{\alpha}(N)\Psi(N)|0,N\rangle}
{\langle 0,N|\Psi(N)^{\dag}\Psi(N)|0,N\rangle}
&=&
\cos 2(\alpha-\beta)p_{\Omega\times(\bm R^3-\Omega)}.
\label{YY}
\end{eqnarray}
The average (\ref{YY}) vanishes if $\Omega=\Omega'=\bm R^3$.
In the most general case, with arbitrary overlap $\Omega_0=\Omega\cap\Omega'$, $\Omega=\Omega_1\cup\Omega_0$, $\Omega'=\Omega_0\cup\Omega_1'$, one finds
\begin{eqnarray}
-
\cos 2(\alpha-\beta)
\Big(
p_{\Omega_1\times\Omega'_1}
+
p_{\Omega_1\times\Omega_0}
+
p_{\Omega_0\times\Omega'_1}
-
p_{\Omega_0\times(\bm R^3-\Omega_0)}
\Big).\label{overlap}
\end{eqnarray}
Let us note that Bell's inequality can be violated only if
\begin{eqnarray}
p_{\Omega_1\times\Omega'_1}
+
p_{\Omega_1\times\Omega_0}
+
p_{\Omega_0\times\Omega'_1}
-
p_{\Omega_0\times(\bm R^3-\Omega_0)}>\frac{1}{\sqrt{2}},
\end{eqnarray}
For example, let
\begin{eqnarray}
\psi(\bm k,\bm k')=f(\bm k)g(\bm k')-f(\bm k')g(\bm k),
\end{eqnarray}
where supports $X_f={\rm supp}(f)$ and $X_g={\rm supp}(g)$ are disjoint (say, photons of two different colors are emitted, like in parametric down-conversion experiments). If $X_f\subset \Omega$, $X_g\subset \Omega'$,
$\Omega\cap\Omega'=\phi$, then
\begin{eqnarray}
p_{\Omega\times\Omega'}
&=&
\frac{
2\int_\Omega dl\,\int_{\Omega'} dl'\,|f(\bm l)g(\bm l')|^2\chi(\bm l) \chi(\bm l')
+
2\int_\Omega dl\,\int_{\Omega'} dl'\,|f(\bm l')g(\bm l)|^2\chi(\bm l) \chi(\bm l')
}
{
\int_\Omega dk\,\int_{\Omega'} dk'\,|f(\bm k)g(\bm k')|^2\chi(\bm k) \chi(\bm k')
+
\int_{\Omega'} dk\,\int_{\Omega} dk'\,|f(\bm k')g(\bm k)|^2\chi(\bm k') \chi(\bm k)}
\nonumber\\
&=&
p_{\Omega_1\times\Omega'_1}
+
p_{\Omega_1\times\Omega_0}
+
p_{\Omega_0\times\Omega'_1}
-
p_{\Omega_0\times(\bm R^3-\Omega_0)}
=
1,
\end{eqnarray}
so that the Bell inequality can be violated. If $X_f\cup X_g\subset \Omega=\Omega'$
\begin{eqnarray}
\int_\Omega dl\,\int_{\bm R^3-\Omega} dl'\,|f(\bm l)g(\bm l')|^2\chi(\bm l) \chi(\bm l')
+
\int_\Omega dl\,\int_{\bm R^3-\Omega} dl'\,|f(\bm l')g(\bm l)|^2\chi(\bm l) \chi(\bm l')
=0,\nonumber
\end{eqnarray}
and no violation is found.
In irreducible representations, where $I(\bm k)=Z\bm 1$, we get identical formulas but with
\begin{eqnarray}
p_{\Omega\times\Omega'}
&=&
2\frac{
\int_\Omega dl\,\int_{\Omega'} dl'\,|\psi(\bm l,\bm l')|^2
}
{\int_{\bm R^3} dk \int_{\bm R^3}dk'\,
|\psi(\bm k,\bm k')|^2 \label{p irred}
}.
\end{eqnarray}
In practice, probabilities (\ref{p red}) and (\ref{p irred}) are not easy to distinguish from one another if $\chi(\bm k)$ is sufficiently flat in the optical-range set of wave vectors. The EPR averages thus do not provide us with any practical clue about the choice of representation of harmonic oscillator Lie algebras appropriate for field quantization.
Relativistic properties of EPR averages cannot be directly inferred from (\ref{Y'Y}) but one has to take into account the remarks made after (\ref{Psi8}) and repeat the whole calculation. I will not pursue the matter further here.
\section{Remarks on two related issues}
One can find in the literature a discussion of `entanglement with vacuum' (where a single-photon state is entangled) and violation of Bell's inequality in vacuum (where the vacuum state is entangled). In light of what I have written above on EPR states an existence of such phenomena may seem weird, but the problem is purely semantic --- different authors have different things in mind when they speak of `vacuum'.
\subsection{Entanglement with vacuum}
The idea of entanglement with vacuum comes from a specific representation of the Lie algebra
\begin{eqnarray}
[a_m,a_n^{\dag}]=\delta_{nm}\bm 1, \label{stand}
\end{eqnarray}
where $m$, $n$ are integers, and
\begin{eqnarray}
a_n=\dots 1\otimes 1\otimes a\otimes 1\otimes 1\dots
\end{eqnarray}
with $a$, $[a,a^{\dag}]=1$, located on an ``$n$th position". $a_n$ acts in the non-separable Hilbert space spanned by infinite tensor products of number sates $|n\rangle$, $a^{\dag}a|n\rangle=n|n\rangle$. Infinite tensor products make the set of basis vectors uncountable (`Orlov states' \cite{Orlov}, such as $\dots |0\rangle\otimes |3\rangle\otimes |1\rangle\otimes |4\rangle\otimes|1\rangle\otimes|5\rangle\otimes|9\rangle\otimes \dots$, with numbers taken from the digits of $\pi$, are indexed by real numbers).
The identity at the right side of (\ref{stand}) is given by the infinite tensor product of identities
\begin{eqnarray}
\bm 1 &=& \dots 1\otimes 1\otimes 1\dots
\end{eqnarray}
while the vacuum is represented by
\begin{eqnarray}
|\bm 0\rangle &=& \dots |0\rangle\otimes |0\rangle\otimes |0\rangle\dots
\end{eqnarray}
A single photon state that exists in a superposition of two different modes, say,
\begin{eqnarray}
\big(a_n^{\dag}+a_m^{\dag}\big)|\bm 0\rangle
&=&
\dots |0\rangle\otimes |1\rangle\otimes \dots \otimes |0\rangle \otimes |0\rangle\dots\nonumber\\
&\pp=&+
\dots |0\rangle\otimes |0\rangle\otimes \dots \otimes |0\rangle \otimes |1\rangle\dots
\nonumber
\end{eqnarray}
resembles, up to all the problems with infinite tensor products, the 2-particle entangled state
\begin{eqnarray}
|1\rangle\otimes |0\rangle
+
|0\rangle\otimes |1\rangle.
\end{eqnarray}
This representation is often treated (especially in the quantum optics literature) as {\it the\/} representation of the harmonic oscillator Lie algebra typical of quantum fields.
We have seen, however, that EPR correlations of photons can be computed at a much more general level and do not rely on specific tensor product structures of Hilbert spaces in question. The reducible representations we have worked with allow us to speak of all the field modes, but involve only $N$th tensor powers. Attempts of interpreting, say,
\begin{eqnarray}
\Big(a(+,\bm k,N)^{\dag}+a(-,\bm k,N)^{\dag}\Big)|0,N\rangle
\end{eqnarray}
as an EPR state do not make much sense. Moreover, one does not have to resort to reducible representations to show that a single-photon 2-mode state may not always be interpretable in terms of entangled states. It is sufficient (see \cite{MPMC}) to take the original (1932) representation of the Fock space \cite{Fock,Berezin}.
\subsection{Nonlocal properties of vacuum states in algebraic quantum field theory}
Summers and Werner \cite{SW0} showed that vacuum states, equipped with all their properties assumed in axiomatic quantum field theory, are enough to violate the Bell inequality. The paper is simultaneously the first published account of relativistic Bell theorem (see, however, \cite{MC84}). The results from \cite{SW0} were subsequently generalized in a number of works (cf. \cite{Summers} for a recent review).
In order to grasp the main idea let us consider a simpler but physically similar problem suggested in \cite{Redhead}. Consider two systems, ${\cal O}_1$, ${\cal O}_2$, equipped with certain algebras ${\cal A}({\cal O}_1)$, ${\cal A}({\cal O}_2)$. We assume commutativity $[A_1,A_2]=0$ if $A_1\in {\cal A}({\cal O}_1)$ and $A_2\in {\cal A}({\cal O}_2)$.
The algebraic structure can be represented in a Hilbert space ${\cal H}={\cal H}_1\otimes {\cal H}_2$ where representatives of $A_1$ and $A_2$ are of the form $A\otimes 1$ and $1\otimes B$, respectively. The crucial assumption about a vacuum state $|\Omega\rangle$ is its cyclicity with respect to both ${\cal A}({\cal O}_1)$ and ${\cal A}({\cal O}_2)$. What it means is that acting on $|\Omega\rangle$ with operators of {\it either\/} ${\cal A}({\cal O}_1)$ or ${\cal A}({\cal O}_2)$ one can generate any vector in ${\cal H}$.
To make our analysis as concrete and simple as possible, assume that ${\cal H}_1$ and ${\cal H}_2$ are 2-dimensional. If vacuum $|\Omega\rangle$ is cyclic in the above sense, then {\it all\/} basis vectors of ${\cal H}$ can be written as $(A\otimes 1)|\Omega\rangle$. Let $|0_j\rangle$, $|1_j\rangle$, span ${\cal H}_j$. It is obvious that $|\Omega\rangle$ cannot be a product state. However, any state from the Bell basis
\begin{eqnarray}
|\Psi_\pm\rangle
&=&
\frac{1}{\sqrt{2}}\Big(|0_1\rangle\otimes |1_2\rangle\pm |1_1\rangle\otimes |0_2\rangle\Big),\\
|\Phi_\pm\rangle
&=&
\frac{1}{\sqrt{2}}\Big(|0_1\rangle\otimes |0_2\rangle\pm |1_1\rangle\otimes |1_2\rangle\Big),
\end{eqnarray}
can play the role of $|\Omega\rangle$. Indeed, take $A=|0_1\rangle\langle 0_1|-|1_1\rangle\langle 1_1|$,
$B=|0_1\rangle\langle 1_1|+|1_1\rangle\langle 0_1|$, and $|\Omega\rangle=|\Psi_+\rangle$. Then
\begin{eqnarray}
(1\otimes 1)|\Omega\rangle &=& |\Psi_+\rangle,\\
(A\otimes 1)|\Omega\rangle &=& |\Psi_-\rangle,\\
(B\otimes 1)|\Omega\rangle &=& |\Phi_+\rangle,\\
(AB\otimes 1)|\Omega\rangle &=& |\Phi_-\rangle.
\end{eqnarray}
The same effect is obtained if one acts on the second qubit.
Moreover, we can replace $|\Psi_+\rangle$ by any entangled state, say, a ground state of some 2-qubit Hamiltonian. $|\Omega\rangle$ is then the lowest energy state which is cyclic with respect to commuting von Neumann algebras ${\cal A}({\cal O}_j)$. So this is precisely the vacuum in the sense of algebraic quantum field theory, but for a trivial toy model. The fact that `vacuum' can maximally violate the Bell inequality is no longer weird.
Our vacuum state $|0,N\rangle$ is different. It belongs to a Poincar\'e invariant subspace which is uniquely defined, but a single $|0,N\rangle$ is neither unique nor relativistically invariant. $|0,N\rangle$ is not cyclic either --- it does not satisfy the axioms employed by Summers and Werner. In spite of that, the formalism leads to a well defined field theory, hopefully with no divergences, which is surprisingly close to the standard one in all the applications considered so far, simultaneously leading to new effects, testable at least in principle \cite{MCKW,MWMC}.
\section*{Appendices}
\section{Dirac delta regular at zero}
\subsection{M-shaped delta-sequences}
Let us consider the function shown in the upper part of Fig.~1. It is a particular example, for $a=1$ and $\epsilon=1/2$, of
\begin{eqnarray}
\delta(k,a,\epsilon)
&=&
\left\{
\begin{array}{crc}
0 & \textrm{for} & k < -\frac{\epsilon}{2} \\
\big(\frac{4k}{\epsilon} +2\big)\big(\frac{2}{\epsilon} - \frac{a}{2}\big) & \textrm{for} & -\frac{\epsilon}{2} \leq k < -\frac{\epsilon}{4}\\
-\frac{4k}{\epsilon}\big(\frac{2}{\epsilon} - \frac{3a}{2}\big) + a & \textrm{for} & -\frac{\epsilon}{4} \leq k < 0\\
\frac{4k}{\epsilon}\big(\frac{2}{\epsilon} - \frac{3a}{2}\big) + a & \textrm{for} & 0\leq k<\frac{\epsilon}{4}\\
\big(-\frac{4k}{\epsilon} +2\big)\big(\frac{2}{\epsilon} - \frac{a}{2}\big) & \textrm{for} & \frac{\epsilon}{4} \leq k < \frac{\epsilon}{2}\\
0 & \textrm{for} & \frac{\epsilon}{2}\leq k,
\end{array}\label{1}
\right.\nonumber
\\
\end{eqnarray}
($a>0$, $\epsilon>0$).
\begin{figure}
\includegraphics[width=8cm]{M-delta.eps}
\caption{The M-shaped function (\ref{1}) with $a=1$, $\epsilon=1/2$ (upper), and its Fourier transform (lower).}
\end{figure}
Its Fourier transform,
\begin{eqnarray}
\hat\delta(x,a,\epsilon)
&=&
\frac{1}{2\pi}
\int_{-\infty}^\infty \delta(k,a,\epsilon) e^{ikx}dk\nonumber\\
&=&
\frac{8}{\pi}
\frac{\epsilon a + (4 - \epsilon a) \cos\frac{\epsilon x}{4}}{\epsilon^2 x^2}\sin^2\frac{\epsilon x}{8},
\\
\lim_{\epsilon\to 0}\hat\delta(x,a,\epsilon) &=& \frac{1}{2\pi},
\end{eqnarray}
is a real function shown in the lower part of Fig.~1 . The sequence $\delta(k,a,\textstyle{\frac{1}{n}})$, with natural $n$ (i.e. $\epsilon=1/n$), is an example of what I call an M-shaped delta-sequence, and is in fact a particular example of a delta-sequence in the sense of \cite{Sikorski}.
Indeed,
\begin{eqnarray}
\int_{-\infty}^\infty \delta(k,a,\textstyle{\frac{1}{n}})dk &=& 1,
\end{eqnarray}
and for a function $f$ left- and right-continuous at 0
\begin{eqnarray}
\lim_{n\to\infty}\int_{-\infty}^\infty f(k)\delta(k,a,\textstyle{\frac{1}{n}})dk &=&\frac{f(0_-)+f(0_+)}{2}.
\end{eqnarray}
A peculiarity of M-shaped delta-sequences is their regularity at 0,
\begin{eqnarray}
\delta(0,a,\textstyle{\frac{1}{n}}) &=& a,
\end{eqnarray}
for all $n$, so that
\begin{eqnarray}
\lim_{n\to\infty}\delta(0,a,\textstyle{\frac{1}{n}}) &=& a.
\end{eqnarray}
The fact that delta-sequences do not have to be divergent at the origin is, perhaps, not widely known but in itself is not new (an example of an analogous `filtering function', vanishing at the origin, can be found in \cite{Pol}).
In what follows we restrict our analysis to M-shaped delta-sequences normalized by $a=1/(2\pi)$,
\begin{eqnarray}
\delta_M(k,\textstyle{\frac{1}{n}})=\delta(k,\textstyle{\frac{1}{2\pi}},\textstyle{\frac{1}{n}}).
\end{eqnarray}
Let us note that, in spite of regularity at 0, one finds
\begin{eqnarray}
\lim_{n\to\infty}\int_{-\infty}^\infty \delta_M(k,\textstyle{\frac{1}{n}})\delta_M(k,\textstyle{\frac{1}{n}})dk
&=&\infty.
\end{eqnarray}
The square of Dirac's delta thus will not exist even if we define delta in terms of M-shaped delta-sequences. However,
\begin{eqnarray}
\lim_{m\to\infty}\lim_{n\to\infty}\int_{-\infty}^\infty \delta_M(k,\textstyle{\frac{1}{m}})\delta_M(k,\textstyle{\frac{1}{n}})dk
&=&
\lim_{m\to\infty}\delta_M(0,\textstyle{\frac{1}{m}})
=\frac{1}{2\pi}.
\end{eqnarray}
\subsection{Plane waves and M-shaped delta-sequences}
We are heading towards an analysis of plane waves in terms of deltas that are regular at 0. To do so, we need delta-sequences that can be represented as scalar products of square-integrable functions. The simplest strategy is to start with the convolution of two M-shaped delta-sequences,
\begin{eqnarray}
\delta_M^*(k,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
\int_{-\infty}^\infty\delta_M(k-k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'
=
\delta_M^*(k,\textstyle{\frac{1}{m}},\textstyle{\frac{1}{n}}),\\
\lim_{m\to\infty}
\delta_M^*(k,\textstyle{\frac{1}{n},\frac{1}{m}})
&=&
\lim_{m\to\infty}
\int_{-\infty}^\infty\delta_M(k-k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'
=
\delta_M(k,\textstyle{\frac{1}{n}}).
\end{eqnarray}
The new sequence is again a delta-sequence,
\begin{eqnarray}
\int_{-\infty}^\infty\delta_M^*(k,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})dk
&=&
\int_{-\infty}^\infty\int_{-\infty}^\infty\delta_M(k-k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'dk
\nonumber\\
&=&\int_{-\infty}^\infty\delta_M(k',\textstyle{\frac{1}{m}})dk'
=1,
\end{eqnarray}
\begin{eqnarray}
{}&{}&
\lim_{n\to\infty}\lim_{m\to\infty}\int_{-\infty}^\infty f(k)\delta_M^*(k,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})dk
\nonumber\\
&{}&\pp{\lim_{n\to\infty}\lim_{m\to\infty}}=
\lim_{n\to\infty}\lim_{m\to\infty}\int_{-\infty}^\infty\int_{-\infty}^\infty f(k)\delta_M(k-k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'dk
\nonumber\\
&{}&\pp{\lim_{n\to\infty}\lim_{m\to\infty}}=
\lim_{n\to\infty}\int_{-\infty}^\infty f(k)\delta_M(k,\textstyle{\frac{1}{n}})dk
=
\displaystyle{\frac{f(0_-)+f(0_+)}{2},}
\end{eqnarray}
but is not exactly M-shaped in the sense of the previous subsection. Indeed,
\begin{eqnarray}
\delta_M^*(0,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
\int_{-\infty}^\infty\delta_M(0-k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'
\nonumber\\
&=&
\int_{-\infty}^\infty\delta_M(k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{m}})dk'
\end{eqnarray}
in general depends on $n$ and $m$. The other properties are nevertheless analogous to M-shaped delta-sequences,
\begin{eqnarray}
\lim_{m\to\infty}
\delta_M^*(0,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
\delta_M(0,\textstyle{\frac{1}{n}})=1/2\pi,\\
\lim_{n\to\infty}\lim_{m\to\infty}
\delta_M^*(0,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
1/2\pi,
\end{eqnarray}
and
\begin{eqnarray}
\lim_{n\to\infty}
\delta_M^*(0,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{n}})
&=&
\int_{-\infty}^\infty\delta_M(k',\textstyle{\frac{1}{n}})\delta_M(k',\textstyle{\frac{1}{n}})dk'=\infty
\nonumber.
\end{eqnarray}
Employing
\begin{eqnarray}
\hat\delta_M^*(x,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
\frac{1}{2\pi}
\int_{-\infty}^\infty\delta_M^*(k,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})e^{ikx}dk
=
2\pi \hat\delta_M(x,\textstyle{\frac{1}{n}})\hat\delta_M(x,\textstyle{\frac{1}{m}})\label{FF}
\end{eqnarray}
we can write
\begin{eqnarray}
\delta_M^*(k-k',\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})
&=&
\int_{-\infty}^\infty \hat\delta_M^*(x,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}}) e^{-i(k-k')x}dx
\nonumber\\
&=&
2\pi\int_{-\infty}^\infty
\overline{\hat\delta_M(x,\textstyle{\frac{1}{n}})e^{ikx}}
\hat\delta_M(x,\textstyle{\frac{1}{m}})e^{ik'x}dx
\nonumber\\
&=&
\frac{1}{2\pi}\langle k,\textstyle{\frac{1}{n}}|k',\textstyle{\frac{1}{m}}\rangle=
\displaystyle{\frac{1}{2\pi}}\langle k,\textstyle{\frac{1}{m}}|k',\textstyle{\frac{1}{n}}\rangle,
\end{eqnarray}
since $\hat\delta_M(x,\textstyle{\frac{1}{n}})$ is real. Of course, $\langle k,\textstyle{\frac{1}{n}}|k',\textstyle{\frac{1}{m}}\rangle<\infty$ for all $k,k'$. Now let us recall that
\begin{eqnarray}
\lim_{n\to\infty}\hat\delta_M(x,\textstyle{\frac{1}{n}}) &=& \frac{1}{2\pi},
\end{eqnarray}
and thus, formally, it is justified to write
\begin{eqnarray}
\lim_{n\to\infty}\lim_{m\to\infty}\langle k,\textstyle{\frac{1}{n}}|k',\textstyle{\frac{1}{m}}\rangle
\textrm{ `$=$' }
\int_{-\infty}^\infty
e^{i(k'-k)x}dx,\label{'delta'}
\end{eqnarray}
while simultaneously we have shown that, for $k=k'$,
\begin{eqnarray}
\lim_{n\to\infty}\lim_{m\to\infty}\langle k,\textstyle{\frac{1}{n}}|k,\textstyle{\frac{1}{m}}\rangle
&=&
2\pi \lim_{n\to\infty}\lim_{m\to\infty}\delta_M^*(0,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})=1.
\end{eqnarray}
There is no contradiction between the above two formulas --- simply, integration does not commute with the limits $n,m\to\infty$ (integration must be performed first).
A similar situation occurs with derivatives. Denoting
\begin{eqnarray}
\langle x|k,\textstyle{\frac{1}{n}}\rangle
&=&
2\pi\hat\delta_M(x,\textstyle{\frac{1}{n}})e^{ikx}
\end{eqnarray}
and taking into account
\begin{eqnarray}
\lim_{n\to\infty}\frac{d^N}{dx^N}\hat\delta_M(x,\textstyle{\frac{1}{n}}) &=& 0, \quad N=1,2\dots,
\end{eqnarray}
we find
\begin{eqnarray}
\lim_{n\to\infty}\Big(\frac{1}{i}\frac{d}{dx}\Big)^N\langle x|k,\textstyle{\frac{1}{n}}\rangle
&=&
k^N\lim_{n\to\infty}\langle x|k,\textstyle{\frac{1}{n}}\rangle=
k^Ne^{ikx}.
\end{eqnarray}
Collecting all the formulas we have derived so far we are ready for generalization.
\subsection{Generalized function $\delta_M^*(k)$}
From the point of view of the sequential approach to distributions \cite{Sikorski} the sequences $\delta_\Lambda(k,\frac{1}{n})=\delta(k,4n,\frac{1}{n})$, $\delta_M(k,\frac{1}{n})=\delta(k,\frac{1}{2\pi},\frac{1}{n})$, and $\delta_M^*(k,\frac{1}{n},\frac{1}{m})$ belong to the same equivalence class and thus define the same distribution.
\begin{figure}
\includegraphics[width=8cm]{N-delta.eps}
\caption{The usual ($\Lambda$-shaped) delta-sequences are special cases of M-shaped delta-sequences --- to see this one puts $a=4/\epsilon$ and defines $\delta_\Lambda(k,\epsilon)=\delta(k,4/\epsilon,\epsilon)$. Here $\epsilon=1/2$, $a=8=4/\epsilon$ (upper), and its Fourier transform (lower).}
\end{figure}
In consequence, any of these delta-sequences can be treated as a representative of the equivalence class and employed in computations involving Dirac deltas.
However, we want to do more, and in particular want to include expressions such as Dirac delta at the origin. To do so,
we will treat the limits $\lim_{n\to\infty}\lim_{m\to\infty}\delta_M^*(k,\frac{1}{n},\frac{1}{m})$ as a new type of generalized function, $\delta_M^*(k)$.
The basic properties are as follows
\begin{eqnarray}
\textstyle{\int_{-\infty}^\infty} f(k)\delta_M^*(k)dk
&=&
\lim_{n\to\infty}\lim_{m\to\infty}\textstyle{\int_{-\infty}^\infty} f(k)\delta_M^*(k,\textstyle{\frac{1}{n}},\textstyle{\frac{1}{m}})dk
=
\displaystyle{\frac{f(0_-)+f(0_+)}{2}},\\
\textstyle{\int_{-\infty}^\infty} \delta_M^*(k)dk
&=&
1,\\
\textstyle{\int_{-\infty}^\infty} \delta_M^*(k)^2dk
&=&
\infty,\\
\delta_M^*(0)
&=&
\frac{1}{2\pi},\\
\hat\delta_M^*(x)
&=&
\frac{1}{2\pi}.
\end{eqnarray}
Plane waves defined by
\begin{eqnarray}
\langle x|k\rangle
&=&
\lim_{n\to\infty}\langle x|k,\textstyle{\frac{1}{n}}\rangle
=
e^{ikx}
\end{eqnarray}
satisfy
\begin{eqnarray}
\Big(\frac{1}{i}\frac{d}{dx}\Big)^N\langle x|k\rangle
&=&
\lim_{n\to\infty}\Big(\frac{1}{i}\frac{d}{dx}\Big)^N\langle x|k,\textstyle{\frac{1}{n}}\rangle\nonumber\\
&=&
k^N\langle x|k\rangle,\\
\langle k|k'\rangle
&=&
\lim_{n\to\infty}\lim_{m\to\infty}\langle k,\textstyle{\frac{1}{n}}|k',\textstyle{\frac{1}{m}}\rangle
\nonumber\\
&=& 2\pi \delta_M^*(k-k')\\
&=&
\lim_{n\to\infty}\lim_{m\to\infty}\textstyle{\int_{-\infty}^\infty}
\langle k,\textstyle{\frac{1}{n}}|x\rangle\langle x|k',\textstyle{\frac{1}{m}}\rangle dx,\nonumber\\
&\pp=&
\\
\langle k|k\rangle
&=&
\lim_{n\to\infty}\lim_{m\to\infty}\langle k,\textstyle{\frac{1}{n}}|k,\textstyle{\frac{1}{m}}\rangle
=
2\pi \delta_M^*(0)=1
\end{eqnarray}
These rules allow for all the standard computations involving the Dirac delta, but one can perform also certain new operations. For example, consider the expression
\begin{eqnarray}
\int\langle x|k\rangle\langle k|k'\rangle\langle k'|y\rangle f(k)dk,
\end{eqnarray}
where $f(k)$ is, say, square-integrable, and which should be understood in the following sense
\begin{eqnarray}
&{}&
\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}
\int\langle x|k,\textstyle{\frac{1}{n_1}}\rangle\langle k,\textstyle{\frac{1}{n_2}}|k',\textstyle{\frac{1}{n_3}}\rangle\langle k',\textstyle{\frac{1}{n_4}}|y\rangle f(k)dk
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}}
=
\lim_{n_2\to\infty}\lim_{n_3\to\infty}
\int\langle x|k\rangle 2\pi\delta_M^*(k-k',\textstyle{\frac{1}{n_2},\frac{1}{n_3}})\langle k'|y\rangle f(k)dk
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}}
=\lim_{n_2\to\infty}
\int\langle x|k\rangle2\pi\delta_M(k-k',\textstyle{\frac{1}{n_2}})\langle k'|y\rangle f(k)dk
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}}
=2\pi\langle x|k'\rangle\langle k'|y\rangle f(k')
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}}
=2\pi\int\langle x|k\rangle\delta(k-k')\langle k'|y\rangle f(k)dk
\end{eqnarray}
where the last formula was written in terms of the `standard' Dirac delta, to make it more familiar, but we could replace $\delta(k)$ by $\delta_M^*(k)$.
Simultaneously,
\begin{eqnarray}
{}&{}&
\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}
\int\langle x|k,\textstyle{\frac{1}{n_1}}\rangle\langle k,\textstyle{\frac{1}{n_2}}|k,\textstyle{\frac{1}{n_3}}\rangle\langle k,\textstyle{\frac{1}{n_4}}|y\rangle f(k)dk
\nonumber\\
&{}&\pp{\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n_3\to\infty}\lim_{n_4\to\infty}}
=
\int\langle x|k\rangle\langle k|y\rangle f(k)dk
\end{eqnarray}
The formulas reduce to
\begin{eqnarray}
|k\rangle\langle k|k'\rangle\langle k'|
&=&
2\pi \delta(k-k')|k\rangle\langle k|,\label{11}
\end{eqnarray}
and
\begin{eqnarray}
|k\rangle\langle k|k\rangle\langle k|
=
|k\rangle\langle k|,\label{22}
\end{eqnarray}
that are mutually consistent if one treats each $|k\rangle$ as a limit of a separate sequence $|k,\frac{1}{n}\rangle$. Eqs. (\ref{11})--(\ref{22})
can be conveniently written as
\begin{eqnarray}
|k\rangle\langle k|k'\rangle\langle k'|
&=&
2\pi \delta_M^*(k-k')|k\rangle\langle k|.
\end{eqnarray}
\subsection{M-shaped deltas with respect to more general measures}
In this section we will concentrate on a generalization that is useful from the point of view of relativistic applications.
Let us treat explicitly only a one-dimensional case. Let us assume that instead of $dp$ we have to use a measure $d\mu(p)=\rho(p)dp$, and an appropriate delta is needed,
\begin{eqnarray}
\int d\mu(p')\delta_{\mu M}(p,p')f(p')=f(p),
\end{eqnarray}
with $\delta_{\mu M}(p,p)=a$, say, where $a$ is a constant.
The standard solution,
\begin{eqnarray}
\delta_\mu(p,p')=\rho(p')^{-1}\delta(p-p'),\label{mu}
\end{eqnarray}
if generalized to M-shaped deltas by
\begin{eqnarray}
\delta_{\mu M}(p,p')=\rho(p')^{-1}\delta_M(p-p'),\label{muM}
\end{eqnarray}
implies $\delta_{\mu M}(p,p)=\rho(p)^{-1}\delta_M(0)$ and will not lead to $a$ independent of $p$. It follows that we have to proceed in a way different from (\ref{muM}).
So, let $\delta(p,a,\frac{1}{n})$, $\delta(0,a,\frac{1}{n})=a$, be an arbitrary M-shaped delta-sequence discussed in the preceding sections. The sequence
\begin{eqnarray}
\delta_\mu(p,p',\textstyle{\frac{1}{n}})
&=&
\rho(p')^{-1}\delta\big(p-p',a\rho(p),\textstyle{\frac{1}{n}}\big),
\end{eqnarray}
$\rho(p)=d\mu(p)/dp$, has the following properties
\begin{eqnarray}
\lim_{n\to \infty}
\int d\mu(p')\delta_\mu(p,p',\textstyle{\frac{1}{n}})f(p')
&=&
\frac{f(p_-)+f(p_+)}{2},\\
\delta_\mu(p,p,\textstyle{\frac{1}{n}})
&=&
a
\end{eqnarray}
and defines distribution $\delta_{\mu M}(p,p')$ satisfying
\begin{eqnarray}
\int d\mu(p')\delta_{\mu M}(p,p')f(p')
&=&
\frac{f(p_-)+f(p_+)}{2},\\
\delta_{\mu M}(p,p)
&=&
a.
\end{eqnarray}
\subsection{The meaning of products of operators occurring in harmonic-oscillator Lie algebra}
\label{subsec 75}
Consider $N=1$ reducible representation of harmonic-oscillator Lie algebra.
Define
\begin{eqnarray}
a_{\bm b}(\bm R,\bm k,1,\textstyle{\frac{1}{n_1},\frac{1}{n_2}})=|\bm R,\textstyle{\frac{1}{n_1}}\rangle
\langle\bm R,\textstyle{\frac{1}{n_1}}|
\otimes
|\bm k,\textstyle{\frac{1}{n_2}}\rangle\langle\bm k,\textstyle{\frac{1}{n_2}}|\otimes a_{\bm b}.\label{XXXX}
\end{eqnarray}
The kets and bras are understood in the sense described in this Appendix. The products such as, say, $a_{\bm b}(\bm R,\bm k,1)a_{\bm c}(\bm R',\bm k',1)^{\dag}$ are understood in the sense of $M$-shaped Dirac deltas:
\begin{eqnarray}
&{}&a_{\bm b}(\bm R,\bm k,1)a_{\bm c}(\bm R',\bm k',1)^{\dag}
\nonumber\\
&{}&
\pp{XX}=
\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n'_1\to\infty}\lim_{n'_2\to\infty}
a_{\bm b}(\bm R,\bm k,1,\textstyle{\frac{1}{n_1},\frac{1}{n_2}})a_{\bm c}(\bm R',\bm k',1,\textstyle{\frac{1}{n'_1},\frac{1}{n'_2}})^{\dag}.
\end{eqnarray}
The order of limits is irrelevant.
\section{Noetherian construction of generators of $N=1$ reducible representation --- example of 4-momentum}
Consider any two fields $A(x)$, $B(x)$ satisfying d'Alembert equation. For example, let $A(x)=A_b(x,1,\textstyle{\frac{1}{n_1},\frac{1}{n_2}})$, $B(x)=A_c(x,1,\textstyle{\frac{1}{n'_1},\frac{1}{n'_2}})$, where the dependence on sequences means that the corresponding amplitude operators are constructed according to the recipe (\ref{XXXX}). d'Alembert equation implies conservation of the energy-momentum tensor
\begin{eqnarray}
T_{ab}(x,\textstyle{\frac{1}{n_1},\frac{1}{n_2},\frac{1}{n'_1},\frac{1}{n'_2}})
&=&
-\frac{1}{2} \Big(\partial_a A(x)\partial_b B(x)+\partial_b A(x)\partial_a B(x)-g_{ab}\partial_c A(x)\partial^c B(x)\Big),\nonumber\\
\end{eqnarray}
i.e. $\partial^aT_{ab}(x,\textstyle{\frac{1}{n_1},\frac{1}{n_2},\frac{1}{n'_1},\frac{1}{n'_2}})=0$, and
\begin{eqnarray}
P_{a}(\textstyle{\frac{1}{n_1},\frac{1}{n_2},\frac{1}{n'_1},\frac{1}{n'_2}})
=
\displaystyle{\int d^3x T_{a0}(x_0,\bm x,\textstyle{\frac{1}{n_1},\frac{1}{n_2},\frac{1}{n'_1},\frac{1}{n'_2}})}
\end{eqnarray}
is independent of $x_0$ for all $n_1,\dots,n'_2$. The limit
\begin{eqnarray}
P_a
&=&
\lim_{n_1\to\infty}\lim_{n_2\to\infty}\lim_{n'_1\to\infty}\lim_{n'_2\to\infty}
P_{a}(\textstyle{\frac{1}{n_1},\frac{1}{n_2},\frac{1}{n'_1},\frac{1}{n'_2}})
\nonumber\\
&=&
-I\otimes {\int} d k{\,}k_a|\bm k\rangle\langle \bm k|\otimes
\frac{1}{2}\Big(a^{\dag}_{\bf a}a^{\bf a}+a^{\bf a}a^{\dag}_{\bf a}\Big)
=P_a(1)+K_a
\end{eqnarray}
is the 4-momentum, vacuum contribution included, corresponding to $N=1$ reducible representation of the harmonic oscillator Lie algebra.
|
\section{Introduction}
The purpose of this paper is to use the homotopical methods for
the description of subgroups determined by certain ideals, here called symmetric ideals, in free
group rings.
\para
Let $F$ be a free group, $\mathbb Z[F]$ its integral group
ring and $I$ a two-sided ideal in $\mathbb Z[F]$. The general
problem of description of the normal subgroup
$$
D(F;I):=F\cap (1+I)
$$
of $F$ is very difficult. As an illustration of the complexity of answers for different particular cases we may mention some examples. Let $R$
be a normal subgroup of $F$, ${\bf r}=(R-1)\mathbb Z[F]$, and ${\bf
f}$ the augmentation ideal of $\mathbb Z[F],$ then \cite{CKG}
\begin{align*}
& F\cap (1+{\bf f}^2{\bf r}^2)=\gamma_3(R\cap [F,F])\gamma_4(R),\\
& F\cap (1+{\bf rf}^2{\bf r})=[R\cap [F,F],R\cap
[F,F],R]\gamma_4(R).
\end{align*} The subtility of the dimension subgroup problem is well-known; this is the case when $I={\bf f}^n+{\bf r}$.
For a survey of problems in this area, see \cite{Passi}, \cite{Gupta}, \cite{MP}.\para
Given a ring $A$ and two-sided ideals $I_1,\dots, I_n\ (n\geq 2)$
in $A$ consider their symmetric product:
$$
(I_1\dots I_n)_S:=\sum_{\sigma \in \Sigma_n}I_{\sigma_1}\dots
I_{\sigma_n},
$$
where $\Sigma_n$ is the symmetric group of degree $n$. For example, in the
case $n=2$, one has $(I_1I_2)_S=I_1I_2+I_2I_1.$ Observe that while
$(I_1\dots I_n)_S\subseteq I_1\cap
\dots \cap I_n$ always, the reverse inclusion does not hold, in general.\para Let $F$ be a free group, and let
$R_1,\dots, R_n$ be normal subgroups of $F$. Consider the induced two-sided ideals in the integral group ring $\mathbb Z[F]$ defined as
${\bf r}_i=(R_i-1)\mathbb Z[F],\
i=1,\dots, n$. The following problems arise naturally:
\begin{quote}
(1) Identify the quotient
$$
Q({\bf r}_1,\dots, {\bf r}_n):=\frac{{\bf r}_1\cap \dots \cap {\bf
r}_n}{({\bf r}_1\dots {\bf r}_n)_S}.
$$
(2) Identify the normal subgroup of $F$, determined by the ideal $({\bf
r_1}\dots {\bf r}_n)_S,$ i.e., the subgroup
$$
D(F; ({\bf r}_1\dots {\bf r}_n)_S):=F\cap (1+({\bf r}_1\dots {\bf
r}_n)_S).
$$\end{quote}
Let $[R_1,\,\dots\,,\,R_n]_S$ denote the symmetric commutator subgroup, namely,\\ $ \prod_{\sigma\in \Sigma_n}[\dots\,
[R_{\sigma_1},\,R_{\sigma_2}],\,\dots\,,\, R_{\sigma_n}]$ of the normal subgroups $R_1,\,\dots\,,\, R_n$:
$$
[R_1,\,\dots\,,\,R_n]_S=\prod_{\sigma\in \Sigma_n}[\dots
[R_{\sigma_1},\,R_{\sigma_2}],\,\dots\,,\, R_{\sigma_n}].
$$
Observe that we have always $$ [R_1,\,\dots\,,\, R_n]_S\subseteq D(F;({\bf r}_1\,\dots\,
{\bf r}_n)_S).
$$The fundamental theorem of free group rings (see \cite{Gupta}, Theorem 3.1, p.\,12) states that, for all $n\geq 1$, the above inclusion is an equality in case $R_i=F$ for $i=1,\,\ldots\,,\,n$. Apart from this case, there is hardly anything else that seems to be available in the literature about the subgroups $D(F;({\bf r}_1\,\dots\,
{\bf r}_n)_S)$, in general. Naturally, one would like to investigate the validity, or otherwise, of the inclusion $D(F;({\bf r}_1\,\dots\,
{\bf r}_n)_S)\subseteq [R_1,\,\dots\,,\,R_n]_S$.
Let
\begin{align*}
& f_{F;R_1,\,\dots\,,\,R_n}: \frac{R_1\cap \dots \cap
R_n}{[R_1,\,\dots\,,\,R_n]_S}\to \frac{{\bf r}_1\cap \dots \cap
{\bf r}_n}{({\bf r}_1\dots {\bf r}_n)_S},\end{align*} be the
natural map defined by \begin{align*} & f_{F;R_1,\,\dots\,,\,R_n}:
g.[R_1,\,\dots\,,\,R_n]_S\mapsto g-1+({\bf r}_1\dots {\bf
r}_n)_S,\ g\in R_1\cap\dots\cap R_n.
\end{align*}
The main idea of this paper is based on the fact that, for a
certain choice of groups $F,\,R_1,\,\dots\,,\, R_n$, there exists
a space $X$, such that the map $f_{F;R_1,\,\dots\,,\, R_n}$ is the
$(n-1)$st Hurewicz homomorphism:
$$
\xyma{ \frac{R_1\cap\, \dots\, \cap R_n}{[R_1,\,\dots\,,\,R_n]_S} \ar@{=}[d]
\ar@{->}[rr]^{\ \ \ \ \ \ \ \ \ \ \ \ \ f_{F;R_1,\,\dots\,,\, R_n}\ \ \
\ \ \ \ \ \ \ \ \ } & & \frac{{\bf r}_1\cap
\dots \cap {\bf r}_n}{({\bf r}_1\,\dots \,{\bf r}_n)_S}\ar@{=}[d]\\
\pi_{n-1}(X) \ar@{->}[rr] & & H_{n-1}(X) }
$$
In that case, the quotient $\frac{D(F;({\bf r}_1\,\dots\, {\bf
r}_n)_S)}{[R_1,\,\dots\,,\, R_n]_S}$ is exactly the kernel of Hurewicz
homomorphism and we are able to use arguments from simplicial homotopy for
the computation of subgroups determined by symmetric product of
ideals. Our analysis also yields an example where the inclusion
$$D(F;({\bf r}_1\,\dots\,
{\bf r}_n)_S)\supseteq [R_1,\,\dots\,,\,R_n]_S$$ is proper.
In Section 2 we prove certain technical results needed for our investigation. Our main results are Theorems 3.1 and 3.2 (see Section 3).
\par\vspace{1cm}
\section{Technical results} We need some preparation for proving our main results. Given a group $G$, let $\Delta(G)$ denote the augmentation ideal of its integral group ring $\mathbb Z[G]$. The following result is well-known.
\para
\begin{lemma}\label{element}
If $N$ is a normal subgroup of a group $G$, then $N\cap
(1+\Delta(N)\Delta(G))=[N,\,N].$
\end{lemma}
\para
For the case of two normal subgroups in the free group $F$, we have the following\para
\begin{prop}
Let $F=R_1R_2$. Then the map
$$
f_{F;\,R_1,\,R_2}: \frac{R_1\cap R_2}{[R_1,\,R_2]}\to \frac{{\bf
r}_1\cap {\bf r}_2}{{\bf r}_1{\bf r}_2+{\bf r_2}{\bf r}_1}
$$
is an isomorphism. In particular, $D(F;\, ({\bf r}_1{\bf r}_2)_S)=[R_1,\,R_2]$.
\end{prop}
\para\noindent{\it Proof.}
Let $T=\{t_i\}_{i\in I}\subseteq R_1$ be a left transversal for
$R_2$ in $F$. Then every element $f\in F$ can be written uniquely
as $f=ts$ with $t\in T$ and $s\in R_2$; in particular, if $f\in
R_1$, then $s\in R_1\cap R_2$. Let $\varphi:\mathbb Z[F]\to
\mathbb Z[R_2]$ be the $\mathbb Z$-linear extension of the map
$F\to R_2$ given by $f\mapsto s$. Observe that ${\bf r}_1 {\bf
r}_2=\Delta(R_1)\Delta(R_2)$ and ${\bf r}_2{\bf
r}_1=\Delta(R_2)\Delta(R_1)$ since $F=R_1R_2$. Furthermore,
$$
\varphi({\bf r}_1{\bf r}_2+{\bf r}_2{\bf r}_1)\subseteq
\Delta(R_1\cap R_2)\Delta(R_2)+\Delta([R_1,\,R_2])\mathbb Z[R_2].
$$
\par\vspace{.25cm}
Consider the map $$\theta:R_1\cap R_2\to \frac{{\bf r}_1\cap {\bf
r}_2}{{\bf r}_1{\bf r}_2+{\bf r}_2{\bf r}_1},\quad f\mapsto
f-1+{\bf r}_1{\bf r}_2+{\bf r}_2{\bf r}_1.$$ Clearly $\theta $ is
a homomorphism and $[R_1,\,R_2]\subseteq \ker \theta$. Let $f\in
R_1\cap R_2$ be an element in $\ker \theta$. We then have
$$f-1=\varphi(f-1)\in \Delta(R_1\cap R_2)\Delta(R_2)+\Delta([R_1,\,R_2])\mathbb Z[R_2]$$ in the group ring $\mathbb Z[R_2]$.
Thus, going modulo $[R_1,\,R_2]$ and invoking Lemma \ref{element}
with $G=R_1/[R_1,\,R_2],\ N=(R_1\cap R_2)/[R_1,\,R_2],$ we must
have $f\in [R_1,\,R_2]$. Consequently $\theta$ induces
a monomorphism
$$f_{R_1,R_2}:\frac{R_1\cap R_2}{[R_1,\,R_2]}\hookrightarrow \frac{{\bf r}_1\cap {\bf
r}_2}{{\bf r}_1{\bf r}_2+{\bf r}_2{\bf r}_1}.$$
\par\vspace{.25cm}
Let $\alpha\in {\bf r}_1$. Then $\alpha=\sum_i(r_i-1)\beta_i$ with
$r_i\in R_1$ and $\beta_i\in \mathbb Z[R_2]$. Now
$r_i=t_{i_j}s_{i_J}$ with $t_{i_j}\in T$ and $s_{i_j}\in R_1\cap
R_2$. Therefore,
$$\alpha\equiv (w-1)+\sum_km_k(t_k-1)\ \mod\ {\bf r}_1{\bf r}_2+{\bf r}_2{\bf r}_1$$ with $m_k\in \mathbb Z$ and $w\in R_1\cap
R_2$. It follows that if $\alpha\in {\bf r}_1\cap {\bf r}_2$, then
$m_k=0$ for all $k$, and we thus conclude that $f_{R_1,R_2}$ is an
epimorphism and hence an isomorphism. $\Box$
\para
\begin{lemma}\label{techlemma} Let $X=X_1\sqcup \dots \sqcup X_n\ (n\geq 2)$ be a disjoint union of
sets. Let \linebreak $p_i: F(X)\to F(X_1\sqcup \ldots \hat{X}_i\ldots\sqcup X_n),\ i=1,\ldots,n$, be the natural projections induced by $$p_i(x)=
\begin{cases}
x, \ for\ x\in X\backslash X_i\\
1,\ for\ x\in X_i
\end{cases}$$
and $R_i=\ker (p_i). $
Then
\begin{quote}
$(i)$ $ R_1\cap \dots
\cap R_n=[R_1,\dots, R_n]_S
$
in $F(X)$;\para\noindent
$ (ii)$ ${\bf r}_1\cap \dots\cap {\bf r}_k=({\bf
r}_1\dots {\bf r}_n)_S
$
in $\mathbb Z[F(X)]$.\end{quote}
\end{lemma}
\begin{proof}
The statement (i) follows from (\cite{Wu}, Corollary 3.5) (see \cite{BMVW}).
\para
For the proof of (ii) observe first that, for each $i,j\in
\{1,2,\dots, n\},\ i\neq j,$ we have
$$
R_i=\langle X_i\rangle^{F(X\setminus X_j)}[R_i,\,R_j].
$$
Since $\mathbb Z[F(X)]=\mathbb Z[F(X\setminus X_j)]+{\bf r}_j$ and
$[R_i,\,R_j]-1\subseteq {\bf r}_i{\bf r}_j+{\bf r}_j{\bf r}_i,$ it
follows that
\begin{equation}\label{02}
{\bf r}_i=(\langle X_i\rangle^{F(X\setminus X_j)}-1)\mathbb
Z[F(X\setminus X_j)]+{\bf r}_i{\bf r}_j+{\bf r}_j{\bf r}_i,
\end{equation}
and consequently, we have
$$
{\bf r}_i\cap {\bf r}_j={\bf r}_i{\bf r}_j+{\bf r}_j{\bf r}_i,\
i\neq j.
$$
Suppose that, for some $k$, $2\leq k<n,$ we have shown that
$$
{\bf r}_{i1}\cap \dots \cap {\bf r}_{ik}=({\bf r}_{i1}\dots {\bf
r}_{ik})_S
$$
for all subsets of $k$ elements from $\{1,\dots,n\},$ and let $j$
be an integer, $1\leq j\leq n,\ j\notin \{i1,\dots, ik\}.$ From
(\ref{02}), we have
$$
{\bf r}_{il}=(\langle X_{il}\rangle^{F(X\setminus X_j)}-1)\mathbb
Z[F(X\setminus X_j)]+{\bf r}_{il}{\bf r}_j+{\bf r}_j{\bf r}_{il},\
l=i1,\dots, ik.
$$
Consequently
$$
({\bf r}_{i1}\dots {\bf r}_{ik})_S\subseteq \mathbb Z[F(X\setminus
X_j)]+({\bf r}_{i1}\dots {\bf r}_{ik} {\bf r}_j)_S.
$$
An application of the natural projection $\mathbb Z[F(X)]\to
\mathbb Z[F(X\setminus X_j)]$ induced by the map which is
identity on $X\setminus X_j$ and trivial on $X_j$ then shows that
$$
{\bf r}_j\cap {\bf r}_{i1}\cap \dots \cap {\bf r}_{ik}\subseteq
({\bf r}_{i1}\dots {\bf r}_{ik}{\bf r}_j)_S.
$$
The reverse inclusion being trivial, it follows that the
intersection of $k+1$ distinct ideals out of ${\bf r}_1,\dots,
{\bf r}_n$ equals the corresponding symmetric sum of their
products, and thus, by induction, assertion (ii) is proved.
\end{proof}
\par\vspace{.5cm}
\section{Simplicial constructions}
\subsection{Milnor's construction.}
Recall that, for a given pointed simplicial set $K$, the Milnor
$F(K)$-construction \cite{Milnor} is the simplicial group with
$F(K)_n=F(K_n\setminus *)$, where $F(-)$ is the free group
functor. Consider the simplicial circle
$S^1=\Delta[1]/\partial\Delta[1]$:
\begin{equation}\label{1-sphere}
S_0^1=\{*\},\ S_1^1=\{*,\,\sigma\},\ S_2^1=\{*,\, s_0\sigma,\,
s_1\sigma\},\,\dots\,,\, S_n^1=\{*,\, x_0,\,\dots\,,\, x_{n-1}\},
\end{equation}
where $x_i=s_{n-1}\dots \hat s_i\dots s_0\sigma$. For the Milnor
construction $F(S^1)$, $F(S^1)_n$ is a free group of rank $n$,
for $n\geq 1:$
$$F(S^1):\ \ \ \ldots\ \begin{matrix}\longrightarrow\\[-3.5mm] \ldots\\[-2.5mm]\longrightarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\ldots\\[-2.5mm]\longleftarrow \end{matrix}\ F_3\ \begin{matrix}\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longleftarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\longleftarrow
\end{matrix}\ F_2\ \begin{matrix}\longrightarrow\\[-3.5mm] \longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\longleftarrow \end{matrix}\ \mathbb Z$$
with face and degeneracy homomorphisms:
\begin{align*}
& \partial_i: F_n\to F_{n-1},\ i=0,\,\dots\,,\, n,\ n=2,\,3,\,\dots\\
& s_i: F_n\to F_{n+1},\ i=0,\,\dots\,,\, n,\ n=1,\,2,\,\dots\,.
\end{align*}
There is a homotopy equivalence \cite{Milnor}:
$$
|F(S^1)|\simeq \Omega S^2.
$$
Hence, for $n\geq 2$, the $n$th homotopy group of $S^2$ can be
described as an intersection of kernels in degree $n-1$ modulo
simplicial boundaries. Following \cite{Wu}, denote the elements
from a basis of $F_{n+1}$ as follows:
\begin{align*}
& y_n=s_{n-1}\dots s_0\sigma,\\
& y_i=s_n\dots \hat s_i\dots s_0\sigma (s_n\dots \hat s_{i+1}\dots
s_0\sigma)^{-1},\ 0\leq i<n.
\end{align*}
Then, it follows from standard simplicial identities that, in the
free group $F_{n+1}$, one has
\begin{align*}
& \ker(\partial_0)=y_0\dots y_n,\\ & \ker(\partial_{i})=\langle
y_{i-1}\rangle^{F_{n+1}},\ 0<i\leq n+1
\end{align*}
Lemma \ref{techlemma} applied to the case $X=\{y_0,\dots, y_n\},$
implies that
$$
\ker(\partial_1)\cap \dots\cap
\ker(\partial_{n+1})=[\ker(\partial_1),\dots, \ker(\partial_{n+1})]_S
$$
Therefore, there is a natural presentation of the $(n+1)$st
homotopy group of $S^2$ given first by Wu \cite{Wu}:
$$
\pi_{n+1}(S^2)\simeq \frac{\ker(\partial_0)\cap \dots \cap
\ker(\partial_n)}{[\ker(\partial_0),\,\dots\,,\,
\ker(\partial_n)]_S},\ n\geq 1.
$$
For similar results obtained without simplicial constructions see
\cite{EllisMikhailov}.
\para
\begin{theorem}
Let $n\geq 3,$ $F_n$ a free group with a basis $\{x_1,\dots,
x_n\}$. Let $R_i=\langle x_i\rangle^{F_n},\linebreak i=1,\,\dots\,,\, n,$
$R_{n+1}=\langle x_1\dots x_n\rangle^{F_n}.$ Then
\begin{quote}
$(i)$ there is an isomorphism $Q({\bf r}_1,\dots, {\bf
r}_{n+1})\simeq \mathbb Z;$\para\noindent $(ii)$
$
D(F; ({\bf r}_1\dots {\bf r}_{n+1})_S)=R_1\cap\dots \cap R_{n+1}.
$\end{quote}
Furthermore, $R_1\cap \dots\cap R_{n+1}\neq [R_1,\,\dots\,,\,
R_{n+1}]_S$ for $n\neq 0\mod 8$.
\end{theorem}
\begin{proof}
First apply the functor $\mathbb Z[-]$ to the Milnor construction
$F(S^1)$:
$$\mathbb Z[F(S^1)]:\ \ \ \ldots\ \begin{matrix}\longrightarrow\\[-3.5mm] \ldots\\[-2.5mm]\longrightarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\ldots\\[-2.5mm]\longleftarrow \end{matrix}\ \mathbb Z[F_3]\ \begin{matrix}\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]\longleftarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\longleftarrow
\end{matrix}\ \mathbb Z[F_2]\ \begin{matrix}\longrightarrow\\[-3.5mm] \longrightarrow\\[-3.5mm]\longrightarrow\\[-3.5mm]
\longleftarrow\\[-3.5mm]\longleftarrow \end{matrix}\ \mathbb Z[\mathbb Z]$$
By definition of homology, we have
$$
\pi_n\mathbb Z[F(S^1)]=H_n(\Omega S^2).
$$
From the classical suspension splitting theorem of loop suspensions~\cite{James}, we have
$$
\Sigma \Omega S^2\simeq \bigvee_{k=2}^\infty S^k
$$
and so $H_n(\Omega S^2)=\mathbb{Z}$ for each $n\geq1$. Thus $\pi_n\mathbb Z[F(S^1)]=\mathbb{Z}, \ n\geq1$.
The kernels of homomorphisms
$$
\bar\partial_i:\mathbb Z[F_{n+1}]\to \mathbb Z[F_n],\ i=0,\,\dots\,,\,
n+1
$$
are ideals
$$
(\ker(\partial_i)-1)\mathbb Z[F_{n+1}],\ i=0,\,\dots\,,\, n+1.
$$
Making the enumeration in the free group $F_n$: $x_i=y_{i+1},\
i=0,\,\dots\,,\, n-1,$ lemma \ref{techlemma} (ii) implies that
$$
H_n(\Omega S^2)\simeq Q({\bf r}_1,\,\dots\,,\, {\bf r}_{n+1})\simeq
\mathbb Z
$$
and the statement (i) is proved.
\para
For proving (ii), observe now that there is a natural diagram
$$
\xyma{ \frac{R_1\cap \dots \cap R_{n+1}}{[R_1,\dots,R_{n+1}]_S}
\ar@{=}[d] \ar@{->}[rr]^{\ \ \ \ \ \ \ \ \ \ \ \ \
f_{F;R_1,\dots, R_{n+1}}\ \ \ \ \ \ \ \ \ \ \ \ } & & \frac{{\bf
r}_1\cap
\dots \cap {\bf r}_{n+1}}{({\bf r}_1\dots {\bf r}_{n+1})_S}\ar@{=}[d]\\
\pi_n(\Omega S^2) \ar@{->}[rr] & & H_n(\Omega S^2) }
$$
The homotopy groups $\pi_n(\Omega S^2)=\pi_{n+1}(S^2)$ are finite
for $n\geq 3$, hence the homomorphism $f_{F; R_1,\dots, R_{n+1}}$
is the zero map and therefore,
$$
R_1\cap \dots\cap R_{n+1}\subseteq D(F; ({\bf r}_1\dots {\bf
r}_{n+1})_S).
$$
The reverse inclusion follows trivially, hence the statement (ii)
follows.
\para
Finally, the remark that $R_1\cap \dots\cap R_{n+1}\neq
[R_1,\dots, R_{n+1}]_S$ for $n\neq 0\mod 8$ is just a reformulation of
the result of Curtis \cite{Curtis} that $\pi_n(S^2)\neq 0,\ n\neq
1\mod 8$.
\end{proof}
\para\noindent
{\bf Remark 3.1.}\label{rem11}
For $n=2$, we have the following situation:\para Let $F=F(x_1,x_2),\
R_1=\langle x_1\rangle^F,\ R_2=\langle x_2\rangle^F,\ R_3=\langle
x_1x_2\rangle^F$. Then the following diagram consists of
isomorphisms
$$
\xyma{\mathbb Z\ar@{=}[d] & \mathbb Z \ar@{=}[d]\\ \pi_2(\Omega S^2)\ar@{->}[r]^\simeq \ar@{=}[d] & H_2(\Omega S^2)\ar@{=}[d] \\
\frac{R_1\cap R_2\cap R_3}{[R_1,\,R_2,\,R_3]_S} \ar@{->}[r]^\simeq &
\frac{{\bf r}_1\cap {\bf r}_2\cap {\bf r}_3}{({\bf r}_1{\bf
r}_2{\bf r}_3)_S}}
$$
and therefore $D(F;\,({\bf r}_1{\bf r}_2{\bf
r}_3)_S)=[R_1,\,R_2,\,R_3]_S$.\para
\subsection{Carlsson's Construction}
For any pointed simplicial set $K$ and any group $G$, the Carlsson construction~\cite{Car,Wu1} is the simplicial group $F^G(K)$ in which $F^G(K)_n$ is the self-free product of the group $G$ indexed by non-identity elements in $K_n$ with the face and degeneracy homomorphisms canonically induced from that of $K$. There is a homotopy equivalence~\cite{Car, Wu1}
$
|F^G(K)|\simeq \Omega (BG\wedge |K|),
$
where $BG$ is the classifying space of the group $G$. We consider the case where $K=S^1$ is the simplicial circle and $G$ is an arbitrary group. Note that the suspension of any path-connected space is (weak) homotopy equivalent to $\Sigma BG$ for certain group $G$ by Kan-Thurston's theorem~\cite{KT}. Thus the construction $F^G(S^1)$ gives the loop space model for the suspensions.
\para
The specific information on the simplicial structure of $F^G(S^1)$ is as follows:
\para\noindent The elements in the simplicial circle $S^1$ are listed in~(\ref{1-sphere}). Thus
$$
F^G(S^1)_{n+1}=G_{x_0}\ast G_{x_1}\ast\cdots \ast G_{x_{n}},
$$
where $G_{x_i}$ is a copy of $G$ labeled by $x_i=s_{n}\dots \hat s_i\dots s_0\sigma\in S^1_{n+1}$. The face $\partial_j\colon S^1_{n+1}\to S^1_n$ is given by the formula:
$$
\partial_j x_i=\left\{
\begin{array}{lcl}
x_i&\textrm{ for }& i<j\\
x_{i-1}&\textrm{ for } & i\geq j,\\
\end{array}\right.
$$
where $x_{-1}=x_n=\ast$ in $S^1_n$. Write $g(x_i)$ for the element $g\in G$ located in the copy $G_{x_i}$ of $G$. The group homomorphism
$$
\partial_j\colon F^G(S^1)_{n+1}=G_{x_0}\ast G_{x_1}\ast\cdots \ast G_{x_{n}}\longrightarrow F^G(S^1)_{n}=G_{x_0}\ast G_{x_1}\ast\cdots \ast G_{x_{n-1}}
$$
is given by the formulae:
\begin{equation}\label{equation3.2}
\begin{array}{c}
\partial_0(g(x_i))=\left\{
\begin{array}{lcl}
1&\textrm{ for }& i=0\\
g(x_{i-1})&\textrm{ for }&0<i\leq n,\\
\end{array}\right.\\
\partial_{n+1}(g(x_i))=\left\{
\begin{array}{lcl}
g(x_i)&\textrm{ for }& 0\leq i\leq n-1\\
1&\textrm{ for }& i=n,\\
\end{array}\right.\\
\end{array}
\end{equation}
and for $0<j<n+1$,
\begin{equation}\label{equation3.3}
\partial_{j}(g(x_i))=\left\{
\begin{array}{lcl}
g(x_i)&\textrm{ for }& i<j\\
g(x_{i-1})&\textrm{ for }& i\geq j.\\
\end{array}\right.\\
\end{equation}
In the free product $G^{\ast n+1}=G_{x_0}\ast G_{x_1}\ast\cdots
\ast G_{x_{n}}$, let $R^G_{n+1,\,0}=\langle g(x_0) \ | \ g\in
G\rangle^{G^{\ast n+1}}$, $R_{n+1,\,n+1}=\langle g(x_n) \ | \ g\in
G\rangle^{G^{\ast n+1}}$ and $R^G_{n+1,\,j}=\langle
g(x_{j-1})^{-1}g(x_j) \ | \ g\in G\rangle^{G^{\ast n+1}}$ for
$0<j<n+1$. Let $\br^G_{n+1,\,j}=(R^G_{n+1,\,j}-1)\Z[G^{\ast n+1}]$.
\para
\begin{theorem}\label{theorem3.2}
Let $G$ be any group. Then there is an isomorphism of groups
$$
Q(\br^G_{n+1,\,0},\br^G_{n+1,\,1},\,\ldots\,,\,\br^G_{n+1,\,n+1})\cong H_{n+1}(\Omega\Sigma BG;\,\Z)\cong \bigoplus_{k=1}^\infty H_{n+1}((BG)^{\wedge k};\,\Z),
$$
where $X^{\wedge k}$ is the $k$-fold self smash product of $X$.
\end{theorem}
\begin{proof}
Let $T=\{t_{\alpha}\ | \ \alpha\in J\}$ be a set of generators for $G$ and let $F$ be the free group generated by $T$. Consider Carlsson's construction $F^{F}(S^1)$. The group $F^F(S^1)_{n+1}$ is the free group generated by $\{t_{\alpha}(x_j) \ | \ \alpha\in J, \ 0\leq j\leq n\}$.
Let $y^{(\alpha)}_j=t_{\alpha}(x_j)t_{\alpha}(x_{j+1})^{-1}$ for $-1\leq j\leq n$ with $t_{\alpha}(x_{-1})=t_{\alpha}(x_{n+1})=1$. Then
$$
\{y^{(\alpha)}_j \ | \ \alpha\in J,\ 0\leq j\leq n\}
$$
is also a basis for $F^F(S^1)_{n+1}$. From formulae~(\ref{equation3.2}) and~(\ref{equation3.3}),
$$
\partial_j(y^{(\alpha)}_k)=\left\{
\begin{array}{lcl}
y^{(\alpha)}_{k-1}&\textrm{ for }& j\leq k\\
1&\textrm{ for }& j=k+1\\
y^{\alpha}_k&\textrm{ for }& j>k+1.\\
\end{array}\right.
$$
Thus
$$
\begin{array}{rcl}
\ker(\partial_j\colon F^F(S^1)_{n+1}\to F^F(S^1)_{n})&=&\langle y^{(\alpha)}_{j-1}\ | \ \alpha\in J\rangle^{F^F(S^1)_{n+1}}\\
&=&\langle t_{\alpha}(x_{j-1})t_{\alpha}(x_{j})^{-1}\ | \ \alpha\in J\rangle^{F^F(S^1)_{n+1}}\\
&=&R^F_{n+1,\,j}.\\
\end{array}
$$
for $0\leq j\leq n+1$.
The canonical epimorphism $\phi\colon F\to G$ induces a simplicial epimorphism
$$\tilde \phi\colon F^F(S^1)\twoheadrightarrow F^G(S^1),$$
which induces the epimorphism
$$
\tilde \phi|\colon \ker(\partial_j\colon F^F(S^1)_{n+1}\to F^F(S^1)_{n})\twoheadrightarrow \ker(\partial_j\colon F^G(S^1)_{n+1}\to F^G(S^1)_{n}).
$$
Thus
$$
\begin{array}{rcl}
\ker(\partial_j\colon F^G(S^1)_{n+1}\to F^G(S^1)_{n})&=&\tilde\phi(R^F_{n+1,\,j})\\
&=&R^G_{n+1,\,j}\\
\end{array}
$$
for $0\leq j\leq n+1$. Note that the faces
$$
\partial_1,\,\ldots\,,\,\partial_n\colon F^F(S^1)_{n+1}\longrightarrow F^F(S^1)_n
$$
are natural projections under the basis $\{y^{(\alpha)}_j \ | \ \alpha\in J,\ 0\leq j\leq n\}$. By Lemma~\ref{techlemma}, the Moore chains of the
simplicial group $\Z[F^F(S^1)]$
$$
N_{n+2}(\Z[F^F(S^1)])=\br^F_{n+2,1}\cap
\br^F_{n+2,\,2}\cap\cdots\cap \br^F_{n+2,\,n+2}=(\br^F_{n+2,\,1}
\br^F_{n+2,\,2}\,\cdots\,\br^F_{n+2,\,n+2})_S.
$$
Thus the Moore boundary
$$
\begin{array}{rcl}
\mathcal{B}_{n+1}(\Z[F^F(S^1)])&=&\partial_0(N_{n+2}(\Z[F^F(S^1)]))\\
&=&\partial_0((\br^F_{n+2,\,1} \br^F_{n+2,\,2}\cdots\br^F_{n+2,\,n+2})_S)\\
&=&(\br^F_{n+1,\,0} \br^F_{n+1,\,1}\,\cdots\,\br^F_{n+1,\,n+1})_S.\\
\end{array}
$$
Now the simplicial epimorphism $\tilde\phi\colon F^F(S^1)\twoheadrightarrow F^G(S^1)$ extends canonically to a simplicial epimorphism
$$
\Z[\tilde\phi]\colon \Z[F^F(S^1)]\twoheadrightarrow \Z[F^G(S^1)],
$$
which induces an epimorphism on the Moore boundaries
$$
\Z(\tilde\phi)|\colon
\mathcal{B}_{n+1}(\Z[F^F(S^1)])\twoheadrightarrow
\mathcal{B}_{n+1}(\Z[F^G(S^1)]).
$$
Thus
$$
\begin{array}{rcl}
\mathcal{B}_{n+1}(\Z(F^G(S^1)))&=&\Z[\tilde\phi]((\br^F_{n+1,\,0} \br^F_{n+1,\,1}\,\cdots\,\br^F_{n+1,\,n+1})_S)\\
&=&(\br^G_{n+1,\,0} \br^G_{n+1,\,1}\,\cdots\,\br^G_{n+1,\,n+1})_S.\\
\end{array}
$$
Note that the Moore cycles
$$
\begin{array}{rcl}
\mathcal{Z}_{n+1}(F^G(S^1))&=&\bigcap\limits_{j=0}^n\ker(\partial_j\colon \Z[F^G(S^1)]_{n+1}\to\Z[F^G(S^1)]_n)\\
&=&\br^G_{n+1,\,0}\cap \br^G_{n+1,\,1}\cap\cdots\cap \br^G_{n+1,\,n+1}.\\
\end{array}
$$
It follows that
$$
\begin{array}{rcl}
Q(\br^G_{n+1,\,0},\,\br^G_{n+1,\,1},\,\ldots\,,\,\br^G_{n+1,\,n+1})&=&\mathcal{Z}_{n+1}(F^G(S^1))/\mathcal{B}_{n+1}(F^G(S^1))\\
&=&\pi_{n+1}(\Z(F^G(S^1)))\\
&\cong&H_{n+1}(F^G(S^1);\Z)\\
&\cong&H_{n+1}(\Omega \Sigma BG;\Z)\\
&\cong&\bigoplus\limits_{k=1}^\infty H_{n+1}((BG)^{\wedge k};\,\Z),\\
\end{array}
$$
where the last isomorphism follows from the classical suspension splitting theorem of loop suspensions~\cite{James}, hence the assertion.
\end{proof}
\para
\begin{cor}
Let $G$ be a group. Then
$$
Q(\br^G_{n+1,\,0},\,\br^G_{n+1,1},\,\ldots\,,\,\br^G_{n+1,n+1})=0
$$
for all $n\geq 0$ if and only if the reduced homology $\tilde H_*(G;\,\Z)=0$.
\end{cor}
\para
\begin{cor}
Let $G$ be a group. Then the groups
$$
Q(\br^G_{n+1,\,0},\,\br^G_{n+1,\,1},\,\ldots\,,\,\br^G_{n+1,n+1})
$$
is torsion-free for all $n\geq 0$ if and only if the integral
homology $\tilde H_*(G;\,\Z)$ is torsion-free.
\end{cor}
\para
\begin{example}
{\rm The group $F=F^{\Z/2}(S^1)_2=\Z/2\ast\Z/2$ is generated by
$x_0,\,x_1$ with defining relations $x_0^2=x_1^2=1$. In this case,
\begin{align*}
& \br^{\Z/2}_{2,\,0}=(\langle x_0\rangle^F-1)\Z[\Z/2\ast\Z/2],\\\ &
\br^{\Z/2}_{2,\,1}=(\langle x_0x_1\rangle^F-1)\Z[\Z/2\ast\Z/2],\\\ &
\br^{\Z/2}_{2,\,2}=(\langle x_1\rangle^F-1)\Z[\Z/2\ast\Z/2]
\end{align*}
with
$$
\begin{array}{rcl}
Q(\br^{\Z/2}_{2,\,0}, \br^{\Z/2}_{2,\,1},\br^{\Z/2}_{2,\,2})&\cong& H_2(\mathbb{R}\mathrm{P}^\infty;\,\Z)\oplus H_2(\mathbb{R}\mathrm{P}^\infty\wedge\mathbb{R}\mathrm{P}^\infty;\,\Z)\\
&=&\Z/2.\quad\Box
\end{array}
$$}
\end{example}
\para\noindent
{\bf Remark 3.2.}
Observe that, for every group $G$, there is the following natural
diagram (see \cite{BL}):
$$
\xyma{H_3(G) \ar@{->}[d] \ar@{->}[r] & \Gamma_2(G_{ab})\ar@{=}[d]
\ar@{->}[r] & \pi_2(\Omega\Sigma BG) \ar@{->>}[r] \ar@{->}[d]& H_2(G)\ar@{->}[d]\\
H_3(G_{ab}) \ar@{->}[r] & \Gamma_2(G_{ab})\ar@{->}[r] &
G_{ab}\otimes G_{ab} \ar@{->>}[r] & H_2(G_{ab}),}
$$
where $\Gamma_2$ is the universal Whitehead quadratic functor. For
a group $G$ with $\ker\{H_2(G)\to H_2(G_{ab})\}=0$ and
torsion-free $G_{ab}$, the natural map $\pi_2(\Omega\Sigma BG)\to
G_{ab}\otimes G_{ab}(\subseteq H_2(\Omega\Sigma BG))$ is a
monomorphism. In particular, this covers the case mentioned in
Remark \ref{rem11}.
|
\section{constant mean curvature annulus meeting spheres
tangentially} In the following, we may assume that the spheres
have radius $1$. Let $\mathcal{A}$ be a compact embedded annulus
with constant mean curvature $H$ and meeting two unit spheres
$S_1$ and $S_2$ tangentially along the boundary curves $\gamma_1$
and $\gamma_2$. We fix the unit normal $N$ of $\mathcal{A}$ to
point away from the centers of the spheres. Let $Y: A(1, R)
\rightarrow \mathbb R^3$ be a conformal parametrization of
$\mathcal{A}$ from an annulus $A(1,R)= \{(x,y)\in \mathbb R^2:
1\le \sqrt{x^2 +y^2} \le R\}$. We define $X$ by $X= Y \circ \exp$
on the strip $B=\{(u,v)\in \mathbb R^2: 0\le u\le \log R\}$. Then
$X$ is periodic with period $2\pi$. Let $z=u+iv$ and $\lambda^2 :=
|X_u| ^2 = |X_v|^2$.
Let $h_{ij}$, $i,j=1,2$, be the coefficients of the second
fundamental form of $X$ with respect to $N$. Note that the Hopf
differential $\phi(z)dz^2=(h_{11}- h_{22} -2i h_{12})dz^2$ is
holomorphic for constant mean curvature surfaces \cite{hopf}. The
theorem of Joachimstahl \cite{dc} says that $\gamma_1$ and
$\gamma_2$ are curvature lines of $\mathcal{A}$. Hence
$h_{12}\equiv 0$ on $u=0$ and $u= \log R$. Since $h_{12}$ is
harmonic and periodic, we have $h_{12}\equiv 0$ on $B$. This
implies that $z$ is a conformal curvature coordinate and $h_{11}-
h_{22}=constant$ \cite{mc}. Let $c=h_{11}-h_{22}$. If
$\mathcal{A}$ is minimal, then we have $K<0$ and $c=2h_{11}>0$ by
the choice of $N$. When $H=-1$, $\mathcal{A}$ is part of the unit
sphere $S_1=S_2$ by the boundary comparison principle for mean
curvature operator \cite{gt}. We assume that $H\not=-1$ in the
following. The principal curvatures of $\mathcal{A}$ are
\begin{equation}\label{pcurv}
\kappa_1= H+ {c \over 2\lambda^2} \mbox{ and } \kappa_2= H- {c
\over 2\lambda^2}.
\end{equation}
We parameterize $\gamma_1$ and $\gamma_2$ by $\gamma_1(v)= X(0,v)$
and $\gamma_2(v)= X(\log R,v)$ for $v\in [0,2\pi)$. In the
following, we assume that $\mathcal{A}$ has non-vanishing Gaussian
curvature.\\
\noindent{\bf Lemma 1.} {\it Each $\gamma_i(v)$, $i=1,2$, has
constant speed $\sqrt{c/ 2(1+H)}$ and $\kappa_2$ is $-1$ on
$\gamma_1$ and $\gamma_2$. As spherical curves, $\gamma_1$ and
$\gamma_2$ are convex. On
$\mathcal{A}\setminus\partial\mathcal{A}$, we have $\lambda^2<{c/
2(1+H)}$ when $K<0$ and $\lambda^2>{c/2(1+H)}$ when $K>0$.} \\
\begin{pf}
The curvature vector of $\gamma_1(v)$ is
\begin{eqnarray}\label{curvature}
\vec{\kappa} &=& {1\over |X_v|}{d \over dv}\left( X_{v}\over |X_v|
\right) = {1\over |X_v|^2}X_{vv} - { X_{v}\over |X_v|^4}(X_v\cdot
X_{vv}) \\
&=& {1 \over \lambda^2} \left( -{\lambda_u \over \lambda} X_u +
h_{22}N\right). \nonumber
\end{eqnarray}
Let the center of $S_1$ be the origin of $\mathbb R^3$. Since
$\mathcal{A}$ is tangential to $S_1$ along $\gamma_1$, we have
$N(0,v)=X(0,v)=\gamma_1(v)$ on $\gamma_1$. Since $\gamma_1$ is on
the unit sphere $S_1$, the curvature vector $\vec{\kappa}$ of
$\gamma_1$ satisfies $(\vec{\kappa}\cdot \gamma_1)(v)=-1$. Hence
we have $\kappa_2={h_{22}\over \lambda^2} =-1$ on $\gamma_1$.
Since $\lambda^2 = |{\gamma_1}_v|^2$ on $\gamma_1$, we have
$|{\gamma_1}_v|=\sqrt{c/ 2(1+H)}$ from \eqref{pcurv}. By choosing
the center of $S_2$ as the origin of $\mathbb R^3$, we get the
results for $\gamma_2$.
The Gaussian curvature $K$ satisfies
\[
\Delta \log \lambda = - K \lambda^2,
\] where $\Delta = {\partial ^2 \over \partial u^2} +{\partial ^2
\over \partial v^2}$. We can rewrite this equation as
\begin{equation}\label{gc}
\lambda \Delta \lambda = |\nabla \lambda|^2 -K \lambda^4.
\end{equation}
Since $\lambda_v(0,v)=0$ and $\lambda_v(\log R,v) =0$ and
$K\not=0$, $\lambda$ does not have interior maximum when $K<0$,
and does not have interior minimum when $K>0$. Since
$\lambda^2={c/2(1+H)}$ on $\gamma_1$ and $\gamma_2$, it follows
that $\lambda^2<{c/2(1+H)}$ on
$\mathcal{A}\setminus\partial\mathcal{A}$ when $K<0$ and
$\lambda^2>{c/2(1+H)}$ when $K>0$. Moreover we have
$\lambda_u\le0$ on $u=0$ and $\lambda_u\ge 0$ on $u=\log R$ when
$K<0$ and $\lambda_u\ge0$ on $u=0$ and $\lambda_u\le 0$ on $u=\log
R$ when $K>0$. Since ${X_u\over |X_u|}\in TS_i$ is perpendicular
to $\gamma_i$, the geodesic curvature of $\gamma_i$ as a spherical
curve is
$\vec{\kappa}\cdot{X_u\over|X_u|}=-{\lambda_u\over\lambda^2}$.
Hence $\gamma_1$ and $\gamma_2$ are convex as spherical curves.
\end{pf}\\
\noindent{\bf Remark 1.} If $\lambda^2\equiv {c/2(1+H)}$ on
$\mathcal{A}$, then $K\equiv0$ and $\mathcal{A}$ is part of a
cylinder.
\section{$-1$-parallel surface}
The $-1$-parallel surface $\tilde{\mathcal{A}}$ of $\mathcal{A}$
is defined by
\[
\tilde{X} = X - N.
\]
The image of $\gamma_1$ (respectively, of $\gamma_2$) in
$\tilde{\mathcal{A}}$ is a point corresponding to the center of
$S_1$ (respectively, of $S_2$). We denote the centers of $S_1$ and
$S_2$ by $O$ and $O_2$ for simplicity. We fix the unit normal
$\tilde{N}$ of $\tilde{\mathcal{A}}$ to be $N$. Since $z=u+iv$ is
a curvature coordinate of $X$, we have
\begin{equation}\label{derivative}
\tilde{X}_u = \left(1+ {h_{11} \over \lambda^2}\right) X_u \mbox{
and } \tilde{X}_v = \left(1+ {h_{22} \over \lambda^2}\right) X_v.
\end{equation}
Since $\kappa_2=-1$ on $\gamma_i$ (Lemma 1), $\tilde{X}$ is
singular for $u=0$ and $u=\log R$. By Lemma 1, we have
$\lambda^2\not= {c/2(1+H)}$ on
$\mathcal{A}\setminus\partial\mathcal{A}$, which implies that
$1+\kappa_2\not=0$ on $\mathcal{A}\setminus\partial\mathcal{A}$.
When $K<0$, we have $\kappa_1> 0$ on
$\mathcal{A}\setminus\partial\mathcal{A}$. Hence $\tilde{X}$ is
regular for $0<u<\log R$ and we have $H>-1$.
Now suppose that $K>0$. Since $\kappa_2=-1$ on $\gamma_i$ (Lemma
1), we have $\kappa_1<0$ and $H<-1/2$. We consider two cases
separately: $H<-1$ and $-1<H<-1/2$. If $H<-1$, then $c<0$
from $\lambda^2 ={c/2(1+H)}>0$ on $\gamma_i$. Hence we have
$\kappa_1<-1$, which implies that $\tilde{X}$ is regular for
$0<u<\log R$. If $-1<H<-1/2$, then we must have $c>0$. This
implies that $1+\kappa_1\not=0$. Otherwise we have
$0<2\lambda^2(1+H) =-c$, which contradicts $c>0$. Hence
$\tilde{X}$ is regular for $0<u<\log R$.\\
\noindent{\bf Remark 2.} When $K<0$ or $K>0$ and $-1<H<-1/2$,
$\mathcal{A}$ stays outside of the balls $B_1$ and $B_2$ bounded
by $S_1$ and $S_2$. If $K>0$ and $H<-1$, then
$\mathcal{A}\subset B_1\cap B_2$.\\
\noindent{\bf Lemma 2.} {\it The mean curvature $\tilde{H}$ and
the Gaussian curvature $\tilde{K}$ of $\tilde{\mathcal{A}}$
satisfies $(1+H)\tilde{K}=(1+2H)\tilde{H} -H$. On
$\tilde{\mathcal{A}}\setminus\{O,O_2\}$, we have
i) if $K<0$ and $H>-1$, then $\tilde{\kappa}_1>0$,
$\tilde{\kappa}_2>1$ and $\tilde{H}>1$,
ii) if $K>0$ and $-1<H<-1/2$, then $0<c/2\lambda^2(1+H) <\min\{1,
-H/(1+H)\}$, $\tilde{\kappa}_1<0$, $\tilde{\kappa}_2<H/(1+H)$ and
$\tilde{H}<H/(1+H)$, and
iii) if $K>0$ and $H<-1$, then $0<c/2\lambda^2(1+H) <1$,
$\tilde{\kappa}_1>(1+2H)/2(1+H)$, $\tilde{\kappa}_2>H/(1+H)$ and
$\tilde{H}>H/(1+H)$.}\\
\begin{pf}
Since
\[
\tilde{h}_{12}=N \cdot \tilde{X}_{uv} = \left(1+ {h_{11} \over
\lambda^2}\right) (N\cdot X_{uv}) =0,
\]
$(u,v)$ is a curvature coordinate (not conformal) for
$\tilde{\mathcal{A}}$ except for $O$ and $O_2$.
We have
\begin{eqnarray*}
\tilde{h}_{11}=N \cdot \tilde{X}_{uu} = \left(1+ {h_{11} \over
\lambda^2}\right)h_{11},\\ \tilde{h}_{22}= N \cdot \tilde{X}_{vv}
=\left(1+ {h_{22} \over \lambda^2}\right)h_{22}.
\end{eqnarray*}
The principal curvatures of $\tilde{\mathcal{A}}$ are
\begin{eqnarray*}
\tilde{\kappa}_1= {\kappa_{1} \over 1+ \kappa_1} =
{{H/(1+H)}+\left(c/2\lambda^2(1+H)\right)\over
{1+\left(c/2\lambda^2(1+H)\right)}},\\
\tilde{\kappa}_2= {\kappa_{2} \over 1+ \kappa_2}=
{{{H/(1+H)}-\left(c/2\lambda^2(1+H)\right)}\over
{1-\left(c/2\lambda^2(1+H)\right)}}.
\end{eqnarray*}
From $\kappa_1+\kappa_2 =2H$, we have
\[
H= {\tilde{H}- \tilde{K} \over 1-2\tilde{H} -\tilde{K}}\mbox{
or } (1+H)\tilde{K}=(1+2H)\tilde{H} -H.
\]
It is straightforward to see that \[
\tilde{H}= {{{H/(1+H)}-\left(c/2\lambda^2(1+H)\right)^2}\over
{1-\left(c/2\lambda^2(1+H)\right)^2}}.
\]
Note that $\kappa_2<0$ on $\mathcal{A}$. First suppose that $K<0$.
Then we have $\kappa_1>0$, which implies that $\tilde{\kappa}_1=
\kappa_1/(1+\kappa_1)>0$. Since $c/2\lambda^2(1+H)>1$ by Lemma 1,
we have $\tilde{\kappa}_2>1$ and $\tilde{H}>1$.
When $K>0$, we have $\kappa_1= H+c/2\lambda^2<0$. If $-1<H<-1/2$,
then we have $c>0$ because $\lambda^2=c/2(1+H)>0$ on $\gamma_i$.
It follows that $c/2\lambda^2(1+H)<-H/(1+H)$. By Lemma 1, we also
have $c/2\lambda^2(1+H)<1$. Therefore we have
$0<c/2\lambda^2(1+H)<\min\{1,-H/(1+H)\}$. It is straightforward to
see that $\tilde{\kappa}_1<0$ and $\tilde{\kappa}_2<H/(1+H)<0$ and
$\tilde{H}<H/(1+H)<0$.
When $K>0$ and $H<-1$, we have $c<0$ and $0<c/2\lambda^2(1+H)<1$.
It is straightforward to see that $\tilde{\kappa}_1>(1+2H)/(1+H)$,
$\tilde{\kappa}_2>H/(1+H)$ and $\tilde{H}>H/(1+H)$.
\end{pf}\\
This lemma says that $\tilde{\mathcal{A}}$ is a linear Weingarten
surface with two singular points $O$ and ${O}_2$ and is positively
curved outside $O$ and $O_2$.\\
\noindent{\bf Lemma 3.} {\it $\tilde{\mathcal{A}}$ is embedded.}\\
\begin{pf} Let $\nu(v)={X_u\over|X_u|}(0,v)$.
Note that $\nu$ is a closed curve in the unit sphere $S_1$. We
claim that $\nu$ is {\it convex as a spherical curve}. Otherwise,
there is a great circle $\eta$ intersecting the image of $\nu$ at
no less than $3$ points $\nu(v_1),\ldots,\nu(v_n)$. ($\nu$ may map
an interval $(v_a, v_b)\subset[0, 2\pi)$ into a single point. We
choose $v_i$'s in such a way that $\nu$ maps no two $v_i$'s to the
same point.) Each $\nu(v_i)$ determines a great circle $\mathbb
S^1_{v_i} \subset S_1$ contained in the plane perpendicular to
$\nu(v_i)$. At each $\gamma_1(v_i)$, $\gamma_1$ is tangent to
$\mathbb S^1_{v_i}$. Since $\eta$ and $\mathbb S^1_{v_i}$ are
perpendicular, $\gamma_1$ cannot be convex when $n\ge3$. Hence
$\nu$ intersect every geodesic of $S_1$ at no more than two
points. This shows that $\nu$ is convex as a spherical curve.
Similarly, ${X_u\over|X_u|}(\log R,v)$ is also convex as a
spherical curve.
Since $\tilde{\mathcal{A}}$ is a parallel surface of
$\mathcal{A}$, the tangent cone $Tan(O,\tilde{\mathcal{A}})$ of
$\tilde{\mathcal{A}}$ at $O$ is the cone formed by rays from $O$
through $\nu$. Since $\nu$ is a convex spherical curve,
$Tan(O,\tilde{\mathcal{A}})$ is convex. This shows that a small
neighborhood of $O$ in $\tilde{\mathcal{A}}$ is embedded and
nonnegatively curved as a metric space \cite{al1}. Similarly,
there is a neighborhood of $O_2$ in $\tilde{\mathcal{A}}$ which is
embedded and nonnegatively curved as a metric space.
Hadamard showed that a closed surface in $\mathbb R^3$ with
strictly positive Gaussian curvature is the boundary of a convex
body \cite{hopf}. In particular, $S$ is embedded. Alexandrov
generalized Hadamard's theorem to nonnegatively curved metric
spaces \cite{al1}. Since $\tilde{\mathcal{A}}$ is a nonnegatively
curved closed metric space, $\tilde{\mathcal{A}}$ is embedded.
\end{pf}\\
\noindent{\bf Remark 2.} We have $\nu_v= {\lambda_u \over
\lambda^2} X_v$. At points where $\lambda_u \not=0$, the curvature
vector of $\nu$ is
\[
\vec{\kappa}_\nu = {1 \over \lambda_u} \left( -{\lambda_u \over
\lambda} X_u + h_{22}N\right).
\]
The geodesic curvature of $\nu$ as a spherical curve
$\vec{\kappa}_\nu\cdot N ={h_{22}\over \lambda_u}$.\\
\section{Main results}
We use the Alexandrov's moving plane argument \cite{al2},
\cite{hopf} to prove the theorems.
\begin{thm}
A compact embedded constant mean curvature annulus $\mathcal{A}$
with non-vanishing Gaussian curvature and meeting two spheres
$S_1$ and $S_2$ of same radius tangentially is part of a Delaunay
surface. In special, if $\mathcal{A}$ is minimal, then
$\mathcal{A}$ is part of a catenoid.
\end{thm}
\begin{pf}
We suppose that the radius of $S_1$ and $S_2$ is $1$. By Lemma 2
and Lemma 3, $\tilde{\mathcal{A}}$ is a compact embedded surface
with two singular points $O$ and $O_2$ and satisfying
$(1+H)\tilde{K}=(1+2H)\tilde{H} -H$ at regular points. A small
neighborhood of a regular point of $\tilde{\mathcal{A}}$ can be
represented as the graph of a function $f(x,y)$ satisfying
\begin{eqnarray}\label{pde}
&& 2(1+H)(f_{xx}f_{yy}-f_{xy}^2) +2H(1+f_x^2+f_y^2)^2
\\ &&= (1+2H)\left((1+f_y^2)f_{xx}-2f_xf_y
f_{xy}+(1+f_x^2)f_{yy}\right)(1+f_x^2+f_y^2)^{1\over2}. \nonumber
\end{eqnarray}
This equation can be rewritten as
\begin{equation}\label{pde2}
\det\left(2(1+H)D^2f + A(Df)\right)=W^4,
\end{equation}
where $A(Df)=-(1+2H)\left(\begin{array}{c}(1+f_x^2)W\\
f_x f_y W\end{array}\begin{array}{c}f_x f_y W\\ (1+f_y^2)W
\end{array}\right)$ and $W=\sqrt{1+f_x^2 +f_y^2}$. The equation
\eqref{pde2} is elliptic with respect to $f$ if $2(1+H)D^2f+A(Df)$
is positive definite. Since
$\det\left(2(1+H)D^2f+A(Df)\right)=W^4>0$, \eqref{pde2} is
elliptic if
\begin{equation}\label{ec1}
\mbox{Tr}\left(2(1+H)D^2f+A(Df)\right)= 2(1+H)\Delta f
-(1+2H)(2+ f_x^2 +f_y^2)W
\end{equation}
is strictly positive.
First we consider the case $K<0$. Since $\tilde{H}>1$ by Lemma 2,
we have
\begin{equation}\label{ec2}
\Delta f + f_y^2f_{xx} -2f_x f_y f_{xy} + f_x^2f_{yy}>2W^{3/2},
\end{equation}
for $f$ representing $\tilde{\mathcal{A}}$. We may assume that $f$
is defined on $B(0,\epsilon) \subset T_p\tilde{\mathcal{A}}$ so
that $\nabla f(0)=\vec{0}$ and $D^2f$ is diagonal. For
sufficiently small $\epsilon=\epsilon(p)$, \eqref{ec2} implies
that \eqref{ec1} is strictly positive. Hence \eqref{pde2} is
elliptic with respect to $f$ representing $\tilde{\mathcal{A}}$.
When $-1<H<-1/2$, \eqref{ec1} is automatically satisfied.
Now we consider the case $K>0$ and $H<-1$. Since
$\tilde{H}>H/(1+H)$ by Lemma 2, we have
\begin{equation}\label{ec3}
\Delta
f+f_y^2f_{xx}-2f_xf_yf_{xy}+f_x^2f_{yy}>{2H\over1+H}W^{3/2}.
\end{equation}
Assuming that $f$ is defined on $B(0,\epsilon)\subset
T_p\tilde{\mathcal{A}}$ with $\nabla f(0)=\vec{0}$ and $D^2f$ is
diagonal, \eqref{ec3} implies that $\Delta f-{1+2H\over2(1+H)}(2+
f_x^2 +f_y^2)W$ is strictly positive for sufficiently small
$\epsilon$. Then $\det\left(-2(1+H)D^2f-A(Df)\right)=W^4$ is
elliptic for $f$ representing $\tilde{A}$. The ellipticity of
\eqref{pde2} for $f$ representing $\tilde{A}$ enables us to use
the maximum principle and the boundary point lemma \cite{gt}.
Since $\tilde{\mathcal{A}}$ is convex and embedded, we can use
Alexandrov's moving plane argument \cite{al2}, \cite{hopf} to show
that $\tilde{\mathcal{A}}$ is rotational as follows. Let
$\Pi_\theta$ be the plane containing the line segment $\overline{O
O}_2\subset \mathbb R^3$ and making angle $\theta$ with a fixed
vector $\vec{E}$ which is perpendicular to $\overline{OO}_2$. Fix
a positive constant $L$ such that each plane $\Pi_\theta^{L}$,
which is parallel to $\Pi_\theta$ with distance $L$ from
$\Pi_\theta$, does not meet $\tilde{\mathcal{A}}$ for all
$\theta$. Let $\Pi_\theta^{l}$ be the plane between
$\Pi_\theta^{L}$ and $\Pi_\theta$ with distance $l$ from
$\Pi_\theta$. When $\Pi_\theta^{l}$ intersects
$\tilde{\mathcal{A}}$, we reflect the $\Pi_\theta^{L}$ side part
of $\tilde{\mathcal{A}}$ about $\Pi_\theta^{l}$. Let us denote
this reflected surface $\tilde{\mathcal{A}}_{l,\theta}^{ref}$. As
we decrease $l$ from $L$, there might be the first $l_\theta\ge 0$
for which $\tilde{\mathcal{A}}_{l_\theta,\theta}^{ref}$ is tangent
to $\tilde{\mathcal{A}}$ at an interior point or at a boundary
point of $\partial\tilde{\mathcal{A}}_{l_\theta,\theta}^{ref}$. We
call this point as the {\it first touch point}. If there is no
nonnegative $l$ with the first touch point, we repeat the process
for $\Pi_{\theta+ \pi}^{L}$ to find $l_{\theta+ \pi}$, which must
be positive. At the first touch point, we apply the comparison
principles for \eqref{pde} to see that the part of
$\tilde{\mathcal{A}}$ in the $\Pi_\theta$ side and
$\tilde{\mathcal{A}}_{l_\theta,\theta}^{ref}$ are identical and,
hence, $l_\theta =0$. This implies that $\Pi_\theta$ is a symmetry
plane for $\tilde{\mathcal{A}}$. Since $\theta$ can be chosen
arbitrarily, $\tilde{\mathcal{A}}$ should be rotational and,
hence, $\mathcal{A}$ is also rotational. Since the Delaunay
surfaces and the catenoid are the only nonplanar rotational
minimal and constant mean curvature surfaces, $\mathcal{A}$ is
part of a a Delaunay surface or part of a catenoid.
\end{pf}\\
We used the embeddedness of $\mathcal{A}$ in proving that
$\tilde{\mathcal{A}}$ is embedded. Whether there is a non-embedded
minimal or constant mean curvature annulus meeting two unit
spheres tangentially is an interesting question. Moreover we raise
the following questions.
1. Is a compact immersed minimal annulus or a compact embedded
minimal or constant mean curvature surface meeting a sphere
perpendicularly or in constant contact angles part of a catenoid
or part of a Delaunay surface? Nitsche showed that an immersed
disk type minimal or constant mean curvature surface meeting a
sphere in constant contact angle is either a flat disk or a
spherical cap \cite{N}.
2. Is a compact immersed minimal annulus or a compact embedded
minimal or constant mean curvature surface meeting two spheres in
constant contact angles part of a catenoid or a plane or part of a
Delaunay surface?
3. Is a compact immersed minimal or constant mean curvature
annulus or a compact embedded minimal or constant mean curvature
surface meeting a sphere and a plane in constant contact angles
part of a catenoid or part of a Delaunay surface? We give an
affirmative answer to this problem in a special case in the
following.
\begin{thm}
A compact embedded constant mean curvature annulus $\mathcal{B}$
with negative (respectively, positive) Gaussian curvature meeting
a sphere tangentially and a plane in constant contact angle $\ge
\pi/2$ (respectively, $\le \pi/2$) is part of a Delaunay surface.
In special, if $\mathcal{B}$ is minimal and the constant contact
angle is $\ge \pi/2$ then $\mathcal{B}$ is part of a catenoid.
\end{thm}
The angle is measured between the outward conormal of
$\mathcal{B}$ and the outward conormal of the bounded domain in
$\Pi$ bounded by the boundary curve. Since the proof of this
theorem is similar to that
of Theorem 1, we omit some details which was previously proved.\\
\begin{pf} Let us denote the sphere by $S_2$ and the plane by
$\Pi$. We may assume that the radius of $S_2$ is $1$. Let $\alpha$
be the constant contact angle between $\mathcal{B}$ and $\Pi$. If
$\alpha =\pi/2$, then we can reflect $\mathcal{B}$ about $\Pi$ to
get a constant mean curvature annulus meeting two unit spheres
tangentially. Hence $\mathcal{B}$ is part of a catenoid or a
Delaunay surface by Theorem 1.
In the following, we assume that $\alpha \not=\pi/2$. As in the
case for $\mathcal{A}$ in $\S 1$, there is a conformal
parametrization $X$ of $\mathcal{B}$ from a strip $\{(u,v)\in
\mathbb R^2:0\le u\le\log R\}$ for which $z=u+iv$ is a curvature
coordinate. We fix the normal $N$ of $\mathcal{B}$ to point away
from the center of $S_2$. Let $c_1(v)=X(0,v)$ be on $\Pi$ and
$c_2(v)=X(\log R,v)$ be on $S_2$ with $\partial X_3/\partial u>0$
along $c_1$. As in Lemma 1, $c_2$ has constant speed
$\sqrt{c/2(1+H)}$ and $\kappa_2=-1$ along $c_2$. Since $K\not=0$
on $\mathcal{B}$ and $z=u+iv$ is a curvature coordinate, we have
$\kappa_2<0$ on $c_1$. The curvature of $c_1$ is
$|\vec{\kappa}|=-\kappa_2/\sin\alpha>0$, which shows that $c_1$ is
locally convex. Since $c_1$ is a Jordan curve, it is convex.
First, we assume that $K<0$ and $\alpha >\pi/2$. Since
${\vec{\kappa}\over |\vec{\kappa}|}\cdot {X_u\over|X_u|}= \cos
\alpha <0$ on $c_1$, it follows from \eqref{curvature} that
$\lambda_u>0$ on $c_1$. Since $\lambda_v(\log R,v)=0$ (cf. Lemma
1), it follows from \eqref{gc} that $\lambda_u\ge0$ on $c_2$.
Otherwise, $\lambda$ will have an interior maximum, which
contradicts \eqref{gc}. Hence we have $\lambda^2<c/2(1+H)$ on
$\mathcal{B}\setminus c_2$. Note that $\kappa_1>0$ and
$\kappa_2<0$ in $\mathcal{B}$. From $\lambda_u\le0$ on $c_2$, we
see that $c_2$ is convex as a spherical curve (cf. Lemma 1).
Arguing as in the proof of Lemma 3, we see that
${X_u\over|X_u|}(\log R,v)$ is also convex as a spherical curve.
When $K>0$ and $\alpha <\pi/2$, we have ${\vec{\kappa}\over
|\vec{\kappa}|}\cdot {X_u\over|X_u|}= \cos \alpha >0$ on $c_1$.
Hence $\lambda_u<0$ on $c_1$. Since $\lambda_v(\log R,v)=0$, it
follows from \eqref{gc} that $\lambda$ does not have interior
minimum. Then we have $\lambda_u\le0$ on $c_2$ and
$\lambda^2>c/2(1+H)$ on $\mathcal{B} \setminus c_2$. Note that
$\kappa_1<0$ and $\kappa_2<0$ in $\mathcal{B}$. From
$\lambda_u\le0$ on $c_2$, it follows that $c_2$ is convex as a
spherical curve. Moreover ${X_u\over|X_u|}(\log R,v)$ is convex as
a spherical curve (cf. Lemma 3).
Let $\tilde{\mathcal{B}}$ be the $-1$-parallel surface of
$\mathcal{B}$. As in $\S 2$, we can show that
$\tilde{\mathcal{B}}$ is regular except for $O_2$: the image of
$c_2$, and $H>-1$ when $K<0$ and $H<-1/2$ when $K>0$. As in Lemma
2, we see that mean curvature $\tilde{H}$ and the Gaussian
curvature $\tilde{K}$ of $\tilde{\mathcal{B}}$ satisfies
$(1+H)\tilde{K}=(1+2H)\tilde{H}-H$ and {\it i) if $K<0$ and
$H>-1$, then $\tilde{\kappa}_1>0$, $\tilde{\kappa}_2>1$ and
$\tilde{H}>1$, ii) if $K>0$ and $-1<H<-1/2$, then
$0<c/2\lambda^2(1+H) <\min\{1, -H/(1+H)\}$, $\tilde{\kappa}_1<0$,
$\tilde{\kappa}_2<H/(1+H)$ and $\tilde{H}<H/(1+H)$, and iii) if
$K>0$ and $H<-1$, then $0<c/2\lambda^2(1+H) <1$,
$\tilde{\kappa}_1>(1+2H)/2(1+H)$, $\tilde{\kappa}_2>H/(1+H)$ and
$\tilde{H}>H/(1+H)$.}
The convexity of ${X_u\over|X_u|}(\log R,v)$ as a spherical curve
implies that there is a neighborhood of $O_2$ in
$\tilde{\mathcal{B}}$ which is embedded and nonnegatively curved
as a metric space. Let $\tilde{\Pi}$ be the plane parallel to
$\Pi$ and containing $\tilde{c}_1$. The curvature of $\tilde{c}_1$
is $|\tilde{\kappa}_2|/\sin\alpha$, which does not vanish. Hence
$\tilde{c}_1$ is locally convex. Using the orthogonal projection
onto $\tilde{\Pi}$, $\tilde{c}_1$ may be considered as a
$\sin\alpha$-parallel curve of $c_1$ in $\tilde{\Pi}$. Hence
$\tilde{c}_1$ is also a convex Jordan curve.
Suppose that $K<0$ and $\alpha>\pi/2$. Since $\kappa_1>0$,
$\tilde{X}_u$ is a positive multiple of $X_u$ by
\eqref{derivative}. The positivity of $\tilde{\kappa}_1$ and
$\tilde{\kappa}_2$ implies that $\tilde{\mathcal{B}}$ meets
$\tilde{\Pi}$ in constant angle $\pi-\alpha$. Suppose that $K>0$
and $\alpha <\pi/2$. If $-1<H<-1/2$, then we have $c>0$ and
$\kappa_1>-1$. Hence $\tilde{X}_u$ is a positive multiple of $X_u$
by \eqref{derivative}. The negativity of $\tilde{\kappa}_1$ and
$\tilde{\kappa}_2$ implies that $\tilde{\mathcal{B}}$ meets
$\tilde{\Pi}$ in constant angle $\alpha$. When $K>0$ and $H<-1$,
we have $c<0$ and $\kappa_1<-1$. Hence $\tilde{X}_u$ is negative
multiple of $X_u$ by \eqref{derivative}. In this case,
$\tilde{\mathcal{B}}$ lies below $\tilde{\Pi}$ and
$\tilde{\kappa}_1$ and $\tilde{\kappa}_2$ are both positive. It is
straightforward to see that $\tilde{\mathcal{B}}$ meets
$\tilde{\Pi}$ in constant angle $\alpha$.
Let $\breve{\mathcal{B}}$ be the singular surface obtained from
$\tilde{\mathcal{B}}$ by attaching the disk in $\tilde{\Pi}$
bounded by $\tilde{c}_1$ to $\tilde{\mathcal{B}}$. Since
$\tilde{\mathcal{B}}$ meets $\tilde{\Pi}$ in acute angle,
$\breve{\mathcal{B}}$ is a nonnegatively curved metric space. By
Alexandrov's generalization of Hadamard's theorem \cite{al1},
$\breve{\mathcal{B}}$ is the boundary of a convex body. Therefore
$\breve{\mathcal{B}}$ is embedded. Note again that $\tilde{H}$,
$\tilde{K}$, $\tilde{\kappa_1}$ and $\tilde{\kappa_2}$ satisfy the
statements of Lemma 2. Hence \eqref{pde} is elliptic for functions
representing $\tilde{\mathcal{B}}$ locally. We can apply
Alexandrov's moving plane argument to $\tilde{\mathcal{B}}$ using
planes perpendicular to $\tilde{\Pi}$ as in the proof of Theorem 1
to see that $\tilde{\mathcal{B}}$ is rotational. Hence
$\mathcal{B}$ is rotational and, as a result, is part of a
Delaunay surface or part of a catenoid.
\end{pf}\\
|
\section{}
\section{Introduction}
\label{sec:intro}
Millisecond pulsars (MSPs) are extremely stable and accurate cosmic
clocks due to their short spin periods and the high degree of
rotational stability. The technique of pulsar timing is the regular
monitoring of the rotation of these objects by tracking the times of
arrival (TOAs) of the radio pulses. Frequent measurements of pulse
arrival times of pulsars provide a powerful tool to determine their
spin and astrometric parameters to a very high degree of precision.
High precision timing of radio millisecond pulsars, most of which are
members of binary systems, can be the key to a number of unanswered
problems in fundamental physics and astronomy, ranging from stellar
evolution to tests of gravitational physics in the strong field regime
and the detection of a cosmological gravitational wave background.
\section{The European Pulsar Timing Array}
\label{sec:EPTA}
The European Pulsar Timing Array (EPTA) network (Table \ref{tab:EPTA})
is a collaboration between the five institutes (ASTRON, JBO, INAF,
MPIfR and Nan\c cay observatory) operating the largest radio
telescopes in Europe. It is consisting of the Effelsberg 100m
radiotelescope of the Max-Planck-Institute for Radioastronomy (MPIfR)
in Germany, the 76m Lovell radiotelescope of the Manchester
University, at Jodrell Bank, UK, the 94m equivalent Westerbork
Synthesis Radio Telescope (WSRT) of ASTRON in the Netherlands, the 94m
equivalent Nan\c cay decimetric radio telescope (NRT) in France and
soon the 64m Sardinia Radio Telescope (SRT) in Italy.
The EPTA is using the available telescopes for high-precision timing
in a coordinated way, in schedules and source lists. This results, in
larger and denser datasets in multiple frequencies. In addition,
exchange of data, people and knowledge between the working groups is
ordinarily happening, resolving swiftly any systematic telescope
problems or other issues. The main aim of the EPTA is to increase the
precision and quality of pulsar timing measurements, to study the
astrophysics of millisecond pulsars and to detect cosmological
gravitational waves from coalescent massive binary black holes in the
nano-Hertz regime.
\begin{table}[!ht]
\center
\caption{The European Pulsar Timing Array collaboration. On the first column the institutes and telescopes and on the second the people participating in the EPTA.}
\begin{tabular}{ll}
\hline
{Institute--Telescope} & {People} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
MPIfR--Effelsberg & M. Kramer, A. Jessner, N. Wex, \\
& P. Freire, D. Champion, K. Lazaridis\\
U. Manchester/Jodrell Bank & B. Stappers, A. Lyne, C. Jordan, \\
--Lovell & G. Janssen, M. Purver, S. Sanidas \\
Nan\c cay observatory--NRT & I. Cognard, G. Theureau, \\
& R. Ferdman, G. Desvignes\\
INAF--SRT & A. Possenti, M. Burgay, M. Pilia \\
ASTRON/U. Amsterdam/ & J. Hessels, Y. Levin, \\
U. Leiden--Westerbork & R. van Haasteren \\
\noalign{\smallskip}
\hline
\end{tabular}
\label{tab:EPTA}
\end{table}
Currently the EPTA is collaborating with the Parkes Pulsar Timing
Array (PPTA) in Australia and the nanoGrav in the USA, forming the
International Pulsar Timing Array (IPTA), a global pulsar timing array
with full sky coverage wishing to lead the way in the detection and
study of gravitational waves.
\subsection{Effelsberg millisecond pulsars}
\label{subsec:Eff}
In the current work, several millisecond pulsars have been chosen from
the Effelsberg source list. On the archival data of those, new
calibration techniques were applied \citep{laz09}, i.e. ephemeris
update, noise removal from individual channels, new time corrections,
summation of individual scans and creation of new synthetic templates
\citep{kwj+94}. This re-calibration resulted in an improved timing
accuracy making the pulsars ideal candidates in the effort of directly
detecting gravitational waves in the nano-Hertz regime. In total
fifteen sources, observed monthly with the Effelsberg radio telescope
at at least two different frequencies, have been chosen. In Table
\ref{tab:14pulsars} the selected sources are shown with the current
post-fit rms of the model fit achieved, in comparison with the
post-fit rms without the improved calibration procedures. In all the
cases the improvement is vast, varying from two times to two orders of
magnitude. In the last column the post-fit rms with the use of only
the 1.4GHz data is shown.
\begin{table}[!ht]
\center
\caption{The fifteen analyzed millisecond pulsars with the post-fit rms with and without calibration. The TOAs are until September 2008 (MJD 54720) and all the available Effelsberg frequencies have been used. On the last column only the 1.4\,GHz TOAs have been used.}
\begin{tabular}{cccc}
\hline
{Source} &{Post-fit rms} & {Post-fit rms} & {Post-fit rms} \\
& {before ($\mu$s)} & {after ($\mu$s)} & {1.4\,GHz ($\mu$s)} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PSR J0030+0451 & ----- $^a$ & 5.7 & 3.8 \\
PSR J0218+4232 & 51.5 & 9.6 & 9.1 \\
PSR J0613$-$0200 & 28.1 & 2.7 & 2.6 \\
PSR J0621+1002 & 23.8 & 6.8 & 6.5\\
PSR J0751+1807 & 15.5 & 4.9 & 4.3\\
PSR J1012+5307 & 3.5 & 2.7 & 2.6\\
PSR J1022+1001 & 16.2 & 3.7 & 3.1\\
PSR J1024$-$0719 & $\sim$1500 & 13.4 & 2.7\\
PSR J1518+4904 & 23.5 & 18.9 & 19.0 \\
PSR J1623$-$2631 & $\sim$2100 & 3.9 & 3.9 \\
PSR J1640+2224 & 15.2 & 1.7 & 1.5\\
PSR J1643$-$1224 & 34.4 & 3.9 & 3.8 \\
PSR J1744$-$1134 & 1.7 & 0.6 & 0.6 \\
PSR J2051$-$0827 & 49.9 & 48.7 & 48.8 \\
PSR J2145$-$0750 & 4.1 & 2.6 & 2.5\\
\noalign{\smallskip}
\hline
\multicolumn{4}{l}{\footnotesize{$^a$: The source was regularly observed from 2008.}}
\end{tabular}
\label{tab:14pulsars}
\end{table}
At least twelve of the analyzed ms pulsars can be good candidates for
the EPTA efforts in detecting the stochastic background of
gravitational radiation. Work is in progress for combination of the
datasets from all the EPTA telescopes for all of these sources, which
will improve the current measurements and take us closer to the
desired detection limit. Until reaching this point the combined
datasets are being used for doing high precision timing analysis of
the individual systems, as it is shown in the following section for
PSR J1012+5307.
\section{PSR J1012+5307}
\label{sec:1012}
\subsection{Introduction}
\label{subsec:intro1012}
PSR J1012+5307 is a 5.3\,ms pulsar in a low eccentricity binary system
with orbital period of $P_b=$14.5\,h \citep{nll+95} and a low mass
helium white dwarf (WD) companion \citep{llfn95}. \cite{cgk98} compared the
measured optical luminosity of the WD to the value expected from WD
models and calculated a distance of $d = 840 \pm 90$\,pc. In addition
they measured, a radial velocity component of $44 \pm 8$\,km\,s$^{-1}$
relative to the solar system barycenter (SSB), and the mass ratio of
the pulsar and its companion $q=m_{p}/m_{c} = 10.5\pm 0.5$. Finally
they derived a companion mass of $m_{c} = 0.16 \pm 0.02\,M_{\odot}$, a
pulsar mass of $m_{p} = 1.64\pm 0.22\,M_{\odot}$ and an orbital
inclination angle of $i=52^\circ \pm 4^\circ$.
\cite{lcw+01} presented a complete precision timing analysis of PSR
J1012+5307 using 4 years of timing data from Effelsberg and 7 years
from Lovell radio telescope. They derived the spin, astrometric and
binary parameters for the system and they discussed the prospects of
future measurements of Post-Keplerian parameters (PK) which can
contribute to the derivation of stringent limits on alternative
gravity theories.
In this work PSR J1012+5307 has been revisited with seven
more years of high-precision timing data and combined
datasets from the EPTA
telescopes. The data have been analyzed using the
timing software TEMPO\footnote{http://www.atnf.csiro.au/research/pulsar/tempo/}.
\subsection{Results}
\label{subsec:results1012}
Combining almost 3000 TOAs, at five different frequencies from the four
telescopes now in use by the EPTA, we have been able to improve on the
timing solution and on all the astrometric, spin and binary
parameters of this system, see Figure \ref{fig:postfit}.
\begin{figure}[!ht]
\plotfiddle{lazaridis_k_fig1.eps}{2.2in}{0}{35}{35}{-120}{-130}
\caption{Post-fit timing residuals using datasets from Effelsberg, Lovell, Nan\c cay and Westerbork.}
\label{fig:postfit}
\end{figure}
For the first time a parallax $\pi = 1.2\pm 0.3$\,mas has been
measured for PSR J1012+5307. This corresponds to a distance of $d =
822 \pm 178$\,pc which is consistent with the $d = 840 \pm 90$ pc
measured from the optical observations. In addition, the proper motion
measurements have been improved by an order of magnitude, with a total
proper motion of $\mu_{t}=25.735\pm0.019$\,mas\,yr$^{-1}$ yielding a
total transverse velocity of $v_t = 102.0\pm 9.8$\,km\,s$^{-1}$. Using
the radial velocity from the optical measurements, the space velocity
of the system $v_{space} = 111.4 \pm 9.5$\,km\,s$^{-1}$ has been
calculated, being consistent and almost three times more precise than
the previous value.
An upper limit of $e < 8.4\times 10^{-7} (95\;{\rm per\;cent}\;{\rm
C.L.})$ has been obtained, being consistent with the evolutionary
scenario of spin-up of PSR J1012+5307 through mass transfer from the
companion, while it is in the red giant phase. A change in the
projected semi-major axis of $\dot{x}_{obs} = 2.3(8)\times 10^{-15}$
has been measured in the current analysis, which has been used to
constrain for the first time the position angle of the ascending node.
As predicted by \cite{lcw+01} a significant measurement of the change
in the orbital period of the system, $\dot{P_b}=5.0(1.4) \times
10^{-14}$, has been obtained for the first time. This is caused by the
Doppler correction, which is the combined effect of the proper motion
of the system \citep{shk70} and a correction term for the Galactic
acceleration and by a contribution due to the quadrupole term of the
gravitational wave emission, as predicted by general relativity
(GR). After subtracting these two contributions from our measured
value, the excess value of $\dot{P}_b^{exc} = (-0.4 \pm 1.6) \times
10^{-14}$ confirms the validity of GR for one more millisecond pulsar
binary system.
All the terms mentioned above are the ones expected to contribute by
using GR as our theory of gravity. However, there are alternative
theories of gravity, that violate the strong equivalence principle (SEP) and
predict extra contributions to the observed orbital period
variation. One is the dipole term of the gravitational wave emission,
which results from the difference in gravitational binding energy of
the two bodies of a binary system. Thus, the case of PSR J1012+5307,
where there is a pulsar-WD system, is ideal for testing the strength
of such emission. One finds for small-eccentricity pulsar-WD systems,
where the sensitivity $s$ (related to the gravitational self-energy of
a body) of the WD is much smaller than the one of the pulsar,
$\dot{P_{b}}^{dipole} = -4\pi^2 \, \frac{T_\odot \mu}{P_b} \,
\kappa_{D} {s_p}^{2}$, \citep{wil01} where $T_\odot = 4.9255\,\mu s$
and $\mu$ is the reduced mass. $s_p$ is the sensitivity of the pulsar
and $\kappa_{D}$ refers to the dipole self-gravitational contribution,
which takes different values for different theories of gravity (zero
for GR).
Another term is predicted by the variation of the locally measured
gravitational constant as the universe expands, $\dot{P}_b^{\dot{G}} =
-2\,\frac{\dot{G}}{G} \left[1 - \left(1+\frac{m_c}{2M}\right)
s_p\right] P_{b} $, \citep{dgt88, nor90} where $M$ is the total mass
of the system. It has been shown that there is no need to add these
extra contributions to explain the variations of the orbital period,
however the excess value has been used to set limits for a wide class
of alternative theories of gravity.
PSR J1012+5307 is an ideal lab for constraining the dipole radiation
term because the WD nature of the companion is affirmed optically, the
mass estimates are free of any explicit strong-field effects and the
mass of the pulsar is rather high, which is important in the case of
strong field effects that occur only above a certain critical mass,
like the spontaneous scalarisation \citep{de93}. Thus, by using the
$\dot{G}/G = (4 \pm 9) \times 10^{-13}$\,yr$^{-1}$ limit from the Lunar
Laser Ranging (LLR) \citep{wtb04} the $\dot{G}/G$ contribution has been
calculated and subtracted from our excess value in order to finally
obtain an improved generic limit for the dipole contribution of
$\kappa_D = (0.2 \pm 2.4) \times 10^{-3}$ (95 per cent C.L.).
A generic test for $\dot G$ cannot be done with a single binary
pulsar, since general theories that predict a variation of the
gravitational constant typically also predict the existence of dipole
radiation. This degeneracy has been broken here in a joint analysis of
PSR J1012+5307 and PSR J0437$-$4715 \citep{vbv+08}, two binary
pulsar-WD systems with tight limits for ${\dot P}_b$ and different
orbital periods. By applying equation $ \frac{\dot{P}_b^{exc}}{P_b} =
- 2 \frac{\dot G}{G} \left[1 - \left(1+\frac{m_c}{2M}\right)
s_p\right] - 4 \pi^2 \frac{T_\odot\mu}{P_b^2} \, \kappa_{\rm D}
s_p^2 \nonumber\\$ to both binary pulsars, and solving in a
Monte-Carlo simulation (Figure \ref{fig:monteKG}) this set of two
equations, stringent and generic limits based purely on pulsar data
and in the strong field regime have been obtained. With a 95 per cent
C.L., $\frac{\dot G}{G} = (-0.7 \pm 3.3) \times 10^{-12} \; {\rm
yr}^{-1}$ and $\kappa_{\rm D} = (0.3 \pm 2.5) \times 10^{-3}$. In
the future, more accurate measurements of $\dot{P_b}$ and distance of
the two pulsars and WDs could constrain even more our derived limits.
\begin{figure}[!ht]
\plotfiddle{lazaridis_k_fig2.eps}{2.2in}{0}{40}{40}{-110}{-20}
\caption{Contour plots of the one and two $\sigma$ confidence regions on $\dot
G/G$ and $\kappa_D$ jointly. The elongation of the regions reflects the
correlation due to the mutual dependence of the two pulsar-WD systems
in this combined test.}
\label{fig:monteKG}
\end{figure}
A more detailed work on this study of
PSR J1012+5307 can be found in \cite{lwj+09}.
\acknowledgements
We are grateful to all the EPTA collaborators for providing data, for
valuable discussions and for their overall help. ln addition, to all
staff at the EPTA radio telescopes for their help with the
observations. Kosmas Lazaridis was supported for this research through
a stipend from the International Max Planck Research School (IMPRS)
for Astronomy and Astrophysics at the Universities of Bonn and
Cologne.
|
\section{Introduction}
In the field of glassy slow dynamics, many experiments and simulations
have been inspired in recent years by the mode-coupling theory for
idealized glass transitions (MCT) \cite{Goetze2009}. Within this theory,
the transition from a liquid to an idealized glass state is described by a
bifurcation in the solutions of certain polynomials in a variable $f$:
Upon smooth variations of a control parameter, say, the temperature $T$,
variable $f$ jumps from $f=0$ to a finite critical value $f=f^c>0$ at
$T=T_c$ and increases further with the distance from the transition
$T_c-T$. Variable $f$ is defined by the long-time limit of some
autocorrelation function $\phi(t)$; states with $f>0$ are identified as
glass states, while states with $f=0$ shall be called fluid or ergodic.
The mentioned bifurcations originate from the equations of motion for
density autocorrelation functions $\phi_q(t)$ of a system of $N$ particles
in a volume $V$ with density $\rho = N/V$. The time-dependent fluctuations
in the density are used to define the canonical normalized correlation
function $\phi_q(t) = \langle \rho(t)^*\rho \rangle/\langle \rho^*\rho
\rangle$ as is well-known in liquid-state theory \cite{Hansen1986}. In
this case, the correlation functions depend on the wave vector modulus $q$
of the corresponding Fourier transform in space. Using projection-operator
techniques, one can derive the following equations of motion
\begin{subequations}\label{eq:eom}
\begin{equation}\label{eq:eom:phi}
\partial_t^2\phi_q(t) + \nu_q\partial_t\phi_q(t) + \Omega_q^2 \phi_q(t)
+ \Omega_q^2\int\,dt'\ m_q(t-t') \partial_{t'}\phi_q(t') = 0\,,
\end{equation}
with characteristic frequencies $\Omega_q$, a white noise $\nu_q$, and the
memory function
\begin{equation}\label{eq:eom:F}
m_q(t) = {\cal F}[\phi_k(t),V] =
\frac{1}{2}\int\frac{d^3k}{(2\pi)^3} V_{\vec{q},\vec{k}}
\phi_k(t)\phi_{|\vec{q}-\vec{k}|}(t)\,,
\end{equation}
where the interaction potential is encoded in the vertex
\begin{equation}\label{eq:eom:V}
V_{\vec{q},\vec{k}} = \rho S_qS_kS_{\vec{q},\vec{k}}\rho
\left[
\vec{q}\vec{k}c_k+\vec{q}(\vec{q}-\vec{k})c_{|\vec{q}-\vec{k}|}
\right]^2/q^4\,,
\end{equation}
through the static structure factor $S_q$ of the fluid and $c_q$ its
direct correlation function. Both functions can be calculated from the
interaction potential together with some closure relation that is known
typically only in some approximation \cite{Hansen1986}.
For $t\rightarrow\infty$, one gets an algebraic equation for the
correlators' long time limits $\phi_q(t)\rightarrow f_q$,
\begin{equation}\label{eq:eom:fq}
\frac{f_q}{1-f_q} = {\cal F}[f_k,V]\,.
\end{equation}
\end{subequations}
It was discovered in 1984 that Eq.~({\ref{eq:eom:fq}) can exhibit
nontrivial solutions, $f_q > 0$, for microscopic interaction potentials
and realistic values of the density in the system \cite{Bengtzelius1984}.
It was shown later that only singularities of type $A_\ell$ can occur in
Eq.~({\ref{eq:eom:fq}) \cite{Goetze1995b}, and these singularities are
equivalent to those emerging for the parameter space of roots of
polynomials upon variation of the coefficients \cite{Arnold1986}. Hence,
an $A_2$ singularity -- also called \textit{fold} -- signals a double root
in the solutions like in the equation $x^2 + t = 0$ for $t_c = 0$. For a
polynomial of high enough order, the variation of a single control
parameter is sufficient to encounter an $A_2$ singularity. Generically,
the variation of $\ell - 1$ control parameters is necessary to identify
singularities of type $A_\ell$. An $A_3$ singularity -- called
\textit{cusp} -- requires two control parameters to adjust a polynomial to
a cubic root; an $A_4$ singularity -- the \textit{swallowtail} -- requires
the variation of three parameters. It was shown in a theorem by Whitney
that only $A_2$ and $A_3$ are robust singularities, all other
singularities can be removed by small perturbations of the control
parameters \cite{Whitney1955,Arnold1986}.
Within MCT, the $A_2$ singularity can be identified with a liquid-glass
transition if $f$ jumps from $f=0$ to a finite value $f^c$ at the
transition. Once such a singularity is identified, asymptotic expansions
of Eq.~(\ref{eq:eom}) can be used to derive the long-time behavior of the
correlator $\phi(t)$. For the $A_2$ singularity, these asymptotic
expansions yield two-step relaxation, time-temperature superposition, and
power-law scaling \cite{Goetze2009}. The unique number characterizing the
leading terms of the asymptotic expansion for an $A_2$ singularity is the
exponent parameter $\lambda$ which is between 0.5 and unity. In addition
to liquid-glass transitions, fold singularities can also describe
glass-glass transitions: In this case an existing first glass state with
$f\geq f_1^c$ transforms into a second distinct glass state with $f\geq
f_2^c > f_1^c$ discontinuously. The endpoint of a line of glass-glass
transition points is the $A_3$ singularity. An $A_4$ singularity signals
the emergence of a glass-glass transition line from an otherwise smooth
surface of liquid-glass transitions. Every $A_\ell$ singularity is
characterized by a unique number $\mu_\ell \geq 0$ that determines the
properties of the asymptotic expansions \cite{Goetze2002}.
$\mu_\ell\rightarrow 0$ signals the emergence of the higher-order
singularity $A_{\ell+1}$ where in turn $\mu_{\ell+1}$ defines the leading
terms of the expansion. An $A_2$ singularity's exponent parameter is
identical to $\mu_2=1-\lambda$. Therefore, a fold gives rise to a cusp
once $\lambda$ approaches unity.
\section{Glass Transitions, Glass-Glass Transitions, and Higher-Order
Singularities}\label{sec:rev}
In the following, the transition singularities of MCT shall be reviewed
briefly for hard spheres (sec.~\ref{sec:HSS}), sticky hard spheres
(sec.~\ref{sec:SHSS}), and the square-well potential (sec.~\ref{sec:SWS}).
The section~\ref{sec:rev} shall be concluded with the discussion of newly
discovered transitions in the square-shoulder system (sec.~\ref{sec:SSS})
where the static structure factors are evaluated in Roger-Young (RY)
approximation. In sec.~\ref{sec:SSSPY}, new results shall be given for the
square-shoulder system with the static structure factors calculated from
the Percus-Yevick (PY) approximation. The comparison between results from
RY- and PY-calculations gives insight into the robustness of the predicted
glass-glass transition phenomena.
\subsection{Glass Transition for the Hard-Sphere System -- a Fold}
\label{sec:HSS}
MCT was first applied to the hard-sphere system (HSS) where a glass
transition was identified upon varying the control parameter packing
fraction $\varphi = \pi\rho d^3/6$ for particles of hard-sphere diameter
$d$ \cite{Bengtzelius1984}. The transition is predicted for a packing
fraction of $\varphi^c = 0.516$ when using the PY approximation for the
calculation of the static structure factor $S_q$ \cite{Hansen1986}. The
hard-sphere interaction can be realized to a very good degree in
experiments performed in colloidal suspensions. In such experiments the
glass transition is found around a packing fraction $\varphi^c = 0.58$,
moreover, the two-step relaxation, scaling laws and several other features
of the theoretical predictions are confirmed \cite{Megen1991}. From a
thorough analysis of Eq.~(\ref{eq:eom:V}) one can derive that for the HSS,
the glass transition is driven by the hard-core repulsion as encoded in
the principal peak of the static structure factor. The HSS has been
investigated by asymptotic expansions \cite{Franosch1997} with the
exponent parameter being around $\lambda = 0.7$ which is close to the
typical value for many other glass forming substances. The full evolution
of glassy dynamics over eight orders of magnitude in time for various
densities has been demonstrated for a tagged particle's mean-squared
displacement (MSD) \cite{Sperl2005}.
\subsection{Glass-Glass Transition in the Sticky Hard-Sphere System -- a
Cusp}
\label{sec:SHSS}
For the investigation of an $A_3$~singularity, more than one control
parameter needs to be varied. Generically the variation of two control
parameters extends an isolated $A_2$~singularity to a line of
$A_2$~singularities, so the existence of an $A_3$ singularity is by no
means certain. However, if such an $A_3$ singularity exists in the given
parameter plane, it can be found as an endpoint of a line of
$A_2$~singularities where close to the $A_3$, these $A_2$~singularities
can be identified as glass-glass transition points. At the $A_3$, the
distinction between the different glass states, the discontinuity between
the $f_q$ on both sides of the glass-glass transition, vanishes. Such
endpoint singularities have been described first for so-called schematic
models of MCT where the microscopic details, i.e., the $q$-dependence, has
been dropped in favor of mathematical simplicity. Rather than a precise
interaction potential these models capture the mathematical structure of
the problem. In these schematic models the transition points can be
calculated analytically and the asymptotic expansions are not affected by
limited numerical accuracy. For such schematic models, the asymptotic
expansions for singularities $A_\ell$ with $\ell > 2$ have been performed:
Two-step relaxation and time-temperature superposition become invalid, and
-- most notably -- logarithmic decay laws emerge
\cite{Goetze1989d,Goetze2002}.
The first microscopic model with an $A_3$~singularity in its parameter
plane was found for Baxter's sticky hard-sphere model (SHSS)
\cite{Baxter1968b}. The SHSS adds a short-ranged attraction to the
hard-sphere potential similar to the square-well interaction, cf. Fig.
\ref{fig:sss_pot}~(a). However, within the SHSS, depth $\Gamma$ and width
$\delta$ can only be changed together while the product $\Gamma\delta$
remains fixed, in addition the limit $\delta\rightarrow 0$ is performed;
this defines a so-called stickiness parameter $\tau$. Hence, the SHSS has
two control parameters, packing fraction $\varphi$ like the HSS and the
stickiness $\tau$. For the SHSS, the MCT predicts glass-glass transitions
and an $A_3$~endpoint singularity \cite{Fabbian1999,Bergenholtz1999}. The
glass-glass transitions take place between a repulsion-dominated glass
state -- as known from the HSS -- and an attraction-dominated glass state
that resembles bond formation as known from the gelation transition
\cite{Bergenholtz1999}. While for a treatment of the SHSS within MCT, a
cutoff wave-vector space needs to be introduced, it is by now well-known
how this cutoff can be interpreted as an inverse length scale, and how the
MCT results stay well-defined \cite{Goetze2003b}. In accordance with
Whitney's theorem, for small changes in the cutoff neither fold nor cusp
singularities change qualitatively.
\begin{figure}[htb]
\centerline{
\includegraphics[width=.45\columnwidth]{sws_pot_mod.eps}
\includegraphics[width=.45\columnwidth]{sss_pot_mod.eps}
}
\caption{\label{fig:sss_pot}(a) Square-well potential with three control
parameters: packing fraction $\varphi = \pi\rho d^3/6$, well depth $\Gamma
= u_0/k_\text{B}T$, and well width $\delta$ for particles of diameter $d$
at density $\rho$.
(b) Square-shoulder potential with shoulder height $\Gamma =
u_0/k_\text{B}T$ and shoulder width $\delta$.
}
\end{figure}
\subsection{Glass-Glass Transitions in the Square-Well System -- a
Swallowtail}\label{sec:SWS}
In order to avoid entirely the introduction of a cutoff and other
peculiarities of the SHSS, one can extend the model attraction to finite
widths in the square-well system (SWS), cf. Fig.~\ref{fig:sss_pot}~(a).
Here, the control-parameter space becomes truly three dimensional, the
parameter triple $(\varphi, \Gamma, \delta)$ defines each state. The MCT
glass-transition scenarios have been worked out for the SWS
\cite{Dawson2001}, and in addition to the cusp scenario of the SHSS there
exists a characteristic well width $\delta^*$ for the SWS above which no
cusp singularity can be found in the $(\varphi, \Gamma, \bar\delta)$
parameter plane for fixed $\bar\delta$ when $\bar\delta > \delta^*$. When
$\bar\delta < \delta^*$, the SWS always exhibits a glass-glass transition
line with an $A_3$~singularity as endpoint. For the exceptional point
$\delta = \delta^*$, the glass-glass transition lines together with the
line of $A_3$ singularities vanish in an $A_4$~singularity, giving rise to
a three-dimensional geometric structure known as swallowtail. In contrast
to the stable $A_2$ and $A_3$ singularities, the $A_4$ is an isolated
point in the parameter space that is sensitive to small numerical
deviations from $(\varphi^*, \Gamma^*, \delta^*)$. But while the details
of the approximations involved and the numerical implementation change the
location of the $A_4$~singularity, its existence is a robust prediction of
MCT for systems with short-ranged attraction regardless of the specific
interaction model \cite{Goetze2003b} or closure relation for the static
structure factors \cite{Dawson2001}.
The occurrence of the glass-glass transitions can be traced to two
different mechanisms of arrest -- resulting from the interplay of
repulsion and attraction -- in the vertex, cf. Eq.(\ref{eq:eom:V}). The
repulsion-dominated glass state within MCT is produced by the local
structure on the wave-vector scale of the peak of $S_q$. The second
mechanism of arrest is given by a $1/q$-tail in $S_q$ for large wave
vectors $q$ \cite{Dawson2001}. Once that tail has enough weight in the
vertex, a second transition -- the glass-glass transition -- can occur.
This possibility of an additional transition is robust regarding the
closure relation as long as the tail can become large enough. This
independence of the results from the closure relation was demonstrated
explicitly for the PY compared with the mean-spherical approximation
\cite{Dawson2001,Sperl2004}.
The asymptotic solutions have been worked out in detail for the SWS and
involve logarithmic decay laws \cite{Sperl2003a}, Vogel-Fulcher-like
divergence of time-scales \cite{Sperl2004b}, unconventional (i.e.,
non-power-law) critical decays at the higher-order singularities
\cite{Goetze2004c}, and the interplay of two $A_2$ singularities at the
crossing of glass- and gel-transition lines \cite{Sperl2004}. Experimental
verifications of the scenarios predicted for the SWS are found in computer
simulations and in colloidal suspensions with attraction among the
particles. Confirmations include the reentrant behavior of the lines of
$A_2$~singularities \cite{Eckert2002,Foffi2002b,Pham2002}, the dynamics at
a crossing of glass- and gel-transition lines for micellar
\cite{Chen2002,Chen2003b} and colloidal suspensions \cite{Pham2004}, and
the logarithmic decay of correlation functions together with novel power
laws for the MSD \cite{Sciortino2003,Sperl2003a}.
\subsection{Glass-Glass Transitions in the Square-Shoulder System}
\label{sec:SSS}
When the attractive well of the SWS is replaced by a repulsive step, one
obtains the square-shoulder system (SSS), cf. Fig.~\ref{fig:sss_pot}~(b).
The SSS also has three control parameters, $(\varphi, \Gamma, \delta)$,
with $\Gamma$ now symbolizing the height of a shoulder. Allowing the
control parameter $\Gamma$ to carry a sign, SWS and SSS may even be
plotted into a single diagram. Recently, the MCT predictions for the SSS
have been worked out using the RY approximation \cite{Rogers1984} for
$S_q$ \cite{Sperl2009}. Further details and references to the numerical
algorithms can be found in recent work \cite{Sperl2009}.
\begin{figure}[htb]
\centerline{
\includegraphics[width=.67\columnwidth]{PDl145.eps}
}
\caption{\label{fig:sssry}Glass-transition diagram (lower panel) and
exponent parameter $\lambda$ for the SSS using the RY approximation
at $\delta = 0.145$. Diamonds show the transition points from an ergodic
state to a glass state; filled circles exhibit the glass-glass transition
points. Dotted lines show the limits for the hard-sphere system with
diameter $1+\delta$.
}
\end{figure}
Figure~\ref{fig:sssry} shows the results of MCT for the SSS at $\delta =
0.145$: For $\Gamma = 0$, the glass transition line (lower panel) emerges
from the HSS value of $\varphi_\text{HSS}^\text{RY} = 0.5206$, increases
in density until around $\Gamma = 3$, bends over in an S-shape towards the
HSS limit for the outer core, $\varphi = 0.5206/(1+0.145)^3 = 0.3468$. The
form of the transition curve can be understood in detail from the
distortions of the local structures by the presence of the repulsion at
distance $1+\delta$ and the resulting changes to the principal peak of the
static structure factor \cite{Sperl2009}. For both reentrant transitions
it can be shown how the weakening of the local structure is compensated by
either higher density (in the case of melting by cooling, i.e., melting at
increasing $\Gamma$) or lower temperature which is equivalent to higher
$\Gamma$ (in the case of melting by compression). In the
pressure-temperature representation, such a reentrant behavior can be
identified with a so-called diffusion anomaly which is known
experimentally e.g. for water \cite{Angell1976}. The exponent parameter
$\lambda$ (upper panel) varies very little along the glass-transition line
and stays around the common value of $\lambda = 0.7$. From this
unremarkable value for $\lambda$ no conclusion can be drawn about any
glass-glass transitions or higher-order singularities in the vicinity.
Moving further into the glass state, $f_q$ experiences an additional jump
marked by the full circles in Fig.~\ref{fig:sssry}. In contrast to the
situation for the SWS, for the SSS the glass-glass-transition line is
located completely within the glassy state. This gives rise to two
endpoints, two $A_3$ singularities. These two endpoints are seen also in
the upper panel of Fig.~\ref{fig:sssry} where $\lambda$ approaches unity
on either side. This novel line originates from the competition of the two
repulsive cores that causes a beating in the static structure factor for
large wave vectors; this beating -- if sufficiently large in amplitude --
can bring additional weight to the MCT vertex in Eq.~(\ref{eq:eom:V}) and
hence cause an additional discontinuous transition \cite{Sperl2009}. In
physical terms, the localization of the already arrested particles drops
drastically when the glass-glass transition line is crossed since upon
crossing the line, the localization now happens at the outer core rather
than at the inner core as before. When varying the control parameters, the
beating is most pronounced when inner and outer core have about the same
impact on $S_q$ -- if either one of the cores is dominant, the beating is
diminished and the discontinuous transition ends in two $A_3$
singularities, respectively. Physically, the glass-glass transition
becomes impossible at the upper endpoint when the density becomes too high
for a transition to the outer core; the glass-glass transition also
vanishes at the lower endpoint because the density becomes too low at the
respective shoulder height to force the particles closer together. For
shoulder widths $\delta$ larger than the value shown in
Fig.~\ref{fig:sssry}, the glass-glass transition line moves towards and
merges with the glass-transition line \cite{Sperl2009}.
\section{Square-Shoulder System using the Percus-Yevick Approximation}
\label{sec:SSSPY}
In this section, it is demonstrated how the scenario shown in
sec.~\ref{sec:SSS} changes when the mentioned beating has not enough
weight in the vertex~(\ref{eq:eom:V}) to cause an additional transition
line.
\begin{figure}[htb]
\centerline{\includegraphics[width=.67\columnwidth]{PDPYall.eps}}
\caption{\label{fig:PDPYall}Glass-transition scenario (transition lines in
the lower and exponent parameter $\lambda$ in the upper one) for the SSS
using the PY approximation for $\delta = 0.25$ (dotted line), 0.2785
(full line), 0.28 (dashed line), and 0.3 (chain line).
In the lower panel, the curves for $\delta = 0.2785$ and 0.28 are almost
on top of each other.
}
\end{figure}
Figure~\ref{fig:PDPYall} shows the glass-transition scenarios for various
shoulder widths of the SSS when the PY approximation is used for the
calculation of the static structure factor. The overall behavior of the
transition diagram is similar to the RY results: The glass-transition
first increases in packing fraction $\varphi$ when starting from the HSS
limit $\varphi_\text{HSS} = 0.516$ at $\Gamma = 0$. Around $\Gamma = 2.5$
the curve bends downwards and reaches the limit of the HSS with an outer
core of $\varphi_\text{HSS}/ (1+\delta)^3$ at around $\Gamma = 5$.
Different from the RY result, the S-shape of the transition curve for
intermediate values of $\Gamma$ only develops at higher values for the
width $\delta$. Also in contrast to the RY result, there appears to be no
indication of additional discontinuities in the $f_q$ or endpoint
singularities inside the glass regime. However, while $\lambda$ in the
upper panel of Fig.~\ref{fig:sssry} stays almost constant at the
glass-transition line for the RY results, for the glass-transition line
within PY approximation the exponent parameter $\lambda$ varies
considerably with $\Gamma$ for any given width $\delta$ in the upper panel
of Fig.~\ref{fig:PDPYall}. In addition, the shape of the
$\lambda$-versus-$\Gamma$ curves varies drastically with $\delta$. For
$\delta = 0.25$, an exponent-parameter maximum is around $\lambda = 0.85$
which is already rather high. For larger $\delta$, the
$\lambda$-versus-$\Gamma$ curves exhibit double maxima over relatively
small intervals in $\Gamma$, like between 1.85 and 2.1 for $\delta$
around 0.28. For $\delta = 0.2785$ and $\delta = 0.28$ the parameter
$\lambda$ approaches unity very closely at the right and the left end of
the interval, respectively. Beyond that regime, $\lambda$ decreases again
for larger $\delta$, and as shown for $\delta = 0.3$, the separation of
both $\lambda$ maxima increases; the $\lambda$ maxima are located at
$\Gamma = 1.65$ and $\Gamma = 2.4$, respectively, while their values are
still as high as $0.9$. In summary, while no glass-glass line can be
detected, the exponent parameter approaches $\lambda = 1$ very closely.
These results indicate a nearby higher-order singularity without the
presence of glass-glass transitions. More puzzling still is the existence
of two maxima in $\lambda$ very close to each other.
\begin{figure}[htb]
\centerline{\includegraphics[width=.67\columnwidth]{PDPY25.eps}}
\caption{\label{fig:PDPY25}Glass-transition scenario for the SSS for
$\delta = 0.25$. Diamonds show the $A_2$ singularities at glass-transition
points. The open circles mark the position of an anomaly in the evolution
of the $f_q$, see text -- a line of hidden glass-glass transition points
emerges.
}
\end{figure}
To clarify the nature of the higher-order singularities on the
glass-transition curves, the glassy region is inspected in more detail for
$\delta = 0.25$ in Fig.~\ref{fig:PDPY25}. The behavior of $\lambda$
indicates that here the higher-order singularities are further away from
the transition line, but a rather broad maximum already hints at these. It
is well-known that after crossing an $A_2$ glass-transition singularity,
the long-time limits $f_q$ increase above their critical value $f_q^c$ in
a square-root in the control parameter; i.e. in the HSS, this increase is
proportional to $\sqrt{\varphi-\varphi^c_\text{HSS}}$ \cite{Franosch1997}.
When a glass-glass transition occurs, this square-root increase after the
first transition is superseded by an additional discontinuity in $f_q$
followed by the square-root increase after this second transition. While
this additional discontinuity was used to identify the glass-glass
transition line in Fig.~\ref{fig:sssry}, such discontinuity is absent for
the PY calculations. Nevertheless, the evolution of the $f_q$ within PY
approximation shows a characteristic deviation from the square-root
behavior in certain parameter regions: First, the range of validity of the
square-root increase is sometimes far smaller than known from the HSS;
second, the apparent square-root increase at larger distances from the
glass-transition indicates a square-root with a different extrapolated
transition point $\varphi^\text{app}$ than the one given by the
discontinuity in $f_q$ at $\varphi^c$. This difference between the actual
and the apparent transition point can be quantified as a difference in the
control parameters, say as a relative difference in packing fraction,
$\Delta\varphi = (\varphi^\text{app}-\varphi^c)/\varphi^c$. When this
relative difference in packing fraction exceeds 1\%, this anomaly is
marked by open circles in Fig.~\ref{fig:PDPY25}. It is seen that this line
of $f_q$ anomalies strongly resembles the line of glass-glass transitions
in Fig.~\ref{fig:sssry}. It can therefore be concluded that within MCT
both approximations, PY and RY, yield similar glass-transition scenarios
with a possible line of glass-glass transitions inside the glassy regime.
\section{Conclusion}
In the present work, the glass-transition diagram for the SSS has been
calculated for the PY approximation. These results can now be used to
estimate the robustness of the results for the SSS obtained with a
different closure relation, e.g. with results from the RY approximation
\cite{Sperl2009}. While the comparison of different closure relations for
the SWS only resulted in shifts of the control parameters, the situation
is more involved in the case of the SSS. The existence of the disconnected
glass-glass transition line depends on two trends that both become more
prominent for larger values of $\delta$. First, the vertex in
Eq.~(\ref{eq:eom:V}) obtains additional weight for higher wave vectors
through a beating phenomenon as described above; this makes a glass-glass
transition possible if the weight becomes strong enough. Second, at the
same time this supposed glass-glass transition line moves closer to the
glass transition line and merges with it. For the RY approximation, the
glass-glass transition line becomes manifest through a discontinuity in
$f_q$ and then moves towards the glass transition line. In contrast for
the PY approximation, the not-yet-manifest glass-glass transition moves
towards the glass transition line and merges with it before it is fully
developed as a discontinuity in the $f_q$. When the endpoints of this
hidden line of glass-glass transitions merge with the regular glass
transition line, very high values of $\lambda$ result since the first
trend of increased weight for a glass-glass transition keeps increasing.
In this sense, the glass-glass transition line emerges extremely close and
on top of the glass transition line and is at the same moment absorbed by
the glass transition line.
The differences between RY and PY approximation with respect to the
higher-order singularities and the corresponding glass-glass transitions
are highly non-trivial in the theoretical calculations. However, for the
experimental test if such a scenario exists, both scenarios need to be
observed in combination: Numerical deviations similar to the ones shifting
the HSS-PY transition from $\varphi^c = 0.516$ to 0.58 in the experiment
can for the SSS switch from the RY scenario shown earlier \cite{Sperl2009}
to the PY scenario described here.
\section*{Acknowledgments}
This work was partially supported by Yukawa International Program for
Quark-Hadron Sciences (YIPQS). Support from BMWi under 50WM0741 is
gratefully acknowledged. I want to thank W.~G\"otze, J.~Horbach, P.~Kumar,
F.~Sciortino H.~E.~Stanley and E.~Zaccarelli for fruitful collaboration
and discussion of the work.
|
\section{Introduction}
Particle polarizability governs the electric response for many inhomogeneous systems ranging from biological cells
to plasmonic nanoparticles and depends strongly on both its dielectric and geometric properties.
Analytical models have been reported \cite{Pauly1959, Asami1980} only for spherical and ellipsoidal geometries,
whereas more complex geometries have been approached by direct numerical solution of the field equations
using, for example, the finite difference methods \cite{Asami2006}, the finite element method \cite{Fear1998},
the boundary element method \cite{Sekine2005,Sancho2003}, or the boundary integral equation (BIE) \cite{Brosseau2003}.
In a simplified representation, biological cells can be regarded as homogeneous particles (cores)
covered by thin membranes (shells) of contrasting electric conductivities and permittivities.
Complex geometries occur when cells are undergoing division cycles (e.g. budding yeasts) or are
coupled in functional tissues (e.g. lining epithelia or myocardial syncytia).
In these cases, the dielectric/impedance analysis of cellular systems is far more complicated than
previous models
\cite{Schwan1996,Gheorghiu1993,Gheorghiu1994},
which considered suspensions of spherical particles.
Intriguing dielectric spectra \cite{Knapp2005}
reveal distinct dielectric dispersions with time evolutions consistently related to
tissue functioning or alteration, identifying a possible role of cell connectors
(gap junctions) in shaping the overall dielectric response.
A direct relation between the microscopic parameters and experimental data can be analytically
derived only for dilute suspensions of particles of simple shapes, and is rather challenging for
system with more realistic shapes,
where only purely numerical solutions have been available.
In this work we demonstrate that a spectral representation of a BIE
provides the analytical structure for the polarizability of particles with a wide range of shapes and structures.
The numerically calculated parameters encode particle's geometry information and are accessible by experiments.
By using single and double-layer potentials \cite{Vladimirov1984}, the Laplace equation for the fields inside
and outside the particle is transformed into an integral equation. A spectral representation
for the solution of this equation is obtained providing the eigenvalue problem for the linear
response operator is solved. Although not symmetric, this operator has a real spectrum bounded by
-1/2 and 1/2 \cite{Ouyang1989,Fredkin2003,Mayergoyz2005} and its eigenvectors are orthogonal to those of the
conjugate double-layer operator. A matrix representation is obtained by using a finite basis of surface
functions.
The true advantage of the spectral method is that the eigenvalues and eigenvectors
of the integral operator provide valuable insight into the dielectric behavior of clusters of biological cells.
The eigenvectors are a measure of surface charge distributions due to a field.
Only eigenvectors with a non-zero dipole moment contribute to the polarizability of the particle.
We call these dipole-active eigenmodes. An effective separation of the geometric and morphologic properties
from dielectric properties is therefore achieved \cite{Vrinceanu1996}.
We also show that for a particle covered by multiple confocal shells,
the relaxation spectrum is a sum of Debye terms with the number of relaxations equal to the number
of interfaces times the number of dipole-active eigenvalues.
This is a generalization of a previous result \cite{Hanai1988} on cells of arbitrary shape.
Our method is related to another spectral approach which uses an eigenvalue differential equation
\cite{Bergman1978,Bergman1992,Stockman2001,Li2003}. This method has been applied to
biological problems by Lei \textit{et al.} \cite{Lei2001} and by Huang \textit{et al.} \cite{Huang2002}.
These authors, however, considered homogeneous cells with much simpler expression for cell polarizability.
The BIE spectral method seeks a solution on the boundary surface defining the particle, as opposed to
the eigenvalue differential equation, where the solution is defined in the entire space.
In a previous study on double (budding) cells it was shown that before cells separation an
additional dispersion occurs \cite{Asami1998}.
Moreover, in recent papers \cite{Asami2006,Gheorghiu2002} numerical experiments have shown that the dielectric
spectra of a suspension of dimer cells connected by tight junctions exhibit an additional, distinct low-frequency
relaxation.
Our numerical calculation shows that the largest dipole-active eigenvalue approaches the value of 1/2 as
the junction become tighter.
Although the coupling of this eigenmode with the electric field stimulus is relatively modest
(the coupling weight is about 1-2 \%), this eigenmode
has a significant contribution to the polarizability of clusters.
Thus the eigenmodes close to 1/2 induce an additional low-frequency
relaxation in the dielectric spectra of clustered biological cells even though the coupling is quite small.
Needle-like objects, such as elongated spheroids or long cylinders, have similar polarizability features.
In this paper we consider rotationally symmetric linear clusters made of up to 4 identical particles
covered by thin insulating membranes and connected by
junctions of variable tightness.
Convenient and flexible representations for the surfaces describing these objects
are provided.
The number of relaxations in the dielectric spectrum of the linear clusters,
their time constants and their relative strengths are analyzed in terms of the eigenmodes of the
linear response operator specific to the given shape.
\section{Theory}
\subsection{Effective permittivity of a suspension}
We consider a suspension of identical, randomly oriented particles of arbitrary shape and
dielectric permittivity, $\varepsilon _{1} $, immersed
in a dielectric medium of dielectric permittivity, $\varepsilon _{0} $.
The dielectric permittivities
are in general complex quantities and the theory described here applies also for time-dependent fields, providing that
the size of a particle is much smaller than the wavelength.
When an applied uniform electric field interacts with the suspension,
the response of the system is linear with the applied field and an
effective permittivity for the whole sample can be measured and is defined by
\cite{Gheorghiu2002,Jackson1975,Prodan1999}:
\begin{equation} \label{eq2}
\varepsilon_{\mbox{\tiny sus}} =\varepsilon_0 + f\frac{\alpha \varepsilon_0 }{1-f\frac{\alpha }{3} } .
\end{equation}
This result is obtained in the limit of low concentration, weak intensity of the stimulus field, and
using an effective medium theory within the dipole approximation.
Here $f = N V_1/V$ is the volume fraction of all $N$ particles, each of volume $V_1$, with respect to the
total volume of the suspension $V$. The averaged normalized polarizability $\alpha$ of a particle is defined as
\cite{Prodan1999,Gheorghiu2002,Sebastian2008}
\begin{equation} \label{eq1}
\alpha =\frac{1}{4\pi {\kern 1pt} V_{1} } \int _{V_{1} } \int _{\Omega _{N} }
\left(\frac {\varepsilon _{1} -\varepsilon _{0}}{\varepsilon _{0}}
\right){\bf E}\left({\bf N}\right) \cdot {\bf N}\;d\Omega _{N}\;dV
\end{equation}
where ${\bf E}\left({\bf N}\right)$ is the electric field perturbation created inside the particle under a
normalized applied electric excitation with direction ${\bf N}$ and $d\Omega _{N} $
is the solid angle element generated by that direction. The above normalized polarizability is
dimensionless and is obtained by multiplying the standard polarizability of a particle with the
factor $4 \pi/V_1$. In the following we will refer only to normalized polarizability,
thus, without any confusion, the normalized polarizability $\alpha$ will be simply called polarizability.
The directional average in \eqref{eq1} is equivalent to the averaged sum
over three orthogonal axes due to the fact that the problem is linear with respect to the applied field.
The latter is more convenient from computational point of view.
The electric field inside a particle is obtained by solving the following Laplace equation for the
electric potential $\Phi$:
\begin{equation} \label{eq3}
\begin{array}{rll}
\Delta \Phi \left( {\bf x}\right) & = 0, & {\bf x} \in \Re ^{3} \backslash \Sigma \\
\left. \Phi \right|_{+} & =\left. \Phi \right|_{-} , & {\bf x} \in \Sigma \\
\left. {\varepsilon}_{0} \frac{\partial \Phi }{\partial \bf n} \right|_{+} & =
\left. {\varepsilon}_{1} \frac{\partial \Phi }{\partial \bf n} \right|_{-} , \quad & {\bf x} \in \Sigma\\
\Phi & \to - {\bf x} \cdot {\bf N} , & \left|{\bf x} \right|\to \infty
\end{array}
\end{equation}
where $\Re^{3}$ is the euclidian 3-dimensional space and $\Sigma$ is the surface of the particle.
The derivatives are taken with respect to the normal vector $\bf n$ to the surface $\Sigma$.
Due to the mismatch between the polarization inside and outside the object, electric charges accumulate at the
interface $\Sigma $ and create an electric potential which counteracts the uniform electric field stimulus.
The solution of the above Laplace problem (\ref{eq3}) is therefore formally given by
\begin{equation} \label{eq4}
\Phi({\bf x})= - {\bf x} \cdot {\bf N}+\frac{1}{{\rm 4\pi }}
\int _{\Sigma }\frac{ \mu({\bf y})}{|{\bf x} - {\bf y}|} \; d\Sigma ({\bf} y)
\end{equation}
The single layer charge distribution $\mu$ induced by the normalized electric field is a solution of the following BIE,
obtained by inserting solution (\ref{eq4}) in equations (\ref{eq3})
\begin{equation} \label{eq5}
\frac{\mu ({\bf x})}{2\lambda } -\frac{1}{4\pi } \int _{\Sigma }{\rm \mu }({\bf y})\;
\frac{{\bf n}({\bf x}) \cdot ({\bf x} - {\bf y})}{|{\bf x} - {\bf y}|^3 } \; d\Sigma ({\bf y})
= {\bf n} ({\bf x} ) \cdot {\bf N}
\end{equation}
Here the parameter $\lambda =({\varepsilon}_{1} -{\varepsilon}_{0} )/({\varepsilon}_{1} +{\varepsilon}_{0} )$
isolates all the information regarding the dielectric properties for this problem.
On using the linear response operator $M$ that acts on the Hilbert space of integrable
functions on the surface $\Sigma$,
\begin{equation} \label{eq6}
M[\mu ]=\frac{ 1}{4\pi} \int _{\Sigma }{\rm \mu }({\bf y})\;
\frac{{\bf n} ({\bf x}) \cdot ({\bf x} - {\bf y})}{|{\bf x} - {\bf y}|^{3} }\; d\Sigma ({\bf y}) ,
\end{equation}
the integral equation \eqref{eq5} is written as
\begin{equation} \label{eq7}
(1/(2\lambda) - M) \mu = {\bf n} \cdot {\bf N} .
\end{equation}
The integral operator \eqref{eq6} is the electric field generated by the single layer charge distribution $\mu$
along the normal to the surface. It encodes the geometric information and has several interesting
properties \cite{Ouyang1989,Fredkin2003,Mayergoyz2005}. Its spectrum is discrete and it is not
difficult to show that all of its eigenvalues are bounded by the [-1/2, 1/2] interval.
Although non-symmetric, the operator \eqref{eq6} has real non-degenerate eigenvalues.
The eigenvectors are biorthogonal, i. e., they are not orthogonal among themselves,
but orthogonal to the eigenvectors of the adjoint operator
\begin{equation} \label{eq8}
M^{\dag} [\mu]=\frac 1{4\pi} \int_{\Sigma } \mu({\bf y})\;
\frac{{\bf n} ({\bf y}) \cdot ({\bf x} - {\bf y})}{|{\bf x} - {\bf y} |^{3} } \;
d\Sigma ({\bf y}) ,
\end{equation}
which is associated with the electric field generated by a surface distribution of electric dipoles
(double layer charge distribution).
Therefore, if $|u_{k}\rangle $ is a right eigenvector of $M$ corresponding to eigenvalue $\chi_k$,
$M|u_{k}\rangle =\chi _{k} |u_{k}\rangle$
and
$\langle v_{k'}|$ is a left eigenvector corresponding to eigenvalue $\chi_{k'}$,
$\langle v_{k'}|M^\dag =\chi _{k'} \langle v_{k'} |$, then
\begin{equation} \label{eq10}
\langle v_{k'} |u_{k} \rangle =\delta _{k'k} ,
\end{equation}
with the scalar product defined as the integral over the interface $\Sigma $,
\begin{equation} \label{eq11}
\left\langle f_{1} |f_{2} \right\rangle =\int _{\Sigma }f_{1}^{*} \left({\bf x}\right)f_{2}
\left({\bf x}\right)d\Sigma ({\bf x}) .
\end{equation}
The value 1/2 is always the largest eigenvalue of the operator $M$, regardless the geometry of the object.
This is immediately seen if the object is considered
to be conductor (${\varepsilon}_{1} \to \infty $), and then the interior electric field has to be zero.
In that case $\lambda =1$, and the charge density that
generates a vanishing internal electric field obeys the equation $(1/2 - M)\mu=0$,
and therefore 1/2 is an eigenvalue of $M$.
However, this eigenmode is not dipole-active and does not contribute to the total
polarization of the object.
The operator \eqref{eq6} is insensitive to a scale transformation, which means that its
eigenvalue and eigenvectors depend only on
the shape of the object and not on its size, or electrical properties.
By employing the spectral representation of the resolvent of the operator $M$
\begin{equation} \label{eq14}
(z-M)^{-1} =\sum _{k} (z-\chi _{k} )^{-1} |u_{k} \rangle \langle v_{k}|,
\end{equation}
the solution of equation \eqref{eq7} is obtained for $z=1/(2\lambda )$ as
\begin{equation} \label{eq15}
{\rm \mu }=\sum _{k} \langle v_k |{\bf n} \cdot {\bf N}
\rangle ( 1/(2\lambda) - \chi_k )^{-1} |u_k\rangle .
\end{equation}
The polarizability of the homogeneous particle is obtained by using the distribution
\eqref{eq15} to build the solution \eqref{eq4} of the Laplace equation and use it in equation \eqref{eq1}.
It has been shown that, operationally, the polarizability is
simply the dipole moment of the distribution \eqref{eq15} over unit volume \cite{Gheorghiu2002,Prodan1999}
\begin{equation} \label{eq_polarizability}
\alpha =\frac{1}{3} \frac{1}{V_{1} } \sum _{i,k} \frac{\langle {\bf x} \cdot {\bf N}_i |u_k \rangle
\; \langle v_{k} |{\bf n} \cdot {\bf N}_{i} \rangle }{1/(2\lambda)-\chi_k } ,
\end{equation}
where ${\bf N}_{i} $ are three mutually orthogonal vectors (directions) of unit norm.
The factor $(1/(2\lambda)-\chi _{k} )^{-1} $ is a generalized Clausius-Mosotti factor.
Each dipole-active eigenmode contributes to $\alpha $ according to its weight
$p_{k} = \frac{1}{3} \frac{1}{V_{1} }\sum _{i} \langle {\bf x} \cdot {\bf N}_i |u_k \rangle
\; \langle v_{k} |{\bf n} \cdot {\bf N}_{i} \rangle $,
which determines the strength of coupling between the uniform electric field and
the $k$-th eigenmode and contains three components
$P_{k,i} =\langle {\bf x} \cdot {\bf {N}_i}|u_{k} \rangle \; \langle v_{k} |{\bf n} \cdot {\bf {N}_i}\rangle /V_{1} $.
Equation \eqref{eq_polarizability} shows a clear separation of the electric properties,
which are included only in $\lambda$,
from the geometric properties expressed by $\chi_k$ and $p_k$.
\subsection{Shelled particles}
The polarizability of an object covered by a thin shell with permittivity ${\varepsilon}_S $
can be calculated in a similar fashion. The electric field is now
generated by two single layer distributions, and boundary conditions are imposed twice,
for $\Sigma_1$ and for $\Sigma_2$. The surface $\Sigma_1$ is the outer surface of the shell
and $\Sigma_2$ is the interface between
the particle and the shell. The solution of a shelled
particle in terms of single layer potentials has the form \cite{Sebastian2008}
\begin{widetext}
\begin{equation} \label{eq16}
\Phi({\bf x})= - {\bf x} \cdot {\bf N} +
\frac 1{4\pi} \int_{\Sigma_1} \frac{\mu_1({\bf y})}{|{\bf x} - {\bf y}|} \; d\Sigma({\bf y}) +
\frac 1{4\pi} \int_{\Sigma_2} \frac{\mu_2({\bf y})}{|{\bf x} - {\bf y}|} \; d\Sigma({\bf y}),
\end{equation}
\end{widetext}
where $\mu_1$ and $\mu_2$ are the densities defined on surface $\Sigma_1$ and $\Sigma_2$, respectively.
Four integral operators $M_{11}$, $M_{12} $, $M_{21} $ and $M_{22} $ are defined,
depending on which surface are variables ${\bf x}$ and ${\bf y}$.
For example $M_{11} $ is defined when ${\bf x}$ and ${\bf y}$ are both on $\Sigma _{1} $,
$M_{12} $ is defined by ${\bf x}$ on $\Sigma _{1} $ and ${\bf y}$ on
$\Sigma _{2} $, and so on. Thus
\begin{equation} \label{eq17}
M_{ij}[\mu_j]=\frac 1{4\pi } \int_{\Sigma_j}
\mu_j ({\bf y})\; \frac{{\bf n}({\bf x})\; \cdot ({\bf x} - {\bf y})}
{|{\bf x} - {\bf y}|^3 }\; d\Sigma({\bf y})
\end{equation}
for $i,j=1,2$. The equations obeyed by $\mu _{1} $ and $\mu _{2} $ are
\begin{equation} \label{eq18}
\begin{array}{l}
\mu_1/(2\lambda_1) - M_{11}[\mu_1] - M_{12}[\mu_2] = {\bf n} \cdot {\bf N}\\
\mu_2/(2\lambda_2) - M_{21}[\mu_1] - M_{22}[\mu_2] = {\bf n} \cdot {\bf N} .
\end{array}
\end{equation}
Here the electric parameters are:
$\lambda_1=(\varepsilon_S - \varepsilon_0)/(\varepsilon_S + \varepsilon_0)$, and
$\lambda_2=(\varepsilon_1 - \varepsilon_S)/(\varepsilon_1 + \varepsilon_S)$.
We further assume a confocal geometry, i. e. the surface $\Sigma_1$
is a slightly scaled version of $\Sigma_2$,
with a scaling factor $\eta $ close to unity.
This assumption does not provide constant thickness for the shell,
but our main results should remain at least qualitatively valid \cite{Asami2006,Sancho2003}.
In the limit of very thin shells, and using the scaling properties of the operator $M$, one can show
\cite{Prodan1999,Gheorghiu2002} that all four $M$ operators are related to $M = M_{11}$
\begin{equation} \label{eq19}
\begin{array}{l}
M_{12}[\mu] = \eta^{-3} ( \mu/2 + M[\mu]) \\
M_{12}[\mu] = -\mu/2 + M[\mu] \\
M_{22}[\mu] = M[\mu]. \end{array}
\end{equation}
Equations \eqref{eq18} can then be arranged in a matrix form as
\begin{equation} \label{eq20}
\left(\begin{array}{cc}
1/(2\lambda_1 - M) & (1/2 + M)/\eta^3 \\
-1/2 + M & 1/(2\lambda_2) - M
\end{array}\right)
\left(\begin{array}{c} \mu_1 \\ \mu_2 \end{array}\right)
= \left(\begin{array}{c} {\bf n} \cdot {\bf N} \\ {\bf n} \cdot {\bf N} \end{array}\right).
\end{equation}
By knowing the eigenvectors and the eigenvalues of $M$ the charge densities $\mu _{1} $ and $\mu _{2} $
can be found by inverting the matrix in \eqref{eq20}. For example, $\mu _{1} $ is
\begin{widetext}
\begin{equation} \label{eq21}
\mu _{1} =\sum _{k} \frac{\eta ^{3} (\frac{1}{2\lambda _{2} } -\chi _{k} )+
\left(\frac{1}{2} +\chi _{k} \right)}{\eta ^{3} \left(\frac{1}{2\lambda _{1} } -
\chi _{k} \right)\left(\frac{1}{2\lambda _{2} } -\chi _{k} \right)+
\left(\frac{1}{2} +\chi _{k} \right)\left(\frac{1}{2} -\chi _{k} \right)} \langle v_k |{\bf n} \cdot {\bf N} \rangle |u_k\rangle.
\end{equation}
\end{widetext}
The field generated by the the two distributions $\mu_1$ and $\mu_2$ outside the particle is the same
as the field generated by an equivalent single layer distribution
\begin{equation} \label{eq25}
\mu_{e} =\sum _{k} \langle v_{k} |{\bf n} \cdot {\bf N}\rangle (1/(2\tilde{\lambda} _{k} )-\chi _{k} )^{-1} |u_{k} \rangle ,
\end{equation}
where $\tilde{\lambda} _{k} =(\tilde{\varepsilon}_{k} -{\varepsilon}_{0} )/(\tilde{\varepsilon}_{k} +{\varepsilon}_{0})$
and the equivalent permittivity $\tilde{{\varepsilon}}_{k}$ is defined for each eigenmode as:
\begin{equation} \label{eq26}
\tilde{\varepsilon}_k ={\varepsilon}_{S} \left(1+\frac{\varepsilon_1 - \varepsilon_S}
{\varepsilon_S + \delta (1/2-\chi_k)\varepsilon_1 +\delta (1/2+\chi_k)\epsilon_S} \right),
\end{equation}
where $\delta =\eta ^{3} -1 \ll 1$.
The distribution \eqref{eq25} is similar with the distribution \eqref{eq15} obtained for a homogeneous particle,
except that $\lambda $ has to be replaced for each mode with an equivalent
quantity $\tilde{\lambda} _k$.
Equation \eqref{eq26} can be applied recursively for a multi-shelled structure.
The strict separation of electric and geometric properties is weakened in this case,
because the shape-dependent eigenvalue $\chi_k $ appears now in the
electric equivalent quantity $\tilde{\lambda} _k $.
The polarizability of the shelled particle is obtained by using the distribution
\eqref{eq25} to build the solution \eqref{eq4} of the Laplace equation and use it in equation \eqref{eq1},
to get
\begin{equation} \label{eq28}
\alpha =\frac{1}{3} \frac{1}{V_{1} } \sum _{i,k} \frac{\langle {\bf x} \cdot {\bf N}_i |u_k \rangle
\; \langle v_{k} |{\bf n} \cdot {\bf N}_{i} \rangle }{1/(2\tilde{\lambda}_k)-\chi_k } .
\end{equation}
Equation \eqref{eq28} is obtained by replacing $\lambda $ with
$\tilde{\lambda} _k $ in Eq. \eqref{eq_polarizability}. The parameter $V_1$ in Eq. \eqref{eq28}
is the total volume of the cell (the core and the shells).
In the limit of a dilute suspension of identical
shelled particles, with a low volume fraction $f$,
the effective permittivity \eqref{eq2} is
\begin{equation} \label{eq30}
\varepsilon_{\mbox{\tiny sus}} = \varepsilon_0 \left(1 + f \sum_k p_k \frac{\tilde{\varepsilon}_k - \varepsilon_0}
{(1/2+\chi_k)\varepsilon_0 + (1/2-\chi_k)\tilde{\varepsilon}_k} \right).
\end{equation}
\subsection{The Debye relaxation expansion}
In general, the effective permittivity ${\epsilon}_{{\rm sus}}$ of a suspension of
objects with $m$ shells will have \textit{m}+1 Debye relaxation
terms for each dipole active eigenmode.
The proof is recursive and
is based on partial fraction expansion
with respect to variable $i\omega $ of equations \eqref{eq28} and \eqref{eq30},
provided that the complex permittivity of various dielectric phases
is ${\epsilon}=\varepsilon -{\rm i\sigma }/(\omega \varepsilon_{vac}) $ where $i=\sqrt{-1} $ and
$\varepsilon_{vac}$ is the permittivity of the free space ($8.85 \times 10^{-12}$ F/m).
Thus the first Debye term comes out from \eqref{eq30} and
the remaining $m$ Debye terms result from \eqref{eq26}
by the homogenization process described for shelled particles.
Hence, a suspension of cells with $m$
shells (and $m+1$ interfaces) has a dielectric spectrum containing a number of Debye terms equal to $m+1$
times the number of dipole-active eigenvalues.
The suspension effective permittivity ${\epsilon}_{{\rm sus}}$ has the expansion
\begin{equation} \label{eq42}
{\epsilon}_{{\rm sus}} ={\epsilon}_{{\rm f}} +\sum _{k,j}\Delta \varepsilon _{kj} /(1+{\rm i\omega T}_{{\rm kj}} )
\end{equation}
where ${\epsilon}_{{\rm f}} =\varepsilon _{{\rm hf}} -{\rm i\sigma }_{{\rm lf}} /(\omega \varepsilon_{vac}) $,
$\varepsilon _{{\rm hf}} $ is the high-frequency permittivity,
and ${\rm \sigma }_{{\rm lf}} $ is the low-frequency conductivity;
$\Delta \varepsilon _{kj} $ and $T_{{\rm kj}}$ are the dielectric decrement
and the relaxation time of the \textit{kj} Debye term, respectively;
index $k$ enumerates the dipole-active eigenmodes and index $j$ enumerates interfaces.
Although the measurable bulk quantities in equation \eqref{eq42} are directly correlated with the microscopic
(electric and shape) parameters, a solution of the inverse problem, which aims at obtaining the microscopic
information non-intrusively, from the effective permittivity, is in general difficult, if not impossible for the
general multi-shell structure. However, biological cell has a thin and almost non-conductive membrane, and several
simplifications and approximations can be made.
Two Debye relaxation terms in the effective permittivity ${\epsilon}_{{\rm sus}}$ are expected
for each dipole-active eigenmode, corresponding to
the two interfaces which define the membrane.
The first relaxation is derived from the equivalent permittivity \eqref{eq26} which can be written also as Debye
relaxation terms:
\begin{equation} \label{eq43}
\tilde{{\epsilon}}_k =\varepsilon +\Delta \varepsilon /(1+ i\omega T ).
\end{equation}
The relaxation time $T$ that is given by the poles of $\tilde\epsilon_k$ in \eqref{eq26} is a quite good approximation
of the first relaxation time $T_{k1}$
\begin{equation} \label{eq44}
T = \varepsilon_{vac} \frac{\left(1+\delta /2+\delta \chi _{k} \right)\varepsilon _{S} +\delta \left(1/2-\chi _{k} \right)\varepsilon _{1} }
{\left(1+\delta /2+\delta \chi _{k} \right)\sigma _{S} +\delta \left(1/2-\chi _{k} \right)\sigma _{1} } \approx T_{k1} .
\end{equation}
The main reason is as follows. At frequencies close to $1/T$ there is a huge change in
$\tilde\epsilon_k$ of order $\varepsilon_{S} /(\delta (1/2-\chi_{k} ))$,
and consequently a significant change in the total permittivity $\epsilon_{\mbox{\tiny sus}}$
given by \eqref{eq30}. Therefore $T$ provides an approximate value for the relaxation time $T_{k1}$ of the suspension effective permittivity
$\epsilon_{\mbox{\tiny sus}}$.
For a non-conductive shell $\sigma_{S}\approx 0$,
or more precisely when $\sigma _{S} {\rm \ll }\delta \left(1/2-\chi _{k} \right)\sigma _{1}$,
the relaxation time \eqref{eq44} is:
\begin{equation} \label{eq45}
T_{k1} \approx \; \varepsilon_{vac} \; \varepsilon_{S} /(\delta \cdot \sigma _{1} (1/2-\chi_{k} ))
\end{equation}
showing a strong dependence on the thickness of the shell and on the shape of the particle, through the
eigenvalue $\chi_k$.
Due to the small parameter $\delta$ in \eqref{eq45} the first relaxation (i. e., membrane relaxation)
tends to have a lower frequency than the second relaxation,
which is present even for particles with no shell (see the discussion below).
In addition, cumbersome but straightforward calculations provide the dielectric decrement $\Delta \varepsilon_{k1} $ in \eqref{eq42}
\begin{equation} \label{eq46}
\Delta \varepsilon_{k1} \approx 4 f p_k \varepsilon_{S} (\delta (1/2+\chi _{k} )^{2} (1/2-\chi _{k} ))^{-1}
\end{equation}
that is very large due to the same strong dependence on the thickness of the shell.
The effect is even more dramatic when
the second largest eigenvalue is very close to the largest eigenvalue,
$\left(1/2-\chi _{2} \right)\to 0$, like in the case of two cells connected by
tight junctions.
For a suspension of shelled spheres $\eta = 1 + \Delta R/R$ and $\delta =\eta ^{3} -1 \approx 3 \Delta R/R$,
where $\Delta R$ is the
thickness of the membrane, $R$ is inner radius, and $R + \Delta R$ is the total radius. Thus both
$T_{k1}$ and $\Delta \varepsilon_{k1} $ are proportional to $R$ and $\varepsilon_{S}$ and inverse proportional
to $\Delta R$ like in the Pauly-Schwan theory \cite{Pauly1959,Schwan1996}.
Moreover, the dielectric decrement $\Delta \varepsilon_{k1} $ in \eqref{eq46} is
a generalization of equation (54a) in Ref. \cite{Schwan1996}. In the same time,
the relaxation time \eqref{eq45} differs with respect to equation (56a) in Ref. \cite{Schwan1996}
only by the conductivity term. We will show elsewhere that a more appropriate treatment of the
relaxation times recovers also the relaxation time given by equation (54a) in Ref. \cite{Schwan1996}.
Thus, a non-conductive and thin shell/membrane produces a large relaxation of the complex
permittivity of the suspension \cite{Fricke1953}. The experimental evidences further
support these theoretical facts:
when attacking the membrane with a membrane disrupting compound ( for example a detergent)
the relaxation almost vanishes as the cellular membrane is permeated \cite{Asami1977}.
For frequencies higher than $1/T_{k1}$ the cell permittivity is essentially determined by the
dielectric properties of the cytoplasm, and does not depend on membrane's properties. The second Debye
relaxation occurs at higher frequencies than the first (membrane) relaxation,
and has the relaxation time
\begin{equation} \label{eq48}
T_{k2} \approx \varepsilon_{vac} \frac{\left(1/2+\chi _{k} \right)\varepsilon _{0} +\left(1/2-\chi _{k} \right)\varepsilon _{1} }
{\left(1/2+\chi _{k} \right)\sigma _{0} +\left(1/2-\chi _{k} \right)\sigma _{1} }
\end{equation}
derived from the pole of equation \eqref{eq30}.
The corresponding dielectric decrement is
\begin{eqnarray*}
&\Delta \varepsilon _{k2} \approx
f p_{k} (1/2-\chi_k)(\varepsilon_1 \sigma_0 - \varepsilon_0\sigma_1)^2 \times\\
&\left((1/2+\chi_k)\varepsilon_0 +(1/2-\chi_k)\varepsilon_1\right)^{-1} \times\\
&\left((1/2+\chi_k)\sigma_0 +(1/2-\chi_k)\sigma_1\right)^{-2}
\end{eqnarray*}
The last two equations are similar to the ones that are given for spherical particles in Ref. \cite{Schwan1996}
(equations (46) and (49) in the aforementioned reference).
The relaxation given by $T_{k2}$ is basically the relaxation of a homogenous particle embedded in a dielectric
environment and was also discussed in Ref. \cite{Lei2001} by a closely related spectral method.
If the conductivity of the cytoplasm is comparable to the conductivity
of the outer medium, the decrement of the second relaxation is small such that it cannot be distinguished
in the spectrum.
On the contrary, if the conductivity of the outer medium is much greater or smaller that that of cytoplasm,
than a second observable relaxation occurs.
Unlike the membrane relaxation, this second relaxation depends only weakly on the shape.
By assuming that $\sigma _{0} \ll \sigma_1$ and by using
a finite-difference method, this resonance was also obtained in \cite{Asami2006}
and it was instrumental in explaining the
experimental data on the fission of yeast cells of Asami \textit{et al.} \cite{Asami1999}
by Lei \textit{et al.} \cite{Lei2001}.
The shape of the particle is important because it affects the number of dipole-active eigenvalues and their
strengths. In principle, each dipole-active eigenvalue introduces a new relaxation in the dielectric spectrum,
providing this relaxation is well separated from the others. A cluster with complex geometry can have several
dipole-active eigenvalues, but unless the cluster is larger in one dimension then in the others, or there are
tight junctions, the relaxations overlap to create broad features in the spectrum. An extra relaxation is
introduced when the particles are covered by thin membranes. In addition, if $\left(1/2-\chi _{k} \right)\to 0$
for that eigenvalue, then the shell induced relaxation has low frequency, large relaxation time and
large dielectric decrement. Based on the spectral BIE method, it is therefore possible to explicitly relate the dielectric
spectra of cell suspensions to cell's geometry and electric parameters, and, even design fitting procedures to
evaluate these parameters from measurements.
\section{Results}
\subsection{Numerical procedure}
The calculation of the effective permittivity for a suspension uses equations \eqref{eq2} or \eqref{eq30},
and reduces then to finding the eigenvalues $\chi_k$ and eigenvectors $|u_k\rangle$ and $|v_k\rangle$ of
the linear response operator $M$. This problem is solved by employing a finite basis of {\em NB} functions
defined on the surface $\Sigma$. A natural basis for a surface not far from a sphere is the
generalized hyperspherical harmonics functions
\begin{equation} \label{eq31}
\tilde{Y}_{lm} ({\bf x})=\frac{1}{\sqrt{s({\bf x})}}\;Y_{lm} \left(\theta({\bf x}),\varphi ({\bf x})\right),
\end{equation}
where $s({\bf x})$ is related to the surface element through
$d\Sigma = s({\bf x}) \; d\Omega_{\bf x} $
and $d\Omega_{\bf x} $ is the solid angle element.
Another choice could be based on Chebyshev polynomials of the first kind \cite{Abramowitz1972}
\begin{equation} \label{eq32}
\tilde{T}_{lm} ({\bf x})=\frac 1{\sqrt{s({\bf x})}} \; T_l\left(\theta ({\bf x})\right)\;
e^{i m \varphi ({\bf x})} .
\end{equation}
Both bases are complete and orthogonal in the Hilbert space of square integrable functions defined on $\Sigma$.
In this paper, we model the linear cluster of particles as an object with axial symmetry. We seek to find a
surface of revolution for which the thickness of the interparticle joints can be varied without perturbing the
overall shape of the object. We use two representations for the surface $\Sigma$:
(A) for clusters of two
particles we use spherical coordinates
$\{x,y,z\} = \{r(\theta)\sin\theta \cos\phi, \; r(\theta)\sin\theta \sin\phi, \; r(\theta) \cos\theta\}$,
and (B) for clusters with more
than two particles we specify the surface in terms of a function $g(z)$ as
$\{x,y,z\} = \{g(z)\cos\phi,\; g(z)\sin\phi, \;z\}$.
In the case B, the surface element is
\begin{equation} \label{eq35}
d\Sigma = g(z) \sqrt{1+g'^2(z)} \; dzd\varphi,
\end{equation}
and the normal to surface $\Sigma $ is
\begin{equation} \label{eq36}
{\bf n}=\frac 1{\sqrt{1+g'^2(z)}}\;
\left(\begin{array}{c} \cos\varphi \\ \sin\varphi \\ -g'(z) \end{array}\right).
\end{equation}
In the basis of generalized hyperspherical harmonics the operator $M$ has matrix elements
given by
\begin{widetext}
\begin{equation} \label{eq38}
M_{lm; l'm'} = \delta_{mm'}
\iint\limits_0^{\kern 10pt 2\pi} \;\iint\limits_{z_{\mbox{\tiny min}}}^{\kern 10pt z_{\mbox{\tiny max}}}
A(z,z', \varphi - \varphi')
P_l^m(\cos\theta(z)) P_{l'}^{m'}(\cos\theta(z')) \; e^{im(\varphi - \varphi')}
\; G(z, z')
\; dz\;dz'\;d\varphi\;d\varphi'
\end{equation}
where
\begin{equation}
G(z,z') = \sqrt{g(z) g(z') \sqrt{(1+g'(z)) (1+g'(z'))^{-1}}} ,
\end{equation}
and
\begin{equation}
A(z,z',\phi) = \frac{(g(z) - g(z')) \cos\phi - (z-z') g'(z)}
{\left[ g^2(z) +g^2(z') - 2 g(z) g(z') \cos\phi + (z-z')^2\right]^{3/2}}
\end{equation}
\end{widetext}
After the angle integration in equation \eqref{eq38} and by using the elliptic
integrals given in the Appendix, the matrix elements are obtained by numerical
evaluation of the resulting ($z$, $z'$) double integral using an {\em NQ}-point
Gauss-Legendre quadrature \cite{Boyd2001,Abramowitz1972}. Because of the integrable singularity apparent in the
kernel of the operator $M$ in equation \eqref{eq6}, the mesh of $z$ must be shifted from the
mesh of $z'$ by a transformation which insures that there is no overlap between the
two meshes.
The delta symbols $\delta_{mm'}$ in equation \eqref{eq38} reflects the fact that we consider only objects with
rotational symmetry in this paper. Moreover, for fields parallel with the cluster axis $m=0$, while
$m=1$ for perpendicular fields.
The convergence of the results is obtained in two steps. First, the number {\em NQ} of quadrature
points is increased until the matrix elements of $M$ converge, and then the size {\em NB} of the
basis set is increased until the relevant eigenvalues $\chi_k$ and their corresponding weights $p_k$
have acquired the desired accuracy. A necessary test for convergence is the fulfillment of the
sum rules $\sum_k P_{k,i} = 1$ and $\sum_{i,k} \chi _k P_{k,i} = 1/2$ with sufficient accuracy \cite{Bergman1978,Bergman1992}.
Usually the convergence is fast in both the number of quadrature points and the size of basis,
unless the system has a tight junction where some care must be considered in order
to achieve the required accuracy of the eigenmodes with the eigenvalues close to $1/2$.
For a sphere there is just one dipole-active eigenmode which has eigenvalue
$\chi =1/6$ and weight $p=1$, while for an ellipsoid there is one dipole-active eigenmode
along each axis.
Fast and accurate solutions are achieved for spheroids with a basis size of {\em NB} =20 and
with {\em NQ} = 64 quadrature points. In general, the size of the basis and the
number of the quadrature points increase with the number of cells in the cluster and with the decreasing of the junction size.
Thus, for our numerical examples a basis with {\em NB}=35-40 and {\em NQ} = 128 quadrature points are enough
for a converged solution in the case of the dimers and {\em NB}=50 and {\em NQ} = 200 quadrature points in the case
of the clusters with up to four cells.
\subsection{Two cells joined by tight junction}
The equation $r(\theta) = (h + \cos^2\theta)/(1 - a\cos^2\theta)$
describes the shape of a two-particle cluster. Parameter $h$ controls the tightness of the inter-particle
junction and parameter $a$ measures the deviation from a spherical shape.
More precisely, $h$ is the radius of the smallest circle at around the thinnest part of the junction.
\begin{figure}[h]
\includegraphics [width=5.0in] {figure1.eps}
\caption{\label{fig:1}
(Color online) The spectrum of the
effective permittivity of a suspension of dimers with various junction thickness $h$, and
with parallel and perpendicular field configurations.
The suspension permittivity for an electric field perpendicular to cluster axis does not depend
on $h$ and is pointed by an arrow.
}
\end{figure}
Figure \ref{fig:1} shows the effective permittivity for a suspension of particles with the following parameters:
$\varepsilon_1 = 70$,
$\sigma _{1} = 0.25$ S/m,
$\varepsilon_S = 6$,
$\sigma_S = 0$,
$\varepsilon_0 = 81$,
$\sigma_0 = 0.374$ S/m,
volume fraction $f=0.05$, membrane thickness $\delta = 0.00947275$, and $a=0.2$.
The effective permittivity does not depend on the thickness parameter $h$ when the stimulus electric field is perpendicular to
the cluster axis. However, a new relaxation becomes apparent as $h \to 0$,
for parallel fields \cite{Asami2006, Gheorghiu2002}.
\begin{figure}
\includegraphics [width=4.0in] {figure2.eps}
\caption{\label{fig:2}
(Color online) The largest seven eigenvalues and their weights for a binary cluster
with parameter $a=0.2$, as a function of $h$. The inset shows the shape of the dimer.
}
\end{figure}
Figure \ref{fig:2} presents the first 7 eigenvalues $\chi _{k} $ and their weights $P_{k,1} $ for a field parallel to the $z$ axis.
As the junction become tighter ($h\to 0$) more eigenvalues become dipole-active. While all eigenvalues are important in shaping the
dielectric spectrum, the second largest eigenvalue $\chi _{2} $ is crucial to explaining the occurrence of an additional relaxation at
low frequencies, as observed for small $h$ in \cite{Asami2006, Gheorghiu2002}.
Although its weight $P_{2,1} $ also decreases for small $h$, this dipole active
eigenmode approaches 1/2 as the junction becomes tighter.
Thus, according to equations \eqref{eq45} and \eqref{eq46} the effect of $\chi _{2} $ is
``enhanced'' due to the presence of a nonconductive shell
(as the case for biological particles analyzed in \cite{Asami2006, Gheorghiu2002}). Moreover,
the decrease of $P_{2,1} $ is compensated by the increase of $1/(1/2-\chi _{2} )$.
\begin{figure}
\includegraphics[width=4.0in]{figure3.eps}
\caption{\label{fig:3}
The first four eigenvectors for a dimer given by equation $r(\theta )$ ; $a=0.2$ and $h=0.5$ (dotted line) and $h=0.01$ (solid line).
}
\end{figure}
The presence of a new relaxation at low frequency along with its relationship with the size of $h$ has been already singled out in
Gheorghiu \textit{et al.} \cite{Gheorghiu2002} by using the same method but without the analysis of dipole-active eigenmodes.
Using a finite discrete model \cite{Asami2006}, the relaxation was observed before the segregation during cell division,
while other papers \cite{Biasio2007,Biasio2009} fail to relate
the size of $h$ to the new relaxation, even though one of them \cite{Biasio2007}
employs essentially the same method as the one outlined in the present work.
Figure \ref{fig:3} shows the charge distribution associated with the first four eigenvalues for two distinct values of $h$.
The second eigenmode is an antisymmetric combination of net charge distributions (monopoles) on each particle of the dimer.
The third charge distribution is an antisymmetric
combination of charge distributions with a dipole moment on each part of the dimer and
the forth distribution is antisymmetric combination of charge distributions with a quadrupole moment on each particle.
At small $h$ (tight junctions), charge accumulates in the vicinity of the junction \cite{Klimov2007}.
\subsection{Clusters of more than two particles}
Smoother yet tight junctions would bring $\left(1/2-\chi _{2} \right)$ closer to 0 than sharp and tight junctions. The reason is simple: smoother
junctions have the two parts of the dimer farther apart.
We have analyzed linear clusters of cells connected by smooth and tight junctions by using a
($z$, $\phi$) parameterization which describes
a surface by $\{ x=g(z)\cos \varphi ,y=g(z)\sin \varphi, z\}$.
The construction starts from a dimer shape that resembles the shape of
the epithelial cells like MDCK (Madin-Darby Canine Kidney) cells. An example of such shape, displayed in
figure \ref{fig:4}, extends
from $-z_{\mbox{\tiny max}}$ to $z_{\mbox{\tiny max}}$ and it can be decomposed in three parts:
the left cap ($-z_{\mbox{\tiny max}} \le z \le -z_1$),
the central part ($ -z_1 \le z \le z_1$), and the right cap ($z_1 \le z \le z_{\mbox{\tiny max}}$).
At position $\pm z_1$ the shape function
has its maximum. An $m$-cell linear cluster is obtained by repeating the central part
$m-1$ times and it extends from $-L_m$ to $L_m$,
where $L_m = z_{\mbox{\tiny max}} + (m-2) z_1$. Mathematically, the shape is described by:
\begin{widetext}
\begin{equation} \label{eq50}
g_m(z)=\left\{
\begin{array}{l}
g(z+(m-2)z_{1} ), \mbox{\quad for}\; -L_m \le z \le -L_m + z_{\mbox{\tiny max}} \\
g\left(\mbox{mod}(z+(1+(-1)^{m})z_1/2, 2z_1) - z_1\right), \mbox{\quad for}\;
-L_m + z_{\mbox{\tiny max}} \le z \le L_m - z_{\mbox{\tiny max}}\\
g(z-(m-2)z_1), \mbox{\quad for}\; L_m \le z \le L_m - z_{\mbox{\tiny max}},
\end{array}\right.
\end{equation}
\end{widetext}
where mod($x$, $y$) is the remainder of the division of $x$ by $y$. For the examples considered here, the dimer shape function is:
\begin{equation} \label{eq51}
\begin{array}{l}
g(z)=0.01+2.32317\;z^2 -11.9862\;z^4 +40.4045\;z^6\\
-74.2226\;z^8 +79.142\;z^{10} -51.8929\;z^{12} +21.3096{\rm \; }z^{14}\\
-5.35113\;z^{16} +0.752147\;z^{18} -0.045375\;z^{20} ,
\end{array}
\end{equation}
with $z_{\mbox{\tiny max}} = 1.77377$ and $z_1 = z_{\mbox{\tiny max}}/2$.
\begin{figure}
\includegraphics [width=4.0in]{figure4.eps}
\caption{\label{fig:4}
Smooth construction of a cluster (lower panel) from a dimer (upper panel).
The parts determined by $z\in \left[-z_{1} ,z_{1} \right]$ are ``glued'' together with the ends of the dimer.
The arrows show where the junctions will be placed in the cluster.
}
\end{figure}
Tables \ref{para} and \ref{perp} list the most representative dipole-active eigenmodes for a trimer in perpendicular and parallel fields.
Only the parallel field configuration has a dipole-active eigenvalue close to 1/2, with a relatively small weight.
\begin{table}
\caption{Most representative dipole-active eigenmodes and their weights for the trimer in parallel field.
}
\label{para}
\begin{ruledtabular}
\begin{tabular}{lcccccc}
& $\chi _{k}$ & $P_{k,1} $ \\
\hline
& 0.4996 & 0.01305 \\
& 0.40642 & 0.07769 \\
& 0.40448 & 0.1391 \\
& 0.37763 & 0.17366 \\
& 0.26409 & 0.20519 \\
& 0.23809 & 0.01839 \\
& 0.17557 & 0.05104 \\
& 0.13522 & 0.05234 \\
& 0.12165 & 0.01066 \\
& 0.07192 & 0.06197 \\
& 0.06649 & 0.01239 \\
& 0.03479 & 0.03899 \\
& 0.03362 & 0.02823 \\
& 0.0281 & 0.04688 \\
& 0.02377 & 0.02519 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}
\caption{Most representative dipole-active eigenmodes and their weights for the trimer in perpendicular field.
}
\label{perp}
\begin{ruledtabular}
\begin{tabular}{lcccccc}
& $\chi _{k}$ & $P_{k,2} $ \\
\hline
& 0.1831 & 0.64102 \\
& 0.06964 & 0.01464 \\
& 0.05492 & 0.01343 \\
& 0.0272 & 0.01065 \\
& 0.01674 & 0.05687 \\
& 0.01479 & 0.13179 \\
& 0.00395 & 0.05029 \\
& -0.01635 & 0.03246 \\
\end{tabular}
\end{ruledtabular}
\end{table}
The results for linear clusters of up to four particles are displayed in Figure \ref{fig:5}.
The electric parameters are the same as ones used for dimers in the previous section.
An additional, distinct low-frequency relaxation emerges for clusters with more then one particles,
only when the stimulus field is parallel with the symmetry axis.
The relaxation frequency decreases, while the intensity of these relaxations increases, as the number of cluster members increases.
This behavior is explained again by the combination of eigenvalues close to 1/2, with thin non-conductive
layers covering the cluster and is consistent
with experimental data on ischemic tissues \cite{Knapp2005}, which reports that the cell separation (closure of gap-junctions)
is responsible for decrease and eventual disappearance of the low-frequency dispersion.
\begin{figure}
\includegraphics [width=5.0in] {figure5.eps}
\caption{\label{fig:5}
(Color online) Effective permittivity for clusters (shown in the inset) of one, two, three, and four cells connected by tight and smooth junctions.
The field is either parallel (solid lines with symbols) or perpendicular (solid lines only) to the cluster axis. The effective
permittivity either increases strongly with the number of cells for parallel geometry, or does not change for a perpendicular geometry.
}
\end{figure}
In Figure \ref{fig:6} we plot $P_{k,1} /(1/2-\chi _{k} )$ versus $(1/2-\chi _{k} )$, which shows that the number of
dipole-active eigenmodes increases with the number of particles in the cluster.
According to \eqref{eq45} and \eqref{eq46}, Figure \ref{fig:6} shows in fact the dielectric
decrement versus its corresponding relaxation frequency for each dipole-active eigenmode of the given clusters.
For clusters of two or three particles, there is one
important active eigenmode close to $1/2$, while for clusters of four particles there are two active eigenmodes.
\begin{figure}
\includegraphics [width=5.0in] {figure6.eps}
\caption{\label{fig:6}
(Color online) $P_{k,1} /(1/2-\chi _{k} )$ versus $(1/2-\chi _{k} )$ for clusters of up to four cells.
The second eigenvalue has the largest contribution to intensity of relaxation.
}
\end{figure}
It can be conjectured that for a general linear cluster made of $m$ particles,
there are $m-1$ eigenvalues close to $1/2$, of which
the largest one is always dipole-active and has the largest weight.
In fact one can show that for two cells connected by smooth and tight
junction characterized by parameter $h$, $\left(1/2-\chi _{2} \right)\propto h^{2} $
when $h\to 0$, or more precisely $\left(1/2-\chi _{2} \right)$ is
proportional with the solid angle encompassed by the missing part of a cell
when it is connected with other cell in the dimer. The proof is based on
the theorem of the solid angle \cite{Vladimirov1984}. The generalization to a finite cluster is
also straightforward to $\left(1/2-\chi _{2} \right)\propto h^{2} /m$ (in that case the solid
angle encompassed by the middle junction is proportional to $h^{2} /m$.
The weight of the second eigenmode is
$P_{2,1} = \langle {\bf x} \cdot {\bf {N}_{1}}|u_{2}\rangle \langle v_{2} |{\bf n} \cdot {\bf {N}_{1}}\rangle /V_{1}$.
If we consider that the surface of the cluster is determined by the function $g(z)$ then, up to a constant
factor, $\langle v_{2} |{\bf n} \cdot {\bf {N}_{1}}\rangle \approx g(0)^2 = h^2$ for two cells connected by
smooth and tight junctions. The proof considers that the second eigenfunction of $M^{\dag }$ is an antisymmetric combination
of constant distributions on each part of the dimer. This assertion is confirmed in Figure \ref{fig:7}.
Moreover, $\langle {\bf x} \cdot {\bf {N}_{1}}|u_{2}\rangle /V_{1} $ is weakly dependent on $h$.
Therefore, for a parallel setting of the field stimulus,
$P_{2,1} /(1/2-\chi _{2} )$, which is the measure of the dielectric decrement of low-frequency relaxation,
is finite and it increases when the number of cells is increased.
The increase of relaxation decrement when $m\to \infty $ is physically limited by
$\sigma _{S} {\rm \ll }\delta \left(1/2-\chi _{k} \right)\sigma _{1} $, since the membrane conductivity is not strictly 0.
\begin{figure}
\includegraphics [width=4.0in] {figure7.eps}
\caption{\label{fig:7}
The first (dotted line) and the second (solid line) eigenfunction of $M$ (upper panel) and $M^{\dag } $ (lower panel) for a dimer whose shape is
depicted by dashed line in the lower panel. The second eigenfunction of $M^{\dag } $ is an antisymmetric combination of almost constant
distributions on each part of the dimer.
}
\end{figure}
Due to cluster's shape and membrane properties, the variation of $P_{k,1} /(1/2-\chi _{k} )$ and $(1/2-\chi _{k} )$
with respect the eigenmode $k$
determines a low frequency relaxation when the dipole-active eigenvalue $\chi _{k}$ is close to 1/2.
We note here that for dipole-active eigenmodes of ellipsoids the
term $(1/2-\chi _{k} )$ is called the depolarization factor
and has analytical expression \cite{Venermo2007}.
Prolate spheroids with longitudinal axis much larger than the transverse axis (needles) have the
longitudinal depolarization factor approaching 0 and the transverse depolarization factor approaching 1/2.
More precisely, for a long prolate spheroid, the longitudinal depolarization factor scales as
$(1/2-\chi _{2} )\propto a_x^2/a_z^2,\; a_z > a_x = a_y$, as $a_x \to 0$.
On the other hand, extensive numerical calculations support the fact that cylinders with the same aspect ratio behave similarly
to prolate spheroids \cite{Venermo2007}.
Thus, it is not hard to observe that the low-frequency relaxation of linear clusters of cells connected
by tight junctions is similar to that of a needle or a thin cylinder as long as the cluster and as thick as the junction.
On the other hand, the high-frequency relaxation of the cluster shows
the relaxation of a suspension of spheroids with the same volume as the volume of
a single cell. Therefore the dielectric spectrum for a suspension of clusters is the same as the spectrum
of a two species suspension made of thin cylinders and spheroids.
\section{Conclusions}
We present a theoretical framework based on a spectral representation of BIE
and able to calculate the dielectric behavior
of linear clusters with a wide range of shapes and dielectric structures.
The theory agrees with the results of Pauly and Schwan for a sphere covered with a shell \cite{Pauly1959,Schwan1996}.
In fact, for spheroids, our theory is the same as the analytical results of Asami \textit{et al.} \cite{Asami1980}.
We present extensive calculations of clusters with shapes resembling MDCK cells.
A practical numerical recipe to compute the effective permittivity of linear clusters with arbitrary number of cells is provided.
Examples are given for cluster with shapes described as $r(\theta)$ in spherical coordinates or
using ($z$, $\varphi$) parameters as $\{ x=g(z)\cos \varphi ,y=g(z)\sin \varphi, z \}$.
Other studies in the literature used only spherical coordinates representation \cite{Gheorghiu2002,Prodan1999,Biasio2009}.
A direct relation between the geometry and dielectric parameters of the cells and
their dielectric behavior described by a Debye representation has been formulated for the first time.
Other work \cite{Lei2001}, which is based on a closely related spectral method
\cite{Bergman1978,Bergman1992,Stockman2001,Li2003},
found a direct relation linking the geometry
and electric parameters to the dielectric behavior only for homogenous particles.
Moreover, the method used in \cite{Lei2001} treats only particles with spheroidal geometry.
We show that the spectral representation provides a straightforward evaluation of
the characteristic time constants and dielectric decrements
of the relaxations induced by cell membrane.
We prove that the effective permittivity is sensitive to the shape of the embedded particles,
specially when the linear response operator has strong dipole active modes (with large weights $p_{k} $).
A low-frequency and distinct relaxation occurs
when the largest dipole-active eigenvalue is very close to $1/2$.
Clusters of living cells connected by tight junctions or very long cells have
such an eigenvalue.
Our results also shed a new light on the understanding of recent numerical calculations \cite{Ron2009}
performed with a boundary element method on clustered cells where the low-frequency relaxation is
attributed to the tight (gap) junctions connecting the cells.
The method used in \cite{Ron2009} does not use the confocal geometry assumption .
The present work has several implications and applications.
We emphasize the capabilities of dielectric spectroscopy to monitor the
dynamics of cellular systems, e.g., cells during cell cycle division,
using synchronized yeast cells \cite{Gheorghiu1998,Asami2006,Knapp2005},
or monolayers of interconnected cells \cite{Wegener2000,Urdapilleta2006}.
Also the method is able to assess the dielectric behavior of linear aggregates or rouleaux of erythrocytes,
where the ellipsoidal or cylindrical approximations are not adequate \cite{Sebastian2005,Asami2007}.
The proposed representation is a powerful alternative to finite element or other purely numerical approaches,
because it provides the analytical framework to explain
and predict the complex dielectric spectra occurring in bioengineering applications.
Extension of this method to other surfaces of revolution, for example linear clusters with more
than 4 particles, is straightforward providing an adequate parametric equation is available.
Finally, in many cases (e.g. shapes with high symmetry)
the method is faster, offers accurate solutions and last but not least can be integrated in
fitting procedures to analyze experimental spectra.
\begin{acknowledgments}
This work has been supported by Romanian Project ``Ideas'' No.120/2007 and FP 7 Nanomagma No.214107/2008.
\end{acknowledgments}
|
\section{Introduction}
The extreme low temperatures, the dryness, the typical high altitude of the
internal Antarctic Plateau (more than 2500~m), joint to the fact that the optical
turbulence seems to be concentrated in a thin surface layer whose thickness is of
the order of a few tens of meters do of this site a place in which, potentially,
we could achieve astronomical observations otherwise possible only by space.
Despite exciting first results (Lawrence \etal~\cite{law2004}; Agabi \etal~\cite{aga2006};
Trinquet \etal~\cite{tri2008}) making the internal Antarctic Plateau a site of potential great interest for astronomical applications, some uncertainties still remain.
Here we studied the Dome C area with a mesoscale meteorological model (Meso-NH, Lafore \etal~\cite{laf1998}).
Numerical simulations offer the advantage to provide volumetric maps of the optical
turbulence ($C_N^2$) extended on the whole internal plateau and, ideally, to
retrieve comparative estimates in a relative short time and homogeneous way on
different places of the plateau.
Fifteen winter nights (the same as from Trinquet \etal~(\cite{tri2008}) were simulated.
Using the forecasted $C_N^2$ profiles, we retrieved the surface layer thicknesses H$_{SL}$ and the free atmosphere
seeing ($\epsilon{_{FA}}$) for all 15 nights.
\par
This study is a short survey of the more detailed study available in Lascaux \etal~(\cite{las2009}).
\section{Surface layer seeing and free atmosphere seeing}
Two different configuration of Meso-NH were chosen: a low horizontal resolution mode, and a high horizontal resolution
mode (with the grid-nesting interactive technique).
To know more about the numerical set-up and the model configuration, the reader can
refer to Lascaux \etal~(\cite{las2009}).
In that paper can also be found a validation of the model with comparisons of meteorological parameters (wind and
temperature) at Dome C between model outputs and observations.
One of the conclusion is that both configurations generated better forecast for wind speed and temperature than
the analysis from the ECMWF, especially near the surface. More over, the grid nested mode gave better results than the
low resolution mode.
\par
In order to verify how well the simulated H$_{SL}$ matches with the measured one we
computed the typical height of the surface layer (averaged each night between 12 UTC and 16 UTC)
using the same criterion as in Trinquet \etal~(\cite{tri2008}):
\begin{equation}
\label{eq:bl1}
\frac{ \int_{8m}^{h_{sl}} C_N^2(h)dh }{ \int_{8m}^{1km} C_N^2(h)dh } < 0.90
\end{equation}
where $C_N^2$ is the refractive index structure parameter.
\par
The observed mean H$_{SL}$ for the 15 winter nights was of 35.3 $\pm$ 5.1~m. The low horizontal resolution
mode gave a result almost twice higher: H$_{SL,LOW}$=65.9$\pm$ 8.7~m.
The grid-nesting mode gave better results, comparable to the observations: H$_{SL,HIGH}$=48.9 $\pm$ 7.6~m.
Using these computed mean H$_{SL}$, we deduced the median free atmosphere seeing using the same method as
in Trinquet \etal~(\cite{tri2008}).
Using H$_{SL,OBS}$=30~m (computed on a larger sample), Trinquet \etal~(\cite{tri2008}) found
$\epsilon{_{FA,OBS}}$=0.3 $\pm$ 0.2~arcsec.
For the low resolution mode (H$_{SL,LOW}$=65.9~m), the corresponding free atmosphere seeing is slightly overestimated:
$\epsilon{_{FA,LOW}}$=0.42$\pm$ 0.28~arcsec.
However, the grid-nested mode (H$_{SL,HIGH}$=48.9~m) gave excellent result: $\epsilon{_{FA,HIGH}}$=0.35 $\pm$
0.24~arcsec, thus confirming the importance of the high horizontal resolution configuration to obtain reliable
forecasts.
The corresponding correlation plots (for all 15 nights) are displayed on Figures \ref{fig_corr1} (surface layer
thickness) and \ref{fig_corr2} (free atmosphere and total seeings).
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{lascaux_1_fig_1.eps}
\end{center}
\caption{Correlation plot between measured and simulated surface layer thicknesses (black: monomodel
configuration; red: grid-nested configuration). For the simulated values only the mean values between 12 UTC
and 16 UTC are considered.
For each configuration of the simulation (high and low horizontal resolution) the error bars are
reported for one point only (and are equal to $\sigma$). Units are in meter (m).}
{\label{fig_corr1}}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{lascaux_1_fig_2.eps}
\end{center}
\caption{Correlation plot between measured and simulated total (top) and free atmosphere (bottom) seeing
(black: monomodel configuration; red: grid-nested configuration).
For the simulated values only the mean values between 12 UTC and 16 UTC are considered.
For each configuration of the simulation (high and low horizontal resolution) the error bars are
reported for one point only (and are equal to $\sigma$). Units are in arcsec.}
{\label{fig_corr2}}
\end{figure}
\par
On Fig. \ref{fig3} is displayed the temporal evolution of the C$_N^2$ profile in the free atmosphere (1-12 km vertical
slab) for one night.
This night is a good example of how the model is active even at such an altitude.
The vertical distribution of the optical turbulence changes in time with a non negligible dynamic from a quantitative
point of view. It is even more visible in the high horizontal resolution mode (grid-nesting).
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{lascaux_1_fig_3.eps}
\end{center}
\caption{Temporal evolution on 18 hours of the C$_N^2$ profiles in the vertical slab (1-12) km related to one winter
night (04 July 2005), in logarithmic scale. On the left, in the low horizontal resolution mode; on the right, in the
high horizontal resolution model.}
{\label{fig3}}
\end{figure}
\section{Conclusion}
We studied the performances of the Meso-Nh mesoscale model in reconstructing optical turbulence profiles, looking at
the Dome C area, in the internal Antarctic Plateau.
This study was focused on the winter season.
The results concerning the optical turbulence computations are resolution dependent.
A high horizontal resolution mode seems to be mandatory to realize realistic optical turbulence forecast.
In high horizontal resolution mode, Meso-Nh gave excellent results: H$_{SL,HIGH}$=48.9$\pm$ 7.6~m, to be compared to
H$_{SL,OBS}$=35.3 $\pm$ 5.1~m.
The resulting free atmosphere seeing is $\epsilon{_{FA,HIGH}}$=0.35 $\pm$ 0.24~arcsec, very close to the observed one,
$\epsilon{_{FA,OBS}}$=0.3 $\pm$ 0.2~arcsec.
\section*{Acknowledgements}
This study has been funded by the Marie Curie Excellence Grant (FOROT) - MEXT-CT-2005-023878.
|
\section{Problem Definitions and Related Work}
We are given a tree $T=(V,E)$ along with edge costs $c\colon
E\rightarrow\mathbb{R}_{\geq 0}$ inducing a distance function $d\colon
V\times V\rightarrow\mathbb{R}_{\geq 0}$. With each node $u$ we associate a
non-negative \emph{penalty} $\pi(u)$. Let $Y$ be a subtree of $T$.
Then $c(Y)$ denotes the \emph{setup cost} of $Y$ and is given by the
sum $\sum_{e\in E(Y)}c(e)$ of its edge costs. A node is \emph{covered
directly by $Y$} if it lies in $Y$. If some node $u$ is not covered
directly it imposes the penalty $\pi(u)$ on $Y$. The \emph{direct
covering subtree problem} \cite{kim-etal:location-tree-shaped} asks
for a subtree $Y$ such that the sum of setup cost $c(Y)$ and the total
penalty $\sum_{u\notin V(Y)}\pi(u)$ is minimized.
The \emph{indirect covering subtree problem} goes one step further. A
node $u$ is said to be covered (indirectly) if it lies within a given
distance from $Y$. Again, a penalty is imposed on $Y$ if it does not
cover $u$. More formally, we assign to each node $u$ some
\emph{radius} $\rho(u)$. The \emph{penalty imposed on $Y$ by $u$} is
given as
\begin{displaymath}
p(u,Y):=
\begin{cases}
0 & \text{ if } d(u,Y)\leq \rho(u)\\
\pi(u) & \text{ otherwise\,.}
\end{cases}
\end{displaymath}
If $U\subseteq V$ is a set of nodes then $p(U,Y):=\sum_{u\in U}
p(u,Y)$ is the penalty imposed on~$Y$ by~$U$. The total penalty
imposed on $Y$ is given by $p(Y):=p(V,Y)$. The indirect covering
subtree problem \cite{kim-etal:location-tree-shaped} asks for a
subtree $Y$ of $T$ such that the total cost $c(Y)+p(Y)$, given by the
sum of setup and penalty cost, is minimum among all subtrees of $T$.
If we require that $Y$ be a node rather than a subtree we obtain the
\emph{single maximum coverage location problem}
\cite{megiddo:maximum-coverage-location,spoerhase+wirth:single-max-coverage}.
It is not hard to see that single maximum coverage location is a
special case of indirect covering subtree. (Scale all edge lengths
and radii with a sufficiently large factor while leaving the penalties
unchanged.)
\subsection{Related Work and Previous Results}
The \emph{multiple} maximum coverage location problem allows the
placement of a an arbitrary set of $r$ nodes. On general graphs this
problem is NP-hard \cite{megiddo:maximum-coverage-location} while it
can be solved in time $O(rn^2)$ on trees \cite{Tamir:p-median-tree}.
This leads to an $O(n^2)$ algorithm for the \emph{single} maximum
coverage location problem on trees by setting $r=1$. Kim et al.\
\cite{kim-etal:location-tree-shaped} provide a faster algorithm
running in $O(n\log^2n)$. Their algorithm works even for the more
general indirect covering subtree problem. Recently a slightly faster
algorithm for single maximum coverage location with time
$O(n\log^2n/\log\log n)$ has been reported
\cite{spoerhase+wirth:single-max-coverage}. Finally, we remark that
\emph{direct} covering subtree can be solved in linear time
\cite{kim-etal:location-tree-shaped}.
\subsection{Contribution and Outline of this Paper}
In this paper we show that indirect covering subtree can be solved in
$O(n\log n)$. This improves upon the previously best algorithms for
this problem and single maximum coverage location on trees. Our
result also implies faster algorithms for the $(1,X)$-medianoid
problem and the $(1,p)$-centroid problem on trees. Specifically, we
obtain an $O(n\log n)$ algorithm for $(1,X)$-medianoid and $O(n^2\log
n\log w(T))$ and $O(n^2\log n\log w(T)\log D)$ algorithms for the
discrete and absolute $(1,p)$-centroid problems on trees,
respectively. Here, $w(T)$ denotes the total weight of the tree and
$D$ is the maximum edge length. The previously best algorithms are
slower by factor of $O(\log n/\log\log n)$
\cite{spoerhase+wirth:rp-centroid}.
Our algorithm employs the same dynamic programming framework used by
Kim et al.\ \cite{kim-etal:location-tree-shaped}. However, we improve
one of their core routines by using a more sophisticated technique to
subdivide trees. This technique, called two-terminal subtree
subdivision (TTST), is a simplification of the recursive coarsening
strategy \cite{spoerhase+wirth:single-max-coverage} used for solving
single maximum coverage location on a tree. The key source of our
improvement is that we manage to avoid explicitly sorting the nodes
according to their distances and radii during the recursion, which has
been necessary in the coarsening approach and also in the original
algorithm of Kim et al. One further advantage of our algorithm is
that it is a lot simpler than the recursive coarsening algorithm.
The two-terminal subtree technique has proved successful also for
other location problems
\cite{spoerhase+wirth:relaxed-mgf-dam,spoerhase+wirth:jda-stackelberg}.
I believe that there are further problem classes where it can be
applied.
The paper is organized as follows. In
Section~\ref{sec:algorithm-kim-et} we briefly outline the algorithm of
Kim et al. This is necessary, since our result relies on an
improvement of a subroutine of that algorithm. The improved routine
is then described in Section~\ref{sec:an-ovlog-v}. In Section
\ref{sec:matching-lower-bound} we provide a matching lower bound on
the running time needed to solve indirect covering subtree. Finally,
we discuss implications on related problems such as competitive
location problems in Section~\ref{sec:impl-relat-probl}.
\section{The Algorithm of Kim et al.}\label{sec:algorithm-kim-et}
In the sequel we will briefly describe the algorithmic approach of Kim
et al.\ \cite{kim-etal:location-tree-shaped} for solving the indirect
covering subtree problem.
Let's first fix some conventions and notations. We assume that the
input tree $T$ is rooted at some distinguished node $s$. For
technical reasons we shall adopt the convention that $s$ is the father
of itself. Let $v$ be an arbitrary node. Then $f(v)$ denotes the
father of $v$. We write $T_v$ for the subtree of $T$ hanging from $v$
and $T_v^+$ for the union of $T_v$ with the edge $(v,f(v))$.
Kim et al.\ reduce the solution of the problem to the computation of
the values $p(v), p(T_v,v)$ and $p(T_v,f(v))$ for all nodes $v$. They
show that one can determine an optimum to the subtree location problem
in linear time once these values have been precomputed for all nodes
$v$.
To convince ourselves, assume that we have computed the values $p(v)$,
$p(T_v,v)$ and $p(T_v,f(v))$ for all $v\in V$. Then define
\begin{equation*}
C(v):=\min\{\,c(Y)+p(T_v,Y)\mid Y\text{ is subtree of } T_v\text{ containing
} v\,\}\, ,
\end{equation*}
and
\begin{equation*}
C^+(v):=\min\{\,c(Y)+p(T_v,Y)\mid Y\textnormal{ is subtree of }
T_v^+ \text{ containing } f(v)\,\}\,.
\end{equation*}
It is not hard to see that the optimum cost can now be expressed by
\begin{equation*}
\min_{v\in V} (C(v)+p(v)-p(T_v,v)\,).
\end{equation*}
Moreover, the $C(\cdot)$- and $C^+(\cdot)$-values can be computed in
linear time by means of a simple bottom-up dynamic programming
approach. To this end assume that $v$ is a leaf of $T$ then
\begin{equation*}
C(v)=0 \text{\quad and\quad } C^+(v)=\min\{p(v,f(v)), c(v,f(v))\}\, .
\end{equation*}
Otherwise, we have
\begin{equation*}
C(v)=\sum_{u\text{ is son of } v} C^+(u)\, ,
\end{equation*}
and
\begin{equation*}
C^+(v)=\min\{C(v)+c(v,f(v)), p(T_v,f(v))\}\, .
\end{equation*}
From this equations it follows that an optimal solution can be
determined in linear time in a bottom-up fashion once the values
$p(v), p(T_v,v)$ and $p(T_v,f(v))$ have been computed for all $v\in
V$. Kim et al.\ suggest an algorithm with running time $O(n\log^2 n)$
to compute these values.
This algorithm is based on the so-called \emph{bitree model}. In this
model, each (undirected) edge $(u,v)$ of the input tree is replaced
with two anti-parallel, directed arcs $(u,v)$, $(v,u)$. We call the
resulting tree $T'$ \emph{bitree} of $T$. With each arc $(u,v)$ of
the bitree we associate a cost $c_{T'}(u,v)$ representing the length
of this arc. But in contrast to the edges of the input tree $T$ we
allow these costs to be negative and asymmetric. This induces a
distance function $d_{T'}\colon V\times V\rightarrow\mathbb{Q}$ where
$d_{T'}(u,v)$ is the length of the unique $u$-$v$-path in $T'$. Now
we define the penalty cost $p'(u,v)$ imposed on $v$ by $u$ to be zero
if $d_{T'}(u,v)\leq\rho(u)$ and $\pi(u)$ otherwise. We set
$p'(v)=\sum_{u\in V}p'(u,v)$.
The algorithm of Kim et al.\ is based on a subroutine for efficiently
computing $p'(v)$ for all nodes $v$ on a given bitree $T'$. By means
of such a subroutine it is then possible to calculate the values
$p(v)$, $p(T_v,v)$ and $p(T_v,f(v))$ for all $v$ in the input tree
$T$. It follows from the above discussion that the knowledge of these
values enables us to identify an optimal tree-shaped facility.
It remains to explain how we can employ such a subroutine to determine
$p(v)$, $p(T_v,v)$ and $p(T_v,f(v))$ for all nodes $v$ of the input
tree $T$ which, in turn, is sufficient to build an optimal tree-shaped
facility.
First we describe how we can determine $p(\cdot)$. For this purpose
we simply set $c_{T'}(u,v):=c_{T'}(v,u):=c_T(u,v)$ for all edges
$(u,v)$ of the input tree $T$. It is then immediately clear that
$p(v)=p'(v)$ for all $v\in V$.
In order to compute $p(T_v,v)$ for all $v\in V$ we set
$c_{T'}(u,f(u)):=c_T(u,f(u))$ and $c_{T'}(f(u),u):=-\infty$ for all
$u\neq s$. This construction ensures that the penalty cost $p'(u,v)$
is always zero if $u$ is \emph{not} a descendant of $v$. Thus
$p(T_v,v)=p'(v)$ holds for this construction.
Finally, we wish to determine $p(T_v,f(v))$. To this end we introduce
on each edge $(v,f(v))$ of $T$ a new node $f'(v)$ such that edge
$(v,f'(v))$ has length $c_T(v,f(v))$ and edge $(f'(v),f(v))$ has
length zero. This increases the number of nodes to $2n-1$. We set
$\pi(f'(v))$ and $\rho(f'(v))$ to zero. It is easy to see that
$p(T_v,f(v))$ in the original tree equals $p(T_u,u)$ in the newly
constructed tree where $u:=f'(v)$. Hence the problem of computing
$p(T_v,f(v))$ for all nodes $v$ can be reduced to the problem of
computing $p(T_v,v)$, which has been described before.
Kim et al.\ provide a subroutine to compute the $p'(\cdot)$-values on
a bitree with $n'$ nodes in $O(n'\log ^2n')$ time which yields
immediately.
\begin{theorem}[\cite{kim-etal:location-tree-shaped}]
The indirect covering subtree problem on a tree can be solved in
$O(n\log^2n)$.\ensuremath{\Box}{}
\end{theorem}
\section{An \emph{O}(\emph{n}$\;$log \emph{n}) Algorithm}\label{sec:an-ovlog-v}
In this section we describe an algorithm for the indirect covering
subtree problem with running time $O(n\log n)$.
Our algorithm uses the algorithmic framework of Kim et al.\ described
in Section~\ref{sec:algorithm-kim-et}. Specifically, we will provide
an improved routine for computing the values $p'(\cdot)$ on a given
bitree in $O(n\log n)$ which can then be extended to an algorithm with
the same asymptotic running time for solving indirect covering
subtree.
The basic approach of the routine of Kim et al.\ for computing
$p'(\cdot)$ is divide-and-conquer. It partitions the node set $V$
into two sets $V_1, V_2$ of bounded size such that both induce
subtrees and have exactly one node (called centroid
\cite{kang-ault:properties-centroid-tree}) in common. Then it sorts
the sets $V_i$ and computes, by means of a clever merge-and-scan
procedure, for all $v\in V_i$ the penalties $p'(v,V_j)$ of the users
in $V_j$ where $j\neq i$. Applying the routine recursively to the
sub-bitrees induced by $V_1,V_2$ one can determine the
$p'(v,V_i)$-values also for each $v\in V_i$. Finally, one obtains the
total penalty $p'(v,V)$ of any node $v\in V$ by adding $p'(v,V_1)$ and
$p'(v,V_2)$.
Our routine proceeds in a similar way but uses a more sophisticated
subdivision, which allows us to avoid the explicit sorting thereby
supressing the additional $\log$-factor. Spoerhase and Wirth
\cite{spoerhase+wirth:relaxed-mgf-dam,spoerhase+wirth:jda-stackelberg}
used an analogous subdivision technique for solving competitive
location problems on undirected trees.
Consider the (undirected) input tree $T=(V,E)$. We may assume that $T$
has maximum degree three. Otherwise, we can split nodes of larger
degree by introducing suitable zero-length edges and zero-weighted
nodes. Let $T'$ be the bitree corresponding to $T$.
If $s$ and $t$ are distinct nodes then $T'_{st}$ denotes the maximal
sub-bitree of $T'$ having $s$ and $t$ as leaves. Let $V_{st}$ be the
node set of $T'_{st}$. We call $s$ and $t$ \emph{terminals} and
$T'_{st}$ \emph{two-terminal sub-bitree} (TTSB).
Our algorithm divides the input bitree recursively into TTSBs. Since
we are dealing with a degree-bounded bitree we can subdivide any TTSB
$S$ into at most five TTSBs, called \emph{child TTSBs}. Each of these
child TTSBs has at most $\frac12|S|+1$ nodes.
\begin{lemma}\label{lem:subdiv-degree-ttsb}
Let $S$ be a TTSB with maximum degree three. Then $S$ can be
partitioned into at most five edge-disjoint TTSBs each of which
having at most $\frac12|S|+1$ nodes. This subdivision can be
computed in $O(|S|)$ time.
\end{lemma}
\begin{proof}
Let $S$ be a TTSB with maximum degree three and terminals $u$ and
$v$. Let $m$ be the unweighted median of $S$, which can be computed
in $O(|S|)$ by means of Goldman's algorithm
\cite{goldman:optimal-center-location}. (All node and arc weights
are temporarily set to one throughout this proof.) It is a
well-known fact that $m$ has the following property: Each of the
connected components of $S-m$ has at most $\frac12|S|$ nodes.
Hence, if $m$ lies on path $P(u,v)$ then $S-m$ contains at most
three components that form the desired subdivision (confer left part
of Figure~\ref{fig:tree-subdivision}). If $m$ does not lie on
$P(u,v)$ then consider the node $m'$ on $P(u,v)$ that is closest to
$m$ (confer right part of Figure~\ref{fig:tree-subdivision}). Then
$S-\{m,m'\}$ has at most five connected components. All of the
child TTSBs obtained this way have clearly at most $\frac12|S|+1$
nodes.
\end{proof}
\begin{figure}[htp]
\centering
\input{fig-tree-subdivision}
\caption{The two cases in the subdivision of a TTSB.}
\label{fig:tree-subdivision}
\end{figure}
Consider a TTSB $T'_{st}$. We introduce the lists $L_{d,s}(T'_{st})$
and $L_{\rho,s}(T'_{st})$. Both lists contain all nodes $v$ of
$T'_{st}$ sorted in increasing order with respect to the values
$d_{T'}(s,v)$ and $\rho(v)-d_{T'}(v,s)$, respectively. The lists
$L_{d,t}(T'_{st})$ and $L_{\rho,t}(T'_{st})$ are defined symmetrically.
The algorithm computes $p'(v,T'_{st})$ for all $v\in T'_{st}$ as well
as the four lists $L_{d,s}(T'_{st})$, $L_{d,t}(T'_{st})$,
$L_{\rho,s}(T'_{st})$ and $ L_{\rho,t}(T'_{st})$ for any TTSB
$T'_{st}$ occurring during the recursion. We shall see that these
information can be propagated inductively from child towards parent
TTSBs such that we will have computed $p'(\cdot, T')=p'(\cdot)$ at the
top of the recursion.
To this end consider an arbitrary TTSB $S=T'_{st}$ being subdivided
into at most five child TTSBs $S_i$ with terminals $s_i$, $t_i$.
Moreover assume that we have already computed $p'(\cdot)$ and the four
lists corresponding to $S_i$ for all $S_i$.
We start with computing $L_{d,s}(S)$. To this end we maintain a list
$L$ which is initialized with an empty list. Now we perform the
following operations for all child TTSBs $S_i$: Assume that $s_i$ is
the terminal of $S_i$ closest to $s$. Then the list $L_{d,s_i}(S_i)$
contains all nodes $v\in S_i$ with associated sorting keys
$d_{T'}(s_i,v)$. Now we create a copy $L'$ of this list and add the
value $d_{T'}(s,s_i)$ to all sorting keys which does not affect its
order. As a result $L'$ contains all nodes $v$ of $S_i$ sorted with
respect to their distance $d_{T'}(s,v)$ from terminal $s$. Finally we
merge $L$ with $L'$. After having carried this out for all child
TTSBs $S_i$ the list $L$ equals the list $L_{d,s}(S)$ we are looking
for. The list $L_{\rho,s}(S)$ is computed very similarly with the
difference that we \emph{subtract} the value $d_{T'}(s_i,s)$ from the
sorting keys when handling the list $L_{\rho,s}(S_i)$. The respective
lists for terminal $t$ are computed symmetrically. The total running
time for computing the four lists associated with $S$ is $O(|S|)$
since we handle a constant number of child TTSBs.
We are now going to explain how $p'(v,S)$ can be determined for all
$v\in S$. To this end assume that $v$ is contained in some $S_i$.
Since we already know $p'(v,S_i)$ by the inductive hypothesis it
suffices to determine $p'(v,S_j)$ for all $S_j\neq S_i$ and to add
these values to $p'(v,S_i)$. Consider an arbitrary $S_j\neq S_i$ and
assume that $s_i,s_j$ are the terminals of these TTSBs closest to each
other. We create a copy $L'$ of list $L_{\rho,s_j}(S_j)$ and
substract the distance $d_{T'}(s_j,s_i)$ from all sorting keys in this
list. As a result $L'$ contains all nodes $u$ of $S_j$ sorted with
respect to the key $\rho(u)-d_{T'}(u,s_i)$. At this point we can
compute $p'(v,S_j)$ for \emph{all} $v\in S_i$ by using the
merge-and-scan procedure of Kim et al. To this end we merge the
sorted list $L'$ with the sorted list $L_{d,s_i}(S_i)$ and store the
result in $L'$. We assume that the nodes in $L'$ are sorted in
increasing order with respect to their numerical sorting keys. Ties
are broken in favor of nodes in $S_i$. This can be achieved in linear
time $O(|S_i|+|S_j|)$. Now recall that a node $u\in S_j$ imposes a
penalty $\pi(u)$ on $v\in S_i$ if $d_{T'}(u,v)>\rho(u)$ or
equivalently $d_{T'}(s_i,v)>\rho(u)-d_{T'}(u,s_i)$. This is
tantamount to that $u$ precedes $v$ in $L'$. Hence, in order to
compute $p'(v,S_j)$ for \emph{all} $v\in S_i$ it suffices to traverse
$L'$ once. In doing so, one can maintain the sum of penalties of all
nodes $u\in S_j$ encountered so far, which equals the penalty
$p'(v,S_j)$ whenever a node $v\in S_i$ is reached.
The running time of this merge-and-scan operation is $O(|S_i|+|S_j|)$
since the necessary sorted lists have already been computed. Thus we
can compute $p'(v,S)$ for all $v\in S$ in total time $O(|S|)$ once we
know the $p'$-values and respective lists for all child TTSBs of~$S$.
Note that the bottom of the recursion, that is, when $T'_{st}$
consists merely of the pair $(s,t)$ and $(t,s)$ of anti-parallel arcs
can trivially be handled constant time.
To sum up, this leads us to an algorithm whose running time $h(|S|)$
can be described by the following recurrence
\begin{displaymath}
h(|S|)=O(|S|)+\sum_{i=1}^{k} h(|S_i|)\, ,
\end{displaymath}
where $k\leq 5$, $\sum_{i=1}^k |S_i|=|S|+4$ and $|S_i|\leq\frac12
|S|+1$. This implies that $h(n)$ is $O(n\log n)$.
\begin{theorem}\label{thm:disc-single-cov-tree}
The indirect covering subtree problem and hence also the single
maximum coverage location can be solved in time $O(n\log n)$. \ensuremath{\Box}{}
\end{theorem}
\section{A Matching Lower Bound}\label{sec:matching-lower-bound}
In this section we complement our algorithm with a lower bound
$\Omega(n\log n)$ on the running time for solving single maximum
coverage location on a tree. This shows that (for certain
computational models) our algorithm is optimal.
We make use of a recent result which is summarized in the following
theorem.
\begin{theorem}[\cite{ben-amram+galil:RAM}]\label{thm:galil}
Let $W\subseteq\mathbb{R}^n$. If $W$ is recognized in time $t(n)$ on a
real-number RAM that supports direct assignments, memory access,
flow control, and arithmetic instructions $\{+,-,\times,/\}$ then
$t(n)=\Omega(\log\beta(W^\circ))$. \ensuremath{\Box}{}
\end{theorem}
Here, $W^\circ$ denotes the interior of $W$ and $\beta(W')$ denotes
the number of connected components of some set $W'\subseteq\mathbb{R}^n$.
To prove our lower bound we introduce a variant of the set
disjointness problem. To this end let $n\in \mathbb{N}$. The set
$W_n\subseteq \mathbb{R}_+^{2n}$ contains all tuples
$(x_1,\ldots,x_n,y_1,\ldots,y_n)$ such that $x_1<\ldots<x_n$ and
$x_i\neq y_j$ for all pairs $i,j$. Consider a permutation $\pi$ on
the set $\{1,\ldots,n\}$ and some tuple
$x_1<y_{\pi(1)}<x_2<y_{\pi(2)}<\ldots<x_n<y_{\pi(n)}$ in $W_n$. It is
easy to see that for different permutations such tuples lie in
different connected components of $W_n^\circ$ so $W_n^\circ$ contains
at least $n!$ connected components. Hence any RAM of the above
described type takes time $\Omega(n\log n)$ to recognize $W_n$.
We establish a linear time reduction from the problem to recognize
$W_n$ to the single maximum coverage location problem on a tree with
$O(n)$ nodes. To this end consider a tuple
$(x_1,\ldots,x_n,y_1,\ldots,y_n)$ for which we want to decide whether
or not it is contained in $W_n$.
First we check if $x_1<\ldots<x_n$. Then we create an edge $(u,v)$ of
some length $c(u,v)>\max\{\,x_i,y_i\mid i=1,\ldots,n\,\}$ and choose
some radius $\rho$ such that $\rho>c(u,v)$. For any $y_i$ we create
two edges $(u,u_i)$ and $(v,v_i)$ of lengths $\rho-y_i$ and
$y_i+\rho-c(u,v)$, respectively. Finally, we create for each $x_i$ a
node $\tilde x_i$ on edge $(u,v)$ with distance $d(u,\tilde
x_i):=x_i$. For each node $z$ in the node set $V:=\{\,u_i,v_i,\tilde
x_i\mid i=1,\ldots,n\,\}\cup\{u,v\}$ we set $\pi(z):=1$ and
$\rho(z):=\rho$, which completes the reduction.
First suppose that we locate a facility outside the path $P(u,v)$.
Assume that the facility is located at some node $u_i$. Then the
distance of $u_i$ to $u$ is positive and $d(u,v_j)\geq \rho$ for any
$j$. Hence, none of the nodes $v_j$ is covered by $u_i$ and the
penalty cost imposed on $u_i$ must be at least $n$. The case where is
the facility is placed at some node $v_i$ is treated analogously.
Now suppose for a moment that we can locate a facility anywhere at the
path $P(u,v)$, that is, also at interior points of edges on $P(u,v)$.
The point $x$ where the facility is located can then be identified
with the distance $d(u,x)$. First, all nodes on $P(u,v)$ are covered
by $x$ since $\rho>d(u,v)$. Due to our construction $x$ covers all
nodes $u_i$ where $x\leq y_i$ and all nodes $v_j$ where $x\geq y_j$.
Thus, the penalty imposed on $x$ is exactly $n$ if $x$ is not
contained in the set $\{y_1,\ldots,y_n\}$. If $x=y_j$ then $x$ covers
both $u_j$ and $v_j$ and hence the penalty is bounded by $n-1$. Since
the facility can only be placed at nodes $\tilde x_j$, that is, at
distances $x_i$ from $u$ we conclude that the minimum penalty cost is
$n$ if the input tuple $(x_1,\ldots,x_n,y_1,\ldots,y_n)$ lies in $W_n$
and $n-1$ otherwise.
\begin{theorem}\label{thm:matching-lower-bound}
Any real-number RAM that complies with Theorem~\ref{thm:galil} takes
at least $\Omega(n\log n)$ time to solve single maximum coverage
location on a tree even for unit penalties and uniform radii. \ensuremath{\Box}{}
\end{theorem}
\section{Implications for Related Problems}\label{sec:impl-relat-probl}
The variant of single maximum coverage location where the facility can
be placed not only at the nodes but also at interior points of edges
is called the \emph{absolute} maximum coverage location problem. Kim
et al.\ show \cite{kim-etal:location-tree-shaped} that a set of $O(n)$
critical points (that is a set of point which is guaranteed to to
contain an optimal point) for absolute single maximum coverage
location can be found in time $O(n\log n)$. We infer that also the
absolute variant can be solved in $O(n\log n)$ on a tree.
Another implication of our result leads us to the realm of
\emph{competitive location}. Let a graph $G=(V,E)$ and $r,p\leq n$ be
given. We assume that the graph is edge and node weighted. Let
$X,Y\subseteq G$ be sets of nodes or interior points of edges. Then
$w(Y\prec X)$ denotes the total weight $\sum\{w(u)\mid u\in V\text{
and }d(u,Y)<d(u,X)\}$ of nodes that are closer to $Y$ than to $X$.
Given some point set $X$ the goal of the \emph{$(r,X)$-medianoid
problem} \cite{Hakimi:competitive-environment} is to identify a set
$Y$ of $r$ points such that $w(Y\prec X)$ is maximized. This maximum
weight is denoted by $w_r(X)$. The goal of the \emph{$(r,p)$-centroid
problem} is to find a $p$-element point set $X$ such that $w_r(X)$
is minimized.
By setting $\rho(u):=d(v,X)-\epsilon$ (where $\epsilon$ is a suitably
small constant) and $\pi(u):=w(u)$ one can easily verify that
$(r,X)$-medianoid is a special case of the multiple maximum coverage
location problem with $r$ servers. On general graphs the problem is
NP-hard \cite{Hakimi:competitive-environment}. It can be solved
efficiently in $O(rn^2)$ on trees \cite{Tamir:p-median-tree}. Our
result leads to an $O(n\log n)$ algorithm for the absolute and the
discrete version of $(1,X)$-medianoid on trees.
\begin{corollary}
The discrete and the absolute $(1,X)$-medianoid problem can be
solved in $O(n\log n)$ on trees. \ensuremath{\Box}{}
\end{corollary}
It is not hard to extend the lower bound provided by
Theorem~\ref{thm:matching-lower-bound} to $(1,X)$-medianoid. Since
the radii of the tree constructed in the reduction are uniform and
hence all equal some number $\rho$, we can furnish each node $z$ on
that tree with a pendant leaf $z'$ at a distance
$d(z,z')=\rho+\epsilon$. The set $X$ contains exactly those pendant
leaves. It is clear that for any node $y$ in this enhanced tree $T'$
the gain $w(y\prec X)$ equals exactly $w(T)-p(y)$ in the original tree
$T$. This implies that both instances lead to the same optimum. Thus
also the algorithm for $(1,X)$-medianoid is optimal in terms of the
running time.
Now let's turn our view to the $(r,p)$-centroid problem. The problem
is known to be $\Sigma^{\text{p}}_2$-complete on general graphs
\cite{spoerhase+wirth:multiple-voting} and NP-hard even on path graphs
\cite{spoerhase+wirth:rp-centroid}. However, both the absolute and
the discrete variant of $(1,p)$-centroid on trees can be solved in
polynomial time $O(n^2\log^2 n\log w(T)/\log\log n)$ and $O(n^2\log^2
n\log w(T)\log D/\log\log n)$, respectively
\cite{spoerhase+wirth:rp-centroid}. Here $D$ is the maximum edge
length of the input tree $T$. Those algorithms rely on $O(n\log
w(T))$ (resp. $O(n\log w(T)\log D)$) calls to a subroutine solving
$(1,X)$-medianoid on a tree.
The algorithm provided here allows us to solve $(1,X)$-medianoid in
$O(n\log n)$ which yields.
\begin{corollary}
The discrete and the absolute $(1,p)$-centroid problem for trees can
be solved in $O(n^2\log n\log w(T))$ and $O(n^2\log n\log w(T)\log
D)$, respectively. \ensuremath{\Box}{}
\end{corollary}
\section{Concluding Remarks}
We have provided an $O(n\log n)$ algorithm for solving the indirect
covering subtree problem which improves upon the previously best
algorithms for this problem and single maximum coverage location on
trees. We have also shown that our algorithm is optimal for certain
unit-cost RAM models. Our result leads also to an optimal algorithm
for $(1,X)$-medianoid and faster algorithms for $(1,p)$-centroid on
trees.
It would be interesting to identify larger problem classes of location
problems on trees where the the two-terminal subtree technique can be
applied. It would also be worth investigating the existence of faster
algorithms on path graphs.
|
\section{Introduction}
Our most stringent constraints on the structure of any hadron follow
from the underlying symmetries of QCD. Translational invariance
dictates that the momentum of a hadron is the sum of the momenta of
all its constituents, giving rise to a powerful sum rule for the
nucleon. Rotational invariance demands that the spins of hadrons have
the values $n/2$ for an integer $n$, hence the nucleon spin is
precisely $1/2$ and not $0.5$ with some experimental error.
Similarly, we know that the electric charge of the nucleon is exactly
$1$ and the net strangeness is $0$. These statements are so
commonplace that they may even appear trivial, but each of these exact
results serves as an entry to different aspects of the dynamics of
QCD.
The momentum of the nucleon is built up from that of the quarks and
gluons. While translational symmetry constrains the entire
contribution, the individual contributions of each quark flavor and
the gluon depend on the details of QCD. Similarly, the spin of the
nucleon arises from the quark spin, quark orbital motion and the gluon
angular momentum, each of which are individually unconstrained by
symmetries but must sum to $1/2$. The charge of the nucleon arises
from the quark charges that yield a total charge of $1$, but the
distribution of this charge in the spatial degrees of freedom probes
the nonperturbative structure of the nucleon. Finally, the strange
quarks in the nucleon occur in precisely matched pairs of quarks and
anti-quarks. Despite the net absence of strangeness, the consequences
of this hidden flavor are felt in many observables. In the following,
I will elaborate on a few of these points and discuss what we
currently know from lattice calculations, phenomenology and
experimental measurements. Along the way, I'll offer my opinion
regarding where our calculations may have to go to provide accurate
results for hadron structure.
The past several years have seen extensive reviews of the calculations
presented each year at the annual Lattice conference. You can find
the few most recent reviews in~\cite{Zanotti:2008zm, Hagler:2007hu,
Orginos:2006zz}, and a very recent and exhaustive collection of
results can be found in~\cite{Hagler:2009ni}. Rather than duplicating
these efforts, I will instead present some of the key examples
mentioned above.\footnote{At the lattice conference I discussed
strangeness in the nucleon. A recent review~\cite{Young:2009ps}
covers much of this, so I will not include these results in this
review.} This will unfortunately prohibit me from discussing all
the hadron structure efforts presented at Lattice 2009, but I hope
these proceedings will still provide a useful overview, nonetheless.
\section{Nucleon Momentum}
The momentum of a nucleon with energy $E$ and mass $m$ is precisely
constrained by Lorentz symmetry as $p^2 = E^2 - m^2$. This is such a
common statement in particle physics that it seems nearly trivial to
even mention it. However, here we want to question how this momentum
arises from the underlying QCD degrees of freedom. In other words,
what is the distribution of this momentum among the nucleon's
constituents? This question concerns the details of the dynamics of
QCD and presents a challenge to lattice calculations of nucleon
structure.
\subsection{Parton Distribution Functions}
The proper field theoretic response to the question above requires
constructing the momentum distributions of the nucleon's constituents,
the parton distribution functions (PDFs). I'll provide a proper
definition of the PDFs shortly, but for the moment we want to focus on
a more intuitive understanding of the quark and gluon distributions,
$q(x)$ and $g(x)$. In the parton model $q(x)$ or $g(x)$ would be the
probability to find a quark, of some flavor, or a gluon with momentum
$p_\mathrm{parton}=x p_\mathrm{nucleon}$. This interpretation is
retained in QCD, however, $q(x,\mu)$ and $g(x,\mu)$ now carry a
renormalization scale $\mu$ that loosely gives the energy resolution
at which the distribution is probed. While this muddies the picture a
bit, the PDFs, nonetheless, are well-defined universal properties of
the nucleon that are probed in many different experiments.
\subsubsection{Phenomenological Results for PDFs}
There is now a well-established industry dedicated to extracting the
PDFs from global analyses. The analyses differ in many details,
including the order of the perturbative expansion used, the treatment
of the quark masses and the handling of the statistical and systematic
errors from the many experimental inputs. These variations have an
impact on the precision ultimately obtained, however, these details do
not impact the discussion at hand. Hence, as merely one example among
several, I show the PDFs from the MSTW collaboration in
\Fig{\ref{x2dist}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=252pt,angle=-90]{plots/pdfs_q2}
\caption{Momentum distribution of quarks and gluons in the nucleon at
$\mu=2~\mathrm{GeV}$. For each parton $xf(x,\mu)$ is plotted where
$f(x,\mu)$ is the corresponding PDF. At large $x$, the up and down
quarks are the largest components of the nucleon momentum. At low
$x$, the gluon dominates and the anti-quark distributions grow. The
curves were generated using the LHAPDF library~\cite{Whalley:2005nh}
and the MSTW2008 NNLO~\cite{Martin:2009iq} dataset at a
renormalization scale of $\mu=2~\mathrm{GeV}$.}
\label{x2dist}
\end{center}
\end{figure}
Field theory effects ultimately handicap any simple interpretation of
the PDFs, however, we can still see that the PDFs at a low scale,
$\mu=2~\mathrm{GeV}$ in \Fig{\ref{x2dist}}, retain features that one
might expect for the nucleon. The up and down quark distributions are
the largest component for high momentum ($x\lesssim 1$). These
distributions in fact have peaks at reasonable values of $x$, as one
might expect for a hadron dominantly composed of two up quarks and one
down quark. However, the most prominent feature in the plot for low
$x$ is clearly the gluon distribution. This is a clear indication of
QCD physics at work. This has a counterpart in the simple observation
that the quark masses directly contribute only about $1\%$ of the
nucleon's mass. Beyond the dynamics of QCD, the presence of the
anti-quarks is a striking field theory effect. This statement may
seem mundane, but we must remember the role played by the anti-quarks
in nucleon structure to appreciate why we go through all the
difficulty to calculate hadronic matrix elements from fully dynamical
lattice QCD rather than being content with models or quenched
calculations.
As mentioned, field theory complicates the interpretation of the PDFs.
To illustrate this, I show the same distributions again in
\Fig{\ref{x100dist}} but now evolved to a scale of
$\mu=100~\mathrm{GeV}$. Any hint of nucleon physics is now well
hidden and all the constituents of the nucleon appear to play nearly
equally important roles. We will return to this issue again when
discussing the evolution of the momentum fraction.
\begin{figure}[t]
\begin{center}
\includegraphics[width=252pt,angle=-90]{plots/pdfs_q100}
\caption{Momentum distribution of quarks and gluons in the nucleon at
$\mu=100~\mathrm{GeV}$. The details are the same as in
\Fig{\protect\ref{x2dist}} but now $\mu=100~\mathrm{GeV}$. At high
scales, the PDFs for the quarks and gluons mix and the simple
non-relativistic nucleon structure, still visible at $\mu=2~\mathrm{GeV}$,
is suppressed at $\mu=100~\mathrm{GeV}$.}
\label{x100dist}
\end{center}
\end{figure}
\subsubsection{Operator Definition of PDFs}
The parton distribution functions can be defined as nucleon matrix
elements of quark and gluon fields separated by light-like distances.
As an example, we record here the operator definition of the
unpolarized quark distribution $q(x,\mu^2)$~\cite{Brock:1993sz},
\begin{equation}
\label{unpol}
q(x,\mu^2) = \frac{1}{2}\int\!\! \frac{d\lambda}{2\pi}\,\, e^{ix p\cdot \lambda n}
\langle p, s | \,\, \overline{q}(-\lambda/2\,\,n)\,\, \slashed{n}\,\,
W_n(-\lambda/2\,\,n,\lambda/2\,\,n)\,\, q(\lambda/2\,\,n)\,\, | p, s \rangle|_{\mu^2}\,.
\end{equation}
For each flavor there are two other twist-two quark distributions, the
helicity and transversity distributions, denoted as $\Delta q
(x,\mu^2)$ and $\delta q (x,\mu^2)$ respectively. The factor
$W_n(-\lambda/2\,\,n,\lambda/2\,\,n)$ is the Wilson line extending
along the arbitrary light-cone direction $n$ from $\lambda/2\,\,n$ to
$-\lambda/2\,\,n$,
\begin{displaymath}
W_n(-\lambda/2\,\,n,\lambda/2\,\,n) =
{\cal P} exp\left( ig \int_{-\lambda/2}^{\lambda/2} d\alpha\,\, A(\alpha n)\cdot n \right)\,.
\end{displaymath}
The scale, $\mu$, and scheme dependence of $q(x,\mu^2)$ comes from the
renormalization of the operator in \Eq{\ref{unpol}}.
The expression in~\Eq{\ref{unpol}} highlights the difficulty of
lattice calculations of PDFs. Lattice QCD calculations are performed
in Euclidean space, however, the PDFs involve quark and gluon fields
separated along the light-cone. These inherently Minkowski-space
observables are difficult to construct explicitly in Euclidean space.
But, as we will see in the next section, the moments in $x$ of these
distributions can be calculated directly in Euclidean space.
\subsection{Moments of Parton Distributions}
As just discussed, the PDFs as a function of $x$ are essentially
Minkowski-space objects. But the moments in $x$ are related to local
operators that can be calculated in Euclidean space on the lattice.
\subsubsection{Mellin Transform:\ from $\mathbf{x}$ to $\mathbf{n}$}
The moments in $x$ are defined as follows,
\begin{displaymath}
\langle x^n \rangle_{q,\mu^2} = \int_{-1}^1\!\!dx\, x^n q(x,\mu^2) = \int_0^1\!\!dx\, x^n \left\{ q(x,\mu^2) - (-1)^n \overline{q}(x,\mu^2) \right\}\,.
\end{displaymath}
The sign in the above equation is determined by the identification of
$q(x)$ for negative $x$ with the anti-quark distribution as
$\overline{q}(x,\mu^2) = - q(-x,\mu^2)$.\footnote{This is a frequent
source of confusion when comparing to phenomenological
determinations of the PDFs. Also note that the sign is different
for the transversity distribution but the same for the helicity
distribution:\ $\delta \overline{q}(x,\mu^2) = \delta q(-x,\mu^2)$
and $\Delta \overline{q}(x,\mu^2) = - \Delta q(-x,\mu^2)$.} These
moments can be related to forward matrix elements of twist-two
operators. First we present the complete result but then sketch the
argument in the following.
\begin{equation}
\label{moments}
\langle p,s | \,\, \overline{q}(0)\,\, \gamma^{\left\{\mu_1\right.}\,\,
iD^{\mu_2} \cdots iD^{\left.\mu_n\right\}} q(0)\,\, | p,s \rangle|_{\mu^2}
= 2 \langle x^n \rangle_{q,\mu^2}\,\, p^{\left\{\mu_1\right.} \ldots p^{\left.\mu_n\right\}}
\end{equation}
The brackets in $T^{\{\mu_1 \cdots \mu_n\}}$ denote symmetrization of
the indices of the tensor $T$ and subtraction of the traces. The
precise meaning of this operation will be clarified shortly.
The derivation of \Eq{\ref{moments}} is almost elementary, but it is
so essential to the method underlying the lattice calculations that we
sketch the arguments using the unpolarized PDFs as an example. The
first step is to introduce light-cone coordinates and gauge. This is
a useful first step in understanding several aspects of the PDFs. In
defining the light-cone coordinates, you introduce two light-cone
directions $n^\mu_\pm = (1,0,0,\pm 1)/\sqrt{2}$. The two new
light-cone coordinates are given by $v\cdot n_\pm = v^{\mp}$ for a
four-vector $v$. (Similarly, $\slashed{n}_\pm = \gamma\cdot n_\pm =
\gamma^{\mp}$.) Light-cone gauge is the choice of $A(x)\cdot n = 0$,
which sets the Wilson line $W_n(-\lambda/2\,\,n,\lambda/2\,\,n)$ to
$1$. Then choosing $n=n_{-}$ in \Eq{\ref{unpol}}, imposing the
light-cone gauge for $n_-$ and relabeling $\lambda$ as $y^-$ gives
\begin{equation}
\label{lcpdfs}
q(x,\mu^2) =
\frac{1}{2}\int\!\! \frac{dy^-\!\!}{2\pi}\,\, e^{ix p^+y^-}
\langle p,s | \,\, \overline{q}(-y^-/2)\,\, \gamma^+\,\,
q(y^-/2)\,\, | p,s \rangle|_{\mu^2}\,.
\end{equation}
The next step in relating the moments to local operators is to use the
known limited support of $q(x)$ to the region $-1\le x\le 1$ to expand
the integration range used to define the moments,
\begin{displaymath}
\langle x^n \rangle_{q,\mu^2} =
\int_{-1}^1\!\!dx\, x^n q(x,\mu^2) =
\int_{-\infty}^\infty\!\!dx\, x^n q(x,\mu^2)\,.
\end{displaymath}
The remaining steps are relatively elementary. The key sequence
follows from combining the above with \Eq{\ref{lcpdfs}},
\begin{displaymath}
\int_{-\infty}^\infty\!\!dx\, x^n e^{ixp^+y^-} =
(ip^+)^{-n} \int_{-\infty}^\infty\!\!dx\, (\partial^+)^n e^{ixp^+y^-} =
\end{displaymath}
\begin{displaymath}
= (-1)^n (ip^+)^{-n} \int_{-\infty}^\infty\!\!dx\, e^{ixp^+y^-} (\partial^+)^n =
(p^+)^{-n} \delta(p^+y^-) (i\partial^+)^n
\end{displaymath}
where $\partial^+ = \partial / \partial y^-$. The above manipulations
must be understood as acting under the $\int dy^-$ in
\Eq{\ref{lcpdfs}}. The final result is
\begin{displaymath}
\langle x^n \rangle_{q,\mu^2} =
2^{-1} (p^+)^{-(n+1)} \langle p,s | \,\, \overline{q}(0)\,\, \gamma^+\,\,
(i\partial^+)^n q(0)\,\, | p,s \rangle|_{\mu^2}\,.
\end{displaymath}
By inspection, this expression can be seen as the light-cone
coordinate, light-cone gauge form of the following
\begin{displaymath}
\langle p,s | \,\, \overline{q}(0)\,\, \slashed{n}\,\,
(in\cdot D)^n q(0)\,\, | p,s \rangle|_{\mu^2}
= 2 \langle x^n \rangle_{q,\mu^2} (n\cdot p)^{n+1}\,.
\end{displaymath}
This expression relates $\langle x^n \rangle_{q,\mu^2}$ to diagonal
matrix elements of local operators. The more familiar form follows
from writing the above as
\begin{displaymath}
n_{\mu_1} n_{\mu_2} \ldots n_{\mu_n}
\langle p,s | \,\, \overline{q}(0)\,\, \gamma^{\mu_1}\,\,
iD^{\mu_2} \cdots iD^{\mu_n} q(0)\,\, | p,s \rangle|_{\mu^2}
= 2 \langle x^n \rangle_{q,\mu^2}\,\,
n_{\mu_1} \ldots n_{\mu_n} p^{\mu_1} \ldots p^{\mu_n}\,.
\end{displaymath}
This form makes it clear precisely what the symmetrization and trace
removal in \Eq{\ref{moments}} means.
\subsubsection{Phenomenology of $\mathbf{\x{u-d}}$}
The global analysis illustrated in \Figs{\ref{x2dist}} and
\ref{x100dist} can also be used to examine the moments of the PDFs.
This time taking as an example the results from the CTEQ
collaboration, I plot the results for $\x{q}$ and $\x{g}$ in
\Fig{\ref{avex}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=252pt,angle=-90]{plots/x_vs_q}
\caption{Running of the momentum fraction of the quarks and gluons in
the nucleon. As $\mu$ increases, the gluon momentum fraction
$\x{g}$ increases towards the asymptotic value of $4/7$ and $\x{q}$,
for each of the quark flavors, approaches the common value of
$3/28$. (These values hold for the $4$ flavor theory.) Despite the
unique limiting values for $x{}$, the non-perturbative input at low
$\mu$ clearly dictates the values of the momentum fraction for each
parton over many orders of magnitude. The results come from the
LHAPDF library~\cite{Whalley:2005nh} using the
CTEQ6.6C~\cite{Nadolsky:2008zw} dataset.}
\label{avex}
\end{center}
\end{figure}
These results follow from numerically integrating curves similar to
those in \Figs{\ref{x2dist}} and \ref{x100dist}. Furthermore, the
state-of-the-art results for the particular quantity of interest to
us, $\x{u-d}$, are collected in \Tab{\ref{phenox}}. The
phenomenological analyses often present results in terms of the
so-called valence distribution $q_v(x) = q(x) - \overline{q}(x)$ and
the anti-quark distribution $\overline{q}(x)$, but \Tab{\ref{phenox}}
gives the correct combination that is comparable to lattice results.
\begin{table}[t]
\begin{center}
\begin{tabular}{|l|l|l|}\hline
& $\left[\,x\,\right]_{u_v-d_v}$ & $\x{u-d}$ \\\hline
ABMK & $0.1790 \pm 0.0023$ & $0.1646 \pm 0.0027$ \\\hline
BBG & $0.1747 \pm 0.0039$ & $0.1603 \pm 0.0041$ \\\hline
JR & $0.1640 \pm 0.0060$ & $0.1496 \pm 0.0062$ \\\hline
MSTW & $0.1645 \pm 0.0046$\footnotemark{} & $0.1501 \pm 0.0048$ \\\hline
AMP06 & $0.1820 \pm 0.0056$ & $0.1676 \pm 0.0058$ \\\hline
BBG & $0.1754 \pm 0.0041$ & $0.1610 \pm 0.0043$ \\\hline
\end{tabular}
\caption{Phenomenological values for $\x{u-d}$ at $\mu=2~\mathrm{GeV}$. For
each calculation we give the moment of the non-singlet valence
distribution, denoted by $\left[\,x\,\right]_{u_v-d_v}$. (The
square brackets denote simply the integral over the region $0\le
x\le 1$ as opposed to the region $-1\le x\le 1$ used in the angular
brackets.) These values were collected in~\cite{Alekhin:2009ni}.
The original references are \cite{Alekhin:2009ni} (ABMK),
\cite{Blumlein:2006be,Blumlein:2004ip} (BBG),
\cite{JimenezDelgado:2008hf} (JR), \cite{Martin:2009bu} (MSTW),
\cite{Alekhin:2006zm} (AMP06) and
\cite{Blumlein:2006be,Blumlein:2004ip} (BBG $\mathrm{N^3LO}$). This
is combined with the result
$\left[\,x\,\right]_{\overline{u}-\overline{d}} = -0.0072 \pm
0.0007$ from \cite{Alekhin:2009ni} to produce a result for
$\x{u-d}$, which can be compared to lattice calculations.}
\label{phenox}
\end{center}
\end{table}
\footnotetext{The error given in~\cite{Martin:2009bu} is asymmetric,
but here we take only the upper error to compare to the lattice
results that are all higher than the phenomenological values.}
What is particularly interesting about the momentum fraction $\x{}$ is
that it obeys a sum rule,
\begin{equation}
\label{momsumrule}
1 = \sum_q\, \x{q,\mu^2} + \x{g,\mu^2}
\end{equation}
where $\sum_q$ is the sum over all relevant flavors. This sum rule
imposes constraints on the scale evolution of $\x{\mu^2}$ such that
$\x{\mu^2}$ asymptotically approaches a perturbatively calculable
limit for large $\mu$.
\begin{displaymath}
\lim_{\mu\rightarrow\infty} \x{q,\mu^2} = \frac{3}{16+3N_\mathrm{F}}
\end{displaymath}
\begin{displaymath}
\lim_{\mu\rightarrow\infty} \x{g,\mu^2} = \frac{16}{16+3N_\mathrm{F}}
\end{displaymath}
For $N_\mathrm{F}=4$, we have $\lim\,\x{q}=3/28\approx 10\%$ and
$\lim\,\x{g}=4/7\approx 60\%$. However, we can clearly see in
\Fig{\ref{avex}} that the asymptotic results for $\x{}$ bear little
resemblance to the results at any reasonable value of $\mu$. In fact,
the hierarchy shown in \Fig{\ref{avex}} is quite clear. The gluons
carry about $40\%$ of the nucleon momentum, the up and down quarks
carry about $35\%$ and $20\%$, and the strange takes most of the
remaining $5\%$. It is precisely this pattern that we eventually hope
to understand from lattice calculations.
\subsection{Lattice Calculation of $\mathbf{\x{u-d}}$}
The lowest non-trivial moments of the unpolarized quark PDFs are
$\x{q}$.\footnote{The lowest moments $\langle 1\rangle_q$ are known in
terms of the quark valence structure of the nucleon.} These are the
quark contributions that enter the momentum sum rule in
\Eq{\ref{momsumrule}}. Here I use the particular combination
$\x{u-d}$ as a benchmark observable to evaluate the lattice
calculation of moments of PDFs.
\begin{displaymath}
\left.\langle p, s| \overline{q}\gamma^{\left\{\mu\right.} iD^{\left.\nu\right\}} \tau^3 q |p,s\rangle\right|_{\mu^2} = 2 \x{u-d,\mu^2} p^{\left\{\mu\right.} p^{\left.\nu\right\}}
\end{displaymath}
Here $q=(u,d)$ is the doublet of the light quarks. The flavor
combination $u-d$ eliminates disconnected diagrams, which would
otherwise require a substantial computational investment.
Additionally, this combination also eliminates any mixing with gluonic
operators, thus greatly simplifying the renormalization of this
quantity.~\footnote{Strictly speaking, the disconnected diagrams and
the mixing with gluons only vanish for lattice actions with an exact
flavor symmetry.} Without the mixing, this observable then only
requires a multiplicative renormalization. Finally, you can calculate
the bare matrix elements using nucleons with $\vec{p}=0$. Hence,
$\x{u-d}$ is basically the most accurate scale-dependent observable in
nucleon structure that we calculate on the lattice.
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/x_all_v1_new}
\caption{World's dynamical lattice QCD results for $\langle
x\rangle_{u-d}$. The lattice calculations all overestimate the
phenomenologically determined value. They also show a fair amount
of scatter amongst themselves. The possible role of finite size and
renormalization effects are discussed in the text. The lattice
results are from~\cite{Allton:2007hx,Allton:2008pn,ohta:email} (RBC
$N_\mathrm{F}=2+1$), \cite{Lin:2008uz,Aoki:2004ht} (RBC
$N_\mathrm{F}=2$), \cite{Hagler:2007xi} (LHPC),
\cite{harraud:email,liu:email} (ETMC) and \cite{james:email}
(QCDSF). The experimental result is generated using the LHAPDF
library~\cite{Whalley:2005nh} and the CTEQ6.6C
dataset~\cite{Nadolsky:2008zw}.}
\label{worldx}
\end{center}
\end{figure}
\Figure{\ref{worldx}} illustrates all the dynamical lattice QCD
results for $\x{u-d}$ for pion masses less than $700~\mathrm{MeV}$. The
values for $\x{u-d}$, $m_\pi$ and $a$ come from a variety of already
published sources and numerous private communications. (The
references are provided in the caption to \Fig{\ref{worldx}}.) The
most striking feature that you observe in \Fig{\ref{worldx}} is that,
despite the community's efforts to calculate with many actions,
several lattice spacings and volumes and a broad range of pion masses
now approaching $250~\mathrm{MeV}$, the calculations still overestimate the
experimental measurement by at least $30\%$ and maybe as much as a
factor of $2$. The spread amongst the groups obviously suggests some
systematic variations, and I will examine two possible explanations
shortly, but it is important to note that the individual calculations
are visibly much more consistent with themselves than with the other
calculations. Additionally, notice that some groups are beginning to
perform calculations ``at the physical point,'' as illustrated by the
QCDSF calculation in \Fig{\ref{worldx}}. This phrase quite often
means simply $m_\pi<200~\mathrm{MeV}$. Nonetheless, the next generation of
lattice calculations will likely shed some much needed light on the
chiral behavior. However, the discrepancy amongst the groups at pion
masses where we should be able to reliably calculate today is an issue
that needs to be addressed. Without resolving these discrepancies, it
will be hard to confidently establish physical results even with
calculations approaching the physical pion mass.
Before discussing finite size and renormalization effects, I want to
make a quick comment on the lattice spacing dependence of $\x{u-d}$.
This is currently poorly studied. Of the five calculations shown in
\Fig{\ref{worldx}}, only QCDSF has calculated beyond a single lattice
spacing. However, even in that case, the range in $a$ that is used to
establish scaling is not large. Their results are an encouraging hint
that lattice artifacts are not a substantial part of the discrepancy
in \Fig{\ref{worldx}}, but there is nothing universal about such
effects and all the groups must make a stronger effort to calculate at
multiple lattice spacings.
A persistent concern in nucleon structure calculations is the role of
finite size effects. In fact, the results at Lattice 2009 have added
much to this issue even if they haven't resolved it. In
\Fig{\ref{fsx}},
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/x_fs_v2_new}
\caption{Finite size studies for $\x{u-d}$. The dependence of
$\x{u-d}$ on $m_\pi L$ is shown for several calculations. Notice
the essentially flat behavior for $m_\pi L >4$ for all but the
lightest pion mass. The QCDSF calculation at the lightest pion mass
may suggest the emergence of a more substantial finite size effect
as $m_\pi$ is decreased. The curve is a very simple scaling to the
$m_\pi=260~\mathrm{MeV}$ results and is meant solely to guide the eye. The
results are from the same references given in
\Fig{\protect\ref{worldx}}.}
\label{fsx}
\end{center}
\end{figure}
I examine several finite size studies by various collaborations.
Excluding the lightest calculations at $m_\pi=260~\mathrm{MeV}$, one observes
no statistically significant finite size effects for any of the
remaining calculations. These results are consistent with the common
rule-of-thumb that $m_\pi L \approx 4$ is sufficient.~\footnote{My use
of $m_\pi L$ to gauge finite size effects is, of course, not
strictly correct. I am loosely assuming a discussion in or near the
chiral limit in which $1/m_\pi$ will be the dominant length scale.}
However, the recent results of QCDSF at $m_\pi=260~\mathrm{MeV}$ potentially
stand in contrast to the finite size dependence observed at higher
pion masses. This calculation suggests that $m_\pi L = 4$ is at best
just barely sufficient to capture the large volume limit of $\x{u-d}$.
It is possible that this observation would vanish with higher
statistics. In fact, the three calculations at $m_\pi=260~\mathrm{MeV}$ do
essentially agree statistically, but the trend in the results is
suggestive, especially in comparison to all the other results at
larger $m_\pi$. Of course, it is also possible that $\x{u-d}$ would
drop even further with still larger $L$.
In a review of this nature it is difficult to perform a detailed
infinite volume limit of all the results presented at the conference.
However, we can illustrate the impact of finite size effects by simply
making the crude restriction to $m_\pi L > 4$. This is illustrated in
\Fig{\ref{largevolx}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/x_all_v2_new}
\caption{Large volume results for $\x{u-d}$. The results are from the
same references given in \Fig{\protect\ref{worldx}}, but only those
calculations with $m_\pi L>4$ are shown. The results from each
calculation are consistent with a smooth, nearly flat $m_\pi$
dependence. However there is a systematic variation between the
calculations.}
\label{largevolx}
\end{center}
\end{figure}
The picture is certainly clearer and one may even be left with the
impression of some downward curvature for some of the lattice
calculations, but we must guard against wishful thinking. Excluding
the QCDSF results, each of the groups is statistically consistent with
a linear dependence on $m_\pi^2$ with $\chi^2/\mathrm{dof} < 0.75$.
The one exception is QCDSF. A linear fit gives $\chi^2/\mathrm{dof} =
2.1$. This is not overwhelming evidence of non-linearity and all the
apparent curvature comes solely from the lightest point at
$m_\pi=260~\mathrm{MeV}$. This point is $2.8\sigma$ less than the second
lightest QCDSF point. Increased statistics in these calculations and
confirmation of this new finite size behavior by other groups will be
necessary to understand what is happening at $m_\pi= 260~\mathrm{MeV}$.
Given the results in \Figs{\ref{fsx}} and \ref{largevolx}, there is no
concrete conclusion that we can draw yet regarding finite size
effects, but I hope that \Fig{\ref{fsx}} will stand as a warning that
the finite size effects can, and likely do, change substantially as we
decrease the pion mass and that collaborations pushing towards the
physical pion mass must keep these effects in mind. At this point one
may argue that we should investigate the finite size effects as
indicated from chiral perturbation theory. It is my personal opinion
that, given an absence of any real evidence for curvature in $\x{u-d}$
and given that such non-analytic behavior is the hallmark of chiral
perturbation theory, it is theoretically questionable to force a fit
to chiral perturbation theory. While this is my opinion, it was clear
at the lattice conference that many presenters also simply chose to
not show chiral fits. Of course, calculations at still lighter pion
masses may show the expected chiral behavior and systematic errors at
the current pion masses may obscure this behavior, but we should keep
in mind that in the end calculations down to the physical point may be
necessary to convincingly establish results for $\x{u-d}$.
One obvious feature of \Fig{\ref{largevolx}} is that the results from
each collaboration have a flat behavior but there are clear shifts
between the groups. One natural concern in this respect is the
renormalization of the operator used to determine $\x{u-d}$. This is
multiplicative and, importantly, quark mass independent. However it
does depend on the lattice action and hence will vary between the
different calculations. By taking the ratio of $\x{u-d}$ to the value
$\x{u-d}^\mathrm{ref}$ at some canonical reference mass, which I
arbitrarily choose to be $m_\pi=500~\mathrm{MeV}$, we can eliminate the
renormalization factor consistently for each calculation. The result
of this is shown in \Fig{\ref{ratiox}},
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/x_all_v3_new}
\caption{Renormalization free ratio:\ $\langle x\rangle_{u-d} /
\langle x\rangle_{u-d}^\mathrm{ref}$. The calculations with $m_\pi
L>4$ from \Fig{\protect\ref{largevolx}} are used to form the ratio
$\langle x\rangle_{u-d} / \langle x\rangle_{u-d}^\mathrm{ref}$,
where the reference scale is chosen as $m_\pi^\mathrm{ref} =
500~\mathrm{MeV}$. This ratio eliminates the renormalization factors from
each calculation. The calculations are now all consistent with each
other. This demonstrates that a systematic error, possibly in the
renormalization, could possibly account for the discrepancy among
the collaborations.}
\label{ratiox}
\end{center}
\end{figure}
and we find that the various calculations appear to collapse onto a
universal curve. As with the finite size effects, we can not make a
specific conclusion, but this result is suggestive of potential
problems in the renormalization of $\x{u-d}$. There are other
possible explanations of the systematic shifts between the groups,
such as the lattice spacing effects mentioned earlier or systematics
due to plateaus that are too short in the bare matrix element
calculations~\cite{Lin:2008uz,Yamazaki:2009zq}. Independent of the
ultimate explanation for these problems, I hope that the difference
between \Fig{\ref{largevolx}} and \Fig{\ref{ratiox}} encourages us to
examine the systematics of our calculations.
\section{Nucleon Spin}
As discussed in the introduction, the nucleon spin is exactly $1/2$
due to rotational symmetry, and similar to the momentum fraction, it
obeys a sum rule~\cite{Ji:1998pc},
\begin{equation}
\frac{1}{2} =
\frac{1}{2} \sum_q\, \1{\Delta q,\mu^2} + \sum_q\, L_{q,\mu^2} + J_{g,\mu^2}\,,
\end{equation}
which relates the nucleon spin to the contributions from quark
helicity $\1{q}$ and orbital angular momentum $L_q$ and a net
contribution from the gluons $J_g$. The asymptotic evolution of the
total quark contribution $J_q=\frac{1}{2} \sum_q\, \1{\Delta q} +
\sum_q\, L_q$ and the total gluon contribution $J_g$ is given by
\begin{displaymath}
\lim_{\mu\rightarrow\infty} J_{q,\mu^2} = \frac{1}{2}\frac{3N_\mathrm{F}}{16+3N_\mathrm{F}}
\end{displaymath}
\begin{displaymath}
\lim_{\mu\rightarrow\infty} J_{g,\mu^2} = \frac{16}{16+3N_\mathrm{F}}\,.
\end{displaymath}
For $N_\mathrm{F}=4$, we find $\lim\, J_q = 2/7\approx 0.60\cdot1/2$ and
$\lim\, J_g = 3/14\approx 0.40\cdot1/2$. Again we find the gluons
playing a substantial role, now in the nucleon spin and earlier in the
nucleon momentum. This is a particularly surprising conclusion in
light of the naive quark model result $\sum_q\, \1{\Delta q}=1$ and
$J_g=0$. However, similar to the asymptotic results for the momentum
fraction, we should be suspicious that non-perturbative QCD dynamics
can alter the picture at low scales.
Unfortunately, much less is known experimentally regarding the
decomposition of the nucleon spin. The one quantity that is known
well is the axial coupling, $g_{A}=\1{\Delta u}-\1{\Delta d}$, that is
measured accurately in neutron beta decay. A recent
review~\cite{Abele:2008zz} gives $g_{A}=1.2750(9)$ and the most recent
PDG~\cite{Amsler:2008zzb} world average is $g_{A}=1.2694(28)$. The
remaining quark contributions are essentially the only other known
pieces. The experimental result from HERMES in
2007~\cite{Airapetian:2007mh} gives $\1{\Delta u} = 0.842(12)$,
$\1{\Delta d} = -0.427 (12)$ and $\1{\Delta s} = -0.085 (17)$. QCD
sum rules~\cite{Balitsky:1997rs} or model
estimates~\cite{Barone:1998dx} indicate that $J_g\approx 0.25$ at low
scales. Taking the experimental results for $\1{\Delta q}$ and this
estimate for $J_g$, we can estimate that $50\%$ of the nucleon spin is
given by the gluons with the remaining divided into $33\%$ from quark
spin and $18\%$ from quark orbital motion. This picture is much less
certain than the momentum sum rule, but again, it provides a challenge
to the lattice QCD effort.
\subsection{Lattice Calculation of $\mathbf{g_{A}}$}
The nucleon axial charge can be defined by
\begin{displaymath}
\langle p, s | \overline{q} \gamma_\mu \gamma_5 \tau_3 q | p , s \rangle =
2 g_{A} s_\mu
\end{displaymath}
or $g_{A}=\1{\Delta u-\Delta d}$. The moments $\1{\Delta q}$ are the
lowest moments of the polarized PDFs. For the unpolarized
distributions, the moments $\1{q}$ are fixed by the valence structure
of the nucleon, but this is not the case for the polarized
distribution. This difference actually makes the moments $\1{\Delta
q}$ even simpler than the momentum fractions, $\x{q}$, discussed in
the previous section. In fact, using a lattice action with chiral
symmetry would eliminate the need to renormalize the lattice operator
used to determine $\1{\Delta q}$ altogether.\footnote{This is only
true for the non-singlet moments, such as $\1{\Delta u-\Delta d}$.
The axial anomaly generates mixing for the singlet moments.}
However, many lattice actions in use today do require this
renormalization, but it is not scale dependent and is in many ways
simpler than the renormalization required for $\x{q}$. Additionally,
since the power of $x$ in $\1{\Delta q}$ is one lower than in $\x{q}$,
the lattice operator does not contain a derivative and is a simple
quark bilinear. Consequently, the bare matrix elements can be
calculated more accurately for $\1{\Delta q}$ than for $\x{q}$. As in
the case of $\x{u-d}$, we focus on the $u-d$ combination again to
eliminate the computationally demanding disconnected diagrams.
Furthermore, this combination gives the axial coupling $g_{A}=\1{\Delta
u - \Delta d}$ that is very accurately measured in neutron beta decay.
In \Fig{\ref{worldga}},
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/ga_all_v1_new}
\caption{World's dynamical results for $g_A$. The lattice results are
from~\cite{Gattringer:2008vj,mohler:email} (BGR),
\cite{Yamazaki:2009zq} (RBC $N_\mathrm{F}=2+1$),
\cite{Lin:2008uz,Aoki:2004ht} (RBC $N_\mathrm{F}=2$),
\cite{negele:email} (LHPC), \cite{korzec:email} (ETMC) and
\cite{james:email} (QCDSF). The experimental result is the PDG 2008
value~\cite{Amsler:2008zzb}. The discrepancy between the lattice
calculations and the experimental measurement and the scatter among
the lattice calculations are both smaller than for $\x{u-d}$ shown
in \Fig{\protect\ref{worldx}}. }
\label{worldga}
\end{center}
\end{figure}
I collect all the full QCD lattice results for $g_{A}$, again with
$m_\pi < 700~\mathrm{MeV}$. Similar to $\x{u-d}$, $g_{A}$ serves as a benchmark
observable for lattice calculations of nucleon structure. As in
\Fig{\ref{worldx}} for $\x{u-d}$, we find that the lattice results
show a degree of scatter and all consistently underestimate the
experimental measurement. However, the discrepancy, between $10\%$
and $20\%$, is more mild and the scatter in the lattice results is
less severe. In fact, as I try to argue next, it appears that this
scatter may be almost entirely accounted for by finite size effects,
at least to the current level of statistical precision.
In \Fig{\ref{fsga}},
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/ga_fs_v3_new}
\caption{Finite size effects in $g_{A}$. The $m_\pi L$ dependence is
shown for several calculations. The results at the heaviest pion
masses indicate a significant finite size effect in $g_{A}$,
suggesting that $m_\pi L$ of $6$ may be necessary. But the light
$m_\pi$ results of QCDSF and ETMC may in fact be showing a weakening
of the finite size effects in $g_{A}$ as measured in terms of $m_\pi
L$. The lattice results and experimental measurement are the same
as in \Fig{\protect\ref{worldga}}, and the curves are simple scaling
fits to guide the eye.}
\label{fsga}
\end{center}
\end{figure}
I examine several finite size studies that are available for $g_{A}$.
The results for the heaviest values of the pion mass show a strong
finite size effect. This has led to the current view that $m_\pi L >
6$ may in fact be needed to reliably determine $g_{A}$. However, we find
that, as quantified by $m_\pi L$, the finite size effects are
diminishing as $m_\pi$ is lowered. This can be seen quite clearly in
\Fig{\ref{fsga}} by examining the lighter pion mass calculations of
both QCDSF and ETMC. It is possible that the volumes for these
calculations at $m_\pi= 310~\mathrm{MeV}$ or $260~\mathrm{MeV}$ may still be too small
to see the asymptotic volume dependence. We can also consult chiral
perturbation theory, which does allow for this sort of behavior;
however, given that $g_{A}$ is essentially flat for the largest volumes,
one might reasonably question the use of chiral perturbation theory at
these pion masses.
In order to attempt to estimate the infinite volume limits for the
lattice calculations presented here, I simply require $m_\pi L > 6$.
This may turn out to be an excessive requirement for low $m_\pi$,
which would actually be good news regarding finite size effects, but
it is clearly required for heavier pion masses. The lattice results
satisfying this are shown in \Fig{\ref{largevolga}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/ga_all_v3_new}
\caption{Results for $g_A$ with $m_\pi L>6$. The lattice results from
\Fig{\protect\ref{worldga}} that satisfy $m_\pi L>6$ are shown. All
the lattice calculations, but one, are in statistical agreement.
The results still underestimate the experimental value, but the
discrepancy is much less severe than in $\x{u-d}$. The QCDSF
results are just slightly low. This may indicate some small
systematic error in the renormalization or it may simply illustrate
the lattice artifacts that are present in all the calculations.}
\label{largevolga}
\end{center}
\end{figure}
As is clear from the figure, this restriction is very severe, however,
the resulting lattice calculations show a strong level of agreement
amongst themselves. The results are still lower than the experimental
measurement, but, unlike for $\x{u-d}$, only a mild curvature is
required to reconcile the current calculations with the physical
limit. In particular, notice that, with one exception, there is no
systematic shift between the various collaborations. As mentioned
earlier, the renormalization of $g_{A}$ is generically easier than that
of $\x{u-d}$, and this lends a bit more support to the hypothesis that
differences in renormalization may be driving part of the variation of
$\x{u-d}$ in \Fig{\ref{largevolx}}. The one small exception is the
result of QCDSF which is just slightly low compared to all the other
calculations. However, this may be consistent with a small discrepancy
in $f_\pi$, which renormalizes with the same factor as $g_{A}$, that is
also present for QCDSF~\cite{Jansen:2008vs}.
\section{Nucleon Charge}
The charge of the nucleon, or any hadron, again appears to be an
essentially trivial topic. There is no doubt that the proton has one
unit of charge and that the neutron is, well, neutral. The certainty
in our understanding of the nucleon's charge arises from the flavor
symmetries of QCD, however, the distribution of this charge in the
spatial degrees of freedom of the nucleon's constituents is not
prescribed by symmetry and provides another probe of the dynamics of
QCD within the nucleon. This idea leads to a broad range of
calculations on the lattice and quickly involves the generalized
parton distributions, but to avoid excessive complications and to
explain the concepts in the simplest setting, I will discuss just the
neutron charge~\cite{Miller:2007uy}.
\subsection{Neutron Transverse Charge Distribution}
Much of the interest in form factors originates with the
interpretations that we can assign to them. There is now a rich field
theoretic discussion dedicated to this issue, but in order to maintain
our footing and retain a strong connection to the experimental
measurements, it is valuable to remember that ultimately, the form
factors parametrize the QCD contributions to cross sections,
independent of any interpretation we attach to them.
\begin{displaymath}
\frac{d\sigma}{d\Omega} = \left(\frac{d\sigma}{d\Omega}\right)_\mathrm{Mott}
\left\{ F_1^2 + \tau F_2^2 + 2\tau(F_1+F_2)^2\tan^2(\theta/2) \right\}
\end{displaymath}
Here $F_1(q^2)$ and $F_2(q^2)$ are the nucleon form factors and
$\tau=q^2/(4m^2)$. For some in our community, this is where the
discussion ends. That is, of course, a perfectly valid point of view.
However, the understanding provided to us by the generalized parton
distributions and the relationship to transverse quark
distributions~\cite{Burkardt:2000za} allows us to go further.
Historically the form factors were interpreted as charge and
magnetization densities. In particular the charge density was
constructed from $G_E = F_1 - \tau F_2$ as
\begin{displaymath}
\rho_{3D}(\vec{r}) =
\int\!\! \frac{d^3\vec{q}}{(2\pi)^3}\,\,\,
e^{-i\vec{r}\cdot\vec{q}}\, \frac{m}{\sqrt{m^2+\vec{q}^2}} G_E(\vec{q}^2)\,.
\end{displaymath}
This is plotted for the neutron in \Fig{\ref{3dcharge}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=252pt,angle=-90]{plots/neutron_3d_charge}
\caption{Naive $3d$ neutron charge distribution. The $3d$ Fourier
transform of the $G_E$ form factor gives the rest frame charge
distribution of the nucleon up to relativistic corrections. The
negative distribution at large $r$ is interpreted as the pion cloud
arising from the $n\rightarrow p\pi^-$ fluctuations. This gives
rise to a negative charge radius for the neutron. The curves were
generated by numerically integrating the Kelly 2004 parametrization
of the nucleon form factors~\cite{Kelly:2004hm}.}
\label{3dcharge}
\end{center}
\end{figure}
There we see that the charge distribution is negative for large $r$.
This is usually understood in terms of the pion cloud generated by
$n\rightarrow p\pi^-$ fluctuations. Overall neutrality of the neutron
forces the core to be positive leading to a negative charge radius.
But this picture is spoiled by relativistic corrections that enter as
$(m\langle r\rangle)^{-1}$, where $\langle r\rangle$ is some relevant
length scale for the charge distribution. These corrections would
vanish in the non-relativistic limit $m\rightarrow\infty$ but are
substantial for the nucleon. The introduction of the transverse
distributions in the infinite-momentum frame (IMF) eliminates these
corrections, which enter now as $(p_z\langle r\rangle)^{-1}$ with
$p_z\rightarrow\infty$ defining the IMF. The transverse
distributions~\cite{Burkardt:2000za} are given by
\begin{displaymath}
\rho_{2D}(\vec{b}) =
\int\! \frac{d^2\vec{q}_\perp}{(2\pi)^2}\,\,\,
e^{-i\vec{b}\cdot\vec{q}_\perp}\, F_1(\vec{q}_\perp^2)\,.
\end{displaymath}
This is shown for the neutron in \Fig{\ref{2dcharge}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=252pt,angle=-90]{plots/neutron_2d_charge}
\caption{Transverse 2d neutron charge distribution. The $2d$ Fourier
transform of the $F_1$ form factor results in the transverse charge
distribution of the nucleon. Relativistic corrections are absent
but the distribution must be understood in the infinite momentum
frame. Notice the negative core, in contrast to
\Fig{\protect\ref{3dcharge}}. Also the charge radius is now
positive in this frame. The distribution was constructed from the
same form factor parametrization in \Fig{\protect\ref{3dcharge}}.}
\label{2dcharge}
\end{center}
\end{figure}
The distribution still has a negative tail for large $r$, but now we
find a negative core. Furthermore, the charge radius is positive for
the transverse charge distribution. Understanding this new view of
the neutron is a challenge onto itself. Additionally, the
introduction of the transverse coordinates in a field-theoretically
clean manner opens many new avenues for investigation of hadron
structure from lattice QCD.
\subsection{Lattice Calculation of $\mathbf{\rsq{u-d}}$}
As for the nucleon momentum and spin structure, I again introduce a
benchmark observable that allows us to summarize the status of lattice
calculations of form factors. Here I choose the isovector charge
radius, $\langle r^2 \rangle_{u-d}$. As before, the flavor structure
is $u-d$, so disconnected diagrams again vanish. Furthermore, this
observable enters matrix elements of the vector current. These two
facts combine to make $\langle r^2 \rangle_{u-d}$ an accurate quantity
to calculate on the lattice. Additionally, all the lattice
calculations can now use an exactly conserved vector current. This
eliminates the need for any renormalization and simplifies the
calculation significantly.
The form of the nucleon's electromagnetic interaction is given by
\begin{displaymath}
\langle p^\prime, s^\prime | J^\mu | p, s \rangle =
\overline{u}(p^\prime,s^\prime) \left\{ \gamma^\mu F_1(q^2)
+ i\sigma^{\mu\nu}\frac{q_\nu}{2m} F_2(q^2) \right\} u(p,s)
\end{displaymath}
and the charge radius is then given by
\begin{displaymath}
\langle r^2\rangle_{u-d} = -6 \left.\frac{dF^{u-d}_1}{dq^2}\right|_{q^2=0}\,.
\end{displaymath}
The factor of $6$ is the conventional choice, but a factor of $4$
would instead give the correct transverse charge radius in the IMF.
To avoid confusion, we stick to the conventional choice, but we should
keep in mind the theoretically clean interpretation that is available
for the transverse distribution.
In \Fig{\ref{worldrsq}}
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/rsq_all_v1}
\caption{World's dynamical results for $\langle r^2
\rangle_{u-d}^{1/2}$. The lattice calculations are
\cite{Yamazaki:2009zq} (RBC $N_\mathrm{F}=2+1$),
\cite{Lin:2008uz,Aoki:2004ht} (RBC $N_\mathrm{F}=2$), \cite{:2010jn}
(LHPC/MILC), \cite{Syritsyn:2009mx} (LHPC/RBC), \cite{korzec:email}
(ETMC) and \cite{james:email} (QCDSF). The experimental result
comes from the Kelly 2004 parametrization of the form
factors~\cite{Kelly:2004hm}. The leading order heavy baryon chiral
perturbation theory prediction (LO HB$\chi$PT) contains no unknown
low energy constants other than $\langle r^2 \rangle_{u-d}^{1/2}$
and is logarithmically divergent as $m_\pi\rightarrow 0$. One
might, therefore, expect substantial finite size effects in this
quantity, but this does not appear to be the case.}
\label{worldrsq}
\end{center}
\end{figure}
I gather all the full QCD results for $\rsq{u-d}$. It is immediately
evident that all the lattice calculations show a remarkable agreement
for this observable. The strong scatter that is present for $\x{u-d}$
and to a lesser extent $g_{A}$ is clearly absent for $\rsq{u-d}$. This
is particularly surprising given the fact that this quantity is
expected to diverge in the chiral limit as indicated by the curve in
\Fig{\ref{worldrsq}}. We might even expect large finite size effects
in such a quantity. However, we must keep in mind that $\rsq{u-d}$ is
calculated from the slope of $F_1(q^2)$ at $q^2=0$ and that the
momentum quantization imposed by a finite volume forces us to extract
this slope using the first non-zero value of $q^2$, which satisfies $q
L\approx 2\pi$. We may, in fact, be extrapolating from a region with
weaker $L$ dependence through the region with stronger $L$ dependence.
Ultimately, whatever the explanation, we find a near absence of volume
dependence for $\rsq{u-d}$.
Lacking any real guidance on the necessary volumes for $\rsq{u-d}$, I
simply follow the conventional wisdom and take $m_\pi L>4$. These
lattice results are shown in \Fig{\ref{largevolr}}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=300pt,angle=0]{plots/rsq_all_v3}
\caption{Large volume results for $\langle r^2 \rangle_{u-d}^{1/2}$.
Due to momentum quantization at finite volume, one can not easily
calculate the form factors at the low $Q^2$ needed to determine
$\langle r^2 \rangle_{u-d}^{1/2}$ while also working at small
volume. Lacking any detailed knowledge of the finite size effects
in $\langle r^2 \rangle_{u-d}^{1/2}$, only the results from
\Fig{\protect\ref{worldrsq}} with $m_\pi L>4$ are shown. There is
general agreement among all the lattice groups and a mild increase
as $m_\pi$ is lowered. However $\langle r^2 \rangle_{u-d}^{1/2}$ is
not yet rising nearly as fast as it needs to match onto the
experimental measurement.}
\label{largevolr}
\end{center}
\end{figure}
An optimist may find a mild increase, but this would be hard to defend
strongly. Additionally, this observable requires much more than a
mild increase to agree with the experimental value at the physical
point. But it does appear that all the calculations seem to agree.
Given that $\rsq{u-d}$ requires no renormalization, this is consistent
with the speculation suggesting that renormalization may be partially
culpable for the systematic variation in $\x{u-d}$ discussed earlier.
Despite the fact that this observable still shows a significant
discrepancy with the experimental result, the agreement amongst all
the lattice groups is encouraging.
\section{Conclusions and Outlook}
In order to illustrate the status and promise of lattice calculations
of hadron structure, I chose to review three benchmark observables
that represent the broad range of hadronic physics that we can address
on the lattice:\ the momentum fraction, axial charge and charge radius
of the nucleon. Nearly all the lattice groups have pushed the pion
mass towards $300~\mathrm{MeV}$ or lower and some are now breaking through
$200~\mathrm{MeV}$. This is an enormous accomplishment, but unfortunately
most hadronic observables are failing to show the long-sought chiral
curvature. Independent of the expectations of chiral perturbation
theory, the three benchmark observables still do not show a convincing
approach to the experimental results.
The obvious conclusion is that yet lighter, and maybe even physical,
pion masses will be required. This may be the case, but we should not
do this at the expense of other systematic errors. Finite size
effects must remain a concern as we lower $m_\pi$. Both $\x{u-d}$ and
$g_{A}$ demonstrate that the volume dependence can have a non-trivial
dependence on $m_\pi$ and our conventional rule-of-thumb that $m_\pi
L\approx4$ may not necessarily be sufficient for all quantities.
Lattice artifacts are still a concern, if only because so few groups
have seriously examined them. The agreement of the groups for
$\rsq{u-d}$ might indicate weak artifacts for some quantities, but the
systematic shifts between the groups for $\x{u-d}$ is a warning that
renormalization, an essential part of the continuum limit for most
quantities of interest, may be partially responsible for these
problems. The role of renormalization is further implicated by the
general agreement of the groups for the renormalization free ratio of
$\x{u-d}/\x{u-d}^\mathrm{ref}$. In fact, the progression from
$\x{u-d}$ (requiring a scale-dependent renormalization) to $g_{A}$
(requiring only a finite renormalization) to $\rsq{u-d}$ (requiring no
renormalization) will be a strong test of our control of
renormalization issues.
These systematic effects will be a part of any critical review of
lattice calculations, but we should remember the rich QCD physics that
provides the motivation for these calculations. Each of the benchmark
observables is a stand-in for a broad range of experimentally relevant
and intellectually interesting observables:\ the decomposition of
momentum among the nucleon's constituents, the nature of the nucleon's
spin and the distribution of the nucleon's charge. Apart from these
observables, Lattice 2009 saw a sizeable effort to study strangeness
in the nucleon, results for the pion, delta and other hadrons, and a
range of exploratory calculations. Clean control of systematics and a
continued push toward the physical pion mass will ultimately allow
detailed calculations that will complement the experimental programs
and advance our insight into hadron structure.
\acknowledgments
I gratefully appreciate the many conversations and email exchanges
with colleagues while preparing for the Lattice 2009 conference. I
would like to thank D.~Alexandrou, G.~Bali, W.~Bietenholz, T.~Blum,
T.~Doi, M.~Engelhardt, W.~Freeman, Ph.~H\"agler, P.-A.~Harraud,
T.~Hemmert, K.~Jansen, T.~Kaneko, T.~Korzec, M.~Lin, Z.~Liu,
D.~Mohler, B.~Musch, J.~Negele, H.~Ohki, S.~Ohta, J.~Osborn,
G.~Schierholz, W.~Schroers, T.~Streuer, K.~Takeda, N.~Ukita,
A.~Walker-Loud and J.~Zanotti. Additionally, this work was supported
by the DFG Sonderforschungsbereich/Transregio SFB/TR9-03
\bibliographystyle{unsrt}
|
\section{Preliminaries}
Recently, the important progress made in the study of evolution
equations had a master role in the developing of a vast
literature, concerning mostly the asymptotic properties of linear
operators semigroups, evolution operators or skew-product
semiflows.
In this paper, the study is led throughout the notion of
skew-evolution semiflow on Banach spaces, defined by means of an
evolution semiflow and an evolution cocycle. As the
skew-evolutions semiflows reveal themselves to be generalizations
of evolution operators and skew-product semiflows, they are
appropriate to study the asymptotic properties of the solutions of
evolution equations having the form
\[
\left\{
\begin{array}{l}
\dot{u}(t)=A(t)u(t), \ t>t_{0}\geq 0 \\
u(t_{0})=u_{0},%
\end{array}%
\right.
\]
where $A:\mathbf{R}\rightarrow \mathcal{B}(V)$ is an operator,
$\textrm{Dom}A(t)\subset V$, $u_{0}\in \textrm{Dom}A(t_{0})$.
The fact that a skew-evolution semiflow depends on three variables
$t$, $t_{0}$ and $x$, while the classic concept of cocycle depends
only on $t$ and $x$, justifies the study of asymptotic behaviors
in a nonuniform setting (relative to the third variable $t_{0}$)
for skew-evolution semiflows.
The basic concepts of asymptotic properties, such as stability,
instability and dichotomy, that appear in the theory of dynamical
systems, play an important role in the study of stable, instable
and central manifolds. We intend to define and exemplify various
concepts of dichotomies, as uniform exponential dichotomy,
Barreira-Valls exponential dichotomy, exponential dichotomy,
uniform polynomial dichotomy, Barreira-Valls polynomial dichotomy,
polynomial dichotomy and to emphasize connections between them. We
have thus considered generalizations of some asymptotic properties
for evolution equations, defined by L. Barreira and C. Valls in
\cite{BaVa_LNM}. Characterizations for the asymptotic properties
in a nonuniform setting are also proved.
Some of the original results concerning the properties of
stability and instability for skew-evolution semiflows were
published in \cite{StMe_NA} and \cite{MeSt_ICNODEA07}.
The exponential dichotomy for evolution equations is one of the
mathematical domains with an impressive development due to its
role in describing several types of differential equations. Its
study led to an extended literature, which begins with the
interesting results due to O. Perron in \cite{Pe_MZ}. The ideas
were continued by J.L. Massera and J.J. Sch\"{a}ffer in
\cite{MaSc_PAM}, with extensions in the infinite dimensional case
accomplished by J.L. Dalecki\u{i} and M.G. Kre\u{i}n in
\cite{DaKr_AMS} and A. Pazy in \cite{Pa_SV}, respectively R.J.
Sacker and G.R. Sell in \cite{sacsel}. Diverse and important
concepts of dichotomy were introduced and studied by S.N. Chow and
H. Leiva in \cite{ChLe_JDE} and by W.A. Coppel in \cite{Co_LNM}.
Some asymptotic behaviors for evolution families were given in the
nonuniform case in \cite{MeSaSa_MIR} by M. Megan, A.L. Sasu and B.
Sasu. The study of the nonuniform exponential dichotomy for
evolution families was considered by P. Preda and M. Megan in
\cite{PrMe_BAMS}.
The property of exponential dichotomy for the case of
skew-evolution semiflows is treated in \cite{MeSt_TP09} and
\cite{StMe_OT}.
\section{Notations. Definitions. Examples}
Let us consider a metric space $(X,d)$, a Banach space $V$,
$V^{*}$ its topological dual and $\mathcal{B}(V)$ the space of all
bounded linear operators from $V$ into itself. $I$ is the identity
operator on $V$. We denote $T =\left\{(t,t_{0})\in \mathbf{R}^{2},
\ t\geq t_{0}\geq 0\right\}$ and $Y=X\times V$.
\begin{definition}\rm\label{def_sfl_ev}
A mapping $\varphi: T\times X\rightarrow X$ is called
\emph{evolution semiflow} on $ X$ if following relations hold:
$(s_{1})$ $\varphi(t,t,x)=x, \ \forall (t,x)\in
\mathbf{R}_{+}\times X$;
$(s_{2})$ $\varphi(t,s,\varphi(s,t_{0},x))=\varphi(t,t_{0},x),
\forall (t,s),(s,t_{0})\in T, x\in X$.
\end{definition}
\begin{definition}\rm\label{def_aplcoc_ev}
A mapping $\Phi: T\times X\rightarrow \mathcal{B}(V)$ is called
\emph{evolution cocycle} over an evolution semiflow $\varphi$ if:
$(c_{1})$ $\Phi(t,t,x)=I$, $\forall (t,x)\in \mathbf{R}_{+}\times
X$;
$(c_{2})$
$\Phi(t,s,\varphi(s,t_{0},x))\Phi(s,t_{0},x)=\Phi(t,t_{0},x),\forall
(t,s),(s,t_{0})\in T, x\in X$.
\end{definition}
\begin{definition}\rm\label{def_coc_ev_1}
The mapping $C: T\times Y\rightarrow Y$ defined by the relation
$$C(t,s,x,v)=(\varphi(t,s,x),\Phi(t,s,x)v),$$ where $\Phi$ is an
evolution cocycle over an evolution semiflow $\varphi$, is called
\emph{skew-evolution semiflow} on $Y$.
\end{definition}
\begin{example}\rm\label{ex_ses}
We denote by
$\mathcal{C}=\mathcal{C}(\mathbf{R}_{+},\mathbf{R}_{+})$ the set
of all continuous functions $x:\mathbf{R}_{+}\rightarrow
\mathbf{R}_{+}$, endowed with the topology of uniform convergence
on compact subsets of $\mathbf{R}_{+}$, metrizable by means of the
distance
\[
d(x,y)=\sum_{n=1}^{\infty}\frac{1}{2^{n}}\frac{d_{n}(x,y)}{1+d_{n}(x,y)},
\ \textrm{where} \ d_{n}(x,y)=
\sup\limits_{t\in[0,n]}{|x(t)-y(t)|}.
\]
If $x\in \mathcal{C}$, then, for all $t\in \mathbf{R}_{+}$, we
denote $x_{t}(s)=x(t+s)$, $x_{t}\in \mathcal{C}$. Let $ X$ be the
closure in $\mathcal{C}$ of the set $\{f_{t},t\in
\mathbf{R}_{+}\}$, where $f:\mathbf{R}_{+}\rightarrow
\mathbf{R}_{+}^{*}$ is a decreasing function. It follows that $(
X,d)$ is a metric space. The mapping $\varphi: T\times
X\rightarrow X, \ \varphi(t,s,x)=x_{t-s}$ is an evolution semiflow
on $X$.
We consider $V=\mathbf{R}^{2}$, with the norm $\left\Vert
v\right\Vert=|v_{1}|+|v_{2}|$, $v=(v_{1},v_{2})\in V$. The mapping
$\Phi: T\times X\rightarrow \mathcal{B}(V)$ given by
\[
\Phi(t,s,x)v=\left(
e^{\alpha_{1}\int_{s}^{t}x(\tau-s)d\tau}v_{1},e^{\alpha_{2}\int_{s}^{t}x(\tau-s)d\tau}v_{2}\right),
\ (\alpha_{1},\alpha_{2})\in\mathbf{R}^{2},
\]
is an evolution cocycle over $\varphi$ and $C=(\varphi,\Phi)$ is a
skew-evolution semiflow.
\end{example}
\begin{remark}\rm
A connection between the solutions of a differential equation
\begin{equation}\label{ec_nuet1}
\dot{u}(t)=A(t)u(t), \ t\in\mathbf{R}_{+}
\end{equation}
and a skew-evolution semiflow is given by the definition of the
evolution cocycle $\Phi$, by the relation $\Phi(t,s,x)v=U(t,s)v$,
where $U(t,s)=u(t)u^{-1}(s)$, $(t,s)\in T$, $(x,v)\in Y$, and
where $u(t)$, $t\in \mathbf{R}_{+}$, is a solution of the
differential equation (\ref{ec_nuet1}).
\end{remark}
The fact that the skew-evolution semiflows are generalizations for
skew-product semiflows is emphasized by
\begin{example}\rm Let $X$ be
the metric space defined as in Example \ref{ex_ses}. The mapping
$\varphi_{0}:\mathbf{R}_{+}\times X\rightarrow X$,
$\varphi_{0}(t,x)=x_{t}$, where $x_{t}(\tau)=x(t+\tau)$, $\forall
\tau\geq 0$, is a semiflow on $X$. Let us consider for every $x\in
X$ the parabolic system with Neumann's boundary conditions:
\begin{equation}\label{sistem_parabolic}
\left\{
\begin{array}{lc}
\displaystyle \frac{\partial v}{\partial
t}(t,y)=x(t)\frac{\partial^{2}v}{\partial y^{2}}(t,y), & t>0, y\in (0,1) \\
v(0,y)=v_{0}(y), & y\in (0,1) \\
\displaystyle\frac{\partial v}{\partial y}(t,0)=\frac{\partial
v}{\partial
y}(t,1)=0, & t>0.%
\end{array}%
\right.
\end{equation}
Let $V=\mathcal{L}^{2}(0,1)$ be a separable Hilbert space with the
orthonormal basis
$\{e_{n}\}_{n\in \mathbf{N}}$, $e_{0}=1$, $e_{n}(y)=\sqrt{2}\cos n\pi y$,
where $y\in (0,1)$, $n\in \mathbf{N}$. We denote
$D(A)=\{v\in \mathcal{L}^{2}(0,1), \ v(0)=v(1)=0\}$ and we define
the operator
\[
A:D(A)\subset V\rightarrow V, \ Av=\frac{d^{2}v}{dy^{2}},
\]
which generates a $\mathcal{C}_{0}$-semigroup $S$, defined by
$S(t)v=\sum\limits_{n=0}^{\infty}e^{-n^{2}\pi^{2}t}\langle
v,e_{n}\rangle e_{n}$, where $\langle \cdot ,\cdot \rangle $
denotes the scalar product in $V$. We define for every $x\in X$,
$A(x):D(A)\subset V\rightarrow %
V$, $A(x)=x(0)A$, which allows us to rewrite system
(\ref{sistem_parabolic}) in $V$ as
\begin{equation}\label{sistem_modif}
\left\{
\begin{array}{lc}
\dot{v}(t)=A(\varphi_{0}(t,x))v(t), & t> 0 \\
v(0)=v_{0}.%
\end{array}%
\right.
\end{equation}
The mapping
\[
\Phi_{0}:\mathbf{R}_{+}\times X\rightarrow \mathcal{B}(V), \
\Phi_{0}(t,x)v=S\left(\int_{0}^{t}x(s)ds\right)v
\]
is a cocycle over the semiflow $\varphi_{0}$ and
$C_{0}=(\varphi_{0}, \Phi_{0})$ is a linear skew-product semiflow
strongly continuous on $Y$. Also, for all $v_{0}\in D(A)$, we have
obtained that $v(t)=\Phi (t,x)x_{0}, \ t\geq 0$, is a strongly
solution of system (\ref{sistem_modif}).
As $C_{0}=(\varphi_{0},\Phi_{0})$ is a skew-product semiflow on
$Y$, then the mapping $C: T\times Y\rightarrow Y$,
$C(t,s,x,v)=(\varphi(t,s,x),\Phi(t,s,x)v)$, where
\[
\varphi(t,s,x)=\varphi_{0}(t-s,x) \ \textrm{and} \
\Phi(t,s,x)=\Phi_{0}(t-s,x), \ \forall (t,s,x)\in T\times X
\]
is a skew-evolution semiflow on $Y$. Hence, the skew-evolution
semiflows generalize the notion of skew-evolution semiflows.
\end{example}
An interesting class of skew-evolution semiflows, useful to
describe some asymptotic properties, is given by
\begin{example}\rm\label{ex_shift}
Let us consider a skew-evolution semiflow $C=(\varphi, \Phi)$ and
a parameter $\lambda \in \mathbf{R}$. We define the mapping
\begin{equation}\label{relcevshift}
\Phi_{\lambda}: T\times X\rightarrow \mathcal{B}(V), \
\Phi_{\lambda}(t,t_{0},x)=e^{\lambda(t-t_{0})}\Phi(t,t_{0},x).
\end{equation}
One can remark that $C_{\lambda}=(\varphi, \Phi_{\lambda})$ also
satisfies the conditions of Definition \ref{def_coc_ev_1}, being
called \emph{$\lambda$-shifted skew-evolution semiflow} on $Y$.
Let us consider on the Banach space $V$ the Cauchy problem
\[
\left\{
\begin{array}{l}
\dot{v}(t)=Av(t), \ t> 0 \\
v(0)=v_{0}%
\end{array}%
\right.
\]
with the nonlinear operator $A$. Let us suppose that $A$ generates
a nonlinear $C_{0}$-semigroup $\mathcal{S}=\{S(t)\}_{t\geq 0}$.
Then $\Phi(t,s,x)v=S(t-s)v$, where $t\geq s\geq 0$, $(x,v)\in Y$,
defines an evolution cocycle. Moreover, the mapping defined by
$\Phi_{\lambda}: T\times X\rightarrow \mathcal{B}(V)$,
$\Phi_{\lambda}(t,s,x)v=S_{\lambda}(t-s)v$, where
$\mathcal{S}_{\lambda}=\{S_{\lambda}(t)\}_{t\geq 0}$ is generated
by the operator $A-\lambda I$, is also an evolution cocycle.
\end{example}
\begin{definition}\rm\label{def_taremas}
A skew-evolution semiflow $C =(\varphi,\Phi)$ is said to be
\emph{strongly measurable} if, for all $(t_{0},x,v)\in T\times Y$,
the mapping $s\mapsto\left\Vert\Phi(s,t_{0},x)v\right\Vert$ is
measurable on $[t_{0},\infty)$.
\end{definition}
\begin{definition}\rm\label{def_neg}
The skew-evolution semiflow $C=(\varphi,\Phi)$ is said to have
\emph{exponential growth} if there exist
$M,\omega:\mathbf{R}_{+}\rightarrow\mathbf{R}_{+}^{*}$ such that:
\[
\left\Vert \Phi(t,t_{0},x)v\right\Vert \leq M(s)e^{\omega
(t-s)}\left\Vert \Phi(s,t_{0},x)v\right\Vert, \forall
(t,s),(s,t_{0})\in T, \forall (x,v)\in Y.
\]%
\end{definition}
\begin{remark}\rm\label{obs_shift}
If $C=(\varphi,\Phi)$ is a skew-evolution semiflow with
exponential growth, as following relations
\[
\left\Vert\Phi_{\lambda}(t,t_{0},x)v\right\Vert=e^{\lambda(t-t_{0})}
\left\Vert\Phi(t,t_{0},x)v\right\Vert \leq
M(t_{0})e^{[\omega(t_{0})+\lambda](t-t_{0})}\left\Vert
v\right\Vert,
\]
hold for all $(t_{0},x,v)\in\mathbf{R}_{+}\times Y$, then
$C_{\lambda}=(\varphi,\Phi_{\lambda})$, $\lambda >0$, has also
exponential growth.
\end{remark}
\begin{remark}\rm\label{obs_eg}
$(i)$ If we consider in Definition \ref{def_neg} the constants
$M\geq 1$ and $\omega>0$, the skew-evolution semiflow $C$ is said
to have \emph{uniform exponential growth};
$(ii)$ If in Definition \ref{def_neg} we consider $M\geq 1$ to be
a constant such that the relation $\left\Vert
\Phi(t,s,x)\right\Vert \leq Me^{\omega (t-s)}$ holds for all
$(t,s)\in T$ and all $x\in X$, the skew-evolution semiflow $C$ is
said to have \emph{bounded exponential growth}.
\end{remark}
\begin{definition}\rm\label{def_nedc}
The skew-evolution semiflow $C=(\varphi,\Phi)$ is said to have
\emph{exponential decay} if there exist
$M,\omega:\mathbf{R}_{+}\rightarrow\mathbf{R}_{+}^{*}$ such that:
\[
\left\Vert \Phi(s,t_{0},x)v\right\Vert \leq M(t)e^{\omega
(t-s)}\left\Vert \Phi(t,t_{0},x)v\right\Vert, \forall
(t,s),(s,t_{0})\in T, \forall (x,v)\in Y.
\]%
\end{definition}
\begin{remark}\rm
If $C=(\varphi,\Phi)$ be a skew-evolution semiflow with
exponential decay, as following relations
\[
\left\Vert\Phi_{-\lambda}(s,t_{0},x)v\right\Vert=e^{-\lambda(s-t_{0})}
\left\Vert\Phi(s,t_{0},x)v\right\Vert \leq
M(t)e^{[\omega(t)+\lambda](t-s)}\left\Vert
\Phi_{-\lambda}(t,t_{0},x)v\right\Vert,
\]
hold for all $(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$, then
$C_{-\lambda}=(\varphi,\Phi_{-\lambda})$, $\lambda>0$, has also
exponential decay.
\end{remark}
\begin{remark}\rm
If in Definition \ref{def_nedc} we consider $M\geq 1$ and
$\omega>0$ to be constants, the skew-evolution semiflow $C$ is
said to have \emph{uniform exponential decay}.
\end{remark}
\section{On various classes of dichotomy}
Let $C: T\times Y\rightarrow Y$,
$C(t,s,x,v)=(\varphi(t,s,x),\Phi(t,s,x)v)$ be a skew-evolution
semiflow on $Y$.
\begin{definition}\rm\label{proiector}
A continuous mapping $P:Y\rightarrow Y$ defined by:
\begin{equation}
P(x,v)=(x,P(x)v), \ \forall (x,v)\in Y,
\end{equation}
where $P(x)$ is a linear projection on $Y_{x}$, is called
\emph{projector} on $Y$.
\end{definition}
\begin{remark}\rm
The mapping $P(x):Y_{x}\rightarrow Y_{x}$ is linear and bounded
and satisfies the relation $P(x)P(x)=P^{2}(x)=P(x)$ for all $x\in
X.$
\end{remark}
For all projectors $P:Y\rightarrow Y$ we define the sets
\[
Im P=\{(x,v)\in Y,P(x)v=v\} \ \textrm{and} \ KerP=\{(x,v)\in
Y,P(x)v=0\}.
\]
\begin{remark}\rm
Let $P$ be a projector on $Y$. Then $ImP$ and $KerP$ are closed
subsets of $Y$ and for all $x\in
X$ we have
\[
ImP(x)+KerP(x)=Y_{x} \ \textrm{and} \ ImP(x)\cap KerP(x)=\{0\}.
\]
\end{remark}
\begin{remark}\rm
If $P$ is a projector on $Y$, then the mapping
\begin{equation}
Q:Y\rightarrow Y, \ Q(x,v)=(x,v-P(x)v)
\end{equation}
is also a projector on $Y$, called \emph{the complementary
projector} of $P$.
\end{remark}
\begin{definition}\rm\label{prinv}
A projector $P$ on $Y$ is called \emph{invariant} relative to a
skew-evolution semiflow $C=(\varphi,\Phi)$ if following relation
holds:
\begin{equation}
P(\varphi(t,s,x))\Phi(t,s,x)=\Phi(t,s,x)P(x),
\end{equation}
for all $(t,s)\in T$ and all $x\in X$.
\end{definition}
\begin{remark}\rm
If the projector $P$ is invariant relative to a skew-evolution
semiflow $C$, then its complementary projector $Q$ is also
invariant relative to $C$.
\end{remark}
\begin{definition}\rm\label{comp_pr_dich}
A projector $P_{1}$ and its complementary projector $P_{2}$ are
said to be \emph{compatible} with a skew-evolution semiflow
$C=(\varphi,\Phi)$ if
$(d_{1})$ the projectors $P_{1}$ and $P_{2}$ are invariant on $Y$;
$(d_{2})$ for all $x\in X$, the projections $P_{1}(x)$ and
$P_{2}(x)$ commute and the relation $P_{1}(x)P_{2}(x)=0$ holds.
\end{definition}
In what follows we will denote
\[
\Phi_{k}(t,t_{0},x)=\Phi(t,t_{0},x)P_{k}(x), \ \forall
(t,t_{0})\in T, \ \forall x\in X, \ \forall k\in \{1,2\}.
\]
We remark that $\Phi_{k}$, $k\in \{1,2\}$ are evolution cocycles
and
\[
C_{k}(t,s,x,v)=(\varphi(t,s,x),\Phi_{k}(t,s,x)v), \ \forall
(t,t_{0},x,v)\in T\times Y, \ \forall k\in \{1,2\},
\]
are skew-evolution semiflows, over all evolution semiflows
$\varphi$ on $X$.
\begin{definition}\rm\label{def_ued}
The skew-evolution semiflow $C=(\varphi,\Phi)$ is called
\emph{uniformly exponentially dichotomic} if there exist two
projectors $P_{1}$ and $P_{2}$ compatible with $C$, some constants
$N_{1}\geq 1$, $N_{2}\geq 1$ and $\nu_{1}$, $\nu_{2}>0$ such that:
\begin{equation}\label{dich_stab}
e^{\nu_{1}(t-s)}\left\Vert \Phi_{1}(t,t_{0},x)v\right\Vert \leq
N_{1}\left\Vert \Phi_{1}(s,t_{0},x)v\right\Vert;
\end{equation}
\begin{equation}\label{dich_instab}
e^{\nu_{2}(t-s)}\left\Vert \Phi_{2}(s,t_{0},x)(x)v\right\Vert \leq
N_{2}\left\Vert \Phi_{2}(t,t_{0},x)(x)v\right\Vert,
\end{equation}
for all $(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$.
\end{definition}
\begin{remark}\rm
Without any loss of generality we can consider
\[
N=\max\{N_{1},N_{2}\} \ \textrm{and} \ \nu=\min
\{\nu_{1},\nu_{2}\}.
\]
We will call $N_{1}$, $N_{2}$, $\nu_{1}$, $\nu_{2}$, respectively
$N$, $\nu$ \emph{dichotomic characteristics}.
\end{remark}
In what follows we will define generalizations for skew-evolution
semiflows of some asymptotic properties given by L. Barreira and
C. Valls for evolution equations in \cite{BaVa_LNM}.
\begin{definition}\rm\label{def_BVed}
The skew-evolution semiflow $C=(\varphi,\Phi)$ is called
\emph{Barreira-Valls exponentially dichotomic} if there exist two
projectors $P_{1}$ and $P_{2}$ compatible with $C$, some constants
$N\geq 1$, $\alpha_{1}$, $\alpha_{2}>0$ and $\beta_{1}$,
$\beta_{2}>0$ such that:
\begin{equation}\label{BV_dich_stab}
\left\Vert \Phi_{1}(t,t_{0},x)v\right\Vert \leq Ne^{-\alpha_{1}
t}e^{\beta_{1} s}\left\Vert \Phi_{1}(s,t_{0},x)v\right\Vert;
\end{equation}
\begin{equation}\label{BV_dich_instab}
\left\Vert \Phi_{2}(s,t_{0},x)v\right\Vert \leq Ne^{-\alpha_{2}
t}e^{\beta_{2} s}\left\Vert \Phi_{2}(t,t_{0},x)v\right\Vert,
\end{equation}
for all $(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$.
\end{definition}
\begin{definition}\rm\label{def_ed}
The skew-evolution semiflow $C=(\varphi,\Phi)$ is called
\emph{exponentially dichotomic} if there exist two projectors
$P_{1}$ and $P_{2}$ compatible with $C$, some mappings $N_{1}$,
$N_{2}:\mathbf{R}_{+}\rightarrow \mathbf{R}_{+}^{\ast }$ and some
constants $\nu_{1}$, $\nu_{2}>0$ such that:
\begin{equation}
\left\Vert \Phi_{1}(t,t_{0},x)v\right\Vert \leq
N_{1}(s)e^{-\nu_{1}t}\left\Vert \Phi_{1}(s,t_{0},x)v\right\Vert;
\end{equation}
\begin{equation}
\left\Vert \Phi_{2}(s,t_{0},x)v\right\Vert \leq
N_{2}(s)e^{-\nu_{2}t}\left\Vert \Phi_{2}(t,t_{0},x)v\right\Vert,
\end{equation}
for all $(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$.
\end{definition}
Some immediate connections concerning the previously defined
asymptotic properties for skew-evolution semiflows are given by
\begin{remark}\rm\label{obs_ued_BVed}
$(i)$ A uniformly exponentially dichotomic skew-evolution semiflow
is Barreira-Valls exponentially dichotomic;
$(ii)$ Barreira-Valls exponentially dichotomic skew-evolution
semiflow is exponentially dichotomic.
\end{remark}
The reciprocal statements are not true, as shown in what follows.
Hence, the next example emphasizes a skew-evolution semiflow which
is Barreira-Valls exponentially dichotomic, but is not uniformly
exponentially dichotomic.
\begin{example}\rm\label{ex_BVed}
Let $f:\mathbf{R}_{+}\rightarrow(0,\infty)$ be a decreasing
function with the property that there exists
$\lim\limits_{t\rightarrow\infty}f(t)=l>0$. We will consider
$\lambda>f(0)$. Let
$\mathcal{C}=\mathcal{C}(\mathbf{R},\mathbf{R})$ be the metric
space of all continuous functions $x:\mathbf{R}\rightarrow
\mathbf{R}$, with the topology of uniform convergence on compact
subsets of $\mathbf{R}$. $\mathcal{C}$ is metrizable relative to
the metric given in Example \ref{ex_ses}. We denote $ X$ the
closure in $\mathcal{C}$ of the set ${\{f_{t}, \ t\in
\mathbf{R}_{+}\}}$, where $f_{t}(\tau)=f(t+\tau)$, $\forall
\tau\in \mathbf{R}_{+}$. Then $( X,d)$ is a metric space. The
mapping $$\varphi: T\times X\rightarrow X, \
\varphi(t,s,x)(\tau)=x_{t-s}(\tau)=x(t-s+\tau)$$ is an evolution
semiflow on $ X$. Let us consider the Banach space
$V=\mathbf{R}^{2}$ with the norm $\left\Vert
v\right\Vert=|v_{1}|+|v_{2}|$, $v=(v_{1},v_{2})\in V$. The mapping
\[
\Phi: T\times X \rightarrow \mathcal{B}(V), \ \Phi(t,s,x)v=
\]
\[
=\left(e^{t\sin t-s\sin
s-2(t-s)-\int_{s}^{t}x(\tau-s)d\tau}v_{1},\ e^{3(t-s)-2t\cos
t+2s\cos s+\int_{s}^{t}x(\tau-s)d\tau}v_{2}\right),
\]
where $t\geq s\geq 0, \ (x,v)\in Y$, is an evolution cocycle over
the evolution semiflow $\varphi$. We consider the projectors
$P_{1}, P_{2}:Y\rightarrow Y$, $P_{1}(x,v)=(v_{1},0)$, $
P_{2}(x,v)=(0,v_{2})$, for all $x\in X$ and all
$v=(v_{1},v_{2})\in V$, compatible with the skew-evolution
semiflow $C=(\varphi, \Phi)$.
We have, according to the properties of function $x$,
$$\left| \Phi(t,s,x)P_{1}(x)v\right|= e^{t\sin t-s\sin
s+2s-2t}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|\leq$$ $$\leq
e^{-t+3s}e^{-l(t-s)}|v_{1}|=e^{-(1+l)t}e^{(3+l)s}|v_{1}|,$$ for
all $(t,s,x,v)\in T\times Y$.
Also, following relations $$\left|
\Phi(t,s,x)P_{2}(x)v\right|=e^{3t-3s-2t\cos t+2s\cos
s+\int_{s}^{t}x(\tau-s)d\tau}|v_{2}|\geq$$ $$ \geq
e^{t-s}e^{l(t-s)}|v_{2}|=e^{(1+l)t}e^{-(1+l)s}|v_{2}|,$$ hold for
all $(t,s,x,v)\in T\times Y$.
Hence, the skew-evolution semiflow $C=(\varphi,\Phi)$ is
Barreira-Valls exponentially dichotomic with $N=1$,
$\alpha_{1}=\alpha_{2}=\beta_{2}=1+l$, $\beta_{1}=3+l$.
Let us suppose now that $C=(\varphi,\Phi)$ is uniformly
exponentially dichotomic. According to Definition \ref{def_ued},
there exist $N\geq 1$ and $\nu_{1}>0$, $\nu_{2}>0$ such that
$$e^{t\sin t-s\sin s+2s-2t}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|\leq Ne^{-\nu_{1}(t-s)}|v_{1}|, \ \forall t\geq s\geq 0$$
and $$Ne^{3t-3s-2t\cos t+2s\cos
s}e^{\int_{s}^{t}x(\tau-s)d\tau}|v_{2}|\geq
e^{\nu_{2}(t-s)}|v_{2}|, \ \forall t\geq s\geq 0.$$ If we consider
$t=2n\pi+\frac{\pi}{2}$ and $s=2n\pi$, we have in the first
inequality
$$e^{2n\pi-\frac{\pi}{2}}\leq Ne^{-\nu\frac{\pi}{2}}e^{\int\limits_{2n\pi}^{2n\pi+\frac{\pi}{2}}x(\tau-2n\pi)d\tau}
\leq Ne^{(-\nu_{1}+\lambda)\frac{\pi}{2}},$$ which, for
$n\rightarrow \infty$, leads to a contradiction. In the second
inequality, if we consider $t=2n\pi$ and $s=2n\pi-\pi$, we obtain
$$Ne^{-4n\pi+3\pi}\geq e^{\nu_{2}\pi}e^{\int\limits_{2n\pi-\pi}^{2n\pi}
x(\tau-2n\pi+\pi)d\tau}\geq e^{(\nu_{2}-\lambda)\pi}.$$ For
$n\rightarrow \infty$, a contradiction is obtained.
We obtain that $C$ is not uniformly exponentially dichotomic.
\end{example}
There exist exponentially dichotomic skew-evolution semiflows that
are not Barreira-Valls exponentially dichotomic, as in the next
\begin{example}\rm\label{ex_ed}
We consider the metric space $(X,d)$, the Banach space $V$, the
evolution semiflow $\varphi$ and the projectors $P_{1}$ and
$P_{2}$ defined as in Example \ref{ex_BVed}. Let us consider a
continuous function
\[
g:\mathbf{R}_{+}\rightarrow[1,\infty) \ \textrm{with} \
g(n)=e^{n\cdot2^{2n}} \ \textrm{and} \
g\left(n+\frac{1}{2^{2n}}\right)=1.
\]
The mapping $\Phi: T\times X\rightarrow \mathcal{B}(V)$, defined
by
\[
\Phi(t,s,x)v=\left(\frac{g(s)}{g(t)}e^{-(t-s)-\int_{s}^{t}x(\tau-s)d\tau}v_{1},
\frac{g(t)}{g(s)}e^{t-s+\int_{s}^{t}x(\tau-s)d\tau}v_{2}\right)
\]
is an evolution cocycle over the evolution semiflow $\varphi$. As
$$\left| \Phi(t,s,x)P_{1}(x)v\right|\leq g(s)e^{-(1+l)(t-s)}|v_{1}|, \ \forall
(t,s,x,v)\in T\times Y$$ and $$g(s)\left|
\Phi(t,s,x)P_{2}(x)v\right|\geq e^{(1+l)(t-s)}|v_{2}|,\ \forall
(t,s,x,v)\in T\times Y,$$ the skew-evolution semiflow
$C=(\varphi,\Phi)$ is exponentially dichotomic, with
$N_{1}(u)=N_{2}(u)=g(u)\cdot e^{(1+l)u}$, $u\geq 0$, and
$\nu_{1}=\nu_{2}=1+l$.
Let us suppose that $C$ is Barreira-Valls exponentially
dichotomic. There exist $N\geq 1$ and $\alpha_{1}$, $\alpha_{2}$,
$\beta_{1}$, $\beta_{2}>0$ such that
$$\frac{g(s)}{g(t)}e^{s}\leq Ne^{t}e^{-\alpha_{1} t}e^{\beta_{1} s}e^{\int_{s}^{t}x(\tau-s)d\tau}$$
and $$e^{\alpha_{2} t}e^{-t}\leq N\frac{g(t)}{g(s)}e^{\beta_{2}
s}e^{-s}e^{\int_{s}^{t}x(\tau-s)d\tau}.$$ Further, if we consider
$t=n+\displaystyle\frac{1}{2^{2n}}$ and $s=n$, it follows that
$$e^{n(2^{2n}+1+\alpha_{1}-\beta_{1})}\leq Ne^{\frac{\lambda-\alpha_{1}}{2^{2n}}}\ \textrm{and}
\ e^{n(2^{n}+\alpha_{2}-\beta_{2})}\leq
Ne^{\frac{1+\lambda-\alpha_{2}}{2^{2n}}}.$$ As, for $n\rightarrow
\infty$, two contradictions are obtained, it follows that $C$ is
not Barreira-Valls exponentially dichotomic.
\end{example}
Let us present some particular classes of dichotomy, given by
\begin{definition}\rm\label{upd}
A skew-evolution semiflow $C =(\varphi,\Phi)$ is \emph{uniformly
polynomially dichotomic} if there exist two projectors $P_{1}$ and
$P_{2}$ compatible with $C$ and some constants $N\geq 1$ and
$\alpha_{1}>0$, $\alpha_{2}>0$ such that:
\begin{equation}
\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq Nt^{-\alpha_{1}
}s^{\alpha_{1}}\left\Vert P_{1}(x)v\right\Vert;
\end{equation}
\begin{equation}
\left\Vert P_{2}(x)v\right\Vert \leq Nt^{-\alpha_{2}
}s^{\alpha_{2}}\left\Vert \Phi_{2}(t,s,x)v\right\Vert;
\end{equation}
for all $(t,s)\in T$ and all $(x,v)\in Y$.
\end{definition}
\begin{definition}\rm\label{BVpd}
A skew-evolution semiflow $C =(\varphi,\Phi)$ is
\emph{Barreira-Valls polynomially dichotomic} if there exist some
constants $N\geq 1$, $\alpha_{1}>0$, $\alpha_{2}>0$ and
$\beta_{1}>0$, $\beta_{2}>0$ such that:
\begin{equation}
\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq Nt^{-\alpha_{1}
}s^{\beta_{1}}\left\Vert P_{1}(x)v\right\Vert;
\end{equation}
\begin{equation}
\left\Vert P_{2}(x)v\right\Vert \leq Nt^{-\alpha_{2}
}s^{\beta_{2}}\left\Vert \Phi_{2}(t,s,x)v\right\Vert,
\end{equation}
for all $(t,s)\in T$ and all $(x,v)\in Y$.
\end{definition}
\begin{definition}\rm\label{pd}
A skew-evolution semiflow $C =(\varphi,\Phi)$ is
\emph{polynomially dichotomic} if there exist a function
$N:\mathbf{R}_{+}\rightarrow [1,\infty)$, some constants
$\alpha_{1}>0$ and $\alpha_{2}>0$ such that:
\begin{equation}
\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq N(s)t^{-\alpha_{1}
}\left\Vert P_{1}(x)v\right\Vert;
\end{equation}
\begin{equation}
\left\Vert P_{2}(x)v\right\Vert \leq N(s)t^{-\alpha_{2}
}\left\Vert \Phi_{2}(t,s,x)v\right\Vert,
\end{equation}
for all $(t,s)\in T$ and all $(x,v)\in Y$.
\end{definition}
Relations between the defined classes of dichotomy are described
by
\begin{remark}\rm\label{obs_upd_BVpd}
$(i)$ A uniformly polynomially dichotomic skew-evolution semiflow
is Barreira-Valls polynomially dichotomic;
$(ii)$ A Barreira-Valls polynomially dichotomic is polynomially
dichotomic.
\end{remark}
The next example shows a skew-evolution semiflow which is
Barreira-Valls polynomially dichotomic but is not uniformly
polynomially dichotomic.
\begin{example}\rm
We consider the metric space $(X,d)$, the Banach space $V$, the
evolution semiflow $\varphi$ and the projectors $P_{1}$ and
$P_{2}$ defined as in Example \ref{ex_BVed}. We will consider the
mapping
$$g:\mathbf{R}_{+}\rightarrow\mathbf{R}, \ g(t)=(t+1)^{3-\sin\ln(t+1)}.$$ We define
$$\Phi(t,s,x)v=\left(\frac{g(s)}{g(t)}e^{-\int_{s}^{t}x(\tau-s)d\tau}v_{1},
\frac{g(t)}{g(s)}e^{\int_{s}^{t}x(\tau-s)d\tau}v_{2}\right),
(t,s)\in T,\ (x,v)\in Y.$$ $\Phi$ is an evolution cocycle over
$\varphi$. Due to the properties of function $x$ and of function
$f:(0,\infty)\rightarrow (0,\infty), \
f(u)=\displaystyle\frac{e^{u}}{u}$, we have
$$\left| \Phi_{1}(t,s,x)v\right| \leq \frac{(s+1)^{4}}{(t+1)^{2}}e^{-l(t-s)}|v_{1}|\leq (s+1)^{2}
\left(\frac{s+1}{t+1}\right)^{2}e^{-lt}e^{ls}|v_{1}|\leq$$ $$\leq
\frac{s(s+1)^{2}}{t}t^{-l}s^{l}|v_{1}|\leq
4t^{-(1+l)}s^{3+l}|v_{1}|,$$ for all $t\geq s\geq t_{0}=1$ and all
$(x,v)\in Y$. Also, following relations $$\left|
\Phi_{2}(t,s,x)v\right| \geq
\frac{(s+1)^{4}}{(t+1)^{2}}e^{-l(t-s)}|v_{2}|\geq
\frac{(t+1)^{2}}{(s+1)^{4}}e^{lt}e^{-ls}|v_{2}|\geq
t^{2+l}s^{-8-l}|v_{2}|,$$ hold for all $t\geq s\geq t_{0}=1$ and
all $(x,v)\in Y$.
Hence, by Definition \ref{BVpd}, the skew-evolution semiflow
$C=(\varphi,\Phi )$ is Barreira-Valls polynomially dichotomic.
We suppose now that $C$ is uniformly polynomially dichotomic.
According to Definition \ref{upd}, there exist $N\geq 1$ and
$\alpha_{1}>0$ such that
$$\frac{(s+1)^{3}}{(t+1)^{3}}\frac{(t+1)^{\sin\ln (t+1)}}{(s+1)^{\sin\ln (s+1)}}
\leq
Nt^{-\alpha_{1}}s^{\alpha_{1}}e^{\int_{s}^{t}x(\tau-s)d\tau}$$ for
all $t\geq s\geq t_{0}$. Let us consider
$$t=e^{2n\pi+\frac{\pi}{2}}-1 \ \textrm{and} \
s=e^{2n\pi-\frac{\pi}{2}}-1.$$ We have, if we consider the
properties of function $x$, that
$$e^{(2n-\lambda-1)\pi}\leq N e^{2\alpha_{1}},$$
which, if $n\rightarrow \infty$, leads to a contradiction.
Also, as in Definition \ref{upd}, there exist $N\geq 1$ and
$\alpha_{1}>0$ such that
$$N\frac{(t+1)^{3}}{(s+1)^{3}}\frac{(s+1)^{\sin\ln (s+1)}}{(t+1)^{\sin\ln (t+1)}}
\geq
t^{\alpha_{2}}s^{-\alpha_{2}}e^{-\int_{s}^{t}x(\tau-s)d\tau}$$ for
all $t\geq s\geq t_{0}$, which implies, for
$t=e^{2n\pi+\frac{\pi}{2}}-1$ and $s=e^{2n\pi-\frac{\pi}{2}}-1$,
$$Ne^{(-2n+\lambda-1)\pi}\geq e^{-2\alpha_{2}},$$ which, for $n\rightarrow \infty$, is a contradiction.
We obtain thus that $C$ is not uniformly polynomially dichotomic.
\end{example}
There exist skew-evolution semiflows that are polynomially
dichotomic but are not Barreira-Valls polynomially dichotomic.
\begin{example}\rm
Let us consider the data given in Example \ref{ex_ed}. We obtain
$$\left| \Phi(t,s,x)P_{1}(x)v\right|\leq g(s)e^{-(1+l)(t-s)}|v_{1}|
\leq g(s)e^{(1+l)s}t^{-(1+l)}$$ and $$g(s)e^{(1+l)s}\left|
\Phi(t,s,x)P_{2}(x)v\right|\geq e^{(1+l)t}|v_{2}|\geq
t^{(1+l)}|v_{2}|,$$ for all $(t,s,x,v)\in T\times Y$, which proves
that the skew-evolution semiflow $C=(\varphi,\Phi)$ is
polynomially dichotomic.
If we suppose that $C$ is Barreira-Valls polynomially dichotomic,
there exist $N\geq 1$, $\alpha_{1}>0$, $\alpha_{2}>0$ and
$\beta_{1}>0$, $\beta_{2}>0$ such that $$\frac{g(s)}{g(t)}\leq
Nt^{-\alpha_{1}}s^{\beta_{1}}e^{t-s+\int_{s}^{t}x(\tau-s)d\tau}\
\textrm{and} \ t^{\alpha_{2}}\leq N\frac{g(t)}{g(s)}s^{\beta_{2}
}e^{t-s+\int_{s}^{t}x(\tau-s)d\tau}.$$ If we consider
$t=n+\displaystyle\frac{1}{2^{2n}}$ and $s=n$, we obtain
$$e^{n\cdot 2^{2n}}\leq N\cdot n^{-\alpha_{1}}\cdot
n^{\beta_{1}}\cdot e^{\frac{1+\lambda}{2^{2n}}}\ \textrm{and}\
e^{n\cdot 2^{2n}} \leq
N\left(n+\frac{1}{2^{2n}}\right)^{-\alpha_{2}}\cdot
n^{\beta_{2}}\cdot e^{\frac{1+\lambda}{2^{2n}}}.$$ For
$n\rightarrow \infty$, two contradictions are obtained, which
proves that $C$ is not Barreira-Valls polynomially dichotomic.
\end{example}
\section{Main results}
The first results will prove some relations between all the
classes of dichotomies.
\begin{proposition}
A uniformly exponentially dichotomic skew-evolution semiflow
$C=(\varphi, \Phi)$ is uniformly polynomially dichotomic.
\end{proposition}
\begin{proof}
Let us consider in Definition \ref{def_ued}, without any loss of
generality, $t_{0}=1$. It also assures the existence of constants
$N\geq 1$ and $\nu_{1} >0$ such that $\left\Vert
\Phi_{1}(t,s,x)v\right\Vert \leq Ne^{-\nu_{1} (t-s)}\left\Vert
P_{1}(x)v\right\Vert.$ As
$$e^{-u}\leq \frac{1}{u+1}, \ \forall u\geq 0 \ \textrm{and }\
\frac{t}{s}\leq t-s+1, \ \forall t\geq s\geq 1,$$ it follows that
$$\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq
N(t-s+1)^{-\nu_{1}}\left\Vert P_{1}(x)v\right\Vert\leq
Nt^{-\nu_{1}}s^{\nu_{1}}\left\Vert P_{1}(x)v\right\Vert,$$ for all
$t\geq s\geq 1$ and all $(x,v)\in Y$.
We also have the property of function $$f:(0,\infty)\rightarrow
(0,\infty), \ f(u)=\frac{e^{u}}{u}$$ of being nondecreasing, which
assures the inequality $$\frac{e^{s}}{e^{t}}\leq\frac{s}{t},\
\forall t\geq s>0$$ and, further, for all $t\geq s\geq 1$ and all
$(x,v)\in Y$, we have
$$\left\Vert P_{2}(x)v\right\Vert\leq Ne^{-\nu_{2}t}e^{\nu_{2}
s}\left\Vert \Phi_{2}(t,s,x)v\right\Vert \leq
Nt^{-\nu_{2}}s^{\nu_{2} }\left\Vert \Phi_{2}(t,s,x)v\right\Vert,$$
where constants $N\geq 1$ and $\nu_{2} >0$ are also given by
Definition \ref{def_ued}.
Thus, according to Definition \ref{upd}, $C$ is uniformly
polynomially dichotomic.
\end{proof}
\vspace{3mm}
We give an example of a skew-evolution semiflow which is uniformly
polynomially dichotomic, but is not uniformly exponentially
dichotomic.
\begin{example}\rm\label{ex_upd}
Let $(X,d)$ be the metric space, $V$ the Banach space, $\varphi$
the evolution semiflow, $P_{1}$ and $P_{2}$ the projectors given
as in Example \ref{ex_BVed}.
Let us consider the function
$g:\mathbf{R}_{+}\rightarrow\mathbf{R}$, given by $g(t)=t^{2}+1$
and let us define
$$\Phi(t,s,x)v=\left(\frac{g(s)}{g(t)}e^{-\int_{s}^{t}x(\tau-s)d\tau}v_{1},
\frac{g(t)}{g(s)}e^{\int_{s}^{t}x(\tau-s)d\tau}v_{2}\right), \
(t,s)\in T,\ (x,v)\in Y.$$ We can consider $t_{0}=1$ in Definition
\ref{upd}. As,
$\displaystyle\frac{s^{2}+1}{t^{2}+1}\leq\frac{s}{t}$, for $t\geq
s\geq 1$ and according to the properties of function $x$, we have
$$\frac{s^{2}+1}{t^{2}+1}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|
\leq t^{-(1+l)}s^{1+l}|v_{1}|$$ and $$\frac{t^{2}+1}{s^{2}+1}
e^{\int_{s}^{t}x(\tau-s)d\tau}|v_{2}| \geq
t^{(2+l)}s^{-(4+l)}|v_{2}|,$$ for all $t\geq s\geq 1$ and all
$v\in V$. It follows that $C=(\varphi,\Phi)$ is uniformly
polynomially dichotomic.
If the skew-evolution semiflow $C=(\varphi,\Phi)$ is also
uniformly exponentially dichotomic, according to Definition
\ref{def_ued}, there exist $N\geq 1$ $\nu_{1}>0$ and $\nu_{2}>0$
such that
$$\frac{s^{2}+1}{t^{2}+1}|v_{1}|\leq Ne^{-\nu_{1}(t-s)}e^{-l(t-s)}|v_{1}|\
\textrm{and}\ N\frac{t^{2}+1}{s^{2}+1}|v_{2}|\geq
e^{\nu_{2}(t-s)}e^{l(t-s)}|v_{2}|,$$ for all $t\geq s\geq t_{0}$
and all $v\in V.$ If we consider $s=t_{0}$ and $t\rightarrow
\infty$, two contradictions are obtained, which proves that $C$ is
not uniformly exponentially dichotomic.
\end{example}
\begin{proposition}
A Barreira-Valls exponentially dichotomic skew-evolution semiflow
$C=(\varphi, \Phi)$ with $\alpha_{i}\geq\beta_{i}>0$, $i\in
\{1,2\}$, is Barreira-Valls polynomially dichotomic.
\end{proposition}
\begin{proof}
According to Definition \ref{def_BVed}, there exist some constants
$N\geq 1$, $\alpha_{1}>0$ and $\beta_{1}>0$ such that $$\left\Vert
\Phi_{1}(t,s,x)v\right\Vert \leq Ne^{-\alpha_{1} t}e^{\beta_{1}
s}\left\Vert P_{1}(x)v\right\Vert, \ \forall (t,s)\in T,\ \forall
(x,v)\in Y.$$ As the mapping $f:(0,\infty)\rightarrow (0,\infty)$,
defined by $f(u)=\displaystyle\frac{e^{u}}{u}$ is nondecreasing,
and as, by hypothesis, we can chose $\alpha_{1}\geq\beta_{1}$, we
obtain that
$$\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq
Nt^{-\alpha_{1}}e^{-\beta_{1} s}s^{\beta_{1}}e^{\beta_{1}
s}\left\Vert
P_{1}(x)v\right\Vert=Nt^{-\alpha_{1}}s^{\beta_{1}}\left\Vert
P_{1}(x)v\right\Vert,$$ for all $t\geq s>0$ and all $(x,v)\in Y$.
Analogously, we obtain $$\left\Vert P_{2}(x)v\right\Vert\leq
Ne^{-\alpha_{2}t}e^{\beta_{2}s}\left\Vert
\Phi_{2}(t,s,x)v\right\Vert\leq
Nt^{-\alpha_{2}}s^{\beta_{2}}\left\Vert
\Phi_{2}(t,s,x)v\right\Vert,$$ for all $t\geq s>0$ and all
$(x,v)\in Y$, where the constants $N\geq 1$, $\alpha_{2}>0$ and
$\beta_{2}>0$ are also assured by Definition \ref{def_ued}, with
the property $\alpha_{2}\geq\beta_{2}$.
Hence, according to Definition \ref{upd}, $C$ is Barreira-Valls
polynomially dichotomic.
\end{proof}
\vspace{3mm}
There exist skew-evolution semiflows that are Barreira-Valls
polynomially dichotomic, but are not Barreira-Valls exponentially
dichotomic.
\begin{example}\rm\label{ex_BVpd}
We consider the metric space $(X,d)$, the Banach space $V$, the
evolution semiflow $\varphi$ and the projectors $P_{1}$ and
$P_{2}$ defined as in Example \ref{ex_BVed}.
Let us consider the function
$g:\mathbf{R}_{+}\rightarrow\mathbf{R}$, given by $g(t)=t+1$ and
let us define an evolution cocycle $\Phi$ as in Example
\ref{ex_upd}. We obtain
$$\frac{s+1}{t+1}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|
\leq \frac{s^{2}}{t}e^{-l(t-s)}|v_{1}|\leq
t^{-1-l}s^{2+l}|v_{1}|$$ and
$$\frac{t+1}{s+1}e^{\int_{s}^{t}x(\tau-s)d\tau}|v_{2}| \geq t^{1+l}s^{-2-l}|v_{2}|,$$
for all $t\geq s\geq 1$ and all $v\in V$. It follows that the
skew-evolution semiflow $C=(\varphi,\Phi)$ is Barreira-Valls
polynomially dichotomic.
Let us suppose that $C$ is also Barreira-Valls exponentially
dichotomic. According to Definition \ref{def_BVed}, there exist
some constants $N\geq 1$, $\alpha_{1},\beta_{1}>0$ and
$\alpha_{2},\beta_{2}>0$ such that
$$\frac{s+1}{t+1}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|
\leq Ne^{-\alpha_{1} t}e^{\beta_{1} s}|v_{1}|$$ and
$$N\frac{t+1}{s+1}e^{\int_{s}^{t}x(\tau-s)d\tau}|v_{2}| \geq
e^{\alpha_{2} t}e^{-\beta_{2} s}|v_{2}|,$$ for all
$(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$. We consider
$s=t_{0}$. We have
$$\frac{e^{\alpha_{1} t}}{t+1}\leq\frac{\overline{N}}{t_{0}+1}\ \textrm{and}
\ \frac{e^{\alpha_{2} t}}{t+1}\leq\frac{\widetilde{N}}{t_{0}+1}, \
\forall t\geq t_{0}.$$ For $t\rightarrow \infty$, we obtain two
contradictions, and, hence, $C$ is not Barreira-Valls
exponentially dichotomic.
\end{example}
\begin{proposition}
An exponentially dichotomic skew-evolution semiflow $C=(\varphi,
\Phi)$ is polynomially dichotomic.
\end{proposition}
\begin{proof}
Definition \ref{def_ed} assures the existence of a function
$N_{1}:\mathbf{R}_{+}\rightarrow[1,\infty)$ and a constant
$\nu_{1}>0$ such that $$\left\Vert \Phi_{1}(t,s,x)v\right\Vert\leq
N_{1}(s)e^{-\nu_{1} t}\left\Vert P_{1}(x)v\right\Vert, \ \forall
(t,s)\in T, \ \forall (x,v)\in Y.$$ As following inequalities
$e^{t}\geq t+1>t$ hold for all $t\geq 0$, we obtain $$\left\Vert
\Phi_{1}(t,s,x)v\right\Vert\leq N_{1}(s)t^{-\nu_{1}}\left\Vert
P_{1}(x)v\right\Vert,$$ for all $t\geq s>0$ and all $(x,v)\in Y$.
As, by Definition \ref{def_ed} there exist a function
$N_{2}:\mathbf{R}_{+}\rightarrow[1,\infty)$ and a constant
$\nu_{2}>0$ such that $$\left\Vert P_{2}(x)v\right\Vert\leq
N_{2}(s)e^{-\nu_{2} t}\left\Vert \Phi_{2}(t,s,x)v\right\Vert, \
\forall (t,s)\in T, \ \forall (x,v)\in Y.$$ Analogously, as
previously, we have $$\left\Vert P_{2}(x)v\right\Vert\leq
N_{2}(s)t^{-\nu_{2}}\left\Vert \Phi_{2}(t,s,x)v\right\Vert,$$ for
all $t\geq s>0$ and all $(x,v)\in Y$.
Hence, according to Definition \ref{pd}, $C$ is polynomially
dichotomic.
\end{proof}
\vspace{3mm}
We present an example of a skew-evolution semiflow which is
polynomially dichotomic, but is not exponentially dichotomic.
\begin{example}\rm
We consider the metric space $(X,d)$, the Banach space $V$, the
evolution semiflow $\varphi$, the projectors $P_{1}$, $P_{2}$ and
function $g$ as in Example \ref{ex_BVpd}. Let
$$\Phi(t,s,x)v=\left(\frac{g(s)}{g(t)}e^{\int_{s}^{t}x(\tau-s)d\tau}|v_{1}|,
\frac{g(t)}{g(s)}e^{-\int_{s}^{t}x(\tau-s)d\tau}|v_{2}|\right)$$
be an evolution cocycle. Analogously as in the mentioned Example,
the skew-evolution semiflow $C$ is Barreira-Valls polynomially
dichotomic, and, according to Remark \ref{obs_upd_BVpd} $(ii)$, it
is also polynomially dichotomic. On the other hand, if we suppose
that $C$ is exponentially dichotomic, there exist $N_{1}$,
$N_{2}:\mathbf{R}_{+}\rightarrow \mathbf{R}_{+}^{\ast }$ and
$\nu_{1}$, $\nu_{2}>0$ such that
$$\frac{s+1}{t+1}|v_{1}| \leq
N_{1}(s)e^{-(\nu_{1}+l)t}e^{l s}|v_{1}|$$ and $$ |v_{2}| \leq
N_{2}(s)e^{-(\nu_{2}+l)t} \frac{t+1}{s+1}|v_{2}|,$$ for all
$(t,s)\in T$ and all $(x,v)\in Y$. If we consider $s=t_{0}$ and
$t\rightarrow \infty$, we obtain two contradictions, which shows
that $C$ is not exponentially dichotomic.
\end{example}
A characterization for the classic and mostly encountered property
of exponential dichotomy is given by the next
\begin{theorem}
Let $C=(\varphi,\Phi)$ be a strongly measurable skew-evolution
semiflow. $C$ is exponentially dichotomic if and only if there
exist two projectors $P_{1}$ and $P_{2}$ compatible with $C$ with
the properties that $C_{1}$ has bounded exponential growth and
$C_{2}$ has exponential decay such that
$(i)$ there exist a constant $\gamma>0$ and a mapping
$D:\mathbf{R}_{+}\rightarrow [1,\infty)$ with the property:
$$\int_{s}^{\infty}e^{(\tau-s)\gamma}\left\Vert
\Phi_{1}(\tau,s,x)v\right\Vert d\tau \leq D(s)\left\Vert
P_{1}(x)v\right\Vert,$$ for all $s\geq 0$ and all $(x,v)\in Y$;
$(ii)$ there exist a constant $\rho>0$ and a nondecreasing mapping
$\widetilde{D}:\mathbf{R}_{+}\rightarrow [1,\infty)$ with the
property:
$$\int_{t_{0}}^{t}e^{(t-\tau)\rho}\left\Vert
\Phi_{2}(\tau,t_{0},x)v\right\Vert d\tau \leq
\widetilde{D}(t_{0})\left\Vert \Phi_{2}(t,t_{0},x)v\right\Vert,$$
for all $t\geq t_{0}\geq 0$ and all $(x,v)\in Y$.
\end{theorem}
\begin{proof}
\emph{Necessity.} As $C$ is exponentially dichotomic, according to
Definition \ref{def_ed}, there exist $N_{1}$,
$N_{2}:\mathbf{R}_{+}\rightarrow \mathbf{R}_{+}^{\ast }$ and
$\nu_{1}$, $\nu_{2}>0$ such that
$$\left\Vert \Phi_{1}(t,t_{0},x)v\right\Vert \leq
N_{1}(s)e^{-\nu_{1}t}\left\Vert \Phi_{1}(s,t_{0},x)v\right\Vert$$
and $$ \left\Vert \Phi_{2}(s,t_{0},x)v\right\Vert \leq
N_{2}(s)e^{-\nu_{2}t}\left\Vert \Phi_{2}(t,t_{0},x)v\right\Vert,$$
for all $(t,s),(s,t_{0})\in T$ and all $(x,v)\in Y$.
In order to prove $(i)$, let us define
$\gamma=-\displaystyle\frac{\nu_{1}}{2}$. We obtain successively
$$\int_{s}^{\infty}e^{(\tau-s)\gamma}\left\Vert
\Phi_{1}(\tau,s,x)v\right\Vert d\tau \leq N_{1}(s)\left\Vert
\Phi_{1}(s,s,x)v\right\Vert\int_{s}^{\infty}e^{-\frac{\nu_{1}}{2}(\tau-s)}e^{-\nu_{1}(s-\tau)}d\tau=$$
$$=N(s)\left\Vert
P_{1}(x)v\right\Vert\int_{s}^{\infty}e^{-\frac{\nu_{1}}{2}(s-\tau)}d\tau=D(s)\left\Vert
P_{1}(x)v\right\Vert,$$ for all $s\geq 0$ and all $(x,v)\in Y$,
where we have denoted $$D(u)=\frac{N_{1}(u)}{\gamma}, \ u\geq 0.$$
To prove $(ii)$, we define $\rho=\displaystyle\frac{\nu_{2}}{2}$.
Following relations $$\int_{t_{0}}^{t}e^{(t-\tau)\rho}\left\Vert
\Phi_{2}(\tau,t_{0},x)v\right\Vert d\tau \leq
N_{2}(t_{0})\left\Vert\Phi_{2}(t,t_{0},x)v\right\Vert
\int_{t_{0}}^{t}e^{\frac{\nu_{2}}{2}(t-\tau)}e^{-\nu_{2}(t-\tau)}d\tau\leq$$
$$\leq \widetilde{D}(t_{0})\left\Vert\Phi_{2}(t,t_{0},x)v\right\Vert$$
hold for all $s\geq 0$ and all $(x,v)\in Y$, where we have
denoted
$$\widetilde{D}(u)=\frac{2N_{2}(u)}{\rho}, \ u\geq 0.$$
\emph{Sufficiency.} According to relation $(i)$, the
$\gamma$-shifted skew-evolution semiflow
$C_{\gamma}^{1}=(\varphi,\Phi^{1}_{\gamma})$, defined as in
Example \ref{ex_shift}, has bounded exponential growth and there
exists $D:\mathbf{R}_{+}\rightarrow [1,\infty)$ such that
$$\int_{s}^{\infty}\left\Vert
\Phi^{1}_{\gamma}(\tau,s,x)v\right\Vert d\tau \leq D(s)\left\Vert
P_{1}(x)v\right\Vert,$$ for all $s\geq 0$ and all $(x,v)\in Y$.
First of all, we will prove that there exists
$D_{1}:\mathbf{R}_{+}\rightarrow [1,\infty)$ such that $\left\Vert
\Phi^{1}_{\gamma}(t,s,x)v\right\Vert \leq D_{1}(s)\left\Vert
P_{1}(x)v\right\Vert$, for all $t\geq s\geq 0$ and all $(x,v)\in
Y$. Let us consider, for $t\geq s+1$,
$$c=\int_{0}^{1}e^{-\omega(u)}d u\leq \int^{t-s}_{0}e^{-\omega(u)}d
u=\int^{t}_{s}e^{-\omega(t-\tau)}d\tau.$$ Hence, for $t\geq s+1$,
we obtain $$c|<v^{*},\Phi^{1}_{\gamma}(t,s,x)v>|\leq
\int^{t}_{s}e^{-\omega(t-\tau)}|<v^{*},\Phi^{1}_{\gamma}(t,s,x)v>|d\tau=$$
$$=\int^{t}_{s}e^{-\omega(t-\tau)}\left\Vert
\Phi^{1}_{\gamma}(t,\tau,\varphi(\tau,s,x))^{*}v^{*}\right\Vert
\left\Vert \Phi^{1}_{\gamma}(\tau,s,x)v\right\Vert d\tau\leq$$
$$\leq M\left\Vert
P_{1}(x)v^{*}\right\Vert\int_{s}^{t}\left\Vert
\Phi^{1}_{\gamma}(\tau,s,x)v\right\Vert d\tau\leq MD(s)\left\Vert
P_{1}(x)v\right\Vert\left\Vert P_{1}(x)v^{*}\right\Vert,$$ where
$v\in V$, $v^{*}\in V^{*}$ and $M$, $\omega$ are given by
Definition \ref{def_neg} and Remark \ref{obs_eg}. Hence,
$$\left\Vert \Phi_{1}(t,s,x)v\right\Vert\leq \frac{MD(s)}{c},
\ \forall t\geq s+1, \ \forall (x,v)\in Y.$$
Now, for $t\in[s,s+1)$, we have
$$\left\Vert \Phi_{1}(t,s,x)v\right\Vert\leq Me^{\omega(1)}\left\Vert P_{1}(x)v\right\Vert,
\ \forall (x,v)\in Y.$$
Thus, we obtain $$\left\Vert \Phi^{1}_{\gamma}(t,s,x)v\right\Vert
\leq D_{1}(s)\left\Vert P_{1}(x)v\right\Vert,$$ for all $t\geq
s\geq 0$ and all $(x,v)\in Y$, where we have denoted $$D_{1}(u)=
M\left[e^{\omega(1)}+\frac{D(u)}{c}\right], \ u\geq 0.$$ Further,
it follows that
$$\left\Vert \Phi_{1}(t,s,x)v\right\Vert \leq
D_{1}(s)e^{-(t-s)\gamma}\left\Vert v\right\Vert,\ \forall t\geq
s\geq 0.$$
According to $(ii)$, there exist a constant $\rho>0$ and a
nondecreasing mapping $\widetilde{D}:\mathbf{R}_{+}\rightarrow
[1,\infty)$ such that
$$\int_{t_{0}}^{t}e^{-(\tau-t_{0})\rho}\left\Vert
\Phi_{2}(\tau,t_{0},x)v\right\Vert d\tau \leq
\widetilde{D}(t_{0})e^{-(t-t_{0})\rho}\left\Vert
\Phi_{2}(t,t_{0},x)v\right\Vert,$$ for all $t\geq t_{0}\geq 0$ and
all $(x,v)\in Y$. Thus,
$$\int_{t_{0}}^{t}\left\Vert
\Phi^{2}_{-\rho}(\tau,t_{0},x)v\right\Vert d\tau \leq
\widetilde{D}(t_{0})\left\Vert
\Phi^{2}_{-\rho}(t,t_{0},x)v\right\Vert,$$ for all $t\geq
t_{0}\geq 0$ and all $(x,v)\in Y$, where $\Phi^{2}_{-\rho}$ is
defined as in Example \ref{ex_shift}. Let functions $M$ and
$\omega$ be given by Definition \ref{def_nedc}. Let us denote
$$c=\int_{0}^{1}e^{-\omega(\tau)}d\tau=\int_{s}^{s+1}e^{-\omega(u-s)}du.$$
Further, for $t\geq s+1$ and $s\geq t_{0}\geq 0$, we obtain
$$c\left\Vert \Phi^{2}_{-\rho}(s,t_{0},x)v\right\Vert=\int_{s}^{s+1}e^{-\omega(u-s)}
\left\Vert\Phi^{2}_{-\rho}(s, t_{0},x)v\right\Vert du\leq$$ $$\leq
\int_{s}^{s+1}M(t_{0})e^{-\omega(u-s)}e^{\omega(u-s)}\left\Vert\Phi^{2}_{-\rho}(u,
t_{0},x)v\right\Vert du\leq$$ $$\leq
M(t_{0})\int_{t_{0}}^{t}\left\Vert\Phi^{2}_{-\rho}(u,
t_{0},x)v\right\Vert du\leq
M(t_{0})\widetilde{D}(t_{0})\left\Vert\Phi^{2}_{-\rho}(t,
t_{0},x)v\right\Vert.$$ We obtain $$\left\Vert
\Phi^{2}_{-\rho}(s,t_{0},x)v\right\Vert\leq
\frac{M(t_{0})\widetilde{D}(t_{0})}{c}\left\Vert\Phi^{2}_{-\rho}(t,
t_{0},x)v\right\Vert,$$ for all $t\geq s\geq t_{0}\geq 0$ with
$t\geq s+1$ and all $(x,v)\in Y.$ Now, for $t\in [s,s+1)$ and
$s\geq t_{0}\geq 0$, we have $$\left\Vert\Phi^{2}_{-\rho}(s,
t_{0},x)v\right\Vert\leq
M(t_{0})e^{\omega(1)}\left\Vert\Phi^{2}_{-\rho}(t,
t_{0},x)v\right\Vert,$$ for all $(x,v)\in Y.$ Finally, we obtain
$$\left\Vert\Phi^{2}_{-\rho}(s,
t_{0},x)v\right\Vert\leq D_{2}(t_{0})\left\Vert\Phi^{2}_{-\rho}(t,
t_{0},x)v\right\Vert,$$ for all $t\geq s\geq t_{0}\geq 0$ and all
$(x,v)\in Y$, where we have denoted
$$D_{2}(u)=M(u)\left[\frac{\widetilde{D}(u)}{c}+e^{\omega(1)}\right], \ u\geq 0.$$
Thus, it follows that $$e^{-(s-t_{0})\rho}\left\Vert\Phi_{2}(s,
t_{0},x)v\right\Vert\leq
D_{2}(t_{0})e^{-(t-t_{0})\rho}\left\Vert\Phi_{2}(t,
t_{0},x)v\right\Vert,$$ which implies
$$\left\Vert\Phi_{2}(s, t_{0},x)v\right\Vert\leq
D_{2}(t_{0})e^{-(t-s)\rho}\left\Vert\Phi_{2}(t,
t_{0},x)v\right\Vert,$$ for all $t\geq s\geq t_{0}\geq 0$ and all
$(x,v)\in Y$, or $$\left\Vert P_{2}(x)v\right\Vert\leq
D_{2}(s)e^{-(t-s)\rho}\left\Vert\Phi_{2}(t,s,x)v\right\Vert,$$ for
all $t\geq s \geq 0$ and all $(x,v)\in Y$.
Hence, the skew-evolution semiflow is exponentially dichotomic,
which ends the proof.
\end{proof}
\vspace{3mm}
\textbf{Acknowledgments.} This work is financially supported from
the Exploratory Research Grant CNCSIS PN II ID 1080 No. 508/2009
of the Romanian Ministry of Education, Research and Innovation.
\vspace{5mm}
\footnotesize{
|
\section{Introduction}
\indent Since the historical work of Arago~\cite{Arago} and
Pasteur~\cite{Pasteur}, chirality (the handedness of nature) has
generally been associated with optical activity, that is the
rotation of the plane of polarisation of light passing through a
medium lacking mirror symmetry~\cite{Hecht,Landau}. Optical
activity is nowadays a very powerful probes of structural
chirality in varieties of system. However, two-dimensional chiral
structures, such as planar molecules, were not expected to display
any chiral characteristics since simply turning the object around
leads to the opposite handedness (we remind that a planar
structure is chiral if it can not be brought into congruence with
its mirror image unless it is lifted from the plane). This
fundamental notion was recently challenged in a pioneering study
where it was shown that chirality has a distinct signature from
optical activity when electromagnetic waves interact with a 2D
chiral structure and that the handedness can be
recognized~\cite{Fedotov}. While the experimental demonstration
was achieved in the giga-Hertz (mm) range for extended 2D
structures, the question remained whether this could be achieved
in the optical range since the laws of optics are not simply
scalable when downsizing to the nanometer level. Here we report
genuine optical planar chirality for a single subwavelength hole
surrounded by left and right handed Archimedian spirals milled in
a metallic film. Key to this finding is the involvement of
surface plasmons, lossy electromagnetic waves at the metal
surfaces, and the associated planar spatial
dispersion~\cite{Barnes,Genet}. Our results reveal how, in a
stringent and unusual way, this optical phenomenon connects
concepts of chirality, reciprocity and broken time symmetry.\\
\indent We remind that partly boosted by practical motivations,
such as the quest of negative refractive lenses~\cite{Pendry} or
the possibility to obtain giant optical activity for applications
in optoelectronics, there is currently a renewed
interest~\cite{Pendry,Papakostas,Schwanecke,Vallius,Gonokami,Canfield,Canfield2,Zhang,Decker,Plum,Rogacheva}
in the optical activity in artificial photonic media with planar
chiral structures. It was shown for instance that planar
gammadionic structures, which have by definition no axis of
reflection but a four-fold rotational
invariance~\cite{Papakostas,Vallius}, can generate optical
activity with giant gyrotropic
factors~\cite{Gonokami,Rogacheva,Plum,Decker}. Importantly, and in
contrast to the usual three dimensional (3D) chiral medium (like
quartz and its helicoidal structure~\cite{Hecht,Bose}), planar
chiral structures change their observed handedness when the
direction of light is reversed through the
system~\cite{Papakostas,Barron1}. This challenged Lorentz
principle of reciprocity~\cite{Landau} (which is known to hold for
any linear non magneto-optical media) and stirred up considerable
debate~\cite{Papakostas,Schwanecke,Gonokami,Barron2} which came to
the conclusion that optical activity cannot be a purely 2D effect
and always requires a small dissymmetry between the two sides of
the system~\cite{Gonokami,Rogacheva,Plum,Decker}. Nevertheless
Zheludev and colleagues did demonstrate in the GHz spectrum that a
pure 2D chiral structure lacking rotational symmetry can have an
optical signature which is distinct from optical
activity~\cite{Fedotov}. They went on to predict that it should be
possible to observe the same phenomena in the optical range by
scaling down their fish-scale structure and playing on localized
plasmons~\cite{Fedotov2}. Following a different strategy, we show
here that SP waves propagating on a 2D metal chiral grating
resonantly excited
by light provide an elegant solution to generate planar optical chirality in the visible.\\
\section{Experiments and Results}
\indent This is a challenging issue as it leads to two fundamental
points which are apparently incompatible. On the one hand, finding
such a 2D chiral effect in the optical domain is not equivalent to
a simple rescaling of the problem from the GHz to the visible part
of the spectrum. Indeed, losses in metal become predominant at the
nanometer scale so that the penetration length of light through
any chiral structure will become comparable to the thickness of
the structure. In-depth spatial dispersion along the propagation
direction of light will hence be induced, corresponding to the
usual 3D optical
activity~\cite{Vallius,Gonokami,Canfield,Canfield2,Zhang,Decker,Plum,Rogacheva}.
One thus expects optical activity, through the losses, to be a
more favorable channel than 2D optical chirality. On the other
hand, losses (i.e., broken time invariance at the macroscopic
scale) are necessary to guarantee planar chiral
behavior~\cite{Fedotov,Fedotov2}. With this in mind, we chose to
make single SP structures such as a single hole in an optically
thick metal film surrounded by an Archimedian spirals (Figure 1)
which can provide all the necessary ingredients for observing 2D
optical chirality. It is a 2D structure lacking point symmetry,
that is rotational and mirror invariances. At the same time, it
resonates due to coupling to surface plasmons which, as lossy
waves, represent a natural way for delocalizing information along
a planar interface, moving in-depth losses to the surface.
Importantly, the thickness of the metal film optically decouples
both interface~\cite{Degiron}, and consequently only the
structured chiral side is involved in the 2D optical chiral effect
reported here. Finally, the structures gives rise to enhanced
transmission~\cite{Genet} enabling high optical throughput for all
the characterization measurements.\\
\indent Using focus ion beam (FIB), we milled in an opaque gold
film a clockwise (right $\mathcal{R}$) or anticlockwise (left
$\mathcal{L}$) Archimedian spiral grooves around a central
subwavelength hole. The polar equation $(\rho,\theta)$ of the left
handed Archimedian spiral is $\rho=P\cdot\theta/(2\pi)$, and the
right handed enantiomeric spiral is obtained by reflection across
the y axis (see Fig.~1). \begin{figure}[h]
\centering\includegraphics[width=8.5cm]{drezet_1.eps}
\caption{Chiral plasmonic metamolecules. On the top panel:
scanning electron micrographs of the left (L) and right (R) handed
enantiomer (mirror image) planar chiral structures investigated.
The scale bare is 3 $\mu$m long. The parameters characterizing the
structure are the following: hole diameter $d=350$ nm, film
thickness $h=310$ nm, grating period $P=760$ nm, groove width
$w=370$ nm, and groove depth $s=80$ nm. The structures are milled,
with a focus ion beam, in a gold film deposited on a glass
substrate. On the bottom panel: transmission spectra at normal
incidence of individual left (blue curve) and right handed (red
curve) Archimede spirals illuminated from the air side. }
\end{figure}
The geometrical parameter $P$ is the radial grating period and we
take its value equal to the SP wavelength $\lambda_{SPP}\simeq
760$ nm (for an excitation at $\lambda\simeq 780$ nm). We recorded
optical transmission spectra at normal incidence with unpolarized
light for both isolated structures (Fig.~1). As it can be seen,
both enantiomers behave like resonant antennas with quasi
identical transmission properties. This resonant behaviour is a
direct indication of the SP excitation by the grating similarly to
what is observed
for circular antennas~\cite{Lezec}.\\
\indent To observe and fully characterize the optical signature of
planar chirality we perform a full polarization
tomography~\cite{Brehonnet,Altewischer} with the aim of
determining the 4$\times$4 Mueller matrix $\mathcal{M}$ associated
with each enantiomer. Experimental results
$\mathcal{M_L}^{\textrm{exp.}}$, and
$\mathcal{\mathcal{M_R}}^{\textrm{exp.}}$ respectively obtained
for left and right handed spirals are given in appendixes A and B.
Here the important point is that the degree of purity $F$ of the
Mueller matrices~\cite{Brehonnet} is near unity with
$F\left(\mathcal{M_L}^{\textrm{exp.}}\right)\simeq 0.967$ and
$F\left(\mathcal{M_R}^{\textrm{exp.}}\right)\simeq 0.939$. This
shows that the coherence in polarization is not degraded by the
structure and that we can therefore restrict our discussion to
Jones matrices~\cite{Brehonnet, Hecht}. In the convenient left
$|L\rangle$ and right $|R\rangle$ circular polarization basis,
these Jones matrices tie the excitation $
[E^{\textrm{in}}_L,E^{\textrm{in}}_R]$ to the transmitted $
[E^{\textrm{out}}_L,E^{\textrm{out}}_R]$ electric fields. In the
case of planar chiral structures displaying 2D chiral activity,
they have the following form~\cite{Fedotov, Fedotov2}:
\begin{eqnarray}
\mathcal{J_L}^{\textrm{th.}}=\left(\begin{array}{cc} A & B
\\ C& A
\end{array}\right),\mathcal{J_R}^{\textrm{th.}}=\left(\begin{array}{cc} A &
C\\B & A
\end{array}\right),
\end{eqnarray}
where $A$, $B$ and $C$ are complex valued numbers such that
$|B|\neq |C|$. This inequality account for chirality. Being non
diagonal, these matrices correspond to polarization converter
elements with no rotational invariance around the $z$ axis
(Fig.~1). They are thus fundamentally different from Jones
matrices associated with optical activity, e.g., gammadions.
Importantly the conditions $|B|\neq |C|$ implies the non unitarity
of $\mathcal{J}^{\textrm{th.}}_{\mathcal{L,R}}$ which means that
reversing the light path through the chiral structures is not
equivalent to reversing the time. From equation (1) we deduce the
associated theoretical forms for the Mueller matrices
$\mathcal{M_L}^{\textrm{th.}}$, $\mathcal{M_R}^{\textrm{th.}}$
(see appendix C) which are used to fit $\mathcal{J_L}$ and
$\mathcal{J_R}$ from experimental results. After normalization by
$A$ we deduce
\begin{eqnarray}
\mathcal{J_L}^{\textrm{fit}}=\left(\begin{array}{cc} 1.000 &
0.166+i0.221\\-0.131+i0.099 &1.000
\end{array}\right),\nonumber \\
\mathcal{J_R}^{\textrm{fit}}=\left(\begin{array}{cc} 1.000 &
-0.129+i0.098\\0.170+i0.230 &1.000
\end{array}\right).
\end{eqnarray}
These matrices indeed satisfy the chirality criteria of equation
(1) within the $\sim 1\%$ uncertainty evaluated from the degree of
purity of the Mueller matrice of the empty
setup.\\\begin{figure}[h]
\centering\includegraphics[width=14cm]{drezet_2.eps}
\caption{Analysis of the polarization states for an input light
with variable linear polarization for both the left (left panel)
and right handed (right panel) individual chiral structures of
Fig.~1. The data points (acquired with a laser light at
$\lambda=780$ nm) are compared to the predictions from equation
(2) (continuous curves) for respectively the transmitted intensity
analyzed along the direction: $|x\rangle$ (green), $|y\rangle$
(yellow), $|+45^{\circ}\rangle$ (cyan), $|-45^{\circ}\rangle$
(magenta), $|L\rangle$ (red), and $|R\rangle$ (blue). The total
transmitted intensity is also shown (black). The symmetries
between both panel expected from group theory (see appendix D) are
observed experimentally. The insets show in each panel the
ellipses of polarization and the handedness (arrow) associated
with the two corotating eingenstates associated with the Jones
matrix $\mathcal{J_L}$ (blue) and $\mathcal{J_R}$ (red). }
\end{figure}
\section{Discussion and Conclusion}
\indent To illustrate the polarization conversion properties of
our chiral structures, we compare in Fig.~2 theory and experiment
when the input state is linearly polarized and when the output
transmitted intensity is analyzed along different orthogonal
directions. A good agreement between the measurements and the
theoretical predictions deduced from the Jones matrices (see
appendix D) is clearly seen, together with the mirror symmetries
between the two enantiomers. Importantly, these symmetries also
imply that for unpolarized light, and in complete consistency with
Fig.~1, the total intensity transmitted by the structures is
independent of the chosen enantiomer. Furthermore, the conversion
of polarization is well (geometrically) illustrated by using the
Poincar\'{e} sphere representation~\cite{Brehonnet}. Indeed, as
shown in Fig.~3, the Mueller matrix defines a geometrical
transformation which projects the unit Poincar\'{e} sphere, drawn
by the input Stokes vector, on an output closed surface with
typical radius $F\left(\mathcal{M}^{\textrm{exp.}}\right)\simeq 1$
in agreement with the absence of net depolarization as already
noticed (from theory, $F\left(\mathcal{M}^{\textrm{th.}}\right)=
1$ exactly). Data shown on Fig.~2 are also plotted on this sphere.
The input state draws a circle in the equator plane while the
output state (for each enantiomer) draws a circle in a different
plane, which center is not located at the center of the sphere.
This is a direct manifestation of planar chirality (see appendix
E). There is clearly an antisymmetrical behaviour between both
enantiomers.\begin{figure}[h]
\centering\includegraphics[width=8.5cm]{drezet_3.eps}
\caption{Full polarization tomography. Poincar\'{e} sphere of unit
radius associated with the input state represented by the Stokes
vector $\mathbf{X}$~\cite{Hecht,Brehonnet}. Also shown are the
results of Fig.~2 for the left (blue) and right handed (red)
structures if the linearly polarized incident state draw the black
circle in the ($X_1$, $X_2$) equator plane of the input sphere.
Data points are compared with the predictions from
$\mathcal{M_{L,R}}^{\textrm{exp.}}$ (continuous curves) and of
equation (2) (dashed curves).}
\end{figure}
The good agreement between the experiment and the prediction of
equations (1,2) shows
the sensitivity of the polarization tomography method and the high reliability of the FIB fabrication.\\
\indent The degree of optical 2D chirality is quantified by
diagonalizing $\mathcal{J_L}^{\textrm{th.}}$ and
$\mathcal{J_R}^{\textrm{th.}}$. For
$\mathcal{J_L}^{\textrm{th.}}$, the eigenstates are
$|\pm_\mathcal{L}\rangle=\sqrt{B}|L \rangle \pm \sqrt{C}|R\rangle$
associated with the eigenvalues
$\lambda_\mathcal{L}(\pm)=A\pm\sqrt{(B\cdot C)}$. The eigenstates
for $\mathcal{J_R}^{\textrm{th.}}$ are obtained by permutation of
$B$ and $C$ with consequently
$\lambda_\mathcal{L}(\pm)=\lambda_\mathcal{R}(\pm)$. The scalar
product $\langle +_\mathcal{L}|
-_\mathcal{L}\rangle^{\textrm{th.}}=-\langle +_\mathcal{R}|
-_\mathcal{R}\rangle^{\textrm{th.}} =(|B|-|C|)/(|B|+|C|)$ is the
eigenstates Stokes parameter $S_3/S_0$ and provides a direct
measurement of the degree of optical chirality. It also evaluates
losses since the non-orthogonality of these two states is related
to the necessary non-unitarity of the Jones matrix for planar
chirality. We have $\langle +_\mathcal{L}|
-_\mathcal{L}\rangle^{\textrm{fit}}\simeq 0.255$ and $\langle
+_\mathcal{R}| -_\mathcal{R}\rangle^{\textrm{fit}}\simeq -0.277$,
which, within experimental uncertainties, are in good agrement
with the theoretical expectations. As shown in the insets of
Fig.~2, both eigenstates of each structure (e.g.,
$|\pm_{\mathcal{L,R}}\rangle$) can be represented by two ellipses
having the same axis ratio and the same handedness, but rotated
$90^{\circ}$ relative to each other. In agreement with the
theoretical predictions, these polarization ellipses for
$|\pm_{\mathcal{L}}\rangle$ and $|\pm_{\mathcal{R}}\rangle$ are
mirror reflections. This behaviour is significantly different from
the results obtained with optically active
media~\cite{Canfield,Canfield2,Gonokami,Rogacheva,Plum,Decker}
where the eigenstates associated with a given enantiomer have
opposite handedness~\cite{Landau}. This point, which reflects
itself in the symmetry property of chiral Jones matrices, namely
$\mathcal{J}_{\mathcal{L,R}}^{\textrm{th.}}(L,L)=\mathcal{J}_{\mathcal{L,R}}^{\textrm{th.}}(R,R)=A$,
has far reaching consequences, as pointed out in
reference~\cite{Fedotov}. It implies that a 2D plasmonic spiral
mimics a Faraday medium when we reverse the light path and this
even if the system, unlike a true Faraday medium, obeys rigorously
to the principle of reciprocity~\cite{Fedotov, Landau} (inversely,
one can show that equation (1) results from both this requirement
and the absence of mirror symmetry). It means that a photon coming
from the second side will probe a structure of opposite chirality.
After going through the structure and retracing back the light
path with a mirror normal to the axis, the polarization state will
be different at the end of journey from the initial one. This
would be impossible for an optically active medium and is solely
due to planar chirality. To summarize, our results therefore
demonstrate that 2D chirality is possible in the visible domain in
the absence of optical activity and add another element to the
promising plasmonic toolkit.
\section{Appendix A: Polarization tomography setup.}
We apply a procedure similar to the one considered in
\cite{Altewischer,Genet2} in order to record the Mueller matrix: a
collimated laser beam at $\lambda=785$ nm is focussed normally on
the structure by using an objective $L_1$ ($\times 50$, numerical
aperture=0.55). The transmitted light is collected and
recollimated by using a second objective $L_2$ ($\times 40$,
numerical aperture=0.6). The input and output states of
polarization are respectively prepared and analyzed in the
collimated part of the light path by using polarizers, half wave
plates and quarter waveplates. A sketch of the setup is provided
below (see Fig.~4).\\
\begin{figure}[h]
\centering\includegraphics[width=12cm]{supplementary.eps}
\caption{Principle of the polarization experiment. (a), Sketch of
the optical set up described in the text. The images are recorded
by using a CMOS camera. (b), A typical image of the transmitting
nanohole showing the Airy spot associated with diffraction by the
optical microscope. The scale bar is $2$ $\mu$m long. (c),
Crosscut of the intensity profile along the yellow dotted line
shown in (b). }
\end{figure}
The Mueller matrix is built by applying an experimental algorithm
equivalent to the one described in \cite{Brehonnet}. More
precisely, in order to write down the full Mueller matrix, we
measured here $6\times 6$ intensity projections corresponding to
the 6 unit vectors $|x\rangle$, $|y\rangle$,
$|+45^{\circ}\rangle$, $|-45^{\circ}\rangle$, $|L\rangle$, and
$|R\rangle$ for the input and the output polarizations. Actually
only 16 measures are needed to determine $\mathcal{M}$~
\cite{Brehonnet}. Our actual procedure is thus more than
sufficient to obtain $\mathcal{M}$.\\ The isotropy of the setup
was first checked by measuring the Mueller matrix
$\mathcal{M}^{\textrm{glass}}$ with a glass substrate. Up to a
normalization constant, we deduced that
$\mathcal{M}^{\textrm{glass}}$ is practically identical to the
identity matrix $\mathcal{I}$ with individuals elements deviating
by no more than 0.02. More precisely, the optical depolarization
(i.~e, the losses in polarization coherence) can be precisely
quantified through the degree of purity of the Mueller matrix
defined by \cite{Brehonnet}
$F\left(\mathcal{M}\right)=\left(\frac{\textrm{Tr}[\mathcal{M}^{\dagger}\mathcal{M}]-\mathcal{M}_{00}^{2}}{3\mathcal{M}_{00}^{2}}
\right)^{1/2}\leq 1$. Here we measured
$F\left(\mathcal{M}^{\textrm{glass}}\right)=0.9851$. It implies
that the light is not depolarized when going through the setup and
that consequently we can rely on our measurement
procedure for obtaining $\mathcal{M}$.\\
Two important points must be noted here: On the one hand we
varied the incident illumination spot size on the sample between 2
and 20 $\mu$m without affecting the matrix, i.~e., without
introducing additional depolarisation. In the rest of the
experiment on chiral structures we consider the case of a large
gaussian spot with FWHM=20 $\mu$m in order to illuminate the whole
individual spiral. On the other hand, it can be observed that in
our experiments the polarization in the Airy spot (see Fig.~4b) is
homogeneous. This implies that we are actually doing the
polarization tomography of the central transmitting hole, i.~e.,
we are dealing only with the SU(2) point symmetry of the Mueller
Matrix. This situation clearly contrasts with previous SOP
tomography measurements on metallic hole arrays in which the
polarization degrees of freedom were mixed with spatial
information responsible for SPP-induced
depolarization~\cite{Altewischer}.
\section{Appendix B:
Experimental Mueller matrices.} The experimental Mueller matrices
deduced from the polarization tomography are after normalization
of $\mathcal{M}^{\textrm{exp.}}_{00}$:
\begin{eqnarray}
\mathcal{M_L}^{\textrm{exp.}}= \left(\begin{array}{cccc}
1.000& 0.031 & -0.107 &-0.029 \\
0.029 & 0.958 &0.044 & -0.251 \\
-0.105 & 0.037 & 0.953 & 0.287\\
0.029 & 0.261 & -0.282 & 0.809
\end{array}\right),\nonumber \\
\mathcal{M_R}^{\textrm{exp.}}= \left(\begin{array}{cccc}
1.000& 0.035 & 0.111 &0.023 \\
0.027 & 0.949 &-0.051 & 0.246 \\
0.096 & -0.034 & 0.943 & 0.267\\
-0.011 & -0.252 & -0.277 & 0.745
\end{array}\right).
\end{eqnarray}
We have $F\left(\mathcal{M_L}^{\textrm{exp.}}\right)\simeq 0.967$
and $F\left(\mathcal{M_R}^{\textrm{exp.}}\right)\simeq 0.939$.\\
We must also note that the normalization used here neglects a
small additional coefficient of proportionality
$|\mathcal{M_L}^{\textrm{exp.}}_{ij}/\mathcal{M_R}^{\textrm{exp.}}_{ij}|\simeq
0.954$ imputed to experimental errors and uncertainties.\\
We also recorded the Mueller matrix of the set up with the glass
substrate only. Up to a normalization factor we deduced
\begin{eqnarray} \mathcal{M}^{\textrm{glass}}=
\left(\begin{array}{cccc}
\underline{1.0000} & 0.0060 & -0.0040 & -0.0070\\
-0.0030 & \underline{0.9851} & -0.0010 & 0.0020\\
-0.0020 & 0.0020 & \underline{0.9965} & 0.0030\\
-0.0050 & -0.0040 & 0.0030 & \underline{0.9821}\\
\end{array}\right)
\end{eqnarray}
which satisfies $\mathcal{M}^{\textrm{glass}}\simeq \mathcal{I}$
with $\mathcal{I}$ the identity matrix. It implies that the
optical set up do not induce depolarization and that consequently
we can rely on our measurement procedure for obtaining
$\mathcal{M}$.
\section{Appendix C: Theoretical Mueller matrices.}
The precise form of the theoretical Mueller matrice
$\mathcal{M_L}^{\textrm{th.}}$ deduced from equation (1) is
\begin{eqnarray}
\mathcal{M_L}^{\textrm{th.}}=\left(\begin{array}{cccc}
\mathcal{M}^{\textrm{th.}}_{00}& \mathcal{M}^{\textrm{th.}}_{01}& \mathcal{M}^{\textrm{th.}}_{02} &\mathcal{M}^{\textrm{th.}}_{03} \\
\mathcal{M}^{\textrm{th.}}_{01}& \mathcal{M}^{\textrm{th.}}_{11} &\mathcal{M}^{\textrm{th.}}_{12} & \mathcal{M}^{\textrm{th.}}_{13} \\
\mathcal{M}^{\textrm{th.}}_{02} & \mathcal{M}^{\textrm{th.}}_{12}& \mathcal{M}^{\textrm{th.}}_{22} & \mathcal{M}^{\textrm{th.}}_{23}\\
-\mathcal{M}^{\textrm{th.}}_{03} & -\mathcal{M}^{\textrm{th.}}_{13} & -\mathcal{M}^{\textrm{th.}}_{23} & \mathcal{M}^{\textrm{th.}}_{33}
\end{array}\right).
\end{eqnarray}
with $\mathcal{M}^{\textrm{th.}}_{00}=(2|A|^2+|B|^2+|C|^2)/2$,
$\mathcal{M}^{\textrm{th.}}_{01}=Re[BA^{\ast}+AC^{\ast}]$,
$\mathcal{M}^{\textrm{th.}}_{02}=Im[AB^{\ast}+CA^{\ast}]$,
$\mathcal{M}^{\textrm{th.}}_{03}=(|C|^2-|B|^2)/2$,
$\mathcal{M}^{\textrm{th.}}_{11}=|A|^2+Re[B^{\ast}C]$,
$\mathcal{M}^{\textrm{th.}}_{12}=Im[B^{\ast}C]$,
$\mathcal{M}^{\textrm{th.}}_{13}=Re[AC^{\ast}-BA^{\ast}]$,
$\mathcal{M}^{\textrm{th.}}_{22}=|A|^2-Re[B^{\ast}C]$,
$\mathcal{M}^{\textrm{th.}}_{23}=Re[CA^{\ast}-AB^{\ast}]$,
$\mathcal{M}^{\textrm{th.}}_{33}=(2|A|^2-|B|^2-|C|^2)/2$. Similar
formula are obtained for $\mathcal{M_R}^{\textrm{th.}}$ after
permuting $B$ and $C$.\\
From the previous relations we deduce the useful equations (valid
for $\mathcal{M_L}^{\textrm{th.}}$)
\begin{eqnarray}
B/A=\frac{\mathcal{M}^{\textrm{th.}}_{01}-\mathcal{M}^{\textrm{th.}}_{13}}{\mathcal{M}^{\textrm{th.}}_{00}+\mathcal{M}^{\textrm{th.}}_{33}}+
i\frac{\mathcal{M}^{\textrm{th.}}_{23}-\mathcal{M}^{\textrm{th.}}_{02}}{\mathcal{M}^{\textrm{th.}}_{00}+\mathcal{M}^{\textrm{th.}}_{33}}\nonumber\\
C/A=\frac{M^{\textrm{th.}}_{01}+M^{\textrm{th.}}_{13}}{M^{\textrm{th.}}_{00}+M^{\textrm{th.}}_{33}}+
i\frac{\mathcal{M}^{\textrm{th.}}_{23}+\mathcal{M}^{\textrm{th.}}_
{02}}{\mathcal{M}^{\textrm{th.}}_{00}+\mathcal{M}^{\textrm{th.}}_{33}}.
\end{eqnarray} Together with equation (3) equation (6) allow us to fit
$B/A$ and $C/A$ if we replace $\mathcal{M}^{\textrm{th.}}$ by
$\mathcal{M_L}^{\textrm{exp.}}$ (a similar procedure is applicable
to $\mathcal{M_R}^{\textrm{exp.}}$ after permuting $B$ and
$C$).\\
The best fit we obtained (see equation (2)) are:
\begin{eqnarray} \mathcal{M_L}^{\textrm{fit}}=
\left(\begin{array}{cccc}
1.000& 0.033 & -0.116 &-0.023 \\
0.033 & 0.951 &0.043 & -0.282 \\
-0.116 & 0.043 & 0.951 & 0.304\\
0.023 & 0.282 & -0.304 & 0.902
\end{array}\right),\nonumber \\
\mathcal{M_R}^{\textrm{fit}}= \left(\begin{array}{cccc}
1.000& 0.0359 & 0.125 &0.026 \\
0.039 & 0.949 &-0.044 & 0.283 \\
0.125 & -0.044 & 0.948 & 0.311\\
-0.026 & -0.283 & -0.311 & 0.897
\end{array}\right).
\end{eqnarray}
From theory we can deduce that
$F\left(\mathcal{M_{L,R}}^{\textrm{th.}}\right)=1$ (i.e., after
normalization by $\mathcal{M}^{\textrm{th.}}_{00})$. We have thus
$F\left(\mathcal{M_{L,R}}^{\textrm{fit}}\right)=1$
\section{Appendix D: Symmetries due to chirality [interpreting figure 2].}
Let $|\Psi_{\textrm{in}}\rangle=E_x|x\rangle+E_y|y\rangle$ and
$|\Psi_{\textrm{out}}\rangle=E'_x|x\rangle+E'_y|y\rangle$ be
respectively the incident and transmitted electric fields when we
consider the left handed planar chiral structure. We have
\begin{equation}
|\Psi_{\textrm{out}}\rangle=\hat{\mathcal{J_L}}|\Psi_{\textrm{in}}\rangle
\end{equation} where $\hat{\mathcal{J_L}}$ is the operator
associated with the Jones matrix $\mathcal{J_L}$. The mathematical
definition of planar chirality is that whatever the mirror
symmetry operation $\hat{\Pi}$ in the plane X-Y we have
$\hat{\mathcal{J}}\hat{\Pi}-\hat{\Pi}\hat{\mathcal{J}}\neq 0$. It
equivalently states that $\hat{\Pi}
\hat{\mathcal{J}}\hat{\Pi}^{-1}\neq \hat{\mathcal{J}}$. If we
consider for example the mirror reflection through the Y axis (see
Fig.~1) we have the matrix representation (in the cartesian basis)
${\Pi}={\Pi}^{-1}=\left(\begin{array}{cc} -1 &
0
\\ 0& 1\end{array}\right)$ and consequently \begin{eqnarray}\hat{\Pi}
\hat{\mathcal{J_L}}\hat{\Pi}^{-1}= \hat{\mathcal{J_R}}\neq
\hat{\mathcal{J_L}}\end{eqnarray} which agrees with equation (1)
and constitutes an other optical definition of
chirality.\\
The previous equations are used in order to interpret the results
of Fig.~3 of the main article. Indeed from equations (8) and (9)
we obtain
\begin{equation}
\hat{\Pi}|\Psi_{\textrm{out}}\rangle=\hat{\mathcal{J_R}}\hat{\Pi}|\Psi_{\textrm{in}}\rangle.
\end{equation}
The input state considered in Fig.~2 is a linearly polarized light
$|\theta\rangle=\sin{(\theta)}|x\rangle+\cos{(\theta)}|y\rangle$
(the angle is measured relatively to the Y axis) and the
transmitted intensity projected along a direction of analysis
$|i\rangle$ (i.e, $|x\rangle$, $|y\rangle$, $|+45^{\circ}\rangle$,
$|-45^{\circ}\rangle$, $|L\rangle$, and $|R\rangle$) is written
$I_{i}^{(\textrm{Left})}(\theta)=|\langle
i|\Psi_{\textrm{out}}\rangle|^{2}=|\langle
i|\hat{\mathcal{J_L}}|\theta\rangle|^{2}$. Similarly we also write
$I_{i}^{(\textrm{Right})}(\theta)=|\langle
i|\hat{\mathcal{J_R}}|\theta\rangle|^{2}$. From equation (10) we
deduce:
\begin{equation}
\langle i'|\hat{\mathcal{J_L}}|\theta\rangle=\langle
i|\hat{\mathcal{J_R}}|-\theta\rangle,
\end{equation} where we used $|i'\rangle=\hat{\Pi}^{-1}|i \rangle=\hat{\Pi}|i \rangle$ and
$|-\theta\rangle=\hat{\Pi}|\theta \rangle$. We consequently have:
\begin{eqnarray}
I_{\textrm{total}}^{(\textrm{Left})}(\theta)=I_{\textrm{total}}^{(\textrm{Right})}(-\theta),\nonumber \\
I_{x,y}^{(\textrm{Left})}(\theta)=I_{x,y}^{(\textrm{Right})}(-\theta),\nonumber \\
I_{\pm 45^{\circ}}^{(\textrm{Left})}(\theta)=I_{\mp
45^{\circ}}^{(\textrm{Right})}(-\theta),\nonumber \\
I_{L,R}^{(\textrm{Left})}(\theta)=I_{R,L}^{(\textrm{Right})}(-\theta).
\end{eqnarray}
Such symmetries are clearly visible in Fig.~2 and correspond to a
direct signature of optical chirality in the planar systems
considered.
\section{Appendix E: Planar chirality on the Poincar\'{e} Sphere [interpreting figure 3]}
We remind that the Stokes parameters associated with a
polarization state of light $|\Psi\rangle$ are defined by
\begin{eqnarray}
S_0=I_x+I_y, &
S_1=I_x-I_y\nonumber\\
S_2=I_{+45^{\circ}}-I_{-45^{\circ}}, &
S_3=I_{\textrm{L}}-I_{\textrm{R}},
\end{eqnarray}
where $I_i$ are projection measurement along the direction $i$,
i.e, $I_{i}=|\langle i|\Psi\rangle|^{2}$. The Stokes vector
$\mathbf{X}$ is a convenient representation of such a state. We
have $\mathbf{X}=X_1\mathbf{x}_1+X_2\mathbf{x}_2+X_3\mathbf{x}_3$
with $X_1=S_1/S_0$, $X_2=S_2/S_0$, $X_3=S_3/S_0$ and with
($\mathbf{x}_1$, $\mathbf{x}_2$, $\mathbf{x}_3$) a cartesian
orthogonal and normalized vector basis.\\
The coherent input state satisfies the
normalization~\cite{Brehonnet} $|\mathbf{X}|=1$, that is the
vector draw a Poincar\'{e} sphere of unit radius in the space
$X_1,X_2,X_3$. The transmitted output state after interaction with
the left or right handed structure is defined by the relation
\begin{eqnarray}
\left(\begin{array}{c}
S_{\mathcal{L,R};0}\\
S_{\mathcal{L,R};1}\\
S_{\mathcal{L,R};2} \\
S_{\mathcal{L,R};3}
\end{array}\right)=\mathcal{M_{L,R}}\left(\begin{array}{c}
S_0\\
S_1 \\
S_2 \\
S_3
\end{array}\right).
\end{eqnarray}
The output state defines a Stokes vector
$\mathbf{X}_{\mathcal{L,R}}$ such that
$|\mathbf{X}_{\mathcal{L,R}}|\leq 1$. A typical value for
this radius is given by $F\left(\mathcal{M_{L,R}}\right)$.\\
If the input state is linearly polarized the input Stokes vector
is:
\begin{eqnarray}
\mathbf{X}^{\textrm{in}}(\theta)= \left(\begin{array}{c}
\cos{(2\theta)} \\
\sin{(2\theta)}\\
0
\end{array}\right),
\end{eqnarray}
and draw a circle $(\sum_\textrm{in})$ along the equator contained
in the plane $X_1,X_2$ of the unit radius Poincar\'{e} sphere.
Using equation (14) the output Stokes vector is now a function of
$\theta$: $\mathbf{X}_{\mathcal{L,R}}(\theta)$ drawing a closed
curve $(\sum_\mathcal{L,R})$ (see Fig.~3) which is the image,
through the Mueller matrix transformation, of the equator circle
$(\sum_\textrm{in})$ above mentioned. Importantly, since the
Mueller matrix $\mathcal{M}$ given by equation (5) represents a
linear relation connecting $\mathbf{X}^{\textrm{in}}$ to
$\mathbf{X}^{\textrm{out}}$, we conclude that the image of the
incident polarization state contained in the equator plane
$X_1,X_2$ through $\mathcal{M}$ must also be contained in a plane
in the space $X_1,X_2,X_3$.\\To analyze this point more in details
we consider the normalized Vector product
\begin{eqnarray}
\mathbf{n}_{\mathcal{L,R}}=\frac{(\mathbf{X}_{\mathcal{L,R}}(0)-\mathbf{X}_{\mathcal{L,R}}(2\pi/3))\times(\mathbf{X}_{\mathcal{L,R}}(0)-\mathbf{X}_{\mathcal{L,R}}(\pi/2))}{|(\mathbf{X}_{\mathcal{L,R}}(0)-\mathbf{X}_{\mathcal{L,R}}(2\pi/3))\times(\mathbf{X}_{\mathcal{L,R}}(0)-\mathbf{X}_{\mathcal{L,R}}(\pi/2))|}
\end{eqnarray} and we write it
\begin{eqnarray}
\mathbf{n}_{\mathcal{L,R}}=\left(\begin{array}{c}
U_{\mathcal{L,R}}\\
V_{\mathcal{L,R}}\\
W_{\mathcal{L,R}}
\end{array}\right),
\end{eqnarray}
with
$|U_{\mathcal{L,R}}|^2+|V_{\mathcal{L,R}}|^2+|W_{\mathcal{L,R}}|^2=1$.
It represents a typical normal to the closed curve
$(\sum_\mathcal{L,R})$. We have
\begin{eqnarray}
\mathbf{n}_{\mathcal{L}}=\left(\begin{array}{c}
0.2845 \\
-0.3065\\
-0.9084
\end{array}\right), \mathbf{n}_{\mathcal{R}}=\left(\begin{array}{c}
0.2861 \\
0.3139\\
0.95053
\end{array}\right).
\end{eqnarray}
Actually, if each curve $(\sum_\mathcal{L,R})$ is contained in a
(different) plane $P_{\mathcal{L,R}}$ we must have
\begin{equation} \mathbf{n}_{\mathcal{L,R}}\cdot
(\mathbf{X}_{\mathcal{L,R}}(\theta)-\mathbf{X}_{\mathcal{L,R}}(0))=0\end{equation}
for every $\theta$. This was indeed checked numerically up to a
precision of $10^{-11}$. It was also checked that
$|\mathbf{X}_{\mathcal{L,R}}(\theta)|=1$ up to the same precision.
This proves that each curve $(\sum_\mathcal{L,R})$ must be a
circle. The equations of the two planes $P_{\mathcal{L,R}}$ are
given by $\mathbf{n}_{\mathcal{L,R}}\cdot
(\mathbf{X}-\mathbf{X}_{\mathcal{L,R}}(0))=0$ where $\mathbf{X}$
is the Stokes vector associated with a running point belonging to
each plane. We write
\begin{eqnarray}
U_{\mathcal{L,R}}X_{1}+V_{\mathcal{L,R}}X_{2}+W_{\mathcal{L,R}}X_{3}+D_{\mathcal{L,R}}=0
\end{eqnarray}
with $D_{\mathcal{L}}=-0.0237$ and $D_{\mathcal{R}}=-0.0266$.
$|D_{\mathcal{L},R}|$ represents the distance separating the
center of the circle $(\sum_\mathcal{L,R})$ to the origin of the
poincar\'{e} sphere. This proves that the planes are not going
through the center of the sphere. It was checked after lengthy
calculations that if $|B|=|C|$ in the Jones matrix (see equation
(1)) then $D=0$. This shows that the property
$|D_{\mathcal{L},R}|\neq 0$ is a characteristic of planar
chirality (i.e, the condition $|B|\neq|C|$). The radius of each
circle $(\sum_\mathcal{L,R})$ is given by
$r_{\mathcal{L,R}}=\sqrt{(1-D_{\mathcal{L},R}^2)}$ and we have
$r_{\mathcal{L}}=0.9997$ and $r_{\mathcal{L}}=0.9996$ which are
slightly smaller than $r=1$ in agreement with the fact that
$P_{\mathcal{L,R}}$ are not going through the center of the
sphere.
|
\section{Introduction}
\label{sec:intro}
Bars are believed to be very important with regard to the dynamical
and secular evolution of disk galaxies. In particular, they represent
the main internal driver of galaxy structure and morphology evolution
within the central $\sim$10 kpc. Stellar bars are also recognized as
the most important internal factor that redistribute the angular
momentum between the baryonic and dark matter components
\citep{debattistasellwood98,debattistasellwood00}. The amount of
angular momentum exchanged is related to specific properties of the
galaxies, such as the bar mass, halo density, and halo velocity
dispersion \citep{athanassoula03,sellwooddebattista06}. Moreover, they
funnel material towards the galaxy center where starbursts can ignite
\citep{sheth05}, contribute to the formation of bulge-like
structures \citep[][]{kormendykennicutt04}, inner star-forming rings
\citep[][]{buta03,munoztunon04}, inner bars \citep[][]{erwin04,debattistashen07}, and
feed the central black hole \citep[][]{shlosman00,corsini03}.
Peanut/boxy bulges in galaxies are also thought to be associated with
bending instabilities and bar vertical resonances
\citep{bureaufreeman99,martinezvalpuesta06,mendezabreu08}
Stellar bars are observed in optical images of roughly half of all the
nearby disk galaxies \citep{barazza08,aguerri09}. This fraction rises
slightly to about 59-62\% when near-infrared images are analysed
\citep{laurikainen04,marinovajogee07, menendezdelmestre07}. It is
established that they appear naturally in most simulations of galaxy
formation once a dynamically cold and rotationally-supported disk is
at place. However, even if bars are ubiquitous in the universe, it is
not clear yet why one galaxy can exhibit a bar structure while another
apparently similar does not.
The mechanisms leading to the formation of bars can be divided into
internals and externals. The most widely accepted internal mechanism
to produce bars in galaxies is based on the $m=2$ mode global
instability in cold, rotationally supported disks
\citep{hohl71,ostrikerpeebles73}. Environmental effects could also be
important, though the existence of competing mechanisms usually
prevents simple interpretations. Depending on the orbital
configuration, interactions can weaken severely bars and even destroy
them or, on the other hand, significantly speed up bar formation
\citep{noguchi87,aguerrigonzalezgarcia09}. Moreover, numerical
simulations of galaxy harassment in clusters have shown that bar
growth is a common and stable process during the evolution of
late-type galaxies in dense environments
\citep[e.g.][]{mastropietro05}.
Only with the recent advent of large galaxy surveys, either at high
\citep[][COSMOS]{sheth08} or low redshift
\citep[][SDSS]{barazza08,aguerri09}, statistical studies of bar
frequencies have been possible. However, bar studies have been
usually restricted to luminous galaxies due to either the lack of
spatial resolution or because images were not deep enough. The present
work attempts to put observational constraints on the internal (mass)
and external (environment) parameters that influence bar formation by
carrying out a comprehensive study of the bar fraction in the Coma
cluster galaxies throughout a wide range of 9 magnitudes, covering
from giant ellipticals (M$_{r} \sim -23$) to dwarf galaxies (M$_{r}
\sim -14$). This research will also provide us the luminosity/mass
interval where cold stellar disks are present in galaxies. For this
purpose we take advantage of the unrivalled resolution of the HST-ACS
Coma cluster Treasury Survey, which provides deep imaging of the core
and infall region of the Coma cluster.
The paper is organised as follows. The galaxy sample, as well as the
selection of the Coma cluster member galaxies is presented in Sect.
\ref{sec:data}; the method adopted to detect bar structures in the
sample galaxies, and the results obtained are explained in Sect.
\ref{sec:results}; the discussion of the results and our conclusions
are given in Sect. \ref{sec:conclusions}. Throughout the paper, we
assume a distance modulus of m-M=35.
\section{Data and cluster membership selection}
\label{sec:data}
The Coma cluster is one of the best-studied galaxy clusters because of
its relative proximity and because it may be a prototype of
dynamically relaxed cluster, even though significant substructures are
present \citep{collessdunn96}.
The HST-ACS Coma Cluster Treasury Survey \citep{carter08}, covers
$\sim$ 230 arcmin$^2$ with 21 ACS pointings ($\sim$3$\times$3 arcmin).
The magnitude limits of the survey at 10 $\sigma$ for 1 arcsec$^2$
extended regions are g'=25.8 mag/arcsec$^2$ and $I_{\rm C}$=25.0
mag/arcsec$^2$. At the distance of the Coma cluster ($\sim$ 100 Mpc),
the resolution of HST-ACS ($0\farcs1$) corresponds to $\sim 50$ pc.
This gives essentially the same physical resolution as ground-based
observations have in Virgo and it will allow us to resolve bars down
to sizes of $r_{\rm bar}\sim 150$ pc.
The Coma cluster is also covered by the SDSS, providing galaxy
magnitudes in five bands ($u, g ,r, i, z$). For the sake of comparison
with recent works on bar fractions \citep{barazza08,aguerri09}, and in
order to have access to galaxy colors, which will help us to determine
cluster memberships, we decided to create our catalogue of sources
using the SDSS-DR6 \citep{adelman08}.
The steps followed to obtain our final sample of Coma cluster members
were the following: from the SDSS-DR6 we downloaded a catalogue of
extended sources within a 5 arcmin radius from the position of every
ACS pointing. This catalogue contains all galaxies with m$_{r}<21$,
which represent approximately the completeness limit of the SDSS
photometric survey for extended sources, and with $b/a>0.5$, $a$ and
$b$ being the semi-major and semi-minor axis lengths of the galaxies,
in order to deal with projection effects. This resulted in a sample
of 477 galaxies (black circles in Fig. \ref{fig:cm}), 104 of them
having recession velocities available from the Nasa Extragalactic
Database (NED). We select galaxies with velocities $\pm 3000$ km/s
with respect to the Coma redshift as cluster members. This range
corresponds to a 3$\sigma$ cut on the velocity distribution of the
Coma cluster galaxies ($\sigma \sim 1000$ km/s; Colless \& Dunn 1996).
We found that all 104 galaxies satisfy this condition and therefore
they are cluster members. At this point, we decided to visually
inspect every galaxy in order to determine its possible cluster
membership based on its morphology. We follow the prescriptions given
by \citet{michardandreon08} where they claim that morphologically
selected cluster members are reliable when compared with redshift
membership, and present a catalogue of 473 Coma members down to
M$_{B}$=-14.25 based on their morphology. The procedure to distinguish
between background objects and cluster members is based in some basic
hypothesis. Typical spirals in the background lie in the magnitude
range of dwarf galaxies in Coma. However, these objects are rare in
the field and even more in dense cluster environments. Moreover, we
exclude from our sample all spiral like galaxies that had either bulge
or disk sizes too small if they were cluster members. With regard to
elliptical galaxies, far-away bright ellipticals are generally more
concentrated that early-type dwarfs in Coma. In addition, they are
more reddened and our further color cut will easily get rid of them.
We found that, from the remaining 373 galaxies without redshift, 127
galaxies followed the morphological criteria to be cluster members,
while 246 did not. A further colour condition was still imposed for
these candidates to be considered as cluster members. To calculate our
color cut, we fit the red sequence of all galaxies in our sample and
imposed that members should have a $g-r$ color less than 0.2
magnitudes above the value of the fit (see Fig. \ref{fig:cm}). With
this further constraint, our sample of Coma {\it secure} members
consists of 188 galaxies (red points in Fig. \ref{fig:cm}) with
magnitudes in the range $-23 <$ M$_r < -14$.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\textwidth]{colormagnitude.ps}
\caption{Color-magnitude diagram for our complete sample of 477
galaxies (black circles). Red points represent the subsample of 188
galaxies considered as members of the Coma cluster, while blue
diamonds point out those Coma galaxies which host a bar. The solid
and dotted black line shows our fit to the red sequence and our
limit of 0.2 magnitudes above the fitted red sequence,
respectively.}
\label{fig:cm}%
\end{figure*}
Several quality checks were done during the morphological membership
classification: at a first step the classification was performed
independently by two of the authors (JMA and RSJ), obtaining
consistent results. Next we compared our cluster memberships with
those given by \citet{michardandreon08}. From the 225 galaxies in
common, both classifications agree for a 78\% of galaxies. In
addition, including colors at this step instead of only as a final
constraint, this percentage grows up to a 85\%, confirming the good
agreement between both classifications. As a final check, we also
compared our morphological membership with the available spectroscopic
redshifts. We found that from the 104 galaxies with redshift, 103
(99\%) share the same class, giving validity again to our
classification.
\section{Method and results}
\label{sec:results}
The presence of a bar can be revealed by visual inspection of the
images \citep{devaucouleurs91}, by analysing the shape and orientation
of the galaxy isophotes
\citep[][]{laurikainen05,menendezdelmestre07,marinovajogee07,sheth08,
barazza08,aguerri09,marinova09}, by studying the Fourier modes of
the light distribution
\citep[][]{ohta90,elmegreenelmegreen85,aguerri98,aguerri00,laurikainen05},
or fitting the different structural components to the surface
brightness distribution \citep[][]{weinzirl09}.
In the present work, we visually classified all galaxies into strong
barred, weakly barred, and unbarred. A caveat regarding this criteria
is that our distinction between strong and weak bars is not directly
related to the contribution of the bar to the total galaxy potential,
but rather they refer to a secure or possible detection of a bar,
respectively. Therefore, the fraction of weak bars could be
understood, in some way, as a measure of our uncertainty in bar
detection. The visual classification was carried out by two of us (JMA
and RSJ) using the redder available filter (F814W) of the HST-ACS
images. Both classifications were in close agreement and only their
mean is reported in the following (see also Fig.
\ref{fig:fractions}). Since the goal of this paper is to understand
where do bars form, we have not defined the bar fraction in the
classic way, i.e., using only disk galaxies, but we have instead used
all galaxies independently of their Hubble type. Using this
definition, our bar fraction turns to be $\sim 9\%$ and $\sim 14\%$
depending if only strong or also weak bars are included, respectively.
Fig. \ref{fig:fractions} shows the bar fraction as a function of the
luminosity and mass of the {\it secure} sample of galaxies in the Coma
cluster. It is clear that independently of the bar strength, bars are
hosted by galaxies in a tight range of luminosities or masses. There
are no strong bars in the Coma cluster out of the luminosity range
between $-21 \lesssim $ M$_{r} \lesssim -18$. These limits become $-22
\lesssim $ M$_{r} \lesssim -17$ if we include also weak bars.
The mass of the galaxies is one of the fundamental parameters
controlling their evolution. We use the prescriptions given by
\citet{bell03} to derive the stellar mass of our galaxies using the
$g-r$ color and the {\it diet} Salpeter initial mass function (IMF).
We find again the same behavior in the bar fraction when using the
galaxy mass: bars exist only in a range of masses between either
$10^{9.5} \lesssim {\cal M_{*}}/{\cal M}_{\sun} \lesssim 10^{11} $ or
$10^{9} \lesssim {\cal M_{*}}/{\cal M}_{\sun} \lesssim 10^{11}$
depending on whether only strong or strong+weak bars are considered,
respectively.
These findings rely on a relatively small number of galaxies and might
be affected by statistical errors. Nevertheless, similar results were
found when comparing the bright/massive side of our bar fraction
distribution with that obtained, also by visual inspection of the
galaxy images, in the large sample of field galaxies studied by
\citet{aguerri09}, indicating that bars are not hosted by very bright
galaxies (see Fig. \ref{fig:fractions}).
\begin{figure*}[!ht]
\centering
\includegraphics[width=0.49\textwidth]{barfrac_lum.ps}
\includegraphics[width=0.49\textwidth]{barfrac_mass.ps}
\caption{Optical bar fraction of strong (solid black line) and
weak+strong (dashed blue line) as a function of the galaxy
absolute magnitude in $r$-band (left panel), and galaxy mass
(right panel). Red points and black circles represent the strong
and weak+strong bar fraction using the field sample of
\citet{aguerri09}, respectively. The number of galaxies per bin is
represented with black triangles. The disks luminosity
distribution by \citet{michardandreon08} is shown with brown
diamonds, the $r$-band magnitudes were derived from the $B$
magnitudes by using the $B-R$ color for every galaxy, and then
calculating the $R$-$r$ differences by taking into account the
equations given by Lupton (2005). Stellar masses were derived
following the prescriptions given by \citet{bell03} using the
$B-R$ color and the {\it diet} Salpeter IMF. }
\label{fig:fractions}%
\end{figure*}
\section{Discussion and conclusions}
\label{sec:conclusions}
We have used HST-ACS images taken in the F814W filter to search for
bars in a sample of 188 galaxies members of the Coma cluster. Bars
were identified based on visual inspection of the images. The
unprecedented spatial resolution provided by HST-ACS in this region
has allowed us to compute the bar fraction throughout a large range of
9 magnitudes, permitting us to explore the presence of bars in the
poorly known region of dwarf galaxies.
We find that bars are not hosted by galaxies in the whole range of
luminosities nor masses covered in this study. On the contrary, they
appear to be well constrained in a tight interval of both luminosities
$-22 \lesssim $ M$_{r} \lesssim -17$ and masses $10^{9} \lesssim {\cal
M_{*}}/{\cal M}_{\sun} \lesssim10^{11}$. This result has several
implications on our current understanding of bar formation and/or
evolution.
If we assume that bars are tracers of cold stellar disks, the presence
of bars is particularly useful to identify galaxies with disks in
clusters \citep{marinova09}. Therefore, it could be logical to think
that the distribution of bar fraction with the galaxy magnitude should
trace the shape of the disk galaxies luminosity distribution.
\citet{binggeli88} showed in his seminal paper the magnitude
distribution of morphological classes for the Virgo cluster. In the
bright side of the distribution ellipticals are the dominant type,
while both dwarf ellipticals and dwarf irregulars are the most common
morphological type in the low luminosity region. Instead, disk
galaxies appear to be constrained in the range $-22 \lesssim $ M$_{B}
\lesssim -16$ in accordance with our results. In order to test this
hypothesis in the Coma cluster, we have computed the luminosity
distribution of morphologically selected disk galaxies (from S0 to Sc)
using the classification carried out by \citet{michardandreon08}. The
resulting luminosity distribution (brown diamonds in Fig.
\ref{fig:fractions}) matches well the shape of the bar fraction
distribution in both luminosity and mass, confirming that bars are
good tracers of disks.
Since simulations suggest that bars form spontaneously in cold and
rotationally-supported disks, the non existence of bars in galaxies
with M$_{r} \lesssim -22$ can be interpreted either as evidence that
cold disks do not form in such massive galaxies, or existing disks are
somewhat heated and they are not able to form/host a bar.
Disk heating should be predominant in high density environments due to
the higher frequency of close encounters, major and minor accretions,
and the presence of tidal forces, however, no differences were found
between field and cluster in the limiting magnitude of galaxies
hosting bars. This constancy could otherwise be interpreted as a
physical limit in the formation of disk galaxies. Numerical
simulations carried out by \citet{dekel06} and \citet{cattaneo06}
suggest that the physics of the gas, which will form galaxy disks,
depends on the galaxy mass. They claim that in halos below a critical
shock-heating mass ${\cal M}_{\rm halo} \le 10^{12} {\cal M}_{\sun}$
disks are built by cold gas streams while for ${\cal M}_{\rm halo} >
10^{12} {\cal M}_{\sun}$ the gas is heated by a virial shock and do
not form disks. This ${\cal M}_{\rm halo} \sim 10^{12} {\cal
M}_{\sun}$ corresponds to a stellar mass of ${\cal M_{*}} \sim
3\times10^{10} {\cal M}_{\sun}$ which roughly coincides with our
limiting mass for galaxies hosting bars.
From the environment perspective external triggers, such as tidal
interactions, can induce bars \citep{noguchi87} but their effects can
be contradictory since they may also heat the disks and thereby make
them less susceptible to bar formation. It is also important to
consider whether our results could be explained in terms of the
evolution and destruction of bars rather than their initial
formation. The scenario in which bars can be dissolved by the presence
of a massive central concentration have been proposed in several
simulations. However, most of them indicate that present-day
supermassive black holes, star clusters or inner parts of bulges are
not massive enough to affect bars significantly
\citep[e.g.,][]{shensellwood04, athanassoula05}. On the other hand,
if bars are hard to be dissolved once they are formed, the presence of
a high density environment such as a cluster should not change
dramatically the bar fraction, as we found in this work when comparing
with the field.
Since our sample galaxies cover both the center and infall regions of
the Coma cluster, we have further tested this issue by dividing our
sample into internal and external galaxies and calculating the bar
fraction for every subsample. As we are introducing large errors due
to small number statistics (especially in the outer regions), we have
repeated this procedure for three values of the separation distance
(0.5, 1, and 1.5 Mpc) from the cluster center ($\alpha$: 12$^{h}$
59$^{m}$ 42$^{s}$, $\delta$: 27$^{\circ}$ 58$^{'}$ 15\farcs6, Godwin
et al. 1983). For the smaller separation distance we found 14\% and
15\% of bars for the internal and external subsamples, respectively.
The bar fractions of both subsamples are 14\% and 17\% when using 1
Mpc, and 14\% and 17\% if we use 1.5 Mpc. Therefore we did not find
differences in the bar fraction between the subsamples for any
separation distance, implying again that the cluster environment plays
a second order role in bar formation/evolution.
In the low luminosity/mass side of the bar fraction distribution we
found also a lack of bars for galaxies with either M$_{r}\gtrsim -17$
or ${\cal M_{*}}/{\cal M}_{\sun} \lesssim 10^{9}$. Few works have
tried to investigate the dwarf realm in order to look for the presence
of bars. \citet{graham03} found spiral structure in two dwarf
galaxies belonging to the Coma cluster. The galaxies magnitudes were
found to be -18.8 and -17.4 in the $R$-band, therefore being in
agreement with our results. \citet{lisker06}, studying a sample of
dwarf galaxies in the Virgo cluster, found the presence of bar
structure in some dwarf galaxies as faint as M$_{B}\sim -16.10$, again
in full agreement with our findings considering a typical $B-r$ color
$\sim1.8$. They also claim that dwarf ellipticals with and without
disks represent two distinct types of galaxies, and show how the
fraction of dwarfs with disk decrease dramatically for galaxies below
M$_{B}\sim -16$. Therefore, even if the non presence of bars in our
low luminosity galaxies could be due to the heating of the disk
component or to its absence, our results support that of
\citet{lisker06} and we suggest that no disks are present in Coma
galaxies below M$_{r}\sim -17$.
The physical mechanisms involved in this case could be different with
respect to that of massive galaxies. The role played by the
environment in the evolution of dwarfs is crucial. For instance,
repeated tidal shocks suffered by a dwarf satellite galaxy can also
remove the kinematic disk signature \citep{mayer01}. Cumulative tidal
fast encounters between galaxies and with the gravitational potential
of the galaxy cluster can produce a dramatic morphological
transformation from spirals to roundish dwarf galaxies
\citep{moore96,mastropietro05,aguerrigonzalezgarcia09}.
We are still far from understanding the mechanisms that drive one
galaxy to host a bar while another apparently similar does not. Even
if a great advance has been done in the last years, we still need to
explore the bunch of observational data already available in order to
provide more inputs to numerical simulations to understand the
formation and evolutionary processes of these important structures at
the center of galaxies.
\acknowledgments
JMA is partially funded by the Spanish MICINN under the
Consolider-Ingenio 2010 Program grant CSD2006-00070: First Science
with the GTC (http://www.iac.es/consolider-ingenio-gtc). JMA and JALA
are partially funded by the project AYA2007-67965-C03-01. We thank V.
Debattista, A. de Lorenzo-Caceres, and I. Martinez-Valpuesta for
useful discussions and suggestions. We also thank the referee,
E. Laurikainen, for constructive comments. Based on observations with
the NASA/ESA Hubble Space Telescope obtained at the STScI, which is
operated by the association of Universities for Research in Astronomy,
Inc., under NASA contractNAS 5-26555. These observations are
associated with program GO10861.
|
\section{Introduction}
\label{introduction}
\indent Nuclear reactions are extensively studied theoretically
and in laboratory experiments. They are important for modelling
many physical phenomena occurring in astrophysical environments,
such as energy production and nucleosynthesis in stars at all
stages of their evolution (e.g., Refs.\
\cite{bbfh57,fh64,clayton83}). Nuclear reactions are important for
the structure and evolution of main sequence and red giant stars,
during the late pre-supernova phase of massive stars and during
core-collapse supernova explosions.
Moreover, nuclear burning is also important in high-density stellar
environments of white dwarfs and neutron stars which are compact
stars at the final stage of their evolutionary development
\cite{st83}. The burning drives nuclear explosions in surface layers
of accreting white dwarfs (nova events), in cores of massive
accreting white dwarfs (type Ia supernovae)
\cite{NiWo97,hoeflich06}, and in surface layers of accreting neutron
stars (type I X-ray bursts and superbursts; e.g., Refs.\
\cite{sb06,schatz03,cummingetal05,Brown06}). The nova events and
type I X-ray bursts are mostly produced by the burning of hydrogen
in the thermonuclear regime, without any strong effect of plasma
screening on Coulomb tunneling of the reacting nuclei. Type Ia
supernovae and superbursts are driven by the burning of carbon,
oxygen, and heavier elements
(e.g., Refs.\ \cite{sb06,schatz03,cummingetal05,Brown06}) at high densities,
where the plasma screening effect can be substantial. It is likely
that pycnonuclear burning of neutron-rich nuclei (e.g.,
$^{34}$Ne+$^{34}$Ne) in the inner crust of accreting neutron stars
in X-ray transients \cite{hz90,hz03,Brown06} (in binaries with
low-mass companions) provides an internal heat source for these
stars. If so, it powers \cite{Brown98} thermal surface X-ray
emission of neutron stars observed in quiescent states of X-ray
transients (see, e.g., Refs.\ \cite{Brown06,pgw06,lh07}).
Nuclear fusion occurs in a wide range of temperatures and densities
of stellar matter (at densities $\rho \;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; 10^{10}$ g~cm$^{-3}$
in white dwarfs and at $\rho \;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; 10^{13}$ g~cm$^{-3}$ in
neutron stars). It can proceed in five regimes as described by
Salpeter and Van Horn \cite{svh69} (also see
\cite{gas2005,yak2006,cdi07,cd09} and references therein). These are
two thermonuclear regimes (with weak and strong plasma screening),
two pycnonuclear regimes (due to zero-point vibrations of atomic
nuclei in crystalline lattice, with and without thermal effects in
nucleus motion), and the intermediate thermo-pycnonuclear regime.
The burning may involve many different, stable as well as very
neutron-rich and unstable nuclei. Neutron-rich nuclei are especially
important in deep neutron star crust, where they can be stabilized
against beta-decay by the presence of the Fermi sea of highly
energetic degenerate electrons \cite{st83,hpy07}.
The nuclear fusion rates depend directly on the reaction cross
section which can be expressed in terms of the astrophysical factor
$S(E)$, with $E$ as the center-of-mass energy of the interacting
nuclei. Stellar burning often proceeds at low temperatures with
sub-picobarn reaction cross sections that are usually not accessible
by direct measurements in laboratory experiments. Therefore, it is
necessary to calculate $S(E)$ for the reactions of interest by a
reliable theoretical method in the important energy range and
approximate the data by analytical expressions for rapid conversion
into thermonuclear or pycnonuclear reaction rates as outlined in our
previous work \cite{yak2006}. In this paper, we calculate (Sec.\
\ref{theory}) the astrophysical factors $S(E)$ for 946 fusion
reactions involving different isotopes of C, N, O, Ne, and Mg
(between the valley of stability and the neutron drip line) using
the barrier penetration model and the S\~ao Paulo
potential~\cite{saoPauloTool}. We approximate our results (Sec.\
\ref{fits}) by analytic expressions convenient for real time
applications. A numerical example is discussed in Sec.\
\ref{example}; calculation of the reaction rates is outlined in
Sec.\ \ref{s:rates}, and we conclude in Sec.\ \ref{s:concl}.
\section{Calculation of astrophysical S-factors}
\label{theory}
\indent The cross section $\sigma(E)$ for a fusion reaction of two
nuclei,
\begin{equation}
(A_1,Z_1) + (A_2,Z_2),
\label{react}
\end{equation}
at the center-of-mass energy $E$ can be expressed in terms of the
astrophysical factor $S(E)$ by
\begin{equation}
\sigma(E) = {1\over E}\, \exp(-2 \pi \eta)\,S(E),
\label{sigma}
\end{equation}
where $\eta= Z_1Z_2e^2/(\hbar v)$ is the Sommerfeld parameter,
$v=\sqrt{2E/ \mu}$ the relative velocity of reacting nuclei at large
separations, and $\mu$ the reduced mass. The factor $\exp(-2 \pi
\eta)$ comes from the probability of penetration through the Coulomb
barrier $V_C(r)=Z_1Z_2e^2/r$ with zero orbital angular momentum,
$\ell=0$, assuming that the barrier extends to $r \to 0$ (for
point-like nuclei); $1/ E$ factorizes out the well-known
pre-exponential low-energy dependence of $\sigma(E)$. The advantage
of this approach is that $S(E)$ is a much more slowly varying
function of $E$ than $\exp(-2 \pi \eta)$ and $\sigma(E)$. It is
easier to extrapolate $S(E)$ to low energies $E$ of astrophysical
interest, than $\sigma(E)$.
We calculate $S(E)$ for a number of fusion reactions using the S\~ao
Paulo potential in the context of the barrier penetration model as
outlined in previous studies \cite{gas2005,yak2006,saoPauloTool}.
The S\~ao Paulo potential is a parameter-free model for the real
part of the nuclear interaction. Underpinning this potential are the
contributions from nonlocality due to quantum effects arising from
the Pauli exclusion principle
\cite{chamon2002,candido1997,chamon1997,chamon1998}. The
interaction takes into account the exchange of nucleons between
reacting nuclei. The potential has been extensively tested in the
description of different nuclear interaction processes (such as
quasi-elastic scattering and fusion reactions), over a broad
spectrum of energies
\cite{gas2004,chamon2004,chamon2002,alvarez2003,alvarez1999,sil2001,
gas2002,rossi2002,gas2003,gascham2003,saoPauloTool}. In the context
of the S\~ao Paulo potential, the real part of the nuclear
interaction is associated with the folding potential through the
relation
\begin{equation}
V_\mathit{SP}(r,E)=V_F(r)\, \exp( - 4 \mathrm{v}^2/c^2),
\label{sp}
\end{equation}
where $r$ is a distance between the centers of the reactants, and
$c$ is the speed of light. The exponential term houses the effects
of Pauli nonlocality. The relative velocity $\mathrm{v}(r,E)$ of the
nuclei in the given model at a separation $r$ is defined as
\begin{equation}
\mathrm{v}^2(r,E) = (2/\mu) \,
\left[ E-V_C(r)-V_\mathit{SP}(r,E) \right] \;,
\label{speed}
\end{equation}
where $V_C(r)$ is the Coulomb potential.
The folding potential used in Eq.\ (\ref{sp}) has a field strength
of $V_0 = -456$ Mev fm$^3$; it is given by
\begin{equation}
V_F(r) = \int \rho_1(\bm{r}_1) \; \rho_2(\bm{r}_2) \;
V_{0} \; \delta(\bm{r}-\bm{r}_{1}+\bm{r}_{2}) \; d\bm{r}_1 \; d\bm{r}_2 .
\label{fold}
\end{equation}
It is convoluted over the
nuclear
matter densities $\rho_1(\bm{r}_1)$ and $\rho_2(\bm{r}_2)$
of the nuclei involved in the reaction. This technique is called the
zero-range approach for the folding potential. It is equivalent
\cite{chamon2002} to adopting the M3Y effective nucleon-nucleon
interaction with nucleon densities of nuclei.
The barrier penetration model calculates the reaction cross section
$\sigma(E)$ using the standard partial wave ($\ell=0,1,\ldots$)
decomposition and the effective potential $V_\mathrm{eff}(r,E)$. The
latter is constructed from the contributions of the Coulomb, nuclear
and centrifugal potentials,
\begin{equation}
V_\mathrm{eff}(r,E)= V_C(r)+V_\mathit{SP}(r,E)+
\frac{ \hbar^2 \ell (\ell+1)}{2 \mu r^2}.
\label{veff}
\end{equation}
At low energies of our interest, the main contribution to
$\sigma(E)$ comes from the $\ell=0$ (s-wave) channel. Once
$\sigma(E)$ is calculated, we determine $S(E)$ from Eq.\
(\ref{sigma}).
The systematics for the
matter densities $\rho_1(\bm{r}_1)$ and
$\rho_2(\bm{r}_2)$ obtained with a two-parameter Fermi (2pF)
\cite{chamon2002} shape might not be appropriate to provide the
nuclear potential for reactions involving neutron-rich nuclei. The
relativistic Hatree-Bogoliubov approach \cite{rhb1,rhb3} has been
shown to perform well in describing nuclear properties across the
nuclear chart, including the region of neutron-rich nuclei
\cite{rhb1,rhb4,rhb5}. In the present manuscript, the density
distributions of all nuclei were obtained within this approach,
employing the NL3 \cite{rhb4} parametrization of the relativistic
mean field Lagrangian. For that reason, the values of $S(E)$
computed for some reactions ($^{12}$C+$^{12}$C, $^{12}$C+$^{16}$O,
$^{16}$O+$^{16}$O), which we considered previously
\cite{gas2005,yak2006} using the 2pF model, are now somewhat
different.
We have already used \cite{gas2004,saoPauloTool} the S\~ao Paulo
potential with the NL3 nucleon density distribution in the context
of the barrier penetration model and calculated $S(E)$ for a
number of reactions involving stable and neutron-rich nuclei. The
results were compared with experimental data and with theoretical
results calculated by other models such as coupled-channels and
fermionic molecular dynamics ones. As detailed in
\cite{saoPauloTool}, the S\~ao Paulo potential gives reasonably
accurate $S(E)$ for non-resonant nuclear reactions; the method is
parameter-free and relatively simple for generating a set of data
for many reactions involving different isotopes.
Here we calculate astrophysical $S$-factors for 946 fusion reactions
involving combinations of carbon, oxygen, neon, and magnesium
isotopes as summarized in Table \ref{tab:reactions}. We consider 10
reaction types, such as C+C and O+Ne, with the range of mass numbers
for both species given in Table \ref{tab:reactions}. For each
reaction, we compute $S(E)$ in the energy range from 2~MeV to a
maximum value $E_\mathrm{max}$ (also given in Table
\ref{tab:reactions}) in energy steps of 0.1~MeV. The value of
$E_\mathrm{max}$ was chosen in such a way that wide energies ranges
below and above the Coulomb barrier are covered. The fifth column in
Table \ref{tab:reactions} presents the number of considered
reactions of a given type and the last column refers to a table
which lists fit parameters of analytic approximation to $S(E)$ for
such reactions (Sec.\ \ref{fits}).
\begin{table}
\caption[]{Fusion reactions $(A_1,Z_1)+(A_2,Z_2)$ under consideration}
\label{tab:reactions}
\begin{center}
\begin{tabular}{c c c c c c c}
\hline \hline
Reaction & $\quad A_1 \quad $ & $\quad A_2 \quad $ &
~$E_\mathrm{max}$~ & Nr.\ of & ~Table~ \\
type & even & even & MeV & cases & of fits \\
\hline
C + C & 10--24 & 10--24 & 17.9 & 36 & \ref{tab:cc} \\
C + O & 10--24 & 12--28 & 17.9 & 72 & \ref{tab:co} \\
C+Ne & 10--24 & 18--40 & 19.9 & 96 & \ref{tab:cne} \\
C+Mg & 10--24 & 20--46 & 19.9 & 112 & \ref{tab:cmg} \\
O + O & 12--28 & 12--28 & 19.9 & 45 & \ref{tab:oo} \\
O+Ne & 12--28 & 18--40 & 21.9 & 108 & \ref{tab:one} \\
O+Mg & 12--28 & 18--46 & 21.9 & 126 & \ref{tab:omg} \\
Ne+Ne & 18--40 & 18--40 & 21.9 & 78 & \ref{tab:nene} \\
Ne+Mg & 18--40 & 20--46 & 24.9 & 168 & \ref{tab:nemg} \\
Mg+Mg & 20--46 & 20--46 & 29.9 & 105 & \ref{tab:mgmg} \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
In this study we focus on systems of even-even nuclei. Due to
pairing, these nuclei are expected to be more stable in the
astrophysical burning environments in question, for instance, in an
accreted crust of a neutron star \cite{H&Z1989}.
The quality of our data can be deduced from the analysis of Ref.\
\cite{saoPauloTool}. We calculate a set of resonance-averaged
$S(E)$-factors. Their values are uncertain due to nuclear physics
effects -- due to using the S\~ao Paulo model with the NL3 nucleon
density distribution. The uncertainties were estimated
\cite{saoPauloTool} in comparison with other theoretical models and
available experimental data. For the reactions involving stable
nuclei, typical uncertainties are expected to be within a factor of
2, with maximum up to a factor of 4. For the reactions involving
unstable nuclei, typical uncertainties can be as large as one order
of magnitude, reaching two orders of magnitude at low energies for
the reactions with very neutron-rich isotopes. These uncertainties
reflect the current state of art in modelling the nuclear structure
and fusion reaction mechanism; they affect $S(E)$ and should be
taken into account while using the data. The advantage of our data
set is that it is wide and uniform. Note that the uncertainties in
$S(E)$, a factor of 2--10, may have no strong effect on the results
of modelling of nuclear burning phenomena \cite{saoPauloTool}.
\section{Analytic approximation of the astrophysical S-factors}
\label{fits}
All $S$-factors calculated here have been approximated by the
analytic expression
\begin{equation}
S(E) = \exp \left\{ B_1+B_2E + B_3 E^2
+ {C_1+C_2 E + C_3 E^2 +C_4 E^3 \over
1+\exp \left[(E_C-E)/D \right ]} \right \}.
\label{sfit}
\end{equation}
In this expression, $E$ is a center-of-mass energy of reacting
nuclei expressed in MeV, and $E_C$, $D$; $B_1$, $B_2$, $B_3$;
$C_1$, $C_2$, $C_3$, and $C_4$ are nine fit parameters for each
reaction. These parameters are described below; their values are
given in Tables \ref{tab:cc}--\ref{tab:mgmg} -- one table for each
of the reaction types listed in Table \ref{tab:reactions}. We
express $S(E)$ in MeV~b.
Equation (\ref{sfit}) generalizes the expression we suggested
earlier (e.g., Ref.\ \cite{saoPauloTool}) with some terms rearranged
for easier use. It contains additional fit parameters to accommodate
a much larger number of fusion reactions and a broader energy range
with one format. The $S(E)$-data calculated for each reaction were
fitted using Eq.\ (\ref{sfit}) over a total energy range (from 2~MeV
to $E_\mathrm{max}$) indicated in Table \ref{tab:reactions}.
The fit parameters in Eq.\ (\ref{sfit}), presented in Tables
\ref{tab:cc}--\ref{tab:mgmg} for different reactions, have clear
physical meaning:
\begin{itemize}
\item The parameter $E_C$ (expressed in MeV) is approximately
equal to the height of the Coulomb barrier. It divides the energy
range into that below the barrier ($E \;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$) and that
above the barrier ($E \;\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$), where $S(E)$-curves show
distinctly different behaviors.
\item
The parameter $D$ (expressed in MeV) characterizes a narrow energy
width ($D \sim 1$ MeV) of the transition region (at $|E- E_C|
\;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; 2 D$) between the energy ranges below and above the Coulomb
barrier. This transition is governed by the function (of the
Fermi-Dirac type)
\begin{equation}
{1 \over 1+\exp [(E_C-E)/D] },
\label{FD}
\end{equation}
which tends to 0 below the barrier and tends to 1 above the
barrier.
\item The parameter $B_1$ determines $S(0)$ (expressed in MeV~b):
\begin{equation}
S(0)= \exp(B_1) \quad \mathrm{MeV~b}.
\label{S(0)}
\end{equation}
The accuracy of extrapolating the approximated $S(E)$ to $E \to 0$
is high, as discussed below.
\item The parameters $B_2$ (expressed in MeV$^{-1}$) and $B_3$
(expressed in MeV$^{-2}$) specify the energy dependence of $S(E)$
at $E \;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$ as
\begin{equation}
S(E)=S(0) \exp (B_2 E+ B_3 E^2) \quad \mathrm{MeV~b}.
\label{below}
\end{equation}
Therefore, $S(E)$ at $E\;\raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$ is actually determined by
\textit{three} parameters $B_1$, $B_2$, and $B_3$ out of the 9. This
sub-barrier energy range is sufficient for the majority of
applications to stellar burning (Sec.\ \ref{s:rates}). The
3-parameter fit remains quite accurate (within few tens percent) at
$E\leq E_C-2 D$ (but becomes divergent at higher $E$; see Sec.\
\ref{example} for an example). The same parameters $B_1$, $B_2$, and
$B_3$ contribute also to the $S(E)$ dependence at $E \;\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$
[see Eq.\ (\ref{above})].
\item
The parameters $C_1$ (dimensionless), $C_2$ (expressed in
MeV$^{-1}$), $C_3$ (expressed in MeV$^{-2}$), and $C_4$ (expressed
in MeV$^{-3}$), together with $B_1$, $B_2$, and $B_3$, specify the
$S(E)$ dependence at $E \;\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$ (say, at $E \geq E_C+2D$),
\begin{equation}
S(E)=S(0)\,\exp[C_1+(B_2+C_2)E+(C_3+B_3)E^2+C_4 E^3]
\quad \mathrm{MeV~b}.
\label{above}
\end{equation}
\item In the last columns of Tables
\ref{tab:cc}--\ref{tab:mgmg} we give $\delta$ (in percentage), which
is the maximum relative error of fitted $S(E)$ values for each
reaction over the energy grid points.
\end{itemize}
Note that the maximum fit error $\delta$ does not exceed 16\% for
all the reactions in this study (and is much lower in many cases).
The maximum errors $\sim$10--16\% occur only for reactions involving
very neutron-rich nuclei. For the majority of reactions, maximum
errors are realized either at $E \sim E_C$ or at $E \sim
E_\mathrm{max}$. Root-mean-square relative errors over the energy
grid points are a factor of 2 to 3 lower than maximum errors
$\delta$. Considering the much larger uncertainties associated with
the nuclear-physics input of the calculated $S(E)$ (Sec.\
\ref{theory}), our fit accuracy can be regarded as unnecessarily
good. We keep it because it is good for a wide uniform data set. We
expect that the same Eq.\ (\ref{sfit}) can be used to approximate
$S(E)$ for other reactions, as well as for the here discussed
reactions when recalculated (in the future) with more advanced
nuclear physics models. As for the present $S(E)$ data, they could
have been approximated by Eq.\ (\ref{sfit}) retaining 7 parameters
out of the 9 (putting $B_3=C_4=0$). The fit errors would be higher;
in some cases the maximum errors would reach 30--40\%.
Our calculations of $S(E)$ are limited by $E \geq 2$ MeV. Because of
technical reasons (to solve the problem of quantum tunneling through
a very thick Coulomb barrier) we cannot directly calculate $S(0)$.
However, our fit expression (\ref{sfit}) is arranged in such a way
that it insures the low-energy dependence $S(E)=\exp(B_1+B_2 E+B_3
E^2)$ that is predicted on theoretical grounds (e.g., Ref.\
\cite{fh64}). Moreover, while fitting the data we have tried to
reach best fit accuracy at low $E$ (that is most important for
astrophysical applications). For that purpose, any $S(E)$ fit was
done in three stages. First, we have fitted all data points by Eq.\
(\ref{sfit}) and determine preliminary values of all fit parameters.
Second, we have fitted only subbarier $S(E)$ points (2 MeV $\leq E
\leq E_C-2.5D$) by Eq.\ (\ref{below}) to find $B_1$, $B_2$, and
$B_3$ (obtaining thus an accurate description at low $E$). Third, we
have refitted the entire $S(E)$ data set by Eq.\ (\ref{sfit}) with
the fixed values of $B_1$, $B_2$, and $B_3$ (as found at the
previous stage) and obtain thus final values of $E_C$, $D$,
$C_1$,\dots $C_4$. Accordingly, we expect that the fit enables one
to correctly determine the expansion terms $B_1$, $B_2$, and $B_3$,
and extrapolate to $E=0$. To check this point, we calculated
$S(1\,\mbox{MeV})=1.732 \times 10^{84}$ MeV~b for the
$^{46}$Mg+$^{46}$Mg reaction; it differs from the extrapolated value
of $1.839 \times 10^{84}$ MeV~b only by 6\%. This indicates that the
extrapolation to $E \to 0$ should be reliable.
\begin{figure}[tbh]
\begin{center}
\includegraphics[width=10.0cm,angle=0,bb=45 180 420 630]{figur1.ps}
\caption{(Color online) Analytic approximations of $S(E)$ for the
C+C reactions involving different isotopes. The curves from bottom
to top refer to ($A_1,A_2$)= (10,10), (12,12), (12,16), (12,20),
(16,16), (12,24), (16,20), (16,24), (20,20), (20,24), and (24,24)
reactions. Solid curves show the 9-parameter approximation
(\ref{sfit}); dashed curves (plotted at $E \leq E_C+0.3$ MeV) show
the 3-parameter approximation (\ref{below}). Filled dots correspond
to $E=E_C$. } \label{fig:cc}
\end{center}
\end{figure}
As an example, Fig.~\ref{fig:cc} presents the approximated $S(E)$
dependence for the C+C reactions with ($A_1,A_2$)= (10,10), (12,12),
(12,16), (12,20), (16,16), (12,24), (16,20), (16,24), (20,20),
(20,24), and (24,24). The solid lines show our 9-parameter
approximation (\ref{sfit}); they coincide with the calculated data
points in the adopted logarithmic $S$-scale. Thus we do not present
the calculated points to simplify the figure. We plot our fit
expression not only in the energy range, where calculations are done
(from 2 to 17.9 MeV, Table \ref{tab:reactions}), but extrapolate
also beyond this range (from $E=2$ MeV to $E=0$, and from 17.9 MeV
to 20 MeV). The thick dots mark the Coulomb barrier threshold,
$E=E_C$. One can see that, indeed, the behavior of $S(E)$ below and
above the Coulomb barrier is distinctly different, and the
transition from one regime to the other takes place within a narrow
energy range at $E \approx E_C$. For heavier isotopes (with larger
nucleus radii), the Coulomb barrier decreases, which increases the
$S(E)$ curve. The value of $S(0)$ for the reaction involving
heaviest isotopes ($^{24}$C+$^{24}$C) is approximately 15 orders of
magnitude larger than for the reaction of lightest isotopes
($^{10}$C+$^{10}$C). Finally, the dashed curves in Fig.~\ref{fig:cc}
show our 3-parameter $S(E)$ approximation (\ref{below}). It is seen
to be fairly accurate below the Coulomb barrier, but becomes
inaccurate in the vicinity of $E=E_C$ and at larger $E$.
\begin{figure}[tbh]
\begin{center}
\includegraphics[width=10.0cm,angle=0,bb=45 180 420 630]{figur2.ps}
\caption{(Color online) Analytic approximation (\ref{sfit}) of
$S(E)$ for 36 C+C reactions, 112 C+Mg reactions, and 105 Mg+Mg
reactions (Table \ref{tab:reactions}). The curves for each reaction
type are enclosed by the (thick) curves for the (indicated)
reactions involving lightest and heaviest isotopes. Filled dots
refer to $E=E_C$. } \label{fig:cmg}
\end{center}
\end{figure}
Figure \ref{fig:cmg} shows the $S(E)$ dependence for all 36 C+C
reactions, 112 C+Mg reactions, 105 Mg+Mg fusion reactions
considered in this study (Table~\ref{tab:reactions}). The curves are
9-parameter fits (\ref{sfit}), and the thick dots again mark the
Coulomb barrier threshold, $E=E_C$. The curves for each fusion type
(C+C, C+Mg, Mg+Mg) are enclosed by the thick curves for the
reactions involving lightest and heavies isotopes (for instance,
$^{10}$C+$^{20}$Mg and $^{24}$C+$^{46}$Mg, in the case of C+Mg
reactions). The general $S(E)$-behavior is seen to be the same as
for C+C reactions. The difference of $S(E)$ values, especially at
low $E$, for different reactions can be extremely large. For
example, the difference of $S(0)$ for the $^{46}$Mg+$^{46}$Mg and
$^{10}$C+$^{10}$C reactions is approximately 70 orders of magnitude.
\section{Numerical example}
\label{example}
\noindent As an example, we calculate $S(E)$ for the
$^{20}$Ne+$^{24}$Mg reaction at $E=$5 MeV. According to Table
\ref{tab:reactions}, fit parameters for the Ne+Mg reactions are
listed in Table \ref{tab:nemg}. The reaction in question corresponds
to $A_1=20$ and $A_2=24$. From line 17
of Table \ref{tab:nemg} we obtain the fit parameters:\\
$E_C=19.019$ MeV,
$D=0.88$ MeV;\\
$B_1=100.480$, $B_2=-0.3985$ MeV$^{-1}$,
$B_3=-0.00583$ MeV$^{-2}$;\\
$C_1=-40.079$, $C_2=6.6560$ MeV$^{-1}$, $C_3=-0.33863$ MeV$^{-2}$,
$C_4=0.005087$ MeV$^{-3}$.\\
Using Eq.\ (\ref{sfit}) at $E=5$ MeV we have
\begin{eqnarray}
S(5~\mathrm{MeV}) & = & \exp \left\{
100.480-0.3985 \times 5 - 0.00583 \times 5^2
{{}\over{}} \right.
\nonumber \\
&+ & \left. {-40.079+6.6560\times 5 -0.33863 \times 5^2 + 0.005087
\times 5^3 \over
1+\exp \left[(19.019-5)/0.88 \right ]} \right \}
\nonumber \\
& = & 5.1223 \times 10^{42}
\quad \mathrm{MeV~b}.
\label{sfit1}
\end{eqnarray}
In this particular example, the fitted value deviates only by
0.006\% from the value
$S(5~\mathrm{MeV})=5.1226$ MeV~b
calculated using the S\~ao Paulo potential.
\begin{figure}[t]
\begin{center}
\includegraphics[width=10.0cm,angle=0]{figur3.ps}
\caption{(Color online) {\em Top}: Analytic approximations of $S(E)$
for the $^{20}$Ne+$^{24}$Mg reaction. The solid and dashed lines
are, respectively, the 9- and 3-parameter approximations
(\ref{sfit}) and (\ref{below}). The filled dot refers to $E=E_C$.
{\em Bottom}: Relative errors of these approximations with respect
to computed data. The dotted line shows zero error to guide the eye.
} \label{fig:nemg}
\end{center}
\end{figure}
Since our chosen energy $E=5$ MeV is lower than the Coulomb barrier
height, $E_C\approx 19$ MeV, we can also calculate $S(E)$ using our
3-parameter fit. From Eq.\ (\ref{S(0)}) we obtain
\begin{equation}
S(0)=\exp(100.480)=4.3459 \times 10^{43}\quad \mathrm{MeV~b},
\label{S(0)1}
\end{equation}
and from Eq.\ (\ref{below}) we have
\begin{eqnarray}
S(5~\mathrm{MeV}) & = & 4.3459 \times 10^{43}
\exp (
-0.3985 \times 5 - 0.00583 \times 5^2)
\nonumber \\
& = & 5.1223 \times 10^{42} \quad \mathrm{MeV~b},
\label{below1}
\end{eqnarray}
in excellent agreement with the result (\ref{sfit1}) of the
9-parameter fit.
The $S(E)$ dependence for the $^{20}$Ne+$^{24}$Mg reaction is
plotted in the upper panel of Fig.~\ref{fig:nemg}. The solid and
dashed lines are the 9- and 3-parameter fits, respectively. In the
lower panel we show relative errors of fitted values of $S(E)$ [not
of $\log S(E)$] with respect to the calculated values. We see that
the 9-parameter fit is accurate over the entire energy range, in
which original $S(E)$ values have been calculated (Table
\ref{tab:reactions}). The maximum fit error of $\approx$1.2\% occurs
at $E=15.8$ MeV (and the root-mean-square fit error over all grid
points is $\approx 0.6$\%). The three-parameter fit (\ref{below})
stays highly accurate below the Coulomb barrier but diverges when
$E$ exceeds $E_C$. For instance, at $E=25$ MeV it overestimates
$S(E)$ by more than two orders of magnitude.
\section{Reaction rates}
\label{s:rates}
In the following we discuss the calculation of fusion reaction rates
$R$ [cm$^{-3}$~s$^{-1}$] in stellar environments. Any reaction rate
$R$ depends on the environmental conditions such as temperature,
density and composition of stellar matter. We cannot present tables
of the reaction rates covering the entire parameter space but give a
short description how the tabulated $S$-factors can be easily
utilized for deriving the reaction rates.
We limit our discussion to the formalism for non-resonant reaction
rates since the calculated $S$-factors presented here are entirely
characterized by non-resonant cross section behavior. If the cross
section is dominated by strong resonances the formalism should be
modified by adding the resonance contribution separately (e.g.,
Ref.\ \cite{itohetal03}).
As a rule, the main contribution to non-resonant rates comes from
nucleus-nucleus collisions in a narrow energy range $E\approx
E_\mathrm{pk}$, and $S(E)$ is a slowly varying function of $E$. Then
$R$ can be expressed through $S(E_\mathrm{pk})$. In particular, in
the thermonuclear regime, neglecting the effects of plasma
screening, $E_\mathrm{pk}$ is the standard Gamow-peak energy and the
reaction rates are given by the well-known classical theory of
thermonuclear burning (e.g., Refs.\ \cite{bbfh57,fh64,clayton83}).
This regime is realized at sufficiently high temperatures of stellar
matter when the nuclei form almost ideal Boltzmann gas. Even in this
regime $E_\mathrm{pk}$ is typically lower than $E_C$, and $S(E)$ can
be approximated by our 3-parameter fit (\ref{below}) without any
loss of accuracy in the reaction rate. If, however, one needs an
accurate reaction rate at
so
high temperatures
that $E_\mathrm{pk} \;\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}}\; E_C$, one should exactly calculate
the rate by integrating over $E$ with the full 9-parameter fit
(\ref{sfit}) for $S(E)$. The same fit can also be used for
evaluating the reaction cross sections $\sigma(E)$ in astrophysical
studies and nuclear physics laboratory experiments.
At lower temperatures the Gamow peak approach remains as a valid
approximation
but one should take into account the
plasma screening effects and a possible transition to the
pycnonuclear burning regime. In this case the equations for reaction
rates should be modified. These modifications are described in the
literature (e.g., Refs.\ \cite{gas2005,yak2006,cdi07,cd09} and
references therein). They are model-dependent, but in any case they
contain the values of $S(E)$ at certain ``Gamow-peak'' energies
$E_\mathrm{pk}$ (that are also modified by the plasma screening and
pycnonuclear burning effects, and depend on $T$ and $\rho$). Such
energies are typically much lower than $E_C$, so that our
3-parameter fits apply.
Specifically, Ref.\ \cite{gas2005} gives the expressions
(Sec.~III.G of \cite{gas2005}) for reaction rates in one-component
ion plasma in all 5 burning regimes. Ref.\ \cite{yak2006}
generalizes these expressions (Sec.\ III.G of \cite{yak2006}) to
the case of multicomponent ion plasma. The authors of Refs.\
\cite{cdi07} and \cite{cd09} present more accurate calculations
and approximations of reaction rates in one-component and
multi-component ion plasma, respectively. They use the WKB Coulomb
tunneling approximation in a radial mean field plasma potential
(that was extracted from extended Monte Carlo simulations of
classical strongly coupled ion plasmas). Their results are valid
in the thermonuclear burning regime and in the intermediate
thermo-pycno nuclear regime (it is possible that they can also be
extended to lower temperatures). In these regimes, the enhancement
factors of nuclear reaction rates due to plasma screening effects,
given in Refs.\ \cite{gas2005,yak2006,cdi07,cd09}, are in a
reasonably good agreement. In the pycnonuclear regimes (at zero
temperature and with thermal enhancement) the results
\cite{gas2005,yak2006} are model dependent due to plasma physics
uncertainties. For that reason, the authors of
\cite{gas2005,yak2006} present optimal (recommended) as well as
maximum and minimum theoretical reaction rates.
Note an omission in Ref.\ \cite{gas2005}: the expression for the
parameter $\lambda$, given by Eq.~(24) in \cite{gas2005}, for
pycnonuclear burning models in face-centered cubic crystals has to
be divided by $2^{1/3}$. This omission did not affect our choice of
optimal, maximum, and minimum reaction rates in \cite{gas2005}, and
our results in \cite{yak2006}. In addition, notice two typos in
\cite{yak2006}. In Eq.\ (32) of \cite{yak2006} there should be the
exponent sign $\exp$ before the last term in brackets in the
expression for $F_\mathrm{pyc}$; in Eq.\ (33) the product $x_i x_j$
must be replaced by $X_i X_j$. Also notice that in Refs.\
\cite{gas2005,yak2006} we neglected the Coulomb shifts of nucleus
energy levels in dense matter in the expression for $E_\mathrm{pk}$
at low temperatures (in pycnonuclear burning regimes). These effects
can influence the reaction rates at low $T$ if $S(E)$ is not a very
slowly varying function of $E$. Such shifts of $E_\mathrm{pk}$ have
been mentioned in the literature (e.g., Ref.\ \cite{itohetal03}) and
were parameterized in \cite{cd09}. We recommend to use Eq.\ (35) of
Ref.\ \cite{cd09} to calculate $E_\mathrm{pk}$ at all densities and
temperatures. The use of Eq.\ (40) of Ref.\ \cite{yak2006} is also
possible (gives correct rates of the reactions of study, within
nuclear physics uncertainties, for those densities and temperatures,
where these reactions are most efficient).
For illustration of the procedure, Fig.\ \ref{fig:nemg1} presents
the rate of the previously discussed example $^{20}$Ne+$^{24}$Mg
as a function of density and temperature of matter composed of
$^{20}$Ne and $^{24}$Mg nuclei with equal number densities of Ne
and Mg nuclei. The rate is calculated using the optimal model of
the reaction rate from Ref.\ \cite{yak2006}. We show
a
wider $\rho-T$ range than the range, where these nuclei can really
exist in stellar matter, in order to demonstrate all features of the
reaction rate; these features are qualitatively the same for all
reactions of our study. One can see a strong temperature dependence
of the rate at high $T$ in the thermonuclear burning regimes, and a
strong density dependence
at high $\rho$ and low $T$ in the pycnonuclear regimes.
\begin{figure}[t]
\begin{center}
\includegraphics[width=10.0cm,angle=0]{figure4.eps}
\caption{(Color online) Density-temperature dependence of the
$^{20}$Ne+$^{24}$Mg reaction rate in an $^{20}$Ne-$^{24}$Mg mixture
with equal number fractions of Ne and Mg. } \label{fig:nemg1}
\end{center}
\end{figure}
Finally, let us note again that nuclear physics uncertainties of our
calculated $S(E)$ are not small (Sec.\ \ref{theory}) and they
translate into uncertainties in the reaction rates. However, as we
analyzed in \cite{saoPauloTool}, such uncertainties are often not
very important for astrophysical applications.
\section{Conclusions}
\label{s:concl}
Using S\~ao Paulo method and the barrier penetration model we have
calculated the astrophysical $S$-factor as a function of energy for
946 fusion reactions involving various isotopes of C, O, Ne, and Mg,
from the stability valley to very neutron-rich nuclei. The
calculations have been performed on a dense grid of center-of-mass
energies $E$, from 2 MeV to 18--30 MeV, covering wide energy ranges
below and above the Coulomb barrier.
We fit calculated $S(E)$ values by a simple and accurate universal
9-parameter analytic formula, and present tables of fit parameters
for all the reactions. The fit error does not exceed 16\%. The
formula allows extrapolating $S(E)$ outside the considered energy
range, particularly, to lower energies $E \to 0$ of astrophysical
interest. We have also shown that the reduced fit expression,
containing 3 fit parameters out of 9, is fairly accurate at energies
below the Coulomb barrier and sufficient for accurate calculations
of the reaction rates at not too high temperatures of stellar
matter.
Our results can be used in computer codes for calculating nuclear
fusion rates and simulating various phenomena associated with
nuclear burning in high temperature and/or high density
astrophysical and laboratory plasmas. In particular, they can be
used for modelling nuclear burning in massive accreting white dwarfs
(type Ia supernovae) or in accreting neutron stars (superbursts,
deep crustal burning in X-ray transient sources). Nuclear burning at
high densities in these compact stars can involve neutron-rich
nuclei considered in the present paper.
The calculated $S(E)$ factors are rapidly varying functions of
energy $E$. Their values for various reactions can differ by many
orders of magnitude. On the other hand, their energy dependence
looks self-similar over a broad spectrum of the reactions. We will
address these findings in a separate publication.
\begin{acknowledgments}
DY is grateful to Andrew Chugunov for critical remarks and to Peter
Shternin for his assistance in the artwork. This work was partly
supported by the Joint Institute for Nuclear Astrophysics
(NSF-PHY-0822648), the U.S. Department of Energy under the grant
DE-FG02-07ER41459, the Russian Foundation for Basic Research (grants
08-02-00837 and 09-02-1208), and by the State Program ``Leading
Scientific Schools of Russian Federation'' (Grant NSh 2600.2008.2).
\end{acknowledgments}
|
\section{Introduction}
It was proved by Mukai in \cite{mukai} that the moduli space
of simple sheaves on an abelian or a projective K3 surface is smooth
and has a symplectic structure.
We will generalize this result to the moduli space
of objects in the derived category of coherent
sheaves, which is introduced in \cite{inaba1}.
By [\cite{inaba2}, Theorem 4.4], the moduli space of
(semi)stable objects with respect to a strict ample sequence
in a derived category of coherent sheaves on an abelian or a projective
K3 surface gives examples of projective symplectic varieties.
In the proof of the main results,
we will use the the trace map that also played a key role
in \cite{mukai}.
More precisely, we will calculate the image
by the trace map of the obstruction class for
the deformation of complexes of coherent sheaves.
So the idea of the proof of this paper is the same
as that of \cite{mukai}.
However, the calculation of the trace map, without any preparation,
seems to be too complicated.
For this reason, we will reconsider in section 2 the definition
of the obstruction class for the deformation of vector bundles.
By virtue of this consideration (Lemma \ref{obstruction=})
in section 2, the calculation of the trace map
becomes clear and the main result can be deduced from it.
The content of this paper was originally written as an appendix
of \cite{inaba2}.
However there was a mistake in the proof of the smoothness of
$\mathop\mathrm{Splcpx}\nolimits_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$.
In this paper the author corrects the mistake.
\section{Obstruction classes for the deformation of vector bundles}
First we recall the obstruction theory of the deformation
of objects in the derived category of bounded complexes
of coherent sheaves.
Let $S$ be a noetherian scheme and $X$ be
a projective scheme flat over $S$.
We fix an $S$-ample line bundle ${\mathcal O}_X(1)$ on $X$.
Let $A$ be an artinian local ring over $S$ with residue field $k=A/m$
and $I$ be an ideal of $A$ such that $mI=0$.
Take a bounded complex $E^{\bullet}$ of $A/I$-flat coherent sheaves on $X_{A/I}$.
Then there are integers $l,l'$ such that $E^i=0$ for $i<l'$ and $i>l$.
We can take a complex $V^{\bullet}=(V^i,d^i)$ of the form
$V^i=V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)$
and a quasi-isomorphism $V^{\bullet}\to E^{\bullet}$,
where $V_i$ are free $A$ modules of finite rank,
$V_i=0$ for $i>l$ and
$1\ll m_l\ll m_{l-1}\ll\cdots\ll m_{i+1}\ll m_i\ll\cdots$.
Take lifts
\[
\tilde{d}^i:V_i\otimes{\mathcal O}_{X_A}(-m_i)\to
V_{i+1}\otimes{\mathcal O}_{X_A}(-m_{i+1})
\]
of the homomorphisms
\[
d^i:V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)\to
V_{i+1}\otimes{\mathcal O}_{X_{A/I}}(-m_{i+1}).
\]
Then we obtain homomorphisms
\[
\delta^i:=\tilde{d}^{i+1}\circ\tilde{d}^i:
V_i\otimes{\mathcal O}_{X_A}(-m_i)\to
I\otimes_A V_{i+2}\otimes{\mathcal O}_{X_A}(-m_{i+2}).
\]
We put
\[
\omega(E^{\bullet}):=[\{\delta^i\}]\in
H^2(\mathop\mathrm{Hom}\nolimits(V^{\bullet},V^{\bullet}\otimes I))\cong
\mathop\mathrm{Ext}\nolimits^2(E^{\bullet}\otimes k,E^{\bullet}\otimes k)\otimes_k I.
\]
\begin{proposition}
$\omega(E^{\bullet})=0$ if and only if $E^{\bullet}$ can be lifted to
an object of $D^b(\mathrm{Coh}(X_A))$
of finite $\mathrm{Tor}$ dimension over $A$.
\end{proposition}
(Proof is in [\cite{inaba1},Proposition 2.3].)
For a vector bundle, there is another definition of the obstruction class.
Let $F$ be a locally free sheaf of rank $r$ on $X_{A/I}$.
Take an affine open covering $\{U_{\alpha}\}$ of $X_A$
such that
$F|_{U_{\alpha}}\cong{\mathcal O}_{U_{\alpha}\otimes {A/I}}^{\oplus r}$
for any $\alpha$.
Let $F_{\alpha}$ be a free ${\mathcal O}_{U_{\alpha}}$-module such that
$F_{\alpha}\otimes A/I \cong F|_{U_{\alpha}}$.
Take a lift
$\varphi_{\beta \alpha}:F_{\alpha}|_{U_{\alpha\beta}}\to
F_{\beta}|_{U_{\alpha \beta}}$
of the composite
\[
F_{\alpha}\otimes A/I|_{U_{\alpha\beta}}\stackrel{\sim}\longrightarrow
F|_{U_{\alpha\beta}}\stackrel{\sim}\longrightarrow
F_{\beta}\otimes A/I|_{U_{\alpha \beta}},
\]
where $U_{\alpha\beta}:=U_{\alpha}\cap U_{\beta}$.
We put
\[
\theta_{\alpha\beta\gamma}:=
\varphi_{\gamma\alpha}^{-1}\circ\varphi_{\gamma\beta}\circ
\varphi_{\beta\alpha}-\mathrm{id}_{F_{\alpha}}:
F_{\alpha}|_{U_{\alpha\beta\gamma}}\longrightarrow
I\otimes F_{\alpha}|_{U_{\alpha\beta\gamma}},
\]
where $U_{\alpha\beta\gamma}:=U_{\alpha}\cap U_{\beta}\cap U_{\gamma}$.
Then the cohomology class
\[
o(F):=[\{\theta_{\alpha\beta\gamma}\}]\in \check{H}^2({\mathcal End}(F)\otimes I)
\cong \mathop\mathrm{Ext}\nolimits^2(F,F\otimes I)
\]
can be defined.
As is stated in [\cite{sga}, III, Proposition 7.1],
we have the following proposition.
\begin{proposition}
$o(F)=0$ if and only if $F$ can be lifted to a locally free sheaf
on $X_A$.
\end{proposition}
A vector bundle $F$ on $X_{A/I}$ can be considered as
the object of $D^b(\mathrm{Coh}(X_{A/I}))$ whose $0$-th component is $F$
and the other components are zero.
We will show that $\omega(F)$ and $o(F)$
are the same element in $\mathop\mathrm{Ext}\nolimits^2(F,F\otimes I)$.
We take a resolution of $F$ by locally free sheaves:
\[
\cdots\longrightarrow V^2 \stackrel{d^2}\longrightarrow V^1
\stackrel{d^1}\longrightarrow V^0 \stackrel{\pi}\longrightarrow
F \longrightarrow 0,
\]
where each $V^i$ is isomorphic to $V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)$
for a free $A$-module $V_i$ of finite rank and
$1\ll m_0\ll m_1\ll\cdots\ll m_i\ll m_{i+1}\ll\cdots$.
Then we have a quasi-isomorphism
${\mathcal Hom}(F,F)\otimes I \to
{\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I$.
Let
\[
{\mathcal Hom}(F,F)\otimes I \to
{\mathcal C}^{\bullet}({\mathcal Hom}(F,F)\otimes I)
\]
be the \v{C}ech resolution of ${\mathcal Hom}(F,F)\otimes I$
with respect to the covering $\{U_{\alpha}\}$
and
\[
{\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I\to
{\mathcal C}^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I)
\]
be that of ${\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I$.
Then we obtain a composition of isomorphisms
\[
f:H^2(\mathop\mathrm{Hom}\nolimits^{\bullet}(V^{\bullet},F))\stackrel{\sim}\longrightarrow
{\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))
\stackrel{\sim}\longrightarrow
\check{H}^2({\mathcal End}(F)\otimes I),
\]
where $C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I)=
\Gamma(X,{\mathcal C}^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$.
\begin{lemma}\label{obstruction=}
Under the above assumption and notation,
we have $f(\omega(F))=o(F)$.
\end{lemma}
\begin{proof}
First note that the element $\omega(F)$ is defined by
\[
\omega(F)=[\{(\pi\otimes\mathrm{id}_I)
\circ(\tilde{d}^1\circ\tilde{d}^2)\}]
\in H^2(\mathop\mathrm{Hom}\nolimits^{\bullet}(V^{\bullet},F\otimes I)),
\]
where $\tilde{d}^i:V_i\otimes{\mathcal O}_{X_A}(-m_i)\to
V_{i+1}\otimes{\mathcal O}_{X_A}(-m_{i+1})$
is a lift of $d^i$.
Replacing $\{U_{\alpha}\}$ by its refinement, we may assume that
$\ker d^2|_{U_{\alpha}}$,
$\mathop\mathrm{im}\nolimits d^2|_{U_{\alpha}}$,
$\mathop\mathrm{im}\nolimits d^1|_{U_{\alpha}}$,
$V_2\otimes{\mathcal O}_{X_{A/I}}(-m_2)|_{U_{\alpha}}$,
$V_1\otimes{\mathcal O}_{X_{A/I}}(-m_1)|_{U_{\alpha}}$
and $F|_{U_{\alpha}}$ are all free sheaves.
Then the exact sequences
\begin{gather*}
0 \longrightarrow \ker d^2|_{U_{\alpha}}
\stackrel{i_2}\longrightarrow
V_2\otimes{\mathcal O}_{X_{A/I}}(-m_2)|_{U_{\alpha}}
\stackrel{p_2}\longrightarrow
\mathop\mathrm{im}\nolimits d^2|_{U_{\alpha}} \longrightarrow 0, \\
0 \longrightarrow \mathop\mathrm{im}\nolimits d^2|_{U_{\alpha}}
\stackrel{i_1}\longrightarrow
V_1\otimes{\mathcal O}_{X_{A/I}}(-m_1)|_{U_{\alpha}}
\stackrel{p_1}\longrightarrow
\mathop\mathrm{im}\nolimits d^1|_{U_{\alpha}} \longrightarrow 0, \\
0 \longrightarrow \mathop\mathrm{im}\nolimits d^1|_{U_{\alpha}}
\stackrel{i_0}\longrightarrow
V_0\otimes{\mathcal O}_{X_{A/I}}(-m_0)|_{U_{\alpha}}
\xrightarrow{\pi|_{U_{\alpha}}}
F|_{U_{\alpha}} \longrightarrow 0
\end{gather*}
split and we can take free ${\mathcal O}_{U_{\alpha}}$-modules
$F_{\alpha}$, $I_1^{\alpha}$, $I_2^{\alpha}$ such that
$F_{\alpha}\otimes A/I\cong F|_{U_{\alpha}}$
and $I_i^{\alpha}\otimes A/I\cong\mathop\mathrm{im}\nolimits d^i|_{U_{\alpha}}$ for $i=1,2$.
Taking lifts $\tilde{i}_0^{\alpha}$, $\tilde{i}_1^{\alpha}$, $\pi_{\alpha}$,
$\tilde{p}_1^{\alpha}$, $\tilde{p}_2^{\alpha}$ of
$i_0$, $i_1$, $\pi|_{U_{\alpha}}$, $p_1$, $p_2$,
we obtain splitting exact sequences
\begin{gather*}
0 \longrightarrow \ker \tilde{p}_2^{\alpha} \longrightarrow
V_2\otimes{\mathcal O}_{X_{A}}(-m_2)|_{U_{\alpha}}
\stackrel{\tilde{p}_2^{\alpha}}\longrightarrow
I_2^{\alpha} \longrightarrow 0, \\
0 \longrightarrow I_2^{\alpha} \stackrel{\tilde{i}_1^{\alpha}}\longrightarrow
V_1\otimes{\mathcal O}_{X_{A}}(-m_1)|_{U_{\alpha}}
\stackrel{\tilde{p}_1^{\alpha}}\longrightarrow I_1^{\alpha} \longrightarrow 0, \\
0 \longrightarrow I_1^{\alpha} \stackrel{\tilde{i}_0^{\alpha}}\longrightarrow
V_0\otimes{\mathcal O}_{X_{A}}(-m_0)|_{U_{\alpha}}
\stackrel{\pi_{\alpha}}\longrightarrow F_{\alpha}\longrightarrow 0.
\end{gather*}
Let
\begin{gather*}
\tilde{s}_2^{\alpha}:I_2^{\alpha}\longrightarrow
V_2\otimes{\mathcal O}_{X_{A}}(-m_2)|_{U_{\alpha}}, \\
\tilde{r}_1^{\alpha}:V_1\otimes{\mathcal O}_{X_{A}}(-m_1)|_{U_{\alpha}}
\longrightarrow I_2^{\alpha}, \quad
\tilde{s}_1^{\alpha}:I_1^{\alpha}\longrightarrow
V_1\otimes{\mathcal O}_{X_{A}}(-m_1)|_{U_{\alpha}}, \\
\tilde{r}_0^{\alpha}:V_0\otimes{\mathcal O}_{X_{A}}(-m_0)|_{U_{\alpha}}
\longrightarrow I_1^{\alpha}, \quad
\nu_{\alpha}:F_{\alpha}\longrightarrow
V_0\otimes{\mathcal O}_{X_{A}}(-m_0)|_{U_{\alpha}}
\end{gather*}
be splittings.
Put
\begin{gather*}
d^2_{\alpha}:V_2\otimes{\mathcal O}_{X_A}(-m_2)|_{U_{\alpha}}
\stackrel{\tilde{p}_2^{\alpha}}\longrightarrow I_2^{\alpha}
\stackrel{\tilde{i}_1^{\alpha}}\longrightarrow
V_1\otimes{\mathcal O}_{X_A}(-m_1)|_{U_{\alpha}}, \\
d^1_{\alpha}:V_1\otimes{\mathcal O}_{X_A}(-m_1)|_{U_{\alpha}}
\stackrel{\tilde{p}_1^{\alpha}}\longrightarrow I_1^{\alpha}
\stackrel{\tilde{i}_0^{\alpha}}\longrightarrow
V_0\otimes{\mathcal O}_{X_A}(-m_0)|_{U_{\alpha}}, \\
\tau_{\alpha}:V_0\otimes{\mathcal O}_{X_A}(-m_0)|_{U_{\alpha}}
\stackrel{\tilde{r}_0^{\alpha}}\longrightarrow I_1^{\alpha}
\stackrel{\tilde{s}_1^{\alpha}}\longrightarrow
V_1\otimes{\mathcal O}_{X_A}(-m_1)|_{U_{\alpha}}, \\
\sigma_{\alpha}:V_1\otimes{\mathcal O}_{X_A}(-m_1)|_{U_{\alpha}}
\stackrel{\tilde{r}_1^{\alpha}}\longrightarrow I_2^{\alpha}
\stackrel{\tilde{s}_2^{\alpha}}\longrightarrow
V_2\otimes{\mathcal O}_{X_A}(-m_2)|_{U_{\alpha}}.
\end{gather*}
We consider the following diagram:
\[
\begin{array}{ccccc}
\mathop\mathrm{Hom}\nolimits(V^0,F\otimes I) & \longrightarrow & \mathop\mathrm{Hom}\nolimits(V^1,F\otimes I) &
\longrightarrow & \mathop\mathrm{Hom}\nolimits(V^2,F\otimes I) \\
\downarrow & & \downarrow & & \downarrow \\
C^0({\mathcal Hom}(V^0,F\otimes I)) & \longrightarrow &
C^0({\mathcal Hom}(V^1,F\otimes I)) & \longrightarrow &
C^0({\mathcal Hom}(V^2,F\otimes I)) \\
\downarrow & & \downarrow & & \downarrow \\
C^1({\mathcal Hom}(V^0,F\otimes I)) & \longrightarrow &
C^1({\mathcal Hom}(V^1,F\otimes I)) & \longrightarrow &
C^1({\mathcal Hom}(V^2,F\otimes I)) \\
\downarrow & & \downarrow & & \downarrow \\
C^2({\mathcal Hom}(V^0,F\otimes I)) & \longrightarrow &
C^2({\mathcal Hom}(V^1,F\otimes I)) & \longrightarrow &
\, C^2({\mathcal Hom}(V^2,F\otimes I)),
\end{array}
\]
where we put $V^i:=V_i\otimes{\mathcal O}_{X_A}(-m_i)$ for $i=0,1,2$.
The image of $\omega(F)$ in
${\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$
can be represented by
\[
\left\{ (\pi\otimes\mathrm{id}_I)\circ
\tilde{d}^1\circ\tilde{d}^2|_{U_{\alpha}}\right\}
\in C^0({\mathcal Hom}(V^2,F)\otimes I),
\]
which defines the same element in
${\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$ as
\[
\left\{ (\pi\otimes\mathrm{id}_I)\circ\tilde{d}^1\circ
\tilde{d}^2\circ(\sigma_{\alpha}-\sigma_{\beta})\right\}
\in C^1({\mathcal Hom}(V^1,F)\otimes I).
\]
On the other hand, the image of the element
\[
\left\{ (\pi\otimes\mathrm{id}_I)\circ
\left(d^1_{\alpha}-\tilde{d}^1\circ
(1-\tilde{d}^2\sigma_{\alpha})\right)\right\}
\in C^0({\mathcal Hom}(V^1,F)\otimes I)
\]
by the homomorphism
$C^0({\mathcal Hom}(V^1,F)\otimes I)\rightarrow
C^0({\mathcal Hom}(V^2,F)\otimes I)$ is
\begin{align*}
\left\{ (\pi \otimes \mathrm{id}_I)\circ\left( d^1_{\alpha}-\tilde{d}^1
\circ(1-\tilde{d}^2\circ\sigma_{\alpha} ) \right)\circ d^2_{\alpha}\right\}
&= \left\{ (\pi\otimes\mathrm{id}_I)(d^1_{\alpha}\circ d^2_{\alpha}-\tilde{d}^1\circ d^2_{\alpha}
+\tilde{d}^1\circ\tilde{d}^2\circ\sigma_{\alpha}\circ d^2_{\alpha}) \right\} \\
&= \left\{ (\pi \otimes \mathrm{id}_I)\left(-\tilde{d}^1\circ d^2_{\alpha}
+\tilde{d}^1\circ\tilde{d}^2\circ\sigma_{\alpha}\circ d^2_{\alpha}\right)\right\} \\
&=\left\{ (\pi\otimes\mathrm{id}_I)\left( -\tilde{d}^1\circ d^2_{\alpha} +
\tilde{d}^1\circ \tilde{d}^2\circ \tilde{s}_2^{\alpha}\circ\tilde{r}_1^{\alpha}\circ\tilde{i}_1^{\alpha}\circ\tilde{p}_2^{\alpha}
\right) \right\} \\
&=\left\{ (\pi\otimes\mathrm{id}_I)\left(-\tilde{d}^1\circ d^2_{\alpha}\circ\tilde{s}^{\alpha}_2\circ\tilde{p}_2^{\alpha}
+\tilde{d}^1\circ \tilde{d}^2\circ\tilde{s}^{\alpha}_2\circ \tilde{p}_2^{\alpha} \right) \right\} \\
&=\left\{ (\pi\otimes\mathrm{id}_I)\circ \tilde{d}^1\circ (\tilde{d}^2-d^2_{\alpha})\circ\tilde{s}_2^{\alpha}\circ\tilde{p}_2^{\alpha}
\right\} \\
&=0.
\end{align*}
Since
\begin{align*}
&\left\{ (\pi\otimes\mathrm{id}_I)\circ\tilde{d}^1
\circ\tilde{d}^2\circ(\sigma_{\alpha}-\sigma_{\beta}) \right\}
+ d \left\{ (\pi\otimes\mathrm{id}_I)\circ
\left(d^1_{\alpha}-\tilde{d}^1\circ
(1-\tilde{d}^2\circ\sigma_{\alpha})\right) \right\} \\
&=\left\{(\pi\otimes\mathrm{id}_I)\circ\tilde{d}^1\circ\tilde{d}^2
\circ(\sigma_{\alpha}-\sigma_{\beta})\right\}
+\left\{ (\pi\otimes\mathrm{id}_I)\circ
\left(d^1_{\beta}-\tilde{d}^1\circ(1-\tilde{d}^2\circ\sigma_{\beta})
\right)|_{U_{\alpha}\cap U_{\beta}}\right\} \\
&\quad -\left\{(\pi\otimes\mathrm{id}_I)\circ
\left(d^1_{\alpha}-\tilde{d}^1\circ(1-\tilde{d}^2\circ\sigma_{\alpha})
\right)|_{U_{\alpha}\cap U_{\beta}}\right\} \\
&=-\left\{ (\pi\otimes\mathrm{id}_I)\circ
(d^1_{\alpha}-d^1_{\beta}) \right\},
\end{align*}
we can see that
$\left\{ (\pi\otimes\mathrm{id}_I)\circ\tilde{d}^1\circ\tilde{d}^2(\sigma_{\alpha}-\sigma_{\beta}) \right\}$
and
$-\left\{ (\pi\otimes \mathrm{id}_I)\circ(d^1_{\alpha}-d^1_{\beta}) \right\}$
define the same element in
$\mathbf{H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$.
We can see that the element
$-\{(\pi\otimes\mathrm{id}_I)\circ(d^1_{\alpha}-d^1_{\beta})\}$
defines the same element as
\begin{align*}
& -\left\{ (\pi\otimes\mathrm{id}_I)\circ
\left( (d^1_{\beta}-d^1_{\gamma})\circ\tau_{\beta}
-(d^1_{\alpha}-d^1_{\gamma})\circ\tau_{\alpha}
+(d^1_{\alpha}-d^1_{\beta})\circ\tau_{\alpha} \right) \right\} \\
&=\left\{ (\pi\otimes\mathrm{id}_I)\circ (d^1_{\beta}-d^1_{\gamma})
\circ (\tau_{\alpha}-\tau_{\beta}) \right\}
\in C^2({\mathcal Hom}(V^0,F)\otimes I)
\end{align*}
in ${\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$.
Thus $\omega(F)$ is equal to the element given by
\[
\left\{ (\pi\otimes\mathrm{id}_I)\circ(d^1_{\beta}-d^1_{\gamma})
\circ (\tau_{\alpha}-\tau_{\beta}) \right\}
\in C^2({\mathcal Hom}(V^0,F)\otimes I)
\]
in ${\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$.
On the other hand, the element $o(F)$ is given by
\[
\{ (\pi_{\gamma}\circ\nu_{\alpha})^{-1}
\circ\pi_{\gamma}\circ\nu_{\beta}
\circ\pi_{\beta}\circ\nu_{\alpha}-\mathrm{id}_{F_{\alpha}} \}
\]
in $\check{H}^2({\mathcal End}(F)\otimes I)$,
whose image in
${\bf H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$
is represented by
\begin{align*}
&\{ (\pi_{\gamma}\circ\nu_{\alpha})^{-1}
\circ\pi_{\gamma}\circ\nu_{\beta}\circ\pi_{\beta}
\circ\nu_{\alpha}\circ\pi_{\alpha}-\pi_{\alpha} \} \\
& = \{ (\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ
(\pi_{\gamma}\circ\nu_{\beta}\circ\pi_{\beta}
\circ\nu_{\alpha}\circ\pi_{\alpha}
-\pi_{\gamma}\circ\nu_{\alpha}\circ\pi_{\alpha}) \} \\
&= \{ (\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ\pi_{\gamma}
\circ(\nu_{\beta}\circ\pi_{\beta}-1)\circ
\nu_{\alpha}\circ\pi_{\alpha} \} \\
&= \{(\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ\pi_{\gamma}
\circ(-d^1_{\beta}\circ\tau_{\beta})\circ(1-d^1_{\alpha}
\circ\tau_{\alpha}) \} \\
&= \left\{(\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ
\left( \pi_{\gamma}\circ d^1_{\beta}\circ
(\tau_{\alpha}-\tau_{\beta})
-\pi_{\gamma}\circ d^1_{\beta}\circ(\tau_{\alpha}-\tau_{\beta})
\circ d^1_{\alpha}\circ\tau_{\alpha}\right) \right\}. \\
\end{align*}
Here we have
\begin{align*}
&\pi_{\gamma}\circ d^1_{\beta}\circ(\tau_{\alpha}-\tau_{\beta})\circ d^1_{\alpha}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ(\tilde{s}_1^{\alpha}\circ\tilde{r}^{\alpha}_0-\tilde{s}_1^{\beta}\circ\tilde{r}_0^{\beta})
\circ d^1_{\alpha}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}^{\alpha}_1\circ\tilde{r}_0^{\alpha}\circ d^1_{\alpha}\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ \tilde{s}^{\beta}_1\circ\tilde{r}_0^{\beta}\circ(d^1_{\alpha}-d^1_{\beta})\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}_1^{\beta}\circ\tilde{r}_0^{\beta}\circ d^1_{\beta}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}_1^{\alpha}\circ\tilde{r}_0^{\alpha}\circ\tilde{i}_0^{\alpha}\circ\tilde{p}_1^{\alpha}\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}_1^{\beta}\circ\tilde{r}_0^{\beta}\circ\tilde{i}_0^{\beta}\circ\tilde{p}_1^{\beta}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}_1^{\alpha}\circ\tilde{p}_1^{\alpha}\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{s}_1^{\beta}\circ\tilde{p}_1^{\beta}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ(\mathrm{id}-\tilde{i}_1^{\alpha}\circ\tilde{r}_1^{\alpha})\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ(\mathrm{id}-\tilde{i}_1^{\beta}\circ\tilde{r}_1^{\beta})\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{i}_1^{\alpha}\circ\tilde{r}_1^{\alpha}\circ\tau_{\alpha}
-\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{i}_1^{\beta}\circ\tilde{r}_1^{\beta}\circ\tau_{\alpha} \\
&=\pi_{\gamma}\circ d^1_{\beta}\circ\tilde{i}_1^{\alpha}\circ\tilde{r}_1^{\alpha}\circ\tilde{s}_1^{\alpha}\circ\tilde{r}_0^{\alpha}
\quad \text{(note that $d^1_{\beta}\circ\tilde{i}_1^{\beta}=0$)} \\
&=0. \quad (\text{note that $\tilde{r}_1^{\alpha}\circ\tilde{s}_1^{\alpha}=0$})
\end{align*}
So the image of $o(F)$ in $\mathbf{H}^2(C^{\bullet}({\mathcal Hom}^{\bullet}(V^{\bullet},F)\otimes I))$ is
\begin{align*}
\left\{(\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ\pi_{\gamma}
\circ d^1_{\beta}\circ(\tau_{\alpha}-\tau_{\beta}) \right\}
&=\left\{(\pi_{\gamma}\circ\nu_{\alpha})^{-1}\circ
\pi_{\gamma}\circ (d^1_{\beta}-d^1_{\gamma})\circ
(\tau_{\alpha}-\tau_{\beta}) \right\} \\
&=\left\{(\pi\otimes\mathrm{id}_I)\circ (d^1_{\beta}-d^1_{\gamma})
\circ(\tau_{\alpha}-\tau_{\beta}) \right\}
\end{align*}
Thus we have the equality
$f(\omega(F))=o(F)$.
\end{proof}
\begin{remark}\rm
Several authors introduced obstruction classes for the deformation
of vector bundles and coherent sheaves.
For example, [\cite{H-L}, Chap 2, Appendix]
is a good reference.
However, it is not so clear that these definitions
are all equivalent.
\end{remark}
\section{Smoothness and symplectic structure}
Let $X$ be a projective scheme over a noetherian scheme $S$,
which is flat over $S$.
We define a functor $\mathrm{Splcpx}_{X/S}$ of the category
of locally noetherian schemes to that of sets by putting
\[
\mathrm{Splcpx}_{X/S}(T):= \left\{ E^{\bullet} \left|
\begin{array}{l}
\mbox{$E^{\bullet}$ is a bounded complex of $T$-flat coherent} \\
\mbox{${\mathcal O}_{X_T}$-modules such that for any $t\in T$,} \\
\mbox{$E^{\bullet}(t)$ satisfies the following condition $(*)$}
\end{array}
\right\} \right/\sim,
\]
where $T$ is a locally noetherian scheme over $S$ and
$E^{\bullet}\sim F^{\bullet}$ if there is a line bundle $L$ on $T$ such that
$E^{\bullet}\cong F^{\bullet}\otimes L$ in $D(X_T)$.
Here $D(X_T)$ is the derived category of
${\mathcal O}_{X_T}$-modules and the condition $(*)$ is
\[
(*)\quad \mathop\mathrm{Ext}\nolimits^i(E^{\bullet}(t),E^{\bullet}(t))\cong
\begin{cases}
0 & \text{if $i=-1$} \\
k(t) & \text{if $i=0$}.
\end{cases}
\]
Note that we denote $E^{\bullet}\otimes^{\mathbf{L}}k(t)$ by $E^{\bullet}(t)$.
Let $\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/S}$ be the \'{e}tale sheafification of
$\mathrm{Splcpx}_{X/S}$.
\begin{theorem}
$\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/S}$ is represented by an algebraic space
over $S$.
\end{theorem}
(Proof is in [\cite{inaba1}, Theorem 0.2].
This result was generalized by Lieblich in \cite{Lieblich}
for $X$ proper over $S$.)
\begin{theorem}
If $X$ is an abelian or a projective K3 surface over an algebraically
closed field $k$,
$\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ is smooth over $k$.
\end{theorem}
\begin{proof}
Take an artinian local ring $A$ over $k$ with residue field $k=A/m$
and an ideal $I$ of $A$ such that $mI=0$.
It is sufficient to show that
$\mathrm{Splcpx}_{X/k}(A)\to\mathrm{Splcpx}_{X/k}(A/I)$ is surjective.
Indeed we can take a scheme $U$ locally of finite type over $k$
and a morphism $p:U\rightarrow\mathop\mathrm{Splcpx}\nolimits_{X/k}$ such that
the composite
$U\stackrel{p}\rightarrow\mathop\mathrm{Splcpx}\nolimits_{X/k}\stackrel{\iota}\rightarrow
\mathop\mathrm{Splcpx}\nolimits_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$
is \'etale and surjective.
Take any artinian local ring $A$ over $k$ with residue field $k=A/m$
and an ideal $I$ of $A$ such that $mI=0$.
Take any member $x\in U(A/I)$.
By the surjectivity of $\mathop\mathrm{Splcpx}\nolimits_{X/k}(A)\rightarrow\mathop\mathrm{Splcpx}\nolimits_{X/k}(A/I)$,
we can take an element $y\in\mathop\mathrm{Splcpx}\nolimits_{X/k}(A)$ such that
$y\otimes A/I=p(x)$.
Then $\iota(y)\in\mathop\mathrm{Splcpx}\nolimits_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}(A)$
and $\iota(y)\otimes A/I=(\iota\circ p)(x)$.
Since $\iota\circ p:U\rightarrow\mathop\mathrm{Splcpx}\nolimits_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$ is \'etale,
there is an element $z\in U(A)$ such that
$z\otimes A/I=x$ and $(\iota\circ p)(z)=y$.
Thus $U$ is smooth over $k$.
Let $E^{\bullet}$ be an $A/I$-valued point of $\mathrm{Splcpx}_{X/k}$.
Put $E^{\bullet}_0:=E^{\bullet}\otimes k$
and
\[
l':=\min \{ i | \text{$H^i( E^{\bullet}_0\otimes^{\mathbf{L}} k(x))\neq 0$ for some $x\in X$}\}.
\]
We may assume that $E^{\bullet}$ is of the form
\[
\cdots \longrightarrow 0 \longrightarrow 0 \longrightarrow E^{l'}
\stackrel{d^{l'}_{E^{\bullet}}}\longrightarrow V^{l'+1} \stackrel{d^{l'+1}}\longrightarrow \cdots \longrightarrow V^l
\stackrel{d^l}\longrightarrow 0 \longrightarrow 0 \cdots,
\]
where $E^{l'}$ is a vector bundle on $X_{A/I}$,
$V^i=V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)$
with $V_i$ a finite dimensional vector space over $k$,
${\mathcal O}_X(1)$ a fixed ample line bundle on $X$ and
$1\ll m_l \ll m_{l-1} \ll \cdots \ll m_{l'+1}$.
We can see that
$d^{l'}_{E^{\bullet}_0}\otimes k(x)$ is not injective for some $x\in X$.
Take a resolution
\[
\cdots\longrightarrow V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)\longrightarrow\cdots
\longrightarrow V_{l'}\otimes{\mathcal O}_{X_{A/I}}(-m_{l'})
\stackrel{\pi}\longrightarrow E^{l'}
\longrightarrow 0,
\]
where each $V_i$ is a vector space over $k$ of finite dimension and
\[
m_{l'+1}\ll m_{l'}\ll \cdots \ll m_i\ll m_{i-1} \ll \cdots.
\]
We put $V^i=V_i\otimes{\mathcal O}_{X_{A/I}}(-m_i)$
for $i\leq l$ and $V^i=0$ for $i>l$.
Let $V^{\bullet}$ be the complex
\[
\cdots\longrightarrow V^i\longrightarrow V^{i+1}
\longrightarrow\cdots\longrightarrow V^{l'}
\xrightarrow{d^{l'}_{E^{\bullet}}\circ\pi}
V^{l'+1}\longrightarrow\cdots\longrightarrow
V^l\longrightarrow 0\longrightarrow\cdots.
\]
Then there is a canonical quasi-isomorphism
\[
V^{\bullet} \longrightarrow E^{\bullet}.
\]
Put $V^{\bullet}_0:=V^{\bullet}\otimes k$.
Let
\[
\mathrm{tr}^{\bullet}:{\mathcal Hom}^{\bullet}(E^{\bullet}_0,E^{\bullet}_0)
\stackrel{\sim}\longrightarrow
{\mathcal Hom}^{\bullet}({\mathcal Hom}
^{\bullet}(E^{\bullet}_0,E^{\bullet}_0),{\mathcal O}_X)
\longrightarrow {\mathcal O}_X
\]
be the dual of the canonical morphism
\[
{\mathcal O}_X \longrightarrow
{\mathcal Hom}^{\bullet}(E^{\bullet}_0,E^{\bullet}_0) ;
\quad 1\mapsto \mathrm{id}_{E^{\bullet}_0}.
\]
Note that
$\mathrm{tr}^p=0$ on ${\mathcal Hom}^p(E_0^{\bullet},E_0^{\bullet})$ for $p\neq 0$
and $\mathrm{tr}^0(\{x^i\})=\sum_i (-1)^i \mathrm{tr}(x^i)$
for $x^i \in {\mathcal Hom}(E^i_0,E^i_0)$.
$\mathrm{tr}^{\bullet}$ is also introduced in [\cite{H-L}, Chapter 10].
There is a commutative diagram
\[
\begin{CD}
\mathop\mathrm{Ext}\nolimits^2_X(E^{\bullet}_0,E^{\bullet}_0) @>H^2(\mathrm{tr}^{\bullet})>>
H^2(X,{\mathcal O}_X) \\
@V s_1 V\cong V @V s_2 V\cong V \\
\mathop\mathrm{Hom}\nolimits_{D(X)}(E^{\bullet}_0,E^{\bullet}_0)^{\vee} @>>>
H^0(X,{\mathcal O}_X)^{\vee},
\end{CD}
\]
where $s_1,s_2$ are the isomorphisms determined by
Grothendieck-Serre duality and the bottom row is the dual of
$k=H^0({\mathcal O}_X)\to \mathop\mathrm{Hom}\nolimits_{D(X)}(E^{\bullet}_0,E^{\bullet}_0)$,
which is bijective since $E^{\bullet}_0$ is simple.
Thus the homomorphism
\[
\mathop\mathrm{Ext}\nolimits^2_X(E^{\bullet}_0,E^{\bullet}_0)
\xrightarrow{H^2(\mathrm{tr}^{\bullet})} H^2(X,{\mathcal O}_X)
\]
is an isomorphism.
Note that there is a commutative diagram
\[
\begin{CD}
{\mathcal Hom}^{\bullet}(E^{l'}[-l'],I\otimes E^{\bullet})
@>>> {\mathcal Hom}(E^{l'},I\otimes E^{l'}) \\
@VVV @VV (-1)^{l'}\mathrm{tr} V \\
{\mathcal Hom}^{\bullet}(E^{\bullet},I\otimes E^{\bullet})
@>\mathrm{tr}>> {\mathcal O}_X\otimes I.
\end{CD}
\]
From the above commutative diagram, we obtain a commutative diagram
\[
(\dag\dag) \quad
\begin{CD}
\mathop\mathrm{Ext}\nolimits^2(E^{l'}[-l'],I\otimes E^{\bullet}) @>\tau>> \mathop\mathrm{Ext}\nolimits^2(E^{l'},I\otimes E^{l'}) \\
@V\sigma VV @V (-1)^{l'}H^2(\mathrm{tr}) VV \\
\mathop\mathrm{Ext}\nolimits^2(E^{\bullet},I\otimes E^{\bullet}) @>H^2(\mathrm{tr})>> H^2({\mathcal O}_X)\otimes I.
\end{CD}
\]
Note that the morphism
\[
\mathop\mathrm{Hom}\nolimits(E^{\bullet}_0,E^{\bullet}_0)\longrightarrow \mathop\mathrm{Hom}\nolimits(E^{\bullet}_0,E^{l'}_0[-l'])
\]
is not zero, since the image of $\mathrm{id}$ by this morphism
is the canonical morphism $\iota:E^{\bullet}_0\rightarrow E^{l'}_0[-l']$
which is not zero because
\[
H^{l'}(\iota\otimes k(x)):\ker(d^{l'}_{E^{\bullet}_0}\otimes k(x))= H^{l'}(E^{\bullet}_0\otimes k(x))\longrightarrow
H^l(E^{l'}_0[-l']\otimes k(x))=E^{l'}_0\otimes k(x)
\]
is not zero.
By Grothendieck-Serre duality, we can see that
\[
\iota^*:\mathop\mathrm{Ext}\nolimits^2(E^{l'}_0[-l'],E^{\bullet}_0)\longrightarrow \mathop\mathrm{Ext}\nolimits^2(E^{\bullet}_0,E^{\bullet}_0)
\]
is not zero.
Since $\mathop\mathrm{Ext}\nolimits^2(E^{\bullet}_0,E^{\bullet}_0)\cong k$, $\iota^*$ is surjective.
So the morphism
\[
\sigma: \mathop\mathrm{Ext}\nolimits^2(E^{l'}[-l'],I\otimes E^{\bullet})\longrightarrow
\mathop\mathrm{Ext}\nolimits^2(E^{\bullet},I\otimes E^{\bullet})
\]
is also surjective.
Take an obstruction class $\omega(E^{\bullet})\in\mathop\mathrm{Ext}\nolimits^2(E^{\bullet},I\otimes E^{\bullet})$
for the lifting of $E^{\bullet}$ to an $A$-valued point of $\mathop\mathrm{Splcpx}\nolimits_{X/k}$.
Then there is a member $\varphi=[(\varphi^i)]\in \mathop\mathrm{Ext}\nolimits^2(E^{l'}[-l'], I\otimes E^{\bullet})$
such that
$\sigma(\varphi)=\omega(E^{\bullet})$.
Here $\varphi^i:V^i\rightarrow I\otimes E^{i+2}$ ($i\leq l'$)
and $\varphi^i=0$ for $i>l'$.
There is an element $\gamma=(\gamma^i)\in \mathop\mathrm{Hom}\nolimits^1(V^{\bullet},I\otimes E^{\bullet})$ such that
\begin{gather*}
\gamma^{i+1}\circ d^i_{V^{\bullet}}+d^{i+1}_{E^{\bullet}}\circ\gamma^i=\tilde{d}^{i+1}\circ\tilde{d}^i-\varphi^i
\quad (\text{for $i\geq l'-1$}) \\
\gamma^{l'-1}\circ d^{l'-2}_{V^{\bullet}}=\pi\circ\tilde{d}^{l'-1}\circ\tilde{d}^{l'-2}-\varphi^{l'-2},
\end{gather*}
where $\tilde{d}^i:V_i\otimes{\mathcal O}_{X_A}(-m_i)\rightarrow V_{i+1}\otimes{\mathcal O}_{X_A}(-m_{i+1})$
is a lift of $d^i_{V^{\bullet}}$.
We can see that the image of $\varphi$ by the morphism
$\tau:\mathop\mathrm{Ext}\nolimits^2(E^{l'}[-l'],I\otimes E^{\bullet})\rightarrow \mathop\mathrm{Ext}\nolimits^2(E^{l'},I\otimes E^{l'})$
is given by
$[\pi\circ\tilde{d}^{l'-1}\circ\tilde{d}^{l'-2}]$,
which is just the obstruction class $\omega(E^{l'})$.
By Lemma \ref{obstruction=}, we have $\omega(E^{l'})=o(E^{l'})$.
We can see that $H^2(\mathrm{tr})(o(E^{l'}))=o(\det(E^{l'}))$.
Since the Picard scheme $\mathop\mathrm{Pic}\nolimits_{X/k}$ is smooth over $k$,
we have $o(\det(E^{l'}))=0$.
So we have
\begin{align*}
H^2(\mathrm{tr})(\omega(E^{\bullet})) &= H^2(\mathrm{tr})(\sigma(\varphi)) \\
&= (-1)^{l'}H^2(\mathrm{tr})(\tau(\varphi)) \\
&= (-1)^{l'}H^2(\mathrm{tr})(\omega(E^{l'})) \\
&=(-1)^{l'}H^2(\mathrm{tr})(o((E^{l'}))) \\
&=(-1)^{l'}o(\det(E'))=0.
\end{align*}
Since the morphism
\[
H^2(\mathrm{tr}):\mathop\mathrm{Ext}\nolimits^2(E,I\otimes E) \longrightarrow
H^2({\mathcal O}_X)\otimes I
\]
is isomorphic, we have
$\omega(E^{\bullet})=0$.
Thus $\mathop\mathrm{Splcpx}\nolimits^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ is smooth over $k$.
\end{proof}
The following theorem is essentially proved in
[\cite{H-L},II-10].
We give a proof again.
\begin{theorem}
Let $X$ be an abelian or a projective K3 surface over
an algebraically closed field $k$.
Then $\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ has a symplectic structure,
that is, there exists a closed $2$-form on
$\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$
which is nondegenerate at every point.
\end{theorem}
\begin{proof}
Note that the tangent bundle $T_{\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}}$
on $\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ can be considered as the sheaf
on the small \'{e}tale site on $\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ defined by
\[
U\mapsto
\left\{ v\in \mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}(U_{k[\epsilon]}) \left|
\begin{array}{l}
\text{the composite
$U\stackrel{i_0}\rightarrow U_{k[\epsilon]}
\stackrel{v}\rightarrow \mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$} \\
\mbox{is the structure morphism $U\to \mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$}
\end{array}
\right\}\right.,
\]
for any algebraic space $U$ \'{e}tale over $\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$,
where $k[\epsilon]$ is the $k$-algebra generated by $\epsilon$
with $\epsilon^2=0$ and
$U\stackrel{i_0}\rightarrow U_{k[\epsilon]}$
is the morphism induced by the ring homomorphism
\[
k[\epsilon]\longrightarrow k; \quad \epsilon\mapsto 0.
\]
There is an \'{e}tale covering
$\coprod_i U_i \to \mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$
such that $U_i\to \mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}$ factors through
$\mathrm{Splcpx}_{X/k}$, that is, there is a universal family
$E^{\bullet}_{U_i}$ on each $X_{U_i}$.
Let $U$ be an affine scheme \'{e}tale over $\coprod_i U_i$
and $E^{\bullet}_U$ be the pull-back of the universal family.
Take any element $v\in T_{\mathrm{Splcpx}^{\mbox{\rm \scriptsize{\'{e}t}}}_{X/k}}(U)$.
Then $U_{k[\epsilon]}\stackrel{v}\longrightarrow\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$
factors through $\coprod_i U_i$,
since $\coprod_i U_i$ is \'etale over $\mathop\mathrm{Splcpx}\nolimits_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$.
Let $E^{\bullet}_{U_{k[\epsilon]}}\in\mathrm{Splcpx}_{X/k}(U_{k[\epsilon]})$
be the pull-back of the universal family.
We can take a complex $\tilde{V}^{\bullet}$ of the form
$\tilde{V}^i=V_i\otimes_{{\mathcal O}_U}
{\mathcal O}_{X_{U_{k[\epsilon]}}}(-m_i)$
and a quasi-isomorphism $\tilde{V}^{\bullet}\to E^{\bullet}_{U_{k[\epsilon]}}$,
where $V_i$ is a locally free sheaf of finite rank on $U$,
$V_i=0$ for $i\gg 0$,
${\mathcal O}_X(1)$ is a fixed ample line bundle on $X$
and $\cdots\gg m_i\gg m_{i+1} \gg\cdots$.
Let $V^{\bullet}$ be the pull-back of $\tilde{V}^{\bullet}$ by
$X\times U\xrightarrow{\mathrm{id}_X\times i_0}
X\times U_{k[\epsilon]}$.
Then we obtain an element
\[
[\{ d^i_{\tilde{V}^{\bullet}}-d^i_{V^{\bullet}}\otimes 1 \}]
\in H^1(\mathop\mathrm{Hom}\nolimits(\tilde{V}^{\bullet},\epsilon k[\epsilon]\otimes \tilde{V}^{\bullet}))
\cong \mathop\mathrm{Ext}\nolimits^1(E^{\bullet}_U,E^{\bullet}_U),
\]
which is independent of the choice of the representative
$\tilde{V}^{\bullet}$.
We can see that the mapping
$v\to [\{ d^i_{\tilde{V}^{\bullet}}-d^i_{V^{\bullet}}\otimes 1 \}]$
defines an isomorphism
\[
T_{\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}}(U)\stackrel{\sim}\longrightarrow
H^0(U,\mathop\mathrm{Ext}\nolimits^1_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U)).
\]
For an affine scheme $U$ \'{e}tale over $\coprod_i U_i$,
there is a canonical pairing:
\[
\begin{array}{ccc}
\alpha_U:\mathop\mathrm{Ext}\nolimits^1_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U)\times
\mathop\mathrm{Ext}\nolimits^1_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U) & \longrightarrow
& \mathop\mathrm{Ext}\nolimits^2_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U) \\
(g,h) & \mapsto & g\circ h.
\end{array}
\]
Note that there are canonical isomorphisms
\[
\mathop\mathrm{Ext}\nolimits^2_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U)\stackrel{\sim}\longrightarrow
\mathop\mathrm{Ext}\nolimits^0_{X_U/U}(E^{\bullet}_U,E^{\bullet}_U)^{\vee}
\stackrel{\sim}\longrightarrow {\mathcal O}_U.
\]
Then we can obtain a pairing
\[
\alpha: T_{\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}}\times T_{\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}}
\longrightarrow {\mathcal O}_{\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}}
\]
by patching $\alpha_U$.
Now we will see that $\alpha$ is skew-symmetric.
Take any $k$-valued point $p$ of
$\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$.
$p$ corresponds to a complex
\[
\cdots \longrightarrow V_i\otimes{\mathcal O}_X(-m_i)
\stackrel{d^i_{V^{\bullet}}}\longrightarrow
V_{i+1}\otimes{\mathcal O}_X(-m_{i+1})
\stackrel{d^{i+1}_{V^{\bullet}}}\longrightarrow\cdots
\longrightarrow V_l\otimes{\mathcal O}_X(-m_l)
\longrightarrow 0\longrightarrow\cdots
\]
We denote this complex by $V^{\bullet}$.
Let us consider the restriction
\[
\alpha(p):\mathop\mathrm{Ext}\nolimits^1(V^{\bullet},V^{\bullet})\times \mathop\mathrm{Ext}\nolimits^1(V^{\bullet},V^{\bullet})
\longrightarrow \mathop\mathrm{Ext}\nolimits^2(V^{\bullet},V^{\bullet})\cong k
\]
of the pairing $\alpha$.
Take any element
$v=[\{v^i\}]\in H^1(\mathop\mathrm{Hom}\nolimits^{\bullet}(V^{\bullet},V^{\bullet}))
\cong\mathop\mathrm{Ext}\nolimits^1(V^{\bullet},V^{\bullet})$
and let $\tilde{V}^{\bullet}$ be a member of
$\mathop\mathrm{Splcpx}\nolimits_{X/k}(k[\epsilon])$
which corresponds to $v$.
$\tilde{V}^{\bullet}$ can be given by the complex
\[
\cdots\longrightarrow
V_i\otimes{\mathcal O}_X(-m_i)\otimes k[\epsilon]
\xrightarrow{d^i_{V^{\bullet}}+\epsilon v^i}
V_{i+1}\otimes{\mathcal O}_X(-m_{i+1})\otimes k[\epsilon]
\longrightarrow\cdots
\]
Consider the surjection
$k[t]/(t^3)\to k[\epsilon];\: t\mapsto \epsilon$
and the extension
\[
d^i_{V^{\bullet}}+tv^i:V_i\otimes{\mathcal O}_X(-m_i)\otimes k[t]/(t^3)
\longrightarrow V_{i+1}\otimes{\mathcal O}_X(-m_{i+1})\otimes k[t]/(t^3)
\]
of the homomorphism
$d^i_{V^{\bullet}}+\epsilon v^i:V_i\otimes{\mathcal O}_X(-m_i)\otimes k[\epsilon]
\rightarrow V_{i+1}\otimes{\mathcal O}_X(-m_{i+1})\otimes k[\epsilon]$.
Then the obstruction class $\omega(\tilde{V}^{\bullet})$
for the lifting of $\tilde{V}^{\bullet}$ to a member of
$\mathop\mathrm{Splcpx}\nolimits_{X/k}(k[t]/(t^3))$
with respect to the surjection
$k[t]/(t^3)\to k[\epsilon];\: t\mapsto \epsilon$
is given by
$[\{(d^{i+1}_{V^{\bullet}}+tv^{i+1})\circ(d^i_{V^{\bullet}}+tv^i)\}]
\in(t^2)\otimes\mathop\mathrm{Ext}\nolimits^2(V^{\bullet},V^{\bullet})$.
However,
\[
(d^{i+1}_{V^{\bullet}}+tv^{i+1})\circ (d^i_{V^{\bullet}}+tv^i)=
d^{i+1}_{V^{\bullet}}\circ d^i_{V^{\bullet}}
+t(d^{i+1}_{V^{\bullet}}\circ v^i+v^{i+1}\circ d^i_{V^{\bullet}})
+t^2v^{i+1}\circ v^i=t^2v^{i+1}\circ v^i.
\]
Then
$\alpha(p)(v,v)=v\circ v=[\{v^{i+1}\circ v^i\}]=\omega(\tilde{V}^{\bullet})=0$
since $\mathrm{Splcpx}_{X/k}^{\mbox{\rm \scriptsize{\'{e}t}}}$ is smooth over $k$.
Next we will see that $\alpha$ is nondegenerate.
The canonical isomorphism
\[
{\bf R}{\mathcal Hom}(E^{\bullet}_U,E^{\bullet}_U)\stackrel{\sim}\longrightarrow
{\bf R}{\mathcal Hom}^{\bullet}(E^{\bullet}_U,E^{\bullet}_U)^{\vee}
\]
induces the composite isomorphism by Grothendieck-Serre duality
\[
\mathop\mathrm{Ext}\nolimits^1(E^{\bullet}_U,E^{\bullet}_U)\stackrel{\sim}\longrightarrow
\mathop\mathrm{Ext}\nolimits^1({\bf R}{\mathcal Hom}^{\bullet}
(E^{\bullet}_U,E^{\bullet}_U),{\mathcal O}_{X_U})
\stackrel{\sim}\longrightarrow
\mathop\mathrm{Hom}\nolimits(\mathop\mathrm{Ext}\nolimits^1(E^{\bullet}_U,E^{\bullet}_U),{\mathcal O}_U),
\]
which is just the homomorphism induced by $\alpha$.
Thus $\alpha$ is nondegenerate.
Finally we will show that $\alpha$ is $d$-closed.
For an affine scheme $U$ \'etale over $\coprod_i U_i $,
take $u,v,w\in T_{\mathop\mathrm{Splcpx}\nolimits_{X/k}^{^{\mbox{\rm \scriptsize{\'{e}t}}}}}(U)$.
Let $E^{\bullet}\in\mathop\mathrm{Splcpx}\nolimits_{X/k}(U)$ be the pullback
of the universal family.
We may assume that there exists a complex
$V^{\bullet}$ of the form
$V^i=V_i\otimes{\mathcal O}_{X_U}(-m_i)$
such that $V^{\bullet}$ is quasi-isomorphic to $E^{\bullet}$
and that $V_i$ are vector spaces of finite dimension over $k$
and $m_i$ are integers.
Take $u\in T_U(U)$.
$u$ can be regarded as a derivation
${\mathcal O}_U\rightarrow{\mathcal O}_U$
over ${\mathcal O}_U$,
which is canonically extended to a derivation
\[
D_u:V_i^{\vee}\otimes V_j\otimes{\mathcal O}_X(m_i-m_j)\otimes{\mathcal O}_U
\longrightarrow
V_i^{\vee}\otimes V_j\otimes{\mathcal O}_X(m_i-m_j)\otimes{\mathcal O}_U
\]
for $i\leq j$.
We have
$d_{V^{\bullet}}^{i+1}\circ D_u(d_{V^{\bullet}}^i)
+D_u(d_{V^{\bullet}}^{i+1})\circ d_{V^{\bullet}}^i=0$
for any $i$.
So we have
$[\{D_u(d_{V^{\bullet}}^i)\}]\in\mathop\mathrm{Ext}\nolimits^1(V^{\bullet},V^{\bullet})$,
which corresponds to $u$ by the isomorphism
$T_U(U)\stackrel{\sim}\rightarrow\mathop\mathrm{Ext}\nolimits^1(V^{\bullet},V^{\bullet})$.
Note that for $u,v\in T_U(U)$ we have
\[
\alpha(u,v)=
\left[\left\{D_u(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i)\right\}\right]
\in\mathop\mathrm{Ext}\nolimits^2(V^{\bullet},V^{\bullet})\cong H^0(U,{\mathcal O}_U).
\]
For $u,v,w\in T_U(U)$, we have
\begin{align*}
d\alpha(u,v,w)&=\left[\{
D_u(\alpha(v,w))+D_v(\alpha(w,u))+D_w(\alpha(u,v))
+\alpha(w,[u,v])+\alpha([u,w],v)+\alpha(u,[v,w]) \}\right] \\
&=\left[\left\{ D_u(D_v(d_{V^{\bullet}}^{i+1})\circ D_w(d_{V^{\bullet}}^i))
+D_v(D_w(d_{V^{\bullet}}^{i+1})\circ D_u(d_{V^{\bullet}}^i))
+D_w(D_u(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i))
\right.\right. \\
& \left.\left. \quad
+D_w(d_{V^{\bullet}}^{i+1})\circ(D_uD_v-D_vD_u)(d_{V^{\bullet}}^i)
+(D_uD_w-D_wD_u)(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i)
\right.\right. \\
& \quad \left.\left.
+D_u(d_{V^{\bullet}}^{i+1})\circ(D_vD_w-D_wD_v)(d_{V^{\bullet}}^i)
\right\}\right] \\
&=\left[\left\{
D_uD_v(d_{V^{\bullet}}^{i+1})\circ D_w(d_{V^{\bullet}}^i)
+D_v(d_{V^{\bullet}}^{i+1})\circ D_uD_w(d_{V^{\bullet}}^i)
+D_vD_w(d_{V^{\bullet}}^{i+1})\circ D_u(d_{V^{\bullet}}^i)
\right.\right. \\
&\quad
+D_w(d_{V^{\bullet}}^{i+1})\circ D_vD_u(d_{V^{\bullet}}^i)
+D_wD_u(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i)
+D_u(d_{V^{\bullet}}^{i+1})\circ D_wD_v(d_{V^{\bullet}}^i) \\
&\quad
+D_w(d_{V^{\bullet}}^{i+1})\circ D_uD_v(d_{V^{\bullet}}^i)
-D_w(d_{V^{\bullet}}^{i+1})\circ D_vD_u(d_{V^{\bullet}}^i)
+D_uD_w(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i) \\
&\quad \left.\left.
-D_wD_u(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i)
+D_u(d_{V^{\bullet}}^{i+1})\circ D_vD_w(d_{V^{\bullet}}^i)
-D_u(d_{V^{\bullet}}^{i+1})\circ D_wD_v(d_{V^{\bullet}}^i)
\right\}\right] \\
&=\left[\left\{
D_uD_v(d_{V^{\bullet}}^{i+1})\circ D_w(d_{V^{\bullet}}^i)
+D_v(d_{V^{\bullet}}^{i+1})\circ D_uD_w(d_{V^{\bullet}}^i)
+D_vD_w(d_{V^{\bullet}}^{i+1})\circ D_u(d_{V^{\bullet}}^i)
\right.\right. \\
& \quad \left.\left.
+D_w(d_{V^{\bullet}}^{i+1})\circ D_uD_v(d_{V^{\bullet}}^i)
+D_uD_w(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i)
+D_u(d_{V^{\bullet}}^{i+1})\circ D_vD_w(d_{V^{\bullet}}^i)
\right\}\right] \\
&=\left[
\left\{ D_uD_vD_w(d_{V^{\bullet}}^{i+1}\circ d_{V^{\bullet}}^i)\right\}
-\left\{ D_uD_vD_w(d_{V^{\bullet}}^{i+1})\circ d_{V^{\bullet}}^i
+d_{V^{\bullet}}^{i+1}\circ D_uD_vD_w(d_{V^{\bullet}})
\right\} \right] \\
&=\left[ \{D_uD_vD_w(0)\}-d\left\{D_uD_vD_w(d_{V^{\bullet}}^i)\right\}\right] \\
&=0.
\end{align*}
Here note that
\begin{align*}
D_uD_vD_w(d_{V^{\bullet}}^{i+1}\circ d_{V^{\bullet}}^i)
&=D_u(D_v(D_w(d_{V^{\bullet}}^{i+1}\circ d_{V^{\bullet}}^i)) )\\
&=D_u(D_v(D_w(d_{V^{\bullet}}^{i+1})\circ d_{V^{\bullet}}^i
+d_{V^{\bullet}}^{i+1}\circ D_w(d_{V^{\bullet}}^i))) \\
&=D_u\left(D_vD_w(d_{V^{\bullet}}^{i+1})\circ d_{V^{\bullet}}^i
+D_w(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i) \right. \\
&\quad \left.
+D_v(d_{V^{\bullet}}^{i+1})\circ D_w(d_{V^{\bullet}}^i)
+d_{V^{\bullet}}^{i+1}\circ D_vD_w(d_{V^{\bullet}}^i)\right) \\
&= D_uD_vD_w(d_{V^{\bullet}}^{i+1})\circ d_{V^{\bullet}}^i
+D_vD_w(d_{V^{\bullet}}^{i+1})\circ D_u(d_{V^{\bullet}}^i)
+D_uD_w(d_{V^{\bullet}}^{i+1})\circ D_v(d_{V^{\bullet}}^i) \\
& \quad +D_w(d_{V^{\bullet}}^{i+1})D_uD_v(d_{V^{\bullet}}^i)
+D_uD_v(d_{V^{\bullet}}^{i+1})\circ D_w(d_{V^{\bullet}}^i)
+D_v(d_{V^{\bullet}}^{i+1})\circ D_uD_w(d_{V^{\bullet}}^i) \\
& \quad +D_u(d_{V^{\bullet}}^{i+1})\circ D_vD_w(d_{V^{\bullet}}^i)
+d_{V^{\bullet}}^{i+1}\circ D_uD_vD_w(d_{V^{\bullet}}^i)
\end{align*}
So $\alpha$ is a closed $2$-form.
\end{proof}
\noindent
{\bf Acknowledgments.}
The author would like to thank Professors
Akira Ishii and K\={o}ta Yoshioka
for giving him the problem solved in this paper.
The author would also like to thank Professor
Fumiharu Kato for teaching him a fundamental
concepts of algebraic spaces.
|
\section*{Acknowledgements}
The authors acknowledge some financial support from the Spanish project
DGICYT-FIS2009-13364-C02-01. J.S. also thanks to the Consejer\'ia de
Econom\'ia, Comercio e Innovaci\'on of the Junta de Extremadura (Spain) for
financial support, Project Ref. GRU09011.
|
\section{Introduction and statement of results}
\label{sec1}
\subsection{Introduction}
In this paper we investigate the integrability properties associated with the second order
nonlinear ordinary differential equation (ODE) of the form
\begin{eqnarray}
\ddot{x}+k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^3+k_5x^2+k_6x=0,
\label{eq6}
\end{eqnarray}
where $k_i$'s, $i=1,2,...,6,$ are arbitrary parameters and the over dot denotes
differentiation with respect to `$t$'. Equation (\ref{eq6}) is a combination of two
different classes of equations, namely Li\'enard type equation ($k_1=0$) and equation with
quadratic friction ($k_2=k_3=0$). Using the transformation $z=\dot{x}$ one can reduce
Eq. (\ref{eq6}) to the Abel equation of the second kind{\footnotesize$^{1,2}$}. However, the
resultant equation is not integrable in general. A set of integrable parametric choices of this
equation has been listed in Refs. 1 \& 2. The importance of the study of Eq. (\ref{eq6})
arises from the fact that this equation contains many physically interesting equations such as the
modified Emden equation, unforced Duffing oscillator, Helmholtz oscillator, etc and is intimately
related to the much analyzed biological model, namely the two dimensional
Lotka-Volterra system{\footnotesize$^{3,4}$} (LV),
\begin{subequations}
\begin{eqnarray}
&&\dot {x}=x(a_{1}+b_{11}x+b_{12}y),\label {eq1a}\\
&&\dot {y}=y(a_{2}+b_{21}x+b_{22}y).\label {eq1b}
\end{eqnarray}
\label {eq1}
\end{subequations}
In order to study the integrable properties of Eq. (\ref{eq6}) we employ the modified Prelle-Singer procedure (PS) which has been intensively used to identify new integrable cases for several equations{\footnotesize$^{5,6}$}. Using this procedure we identify several new integrable
parametric choices and deduce their corresponding integrals of motion. In order to prove the integrability we either associate a conservative Hamiltonian structure to the equation constructed from the time independent integrals of motion, thereby proving Liouville sense of integrability, or explicitly integrate the integrals of motion to obtain general solution, proving complete integrability.
During the past three decades
or so several attempts have been made to explore the integrable cases in the LV system
because of the immense importance of the problem in mathematical ecology and biology{\footnotesize$^{7-12}$}. Integrals of motion of the LV system have been constructed for several parametric choices{\footnotesize$^{9,10,12-19}$}. However, we find that all these parametric choices effectively reduce to one of the three general parametric choices for which integrals of motion have been reported{\footnotesize$^{9}$}. Interestingly,
we recover all these three integrable cases in addition to many subcases by comparing and expressing the parameters
appearing in (\ref{eq6}) in terms of the LV system parameters. The results exactly coincide with the
reported ones in the literature. Thus, as a by-product, we recover all the known integrable cases
of LV by investigating the integrability properties of Eq. (\ref{eq6}). At this point
we stress the fact that as far as integrability is concerned one can extract complete information
about (\ref{eq1}) from (\ref{eq6}) and not vice-versa. In other words when one tries to deduce
the integrable cases of (\ref{eq6}) from (\ref{eq1}) only partial results can be extracted. That is one will loose a major portion of the results when we proceed in the reverse way (see
Sec. VI for more details).
\subsection{Results}
Our main results can be summarized as follows.
Using the modified PS method we identify seven integrable parametric choices of which five seem to be new as far as our knowledge goes. The underlying forms of the equations are
\begin{eqnarray}
&&\hspace{-1.2cm}\ddot{x}+k_1
\frac{\dot{x}^2}{x}+(k_2+ k_3x)\dot{x} +
k_4x^3+\frac{k_2k_4(3+2k_1)}{k_3(1+k_1)}x^
+\frac{k_4k_2^2(2+k_1)}{k_3^2(1+k_1)}x=0,\label{tid5}\\
&&\hspace{-1.2cm}\ddot{x}+k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}
+\frac{(1+k_1)k_3^2}{(3+2k_1)^2}x^3+\frac{k_2k_3}{(3+2k_1)}x^2+k_6x=0,\label{caseia,b}\\
&&\hspace{-1.2cm}\ddot{x}+k_1 \frac{\dot{x}^2}{x}+\left(k_2+k_3x\right)\dot{x}+\frac{k_3k_2(1+k_1)}{(3+2k_1)} x^2
+\frac{(2+k_1)k_2^2}{(3+2k_1)^2} x=0,\label{td2-1}\\
&&\hspace{-1.2cm}\ddot{x}+k_1 \frac{\dot{x}^2}{x}+k_2\dot{x}+k_5 x^2+\frac{2(3+2 k_1)k_2^2}{(5+4k_1)^2} x=0,\label{td3eq}\\
&&\hspace{-1.2cm}\ddot{x}+k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^3+\frac{k_2k_3}{(3+2k_1)} x^2
+\frac{(2+k_1)k_2^2}{(3+2k_1)^2}x=0.\label{eqiv}
\end{eqnarray}
For these five cases we prove integrability either in the Liouville sense by constructing
the time independent Hamiltonian structure or deduce the
solution by explicitly integrating the time dependent integrals of motion.
The remaining two cases which are already known in the literature are
\begin{eqnarray}
&&\ddot{x}+k_1 \frac{\dot{x}^2}{x}+k_4 x^3+k_5 x^2+k_6 x=0\,\label{bern1}\\
&&\hspace{-5cm}\mbox{and}\nonumber\\
&&\ddot{x}+k_3 x\dot{x}+k_4 x^3+k_6x=0\,\label{tid4}.
\end{eqnarray}
We note that equation (\ref{bern1}) can be reduced to the Bernoulli's equation, which can then be integrated
to give the solution in terms of quadratures. Equation (\ref{tid4}) is the modified Emden type equation with linear forcing term which has been studied in some detail in Refs. 20 \& 21. The Hamiltonian
structure for this equation for the parametric choices $k_3^2\ge8k_4$ has been given in Ref. 20.
Here we present the Hamiltonian structure for the parametric choice $k_3^2<8k_4$ as well. In addition to these seven integrable parameteric choices we find the following equation
\begin{eqnarray}
\ddot{x}+k_1 \frac{\dot{x}^2}{x}+(k_2+k_3x)
\dot{x}+\frac{k_3(k_2\pm\omega)}{2(2+k_1)} x^2+k_6
x=0,\label{eq87}
\end{eqnarray}
where $\omega=\sqrt{k_2^2-4(1+k_1)k_6}$, for which we obtain a
time-dependent integral of motion.
However, we are able to prove its complete integrability only with an additional parametric restriction $k_5=\frac{k_3k_2(1+k_1)}{(3+2k_1)}$, and $k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2}$, which reduces Eq. (\ref{eq87}) to Eq. (\ref{td2-1}). In this parametric choice we are able to explicitly integrate the time-dependent integral of motion to find the general solution.
The plan of the paper is as follows. In the following section we briefly describe the
extended Prelle-Singer procedure{\footnotesize$^{5,6}$} applicable to second-order ODEs. In Sec. \ref{sectid}
we identify the integrable parametric choices of Eq. (\ref{eq6}) which admit time independent integral of motion through the extended
PS procedure. In Sec. IV we construct explicit conservative Hamiltonians from the time independent integral of motion. Further we transform these Hamiltonians to simpler forms using canonical transformations in order to explicitly integrate the canonical equations of motion. In Sec. \ref{sectd}, we then identify the integrable cases of (\ref{eq6}) which admit explicit time-dependent
integrals of motion. To establish the complete integrability of these cases
we transform the time-dependent integrals of motion
into time-independent integrals of motion and integrate the latter and derive the
general solution. In Sec. \ref{connection} we present the connection between Eq. (\ref{eq6}) and the LV equation. In Sec. VII we rewrite the results in terms of the LV parameters
and point out the integrable equations. Finally, we present our conclusions in Sec. \ref{seccon}.
\section{Extended Prelle-Singer (PS) procedure}
\label{sec2}
\noindent Let us rewrite Eq.~(\ref{eq6}) in the form
\begin{eqnarray}
\ddot{x}=-(k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^{3}+k_5x^{2}+k_6x)
\equiv \phi(x,\dot{x}).
\label{met1}
\end{eqnarray}
Further, we assume that Eq. (\ref{met1})
admits a first integral $I(t,x,\dot{x})=C,$ with $C$ constant on the
solutions, so that the total differential becomes
\begin{eqnarray}
dI={I_t}{dt}+{I_{x}}{dx}+{I_{\dot{x}}{d\dot{x}}}=0,
\label{met3}
\end{eqnarray}
where subscript denotes partial differentiation with respect
to that variable. Rewriting Eq.~(\ref{met1}) in the form
$\phi dt-d\dot{x}=0$ and adding a null term
$S(t,x,\dot{x})\dot{x}dt - S(t,x,\dot{x})dx$ to the latter, we obtain that on
the solutions the 1-form
\begin{eqnarray}
(\phi +\dot{x}S) dt-Sdx-d\dot{x} = 0.
\label{met6}
\end{eqnarray}
Hence, on the solutions, the 1-forms (\ref{met3}) and
(\ref{met6}) must be proportional. Multiplying (\ref{met6}) by the
function $ R(t,x,\dot{x})$ which acts as the integrating factor
for (\ref{met6}), we have on the solutions that
\begin{eqnarray}
dI=R(\phi+S\dot{x})dt-RSdx-Rd\dot{x}=0.
\label{met7}
\end{eqnarray}
Comparing Eq.~(\ref{met3})
with (\ref{met7}), we have the relations
\begin{eqnarray}
I_{t} = R(\phi+\dot{x}S),\quad
I_{x} = -RS, \quad
I_{\dot{x}} = -R.
\label{met8}
\end{eqnarray}
Then the compatibility conditions,
$I_{tx}=I_{xt}$, $I_{t\dot{x}}=I_{{\dot{x}}t}$, $I_{x{\dot{x}}}=I_{{\dot{x}}x}$,
between the different equations of (\ref{met8}), provide us the relations
\begin{eqnarray}
S_t+\dot{x}S_x+\phi S_{\dot{x}} &=&
-\phi_x+\phi_{\dot{x}}S+S^2,\label {lin02}\\
R_t+\dot{x}R_x+\phi R_{\dot{x}} & =&
-(\phi_{\dot{x}}+S)R,\label {lin03}\\
R_x-SR_{\dot{x}}-RS_{\dot{x}} &= &0.
\qquad \qquad\qquad \;\;\;\label {lin04}
\end{eqnarray}
Solving Eqs.~(\ref{lin02})-(\ref{lin04}) one can obtain expressions for $S$ and
$R$. It may be noted that two sets of independent special solutions $(S,R)$ are sufficient for
our purpose. Once these forms are determined the integral of motion
$I(t,x,\dot{x})$ can be deduced from the expression
\begin{eqnarray}
I= r_1
-r_2 -\int \left[R+\frac{d}{d\dot{x}} \left(r_1-r_2\right)\right]d\dot{x},
\label{met13}
\end{eqnarray}
where
\begin{eqnarray}
r_1 = \int R(\phi+\dot{x}S)dt,\quad
r_2 =\int (RS+\frac{d}{dx}r_1) dx. \nonumber
\end{eqnarray}
Equation~(\ref{met13}) can be derived straightforwardly by
integrating Eq.~(\ref{met8}).
We solve Eq. (\ref{eq6}) through the extended PS procedure in the
following way. For the given second-order ODE (\ref{eq6}), the first
integral $I$ should be either a time-independent or time dependent one. In the
former case, it is a conservative system and we have $I_t=0$ and in the latter
case we have $I_t\ne 0$. So, let us first consider the case $I_t=0$ and
determine the null forms and the corresponding integrating factors, and from
these we construct the integrals of motion and then we extend the analysis
to the case $I_t\ne 0$.
\section{Time independent integrals : Integrable parametric choices}
\label{sectid}
In this section, we identify a set of parametric choices of (\ref{eq6}) for which time
independent integrals exist. For this purpose we first find the null forms and
integrating factors corresponding to equation (\ref{eq6}) and using these functions we
construct time independent integrals of (\ref{eq6}) through the relation (\ref{met13}).
\subsection{Null forms and integrating factors}
Since $I_t=0$, one can easily fix the null form $S$ from the first equation in (\ref{met8}) as
\begin{eqnarray}
S=\frac{-\phi}{\dot{x}}=\frac{(k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^{3}+k_5x^{2}+k_6x)}{\dot{x}}.
\label{timeinds}
\end{eqnarray}
Of course one can easily check that the $S$ form given above satisfies Eq. (\ref{lin02}).
Substituting this form of $S$, given in (\ref{timeinds}), into (\ref{lin03}) we get
\begin{eqnarray}
\dot{x}R_x-(k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^{3}+k_5x^{2}+k_6x) R_{\dot{x}}
=\bigg(\frac{k_1\dot{x}}{x}-\frac{k_4x^{3}+k_5x^{2}+k_6x}{\dot{x}}\bigg)R.
\label{timeindreq}
\end{eqnarray}
Since we are interested in time independent integrals we take $R_t=0$.
As we noted earlier any particular solution of the above equation along with the
null form $S$ is sufficient to construct an integral of motion. To derive a particular
solution of (\ref{timeindreq}) we make an ansatz for $R$ of the form
\begin{eqnarray}
R=\frac{\dot{x}}{(A(x)+B(x)\dot{x}+C(x)\dot{x}^2)^r},\label{timeindr}
\end{eqnarray}
where $r$ is a constant and $A(x)$, $B(x)$ and $C(x)$ are arbitrary functions
of their argument.
The reason for choosing the above form of ansatz is as follows. To deduce
the time independent first integral $I$ we assume a rational form for $I$, that is,
$I=\frac{f(x,\dot{x})}{g(x,\dot{x})}$, where $f$ and $g$ are arbitrary functions
of $x$ and $\dot{x}$. Using (\ref{met3}) we obtain
$S=I_x/I_{\dot{x}}=(f_xg-fg_x)/(f_{\dot{x}}g-fg_{\dot{x}})$.
and $R=I_{\dot{x}}=(f_{\dot{x}}g-fg_{\dot{x}})/g^2$ and from these two expressions
we find that the numerator of $R$ should be the
denominator of $S$ and so we fixed the numerator of the $R$ in Eq. (\ref{timeindr}) as $\dot{x}$ (see expression (\ref{timeinds}) for $S$), that is $R=\frac{\dot{x}}{h(x,\dot{x})}$, where $h$ is an arbitrary function. However, it is difficult to proceed with this choice of $h$. So, we further assume that $h(x ,\dot{x})$ is a polynomial in $\dot{x}$.
To begin with we consider the case in which $h$ is a quadratic function in $\dot{x}$,
that is $h = A(x) + B(x) \dot{x}+C(x)\dot{x}^2$. Since $R$ is in rational form while taking differentiation or integration the form of the denominator remains the same but the power of the denominator increases or decreases by a unit order from that of the initial one. So, instead of considering
$h$ to be of the form $h = A(x) + B(x) \dot{x}+C(x)\dot{x}^2$, we consider a more general form
$h =( A(x) + B(x) \dot{x}+C(x)\dot{x}^2) ^r$, where $r$ is a constant to be determined. We
note here that this form of ansatz played a crucial role in deducing the time independent
integrals of certain dissipative systems, see for example Ref. 20.
Substituting (\ref{timeindr}) into (\ref{timeindreq}) and
solving the resultant equation, we find that the solution exists only for certain
specific choices of $k_i$'s. In the following we provide these parametric restrictions
with the resultant forms of R:
\begin{eqnarray}
&&\mbox{\emph{Case \bf (i) :}}\nonumber\\
&& k_5=k_4(3+2k_1)\rho_1,\,
k_6=k_4(1+k_1)(2+k_1)\rho_1^2,\,\,\rho_1=\frac{k_2}{k_3(1+k_1)},\,k_1\ne-1\nonumber\\
&&R=\left\{
\begin{array}{ll}
\displaystyle\frac{\dot{x}}
{((2+k_1)\dot{x}+k_3x(\rho_1(2+k_1)+x))\dot{x}
+k_4x^2(\rho_1(2+k_1)+x)^2},&k_3^2<4k_4(2+k_1)\\
\vspace{-0.5cm}\\
\displaystyle \frac{\dot{x}x^{(2-r)k_1}}
{\left[k_3(r-1)(\rho_1(2+k_1)+x)x+(2+k_1)r\dot{x}\right]^{r}},&k_3^2\ge4k_4(2+k_1)\\
\end{array}
\right.
\end{eqnarray}
\begin{eqnarray}
&&\hspace{-0.5cm}\mbox{where }r=\frac{k_3^2\pm k_3\sqrt{k_3^2-4k_4(2+k_1)}}{2k_4(2+k_1)}.\nonumber\\
&&\hspace{-0.5cm}\mbox{For the choice }k_1=-1,\mbox{ we find }k_4=0,k_6=\frac{k_2k_5}{k_3},\,\,R=\frac{\dot{x}}{k_3x\dot{x}+k_5x^2}\\
&&\mbox{\emph{Case \bf (ii) :}}\nonumber\\
&&k_2=0,\,k_3=0, \,\,R=\dot{x}x^{2k_1}\\
&&\mbox{\emph{Case \bf (iii) :}}\nonumber\\
&&k_1=0,\,k_2=0,\,k_5=0,\nonumber\\
&&\displaystyle R=\left\{
\begin{array}{ll}
\displaystyle\frac{\dot{x}}{2k_4\dot{x}^2+k_3\dot{x}(k_4x^2+k_6)+(k_4x^2+k_6)^2},&k_3^2<8k_4\\
\displaystyle\frac{\dot{x}}{(\dot{x}+\frac{(r-1)}{2r}k_3x^2+\frac{rk_6}{k_3})^{r}},&k_3^2\ge8k_4,
\,\,r=\frac{k_3^2\pm k_3\sqrt{k_3^2-8k_4}}{4k_4}.
\end{array}
\right.
\end{eqnarray}
The question now remains as to whether the functions $S$ and $R$ given
above satisfy the third of the determining equations (vide Eq. (\ref{lin04})) or not. One can
easily check that all the above three sets of functions $S$ and $R$ do indeed satisfy the equation
(\ref{lin04}). The parametric restrictions given above fix the equation of
motion (\ref{eq6}) to the following specific forms:
\begin{eqnarray}
&&\mbox{\underline{\emph{Case {\bf (i)} Eq. (\ref{tid5}):}} }\nonumber\\
&&\ddot{x}+k_1
\frac{\dot{x}^2}{x}+k_3(\rho_1(1+k_1)+ x)\dot{x} +
k_4x^3+k_4\rho_1(3+2k_1)x^2
+k_4(1+k_1)(2+k_1)\rho_1^2x=0,\nonumber\\
&&\mbox{where }\rho_1=\frac{k_2}{k_3(1+k_1)}\nonumber\\
&&\mbox{\underline{\emph {Case {\bf (ii)} Eq. (\ref{bern1}):}}}\nonumber\\
&&\ddot{x}+k_1\frac{\dot{x}^2}{x}+k_4 x^3+k_5 x^2+k_6 x=0.\,\nonumber%
\end{eqnarray}
\begin{eqnarray}
&&\hspace{-5cm}\mbox{\underline{\emph{Case {\bf (iii)} Eq. (\ref{tid4}):}}}\nonumber\\
&&\hspace{-2cm}\ddot{x}+k_3 x\dot{x}+k_4x^3+k_6x=0.\nonumber
\end{eqnarray}
In the next section, we construct time independent integrals
for the above equations. We mention here that the last equation (\ref{tid4}) is nothing
but the modified Emden equation with additional linear force term which admits a conservative Hamiltonian
description for all values of $k_3$, $k_4$ and $k_6$. For the special choice,
$k_4=\frac{k_3^2}{9}$, this equation has been shown to exhibit unusual phenomena like
amplitude independent frequency of oscillations and conservative Hamiltonian structure (for more details one may refer 21).
\subsection{Integrals of motion}
Having determined the explicit forms of $S$ and $R$, one can
proceed to construct the integrals of motion using the expression
(\ref{met13}) for the above cases. Substituting the corresponding forms of $S$ and $R$ into the general
form of the integral of motion (\ref {met13}) and evaluating the
resultant integrals, we obtain the following time independent
integrals for the above three cases:
\begin{eqnarray}
&&\hspace{-1cm}\mbox{\emph{Case \bf (i)}: }\nonumber\\
&& I_1=\left\{
\begin{array}{ll}
\displaystyle\log\left[x^{2k_1}\{(2+k_1)\dot{x}^2+[(2+k_1)\rho_1+x]
[k_3x\dot{x}\right.\\
\left.\displaystyle\qquad+ k_4x^2((2+k_1)\rho_1+x)]\}\right]
-\frac{2k_3\tan^{-1}\left[\frac{\Phi(x,\dot{x})}{\Omega(x)}\right]}
{\sqrt{4k_4(2+k_1)-k_3^2}}
\,,&k_3^2<4k_4(2+k_1)\label{integral6}\\
\vspace{-0.3cm}\\
\displaystyle\frac{x^{k_1(2-r)}[k_3x((2+k_1)\rho_1+x)+(2+k_1)r\dot{x}]}
{\left\{k_3(r-1)x((2+k_1)\rho_1+x)+(2+k_1)r\dot{x}\right\}^{r-1}}
\,,\,&k_3^2>4k_4(2+k_1)\\\vspace{-0.3cm}\\
\displaystyle\log\bigg[\bigg(k_3x[(2+k_1)\rho_1+x]+2(2+k_1)\dot{x}\bigg)x^{k_1}\bigg]\\
\displaystyle\qquad-\frac{2(2+k_1)\dot{x}}
{k_3x((2+k_1)\rho_1+x)+2(2+k_1)\dot{x}},&k_3^2=4k_4(2+k_1)\\\vspace{-0.3cm}\\
k_2\frac{\dot{x}}{x}+k_2k_3x+\left(k_2^2+\frac{k_2k_5}{k_3}\right)
\log[x]-\frac{k_2k_5}{k_3}\log\left[k_2\dot{x}+\frac{k_2k_5}{k_3}\right],&k_1=-1
\end{array}
\right.\\\nonumber
&&\hspace{-1cm}\mbox{where }\Phi(x,\dot{x})= k_3x^2+(2+k_1)
(k_3x\rho_1+2\dot{x}),\nonumber\\
&&\hspace{-1cm}\mbox{and }\Omega(x)= \sqrt{4(2+k_1)k_4-k_3^2}
\bigg[(2+k_1)\rho_1+x\bigg]x.\nonumber\\
&&\hspace{-1cm}\mbox{\emph{Case \bf (ii)}: }\nonumber\\
&&I_1=
\begin{array}{ll}
\displaystyle\bigg (\frac {\dot{x}^2} {x^2} +\frac {k_6}{1+k_1}
+\frac
{2(2+k_1)k_5x+(3+2k_1)k_4x^2}{(2+k_1)(2k_1+3)}\bigg)x^{2(1+k_1)},
\label{integral8}
\end{array}
\end{eqnarray}
\begin{eqnarray}
&&\hspace{-1cm}\mbox{\emph{Case \bf (iii)}: }\nonumber\\
&&I_1=\left \{
\begin{array}{ll}
\log[2k_4\dot{x}^2+k_3(k_4x^2+k_6)\dot{x}+(k_4x^2+k_6)^2]
\\
\qquad+\frac{2k_3\tan^{-1}\left[\frac{k_3\dot{x}+2(k_4x^2+k_6)}{\dot{x}
\sqrt{8k_4-k_3^2}}\right]}{\sqrt{8k_4-k_3^2}},&k_3^2<8k_4\\\vspace{-0.3cm}\\
\bigg(\dot{x}+\frac{(r-1)}{2r}k_3x^2+\frac{rk_6}{k_3}\bigg)^{-r}
\bigg\{\dot{x}\bigg(\dot{x}+\frac{k_3}{2}x^2+\frac{r^2k_6}{(r-1)k_3}\bigg)\\
\qquad+\frac{(r-1)}{r^2}\bigg(\frac{k_3}{2}x^2\label{tid3integral}
+\frac{r^2k_6}{(r-1)k_3}\bigg)^2\bigg\},\,&k_3^2>8k_4\\\vspace{-0.3cm}\\
\displaystyle\frac{4k_3\dot{x}}{k_3^2x^2+4k_3\dot{x}+8k_6}
-\log[k_3^2x^2+4k_3\dot{x}+8k_6],\,&k_3^2=8k_4,\\
\end{array}
\right.
\end{eqnarray}
Note that in the above the ranges of $x$ and $\dot{x}$ should be so restricted that no
multivaluedness occurs.
In a recent paper{\footnotesize$^{22}$} we have shown that the general equation of the form
\begin{eqnarray}
\ddot{x}+\frac{g'(x)}{g(x)}\dot{x}^2+\alpha \frac{f(x)}{g(x)}\dot{x} +\lambda \frac{f(x)}{g(x)^2}\int f(x)dx=0,\label{lienard}
\end{eqnarray}
is related
to the damped harmonic oscillator equation
\begin{eqnarray}
y''+\alpha y'+\lambda y=0,\qquad \left('=\frac{d}{d\tau}\right)\label{damp}
\end{eqnarray}
through the nonlocal connection
\begin{eqnarray}
y=\int f(x)dx,\,\,\, d\tau=\frac{f(x)}{g(x)}dt.
\end{eqnarray}
By suitably choosing $f(x)$ and $g(x)$ one can show that the damped harmonic oscillator gets transformed to Eqs. (\ref{tid5}), (\ref{bern1}) and (\ref{tid4}). By applying this transformation to the time independent integrals of motion of the damped harmonic oscillator one can also obtain the
above results. However, it is not possible to obtain the time dependent integrals of motion of Eq. (\ref{eq6}) using the above form of nonlocal transformation. We mention here that the time independent
integrals for the specific equation (\ref{tid4}) (Li\'enard type equation) can also be deduced by the procedure described in
Ref. 23.
\section{Time independent integrals of motion: integrability \& general solution}
\subsection{Hamiltonian description and integrability}
\label{hamdes}
In the previous subsection we showed that the equations (\ref{tid5}), (\ref{bern1}) and (\ref{tid4})
admit time independent integrals of motion. Interestingly one can
interpret these integrals as time independent (but nonstandard) Hamiltonians for the respective systems and
they can be treated as conservative systems. In the following we deduce the underlying Hamiltonian
structure for the Eqs. (\ref{tid5})-(\ref{tid4}) from the first integrals (\ref{integral6})-(\ref{tid3integral}). To do
so we assume a Hamiltonian of the form
\begin{eqnarray}
H(x,p)=I(x,\dot{x})=p\dot{x}-L(x,\dot{x}),\label{ham1}
\end{eqnarray}
where $L(x,\dot{x})$ is the Lagrangian and $p$ is the canonically conjugate
momentum. From (\ref{ham1}) we get
\begin{eqnarray}
\frac{\partial I}{\partial \dot{x}}=\frac{\partial p}{\partial
\dot{x}}\dot{x},\label{ham}
\end{eqnarray}
from which we identify
\begin{eqnarray}
p=\int\frac{I_{\dot{x}}}{\dot{x}}d\dot{x}.\label{momentum}
\end{eqnarray}
It is clear from (\ref{momentum}) that once $I$ is known $p$ can be
determined in terms of $\dot{x}$ and $x$ and inverting one can express $\dot{x}$ in terms of $x$ and $p$. Substituting the expression for $\dot{x}$ in terms of $p$ into the expression for $I$ and from (\ref{ham1}) one can
deduce $L$. Once $p$ and $L$ are known the same expression (\ref{ham1}) can
be utilized to derive $H$. Using this procedure
one can deduce the Hamiltonian from the first integrals for all the three Cases (i)-(iii).
The Hamiltonians and the corresponding canonical
conjugate momenta read as follows:
\begin{eqnarray}
&&\hspace{-1cm}\mbox{{Case {\bf (i)}:} }\nonumber\\
&&H=\left\{
\begin{array}{ll}
\displaystyle\log\left[x^{k_1}(\rho_1(2+k_1)x+x^2)\sec\left[\frac{\Omega p}{2(2+k_1)}\right]\right]\\
\qquad-\frac{k_3p}{2(2+k_1)}(\rho_1(2+k_1)x+x^2),
&k_3^2<4k_4(2+k_1)\\\vspace{-0.6cm}\\
\mbox{where }\Omega(x)= \sqrt{4(2+k_1)k_4-k_3^2}
\bigg[(2+k_1)\rho_1+x\bigg]x.\nonumber\\
\displaystyle\frac{(r-1)}{(r-2)}\left(px^{-k_1}\right)^{\frac{(r-2)}{(r-1)}}-\frac{k_3p(r-1)}{r(2+k_1)}\label{tidham1}
(\rho_1(2+k_1)x+x^2),
&k_3^2>4k_4(2+k_1)\\\vspace{-0.4cm}\\
\displaystyle\log\left[\frac{x^{k_1}}{p}\right]+\frac{k_3px}{2(2+k_1)}((2+k_1)\rho_1+x),&k_3^2=4k_4(2+k_1)\\\vspace{-0.4cm}\\
\end{array}
\right .
\end{eqnarray}
where
\begin{eqnarray}
&&\hspace{-1.5cm}\quad p=\left\{
\begin{array}{ll}
\frac{2(2+k_1)}{\Omega}\tan^{-1}\bigg[\frac{k_3x((2+k_1)\rho_1+x)+2(2+k_1)\dot{x}}
{\Omega}\bigg],
&\quad k_3^2<4k_4(2+k_1)
\\\vspace{-0.4cm}\\
\frac{x^{(2-r)k_1}}
{\left(\dot{x}+\frac{k_3(r-1)}{r(2+k_1)}((2+k_1)\rho_1x+x^2)\right)^{r-1}},
&\quad k_2^2\ge4k_4(2+k_1).\\\vspace{-0.4cm}\\
\vspace{-0.4cm}\\
\end{array}
\right.\\\nonumber\\
&&\hspace{-1cm}\mbox{{Case {\bf (ii)}:} }\nonumber\\
&& H=
x^{2(k_1+1)}\bigg (\frac{p^2}{4x^{2(2k_1+1)}}+\frac
{k_6}{k_1+1}+\frac{x(2(k_1+2)k_5+(2k_1+3)k_4x)}{2k_1^2+7k_1+6}\bigg ),\,p=2x^{2k_1} \dot{x}.\label{tid2ham
\end{eqnarray}
\begin{eqnarray}
&&\hspace{-1cm}\mbox{{Case {\bf (iii)}:} }\nonumber\\
&& H=\left\{
\begin{array}{ll}
\displaystyle\frac{k_3}{4}p\left(x^2+\frac{k_6}{k_4}\right)
-\log\left[\left(x^2+\frac{k_6}{k_4}\right)\sec[\frac{\gamma p}{2}
(x^2+\frac{k_6}{k_4})]\right],\label{tid3ham}
&k_3^2<8k_4 \\\vspace{-0.4cm}\\
\displaystyle\frac{(r-1)}{(r-2)}
p^{\frac{(r-2)}{(r-1)}}-\frac{(r-1)}{2r}k_3px^2-\frac{rpk_6}{k_3},& k_3^2>8k_4\\\vspace{-0.4cm}\\
\displaystyle\log[p]-\frac{k_3}{2}p\left(\frac{x^2}{2}+\frac{4k_6}{k_3^2}\right),&k_3^2=8k_4\\
\end{array}
\right .\\\nonumber
&&\hspace{-1.5cm}\mbox{where}\\
&&\hspace{-1.5cm}\qquad\quad p=\left\{
\begin{array}{ll}
\frac{2}{\gamma( x^2+\frac{k_6}{k_4})}\tan^{-1}\left[\frac{4\dot{x}+k_3(x^2
+\frac{k_6}{k_4})}{2\gamma (x^2+\frac{k_6}{k_4})}\right],&k_3^2<8k_4,\,\,\gamma=\frac{1}{2}\sqrt{8k_4-k_3^2}
\\\vspace{-0.4cm}\\
\dot{x}+\frac{(r-1)}{2r}k_3x^2+\frac{rk_6}{k_3},\qquad\quad& k_3^2\ge8k_4.
\end{array}
\right .
\end{eqnarray}
One can check that in all the above cases the second order
equivalence of the Hamilton's equation of motion coincides exactly
with the associated equations (\ref{tid5}), (\ref{bern1}) and
(\ref{tid4}) for the appropriate parametric choices. The
existence of the time independent Hamiltonian for the Eqs.
(\ref{tid5}), (\ref{bern1}) and (\ref{tid4}) assures us that they
are Liouville integrable{\footnotesize$^{4}$}. However, our target is to
go beyond this statement and construct the solutions for these
equations. On the other hand, we find that it is difficult to integrate the
Hamilton's equations of motion corresponding to the above Hamiltonians except for the second case (vide
Eq. (\ref{tid2ham})) whose solution can be deduced using the
standard methods. In order to deduce the solutions we introduce
suitable canonical transformations so that the new Hamilton's
equations of motion can be integrated. However, we are able to
construct the explicit solution of Eq. (\ref{tid5}) only for the
parametric choice $\rho_1=0$, making the resultant equation
equivalent to Eq. (\ref{eqiv}) whose solution is constructed in
the subsection \ref{caseivgensol}. Hence we do not discuss the
solution of case (i) separately here.
\subsection{General solution}
In this sub-section we discuss the method
of finding the general solution for the case (ii) and case (iii).
\noindent
{\bf{Case {\bf (ii)}:}}
We find that one can rewrite the first integral (\ref{integral8})
straightforwardly into the following quadrature
\begin{eqnarray}
t-t_0=\int\frac{dx}{\sqrt{I_1x^{-2k_1}
-\frac{k_4(3+2k_1)x^4+2k_5(2+k_1)x^3}{2k_1^2+7k_1+6}-\frac{k_6x^2}{1+k_1}}}.\label{solbern1}
\end{eqnarray}
On the other hand the
solution of (\ref{bern1}) can also be obtained by transforming it to the Bernoulli equation
through the transformation $\dot{x}=z(x)$, where $x$ is the new independent variable. The solution of this reduced Bernoulli
type equation is again given in terms of quadratures{\footnotesize$^{2}$}.
\noindent
{\bf{Case {\bf (iii) a)}: Parametric choice $k_3^2<8k_4$}}
In order to obtain the solution for the under damped case ($k_3^2<8k_4$) we use the following
canonical transformation,
\begin{eqnarray}
x=-\sqrt{\frac{k_6}{k_4}}\tan\left[\sqrt{\frac{k_6}{k_4}}P\right],\quad
p=\frac{k_4U}{k_6}\cos^2\left[\sqrt{\frac{k_6}{k_4}}P\right],
\end{eqnarray}
and transform the Hamiltonian
$
H=\frac{k_3}{4}p\left(x^2+\frac{k_6}{k_4}\right)
-\log\left[\left(x^2+\frac{k_6}{k_4}\right)\sec[\frac{\gamma p}{2}
(x^2+\frac{k_6}{k_4})]\right]\nonumber
$
into
\begin{eqnarray}
H=\log\left[\frac{k_6}{k_4}\sec^2\left[\sqrt{\frac{k_6}{k_4}}P\right]
\sec\left[\frac{\gamma}{2}U\right]\right]-\frac{k_3}{4}U.
\label{tid3canham1}
\end{eqnarray}
The corresponding Hamilton equations of motion are
\begin{eqnarray}
\dot{U}=2\sqrt{\frac{k_6}{k_4}}\tan\left[\sqrt{\frac{k_6}{k_4}}P\right],\quad
\dot{P}=\frac{k_3}{4}-\frac{\gamma}{2}\tan\left[\frac{\gamma}{2}U\right].
\end{eqnarray}
From the first equation we can express $P$ in terms of $U$, that is,
$P=\sqrt{\frac{k_4}{k_6}}\tan^{-1}\left[\frac{\sqrt{k_4}\dot{U}}{2\sqrt{k_6}}\right]$ and
substituting this form of $P$ into (\ref{tid3canham1}) we obtain
\begin{eqnarray}
H=\log\left[\frac{(4k_6+k_4\dot{U}^2)}{4k_4}\sec\left[\frac{\gamma}{2}U\right]\right]
-\frac{k_3U}{4}\equiv E.
\label{eq59a}
\end{eqnarray}
Rewriting the last equation (\ref{eq59a}) for $\dot{U}$ as
\begin{eqnarray}
\dot{U}=2\sqrt{\exp\left[E+\frac{k_3}{4}U\right]\cos\left[\frac{\gamma}{2}U\right]
-\frac{k_6}{k_4}}
\end{eqnarray}
and integrating we arrive at the following quadrature
\begin{eqnarray}
t-t_0=\int\frac{dU}
{2\sqrt{\exp\left[E+\frac{k_3}{4}U\right]\cos\left[\frac{\gamma}{2}U\right]
-\frac{k_6}{k_4}}}.
\end{eqnarray}
\noindent
{\bf{Case {\bf (iii) b)}: Parametric choice $k_3^2>8k_4$}}
In order to obtain the solution for the over damped case ($k_3^2>8k_4$), we use the following canonical transformation
$x=\frac{P}{U},\quad p=\frac{U^2}{2}$ and obtain a new Hamiltonian of the form
\begin{eqnarray}
H=\sigma_1U^{\frac{2}{r_{12}}}
-\left(\frac{k_3(r-1)P^2}{4r}+\frac{k_6rU^2}{2k_3}\right),
\end{eqnarray}
where $\sigma_1=\frac{r_{12}}{2^{\frac{1}{r_{12}}}}$.
The canonical equations in the new variable assume the form
\begin{eqnarray}
\dot{U}=\frac{k_3P(1-r)}{2r},\quad \dot{P}=\frac{k_6rU}{k_3}-\frac{2U^{\frac{2}{r_{12}}}}
{r_{12}U}.
\end{eqnarray}
From the first relation one can express $P$ in terms of $\dot{U}$, that is
$P=\frac{2r\dot{U}}{k_3(1-r)}$. Substituting the latter in the Hamiltonian we get
\begin{eqnarray}
H\equiv E=\sigma_1U^{\frac{2}{r_{12}}}-\frac{k_6r}{2k_3}U^2
+\frac{r\dot{U}^2}{k_3(1-r)}
\end{eqnarray}
which in turn leads us to the quadrature of the form
\begin{eqnarray}
t-t_0=\left(\frac{k_3(1-r)}{r}\right)^{\frac{1}{2}}\int
\frac{dU}{\sqrt{E-\sigma_1U^{\frac{2}{r_{12}}}+\frac{k_6rU^2}{2k_3}}}
\end{eqnarray}
\noindent
{\bf{Case {\bf (iii) c)}: Parametric choice $k_3^2=8k_4$}}
Finally, now we focus our attention on the critically damped case ($k_3^2=8k_4$).
Using the same canonical transformation $x=\frac{P}{U}$, $p=\frac{U^2}{2}$ we rewrite the
underlying Hamiltonian (\ref{tid3ham}) as
\begin{eqnarray}
H=2\log[U]-\frac{k_3}{4}\left(\frac{P^2}{2}+k_6U^2\right).
\end{eqnarray}
The corresponding canonical equations turn out to be
\begin{eqnarray}
\dot{U}=\frac{-k_3P}{4},\quad \dot{P}=\frac{k_3k_6}{2}-\frac{2}{U}.
\end{eqnarray}
Substituting $P=-\frac{4\dot{U}}{k_3}$ in the Hamiltonian, we get
\begin{eqnarray}
H\equiv E=2\log[U]-\frac{1}{4}k_3k_4U^2-\frac{2}{k_3}\dot{U}^2
\end{eqnarray}
Rearranging we get,
\begin{eqnarray}
\dot{U}=
\left(\frac{8k_3\log[U]-k_3^2k_6U^2-4Ek_3}{2}\right)^{\frac{1}{2}},
\end{eqnarray}
which upon integrating reduces to the following form of quadrature
\begin{eqnarray}
t-t_0=\sqrt{2}\int\frac{dU}{\sqrt{8k_3\log[U]-k_3^2k_6U^2-4Ek_3}}.
\end{eqnarray}
Summarizing the results, we find that the nonlinear equations (\ref{tid5}), (\ref{bern1}) and (\ref{tid4}) admit time
independent Hamiltonians for all values of the system parameters and can be classified
as integrable ones in the Liouville sense. While constructing the solution for these
three equations
we find that the nonlinear system (\ref{bern1}) can be transformed into Abel equation which
in turn can be integrated into a quadrature. Solution of Eq. (\ref{tid4}) is
given in terms of quadrature by applying suitable canonical transformation to its Hamiltonian. By using again canonical transformations the Hamiltonian corresponding to Eq. (\ref{tid5}) can be transformed and the canonical equations corresponding to this transformed Hamiltonian can be integrated and the solution is given in terms
of quadratures for the parametric choice $\rho_1=0$ (see Sec. \ref{caseivgensol}).
\section{Time dependent integrals $(I_t\ne0)$}
\label{sectd}
In this section, we explore the parametric choices for which (\ref{eq6}) admits
time dependent integrals. The underlying procedure is same as that of the time
independent integral case but involves somewhat lengthy calculations. As a first step
in this process we derive the null forms and integrating factors corresponding to Eq. (\ref{eq6})
by solving the determining Eqs. (\ref{lin02})-(\ref{lin04}).
\subsection{Null forms and integrating factors}
\noindent In the previous section we considered the case $I_t=0$. As a
consequence $S$ turns out to be $\frac{-\phi}{\dot{x}}$. However, in
the case $I_t\ne0$, the function $S$ has to be determined from
Eq. (\ref{lin02}), that is
\begin{eqnarray}
&&S_t+\dot{x}S_x-\bigg(\frac{k_1
\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^3+k_5x^2+k_6x\bigg)S_{\dot{x}}=-\frac{k_1\dot{x}^2}{x^2}\nonumber
\\&&\qquad+k_3\dot{x}+3k_4x^2+2k_5x+k_6-S\left(\frac{2k_1\dot{x}}{x}+(k_2+k_3x)\right)+S^2.
\label{sequation}
\end{eqnarray}
Since it is too difficult to solve Eq. (\ref{sequation}) for its general
solution, we seek a particular solution for S, which is sufficient for our
purpose. The time independent integral case clearly indicates that
$S$ should be in a rational form. To begin with one may consider $S=\frac{f(t,x,\dot{x})}
{g(t,x,\dot{x})}$, where $f$ and $g$ are arbitrary functions. However, it is difficult to solve Eq. (\ref{sequation}) with this
form of $S$. So one may assume that $f$ and $g$ are simple polynomials in $\dot{x}$
with coefficients which are arbitrary functions of $t$ and $x$. We seek a simple rational
expression for $S$ in the form
\begin{eqnarray}
S=\frac{a(x,t)+b(x,t)\dot{x}+c(x,t)\dot{x}^2}{d(x,t)+e(x,t)\dot{x}+f(x,t)\dot{x}^2},
\label{sform}
\end{eqnarray}
where $a$, $b$, $c$, $d$, $e$ and $f$ are arbitrary functions of
$x$ and $t$ which are to be determined. One may also consider a
cubic polynomial in $\dot{x}$ both in the numerator and in the
denominator. However, the resultant analysis did not yield any
new result. So we confine our presentation here to the form (\ref{sform}) only. Substituting (\ref{sform}) into (\ref{sequation}) and
equating the coefficients of different powers of $\dot{x}$ to
zero, we get
\begin{eqnarray}
&&-a^2+a_td-ad_t+adk_2-d^2k_6+adk_3x-2d^2k_5x-bdk_6x+aek_6x-3d^2k_4x^2
\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad\qquad-bdk_5x^2+aek_5x^2-bdk_4x^3+aek_4x^3=0,\nonumber\\
&&a_xd-2ab-ad_x+b_td-bd_t+a_te-ae_t+2aek_2-d^2k_3-2dek_6+\frac{2adk_1}{x}
+2aek_3x
\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\qquad-4dek_5x-2cdk_6x-6dek_4x^2-2cdk_5x^2-2cdk_4x^3=0,\nonumbe
\end{eqnarray}
\begin{eqnarray}
&&b_xd-b^2-2ac-bd_x+c_td-cd_t+a_xe-ae_x+b_te-be_t-cdk_2
+bek_2-2dek_3-e^2k_6
\nonumber\\
&&\qquad\qquad\qquad\qquad+\frac{d^2k_1}{x^2}+\frac{bdk_1}{x}+\frac{3aek_1}{x}-cdk_3x
+bek_3x-2e^2k_5x-cek_6x\nonumber\\
&&\qquad\quad\qquad\qquad\qquad\qquad\qquad\qquad\qquad-3e^2k_4x^2-cek_5x^2-cek_4x^3=0,\nonumber\\
&&c_xd-cd_x-2bc+b_xe-be_x+c_te-ce_t-e^2k_3+\frac{2dek_1}{x^2}+\frac{2bek_1}{x}=0,\nonumber\\
&&c_xe-c^2-ce_x+\frac{e^2k_1}{x^2}+\frac{cek_1}{x}=0.
\label{detereq}
\end{eqnarray}
We obtain four parametric choices for which we are able to find
nontrivial particular solutions of the above set of coupled
partial differential equations. We present the parametric choices
and the null forms $S$ in Table-I. We find that the null forms $S$
obtained are incidently independent of time.
Next, the above identified forms of $S$ along with their corresponding parametric
restrictions are substituted in the determining equation (\ref{lin03}) for $R$. To solve
the resultant equation for $R$ we make use of the ansatz
\begin{eqnarray}
R=\frac{F(t)S_d }{(A(x)+B(x)\dot{x}+C(x)\dot{x}^2)^r},\label{ransatz}
\end{eqnarray}
where $S_d$ is the denominator of $S$. We demand
the above form due to the following reason. To deduce
the first integral $I$ we assume a rational form for $I$, that is,
$I=\frac{f(t,x,\dot{x})}{g(t,x,\dot{x})}$, where $f$ and $g$ are
arbitrary functions of $t$, $x$ and $\dot{x}$. From (\ref{met8}), we know that
$S=I_x/I_{\dot{x}}=(f_xg-fg_x)/(f_{\dot{x}}g-fg_{\dot{x}})$ and as
we find the deduced $S$ forms are independent of time, the
numerator and denominator of $S$ should share a common factor
which is a function of time alone, that is, the numerator and
denominator of $S$ should be of the form
$(f_xg-fg_x)=F(t)f_1(x,\dot{x})$ and
$(f_{\dot{x}}g-fg_{\dot{x}})=F(t)g_1(x,\dot{x})$, respectively.
Moreover, from the relation
$R=I_{\dot{x}}=(f_{\dot{x}}g-fg_{\dot{x}})/g^2$, we find that the
numerator of $R$ should be the denominator of $S$.
On solving the resultant equation,
we obtain the integrating factor $R$. We present the forms of $R$
along with the $S$ forms in Table 1. Once $S$ and $R$ are determined then one
has to verify the compatibility of this set $(S,R)$ with the extra
constraint Eq. (\ref{lin04}). Having verified the compatibility of
$(S,R)$, we substitute $S$ and $R$ into Eq. (\ref{met13}) and
construct the associated integral of motion. In this way we identify
four sets of integrable parametric choices. We present the explicit form
of these integrals of motion in Sec \ref{application}.
\begin{sidewaystable}
\caption{Parametric restrictions, null forms $S$ and integrating
factors $R$ of\\
$\ddot{x}+k_1 \frac{\dot{x}^2}{x}+(k_2+k_3
x)\dot{x}+k_4 x^3+k_5 x^2+k_6 x=0$}
\noindent\begin{tabular}{|l|l|l|l|}
\hline Case & Parametric restriction &\qquad\qquad\qquad Form of $S$ &\qquad\qquad\qquad Form of $R$ \\
\hline & & & \\ (i) &$k_4=\frac{(1+k_1) k_3^2}{(3+2 k_1)^2},$
& & \\
& $k_5=\frac{k_2 k_3}{3+2k_1}.$&$\qquad\qquad\quad\frac{k_3 x}{(3+2 k_1)}-\frac{\dot{x} }{x}$ &$\qquad\frac{\pm e^{\pm\omega t}x}
{\bigg(\dot{x}+\frac{1}{2}\frac{k_2\pm\omega}{1+k_1}x+\frac{k_3}{3+2k_1}x^2\bigg)^2}$\\
&$k_1$, $k_2$, $k_3$ arbitrary & &$\qquad\omega=\sqrt{k_2^2-4(1+k_1)k_6}$ \\
\hline (ii) &$k_5=\frac{k_3(k_2\pm\omega)}{2 (2+k_1)},$
& & \\
&$k_4=0.$ &$\qquad\quad\frac{1}{2}(k_2\mp\omega+2 k_3 x)
+k_1\frac{\dot{x}}{x}$ &$\qquad\qquad e^{\frac{1}{2}(k_2\pm\omega)t}
x^{k_1}$\\
& $k_1$, $k_2$, $k_3$ arbitrary& & \\
\hline (iii) &$k_6=\frac{2(3+2 k_1)k_2^2}{(5+4 k_1)^2},$
&& \\
&$k_3=0,k_4=0.$ &$\frac{\frac{4(1+k_1)k_2^2}{5+4k_1} x^2+(5+4k_1)k_5 x^3
+2(1+2k_1)k_2 x\dot{x}+k_1(5+4 k_1)\dot{x}^2)}{2 k_2x^2+(5+4 k_1)x\dot{x}}$ &
$e^{\frac{2(3+2k_1)k_2 t}{5+4 k_1}}x^{2 k_1}\bigg(2 k_2 x+(5+4k_1)\dot{x}\bigg)$\\
& $k_1$, $k_2$ arbitrary& & \\
& & & \\
\hline
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}
\noindent\begin{tabular}{|l|l|l|l|}
\hline Case \,\, & Parametric restriction &\qquad\qquad\qquad Form of $S$ & \qquad\qquad\quad\qquad\qquad
Form of $R$ \\
\hline & & & \\
(iva) &$k_5=\frac{k_2 k_3}{3 + 2k_1},$
& &
$\bigg(k_2x+(3+2k_1)\dot{x}\bigg)
\bigg[(2+k_1)k_2^2x^2+(3+2k_1)k_2x(k_3x^2+2(2+k_1)\dot{x})$ \\
&$k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2}$. &$k_1\frac{\dot{x}}{x}+k_3x+\frac{(3+2k_1)k_4x^3}{(3+2k_1)\dot{x}+k_2x}+
\frac{(1+k_1)k_2}{3+2k_1}$ &$+(3+2k_1)^2
(k_4x^4+\dot{x}(k_3x^2+(2+k_1)\dot{x}))\bigg]^{-1}$\\
&$k_1$, $k_2$, $k_3$, $k_4$ arbitrary & &
\\
&&&\\
\hline
(ivb) &$k_5=\frac{k_2k_3}{3+2k_1}$,&&\\
&$k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2},$ & $k_1\frac{\dot{x}}{x}+k_3x+\frac{(3+2k_1)k_4x^3}{(3+2k_1)\dot{x}+k_2x}$ &
\\
&$k_4=\frac{(r-1)k_3^2}{(2+k_1)r^2}. $ & &$\qquad
\frac{e^{\frac{(2-r)(2+k_1)k_2 t}{3+2k_1}}(k_2x+(3+2k_1)\dot{x})
x^{k_1(2-r)}}{\bigg((2+k_1)r\left(\frac{k_2x}{3+2k_1}+\dot{x}\right)+k_3(r-1)x^2\bigg)^r}$ \\
&$k_1$, $k_2$, $k_3$ arbitrary & & \\
\hline
\end{tabular}
\end{sidewaystable}
\subsection{Integrals of motion and general solution}
\label{application}
\noindent {\bf Case (i)} $k_1,\,k_2,\,k_3,\,k_6$ : arbitrary, $\displaystyle k_4=\frac{(1+k_1) k_3^2}{(3+2
k_1)^2},\,k_5=\frac{k_2 k_3}{3+2k_1}$
The parametric restriction given above fixes the equation of
motion (\ref{eq6}) as
\begin{eqnarray}
\ddot{x}+k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+\frac{k_3^2(1+k_1)}{(3+2k_1)^2}x^3+\frac{k_2k_3}{(3+2k_1)}x^2+k_6x=0.\nonumber
\qquad\qquad\qquad(\ref{caseia,b})
\end{eqnarray}
Equation (\ref{caseia,b}) reduces to the generalized modified Emden
equation for the parametric choice $k_1=0$ whose integrability has been
studied in detail in Ref. 5. The integral of motion
associated with Eq. (\ref{caseia,b}) for $k_1\ne0$ turns out to be
\begin{eqnarray}
&&\mbox{\bf(ia)}\,\,\,\,\,I_1=e^{\pm \omega t}
\bigg(\frac{\dot{x}+
\frac{(k_2\mp\omega)}{2(1+k_1)}x+\rho_2x^2}{\dot{x}
+\frac{(k_2\pm\omega)}{2(1+k_1)}x+\rho_2x^2}\bigg),\,\,\,
\,\qquad\qquad\qquad\qquad\,\,\,\, k_1\ne-1,\omega\ne0\label{integral1}\\\vspace {-0.5 cm}\nonumber\\
&&\mbox{\bf(ib)}\,\,\,\,\,I_1=e^{k_2 t}\bigg (
\frac{k_2(\dot{x}+k_3x^2)+k_6x}{x}\bigg ),\,\,\,\,\,\,\,\,\,\,\qquad\qquad\qquad\qquad
k_1=-1\label{integral2}\\\vspace {-0.5 cm}\nonumber\\
&&\mbox{\bf(ic)}\,\,\,\,\,I_1=t-\frac{2(3+2k_1)x}{(3+2k_1)k_2x+2(1+k_1)(k_3x^2+(3+2k_1)\dot{x})},\,\,\,\,\,\omega=0\label{int3}
\end{eqnarray}
where $\rho_2=\frac{k_3}{3+2k_1}$, $\omega=\sqrt{k_2^2-4(1+k_1)k_6}$.
Rewriting (\ref{integral1}) - (\ref{int3}) as first order ODEs we find that
the resultant equations are of Bernoulli type
which can be solved using the standard method{\footnotesize$^{2}$}. The general solution to (\ref{caseia,b}) in each of these cases
turns out to be
\begin{eqnarray}
&&\hspace{-0.8cm}\mbox{\bf (ia) }x(t)=\left\{c_2e^{\frac{c_2t}{2}}
\left[\frac{I_1-e^{\pm\omega t}}{I_1}\right]^\frac{\pm c_2}
{2\omega}(e^{\pm\omega t}-I_1)^\frac{\pm c_1}{2\omega}I_2
-2\rho_2\left[\frac{I_1-e^{\pm\omega t}}{I_1}\right]^
\frac{\pm c_1}{2\omega}(e^{\pm\omega t}-I_1)^{
\frac{\pm c_2}{2\omega}}\right.\nonumber\\
&&\qquad\quad\quad\left.\times F\left[\frac{\pm(c_1-c_2)}{2\omega},
\frac{\mp c_2}{2\omega},1\mp\frac{c_2}{2\omega},\frac{e^{\pm\omega t}}
{I_1}\right]\right\}^{-1}c_2\left\{\frac{-(I_1-e^{\pm\omega t})^2}
{I_1}\right\}^{\frac{\pm c_2}{2\omega}}\hspace{-0.5cm},
k_1\ne-1\label{solia}\\\vspace {-0.5 cm}\nonumber\\
&&\hspace{-0.8cm}\mbox{\bf (ib)}\,x(t)=(I_2 e^{\Phi}+k_3e^\Phi\int e^{-\Phi}dt)^{-1},
\hspace{6.5cm}\,k_1=-1\\\vspace {-0.5 cm}\nonumber\\
&&\hspace{-0.8cm}\mbox{\bf (ic)}\,x(t)=\frac{k_2(3+2k_1)\left(-c_3\right)^{\frac{1}{1+k_1}}
\mbox{exp}\left[\frac{c_3(I_1-t)}{2}\right]}
{(I_1-t)^{\frac{1}{1+k_1}}}\left(2^{\frac{2+k_1}{1+k_1}}(1+k_1)k_3
\Gamma\left[\frac{2+k_1}{1+k_1},\frac{c_3(t-I_1)}{2}\right]\right.
\nonumber\\
&&\qquad\qquad\quad\left.-I_2(3+2k_1)k_2(
-c_3)^{\frac{1}{1+k_1}}\right)^{-1},\,\omega=0
\end{eqnarray}
where $F$ is the hypergeometric function, $\Gamma$ is the gamma function, $c_1=\frac{k_2\mp\omega}{1+k_1}$,
$c_2=\frac{k_2\pm\omega}{1+k_1}$, $c_3=\frac{k_2}{1+k_1}$, $\Phi=\frac{k_2k_6t+e^{-k_2t}I_1}{k_2^{2}}$, $I_1$ and
$I_2$ are the integration constants.\\
{\bf{\bf Case (ii)} $k_1$, $k_2$, $k_3$ }: arbitrary,
$k_5=\frac{k_3(k_2\pm\omega)}{2(2+k_1)},\,k_4=0,\,
\omega=\sqrt{k_2^2-4(1+k_1)k_6}
$
The parametric restriction given above fixes the equation of
motion (\ref{eq6}) as
\begin{eqnarray}
&&\ddot{x}+k_1 \frac{\dot{x}^2}{x}+(k_2+k_3x)
\dot{x}+\frac{k_3(k_2\pm\omega)}{2(2+k_1)} x^2+k_6
x=0.\hspace{3cm}(\ref{eq87})\nonumber
\end{eqnarray}
The system (\ref{eq87}) possesses a
first integral of the form
\begin{eqnarray}
&&I_1=e^{\frac{1}{2}(k_2\pm\omega)t}x^{k_1}\bigg(\dot{x}+\frac{k_2\mp\omega}{2(1+k_1)}x+\rho_3x^2\bigg),\,\,k_1\ne-1,
\label{riccati1}
\end{eqnarray}
where $\rho_3=\frac{k_3}{(2+k_1)}$.
For the parametric choice $k_1=-1$, we obtain the same first integral as we have obtained in Case (ib) above.
On the other hand for the choice $k_1=-2$, Eq. (\ref{riccati1}) gets reduced
to the Riccati equation which in turn can
be integrated by the standard methods and the solution can be obtained of the form{\footnotesize$^{2}$}
\begin{eqnarray}
x(t)=\frac{2c_1e^{c_1t}}{I_1+2I_2c_1e^{2c_1t}-2\rho_3e^{c_1t}},
\end{eqnarray}
where $c_1=\frac{1}{2}(k_2\pm\omega)$, $I_1$ and $I_2$ are the integration constants.
For other choices of $k_1$ one is able to integrate the first integral (\ref{riccati1}) only by
imposing additional parametric restriction. For example, choosing
$\displaystyle k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2}$, and $\displaystyle k_5=\frac{k_3k_2(1+k_1)}{(3+2k_1)}$, one
is able to transform the time dependent integral of motion (\ref{riccati1}) into the following
time independent integral of motion
\begin{eqnarray}
I_1=w'+\frac{k_3(1+k_1)}{2+k_1}w^{\frac{2+k_1}{1+k_1}},\quad '=\frac{d}{dz},
\label{transintcii}
\end{eqnarray}
where $w$ and $z$ are the new dependent and independent variables, respectively and are given by
\begin{eqnarray}
w=\displaystyle e^{\frac{(1+k_1)k_2t}{3+2k_1}}x^{1+k_1},\qquad\,z=-\frac{(3+2k_1)e^{-\frac{k_2t}{3+2k_1}}}{k_2}.
\label{transformii}
\end{eqnarray}
The additional parametric restrictions fix Eq. (\ref{eq87}) to the specific form
\begin{eqnarray}
&&\ddot{x}+k_1 \frac{\dot{x}^2}{x}+\left(k_2+k_3x\right)\dot{x}+\frac{k_3k_2(1+k_1)}{(3+2k_1)} x^2
+\frac{(2+k_1)k_2^2}{(3+2k_1)^2} x=0,\hspace{1cm}(\ref{td2-1})\nonumber
\end{eqnarray}
Integrating (\ref{transintcii}) we get,
\begin{eqnarray}
z-z_0&=&\int\frac{dw}{I_1-\frac{k_3(1+k_1)}{2+k_1}w^{\frac{2+k_1}{1+k_1}}},\nonumber\\
&=&
wF\bigg[\frac{1+k_1}{2+k_1},1,\frac{3+2k_1}{2+k_1},
\frac{\hat{k}w^{\frac{2+k_1}{1+k_1}}}{I_1}\bigg]I_1^{-1},\label{solii}
\end{eqnarray}
where $F$ is the hypergeometric function{\footnotesize$^{24}$} and
$\displaystyle\hat{k}=\frac{k_3(1+k_1)}{2+k_1}$.
We also mention here that Eq. (\ref{td2-1})
contains
several sub-cases which are already known to be integrable, see for example Ref. 1.
\\
\noindent{\bf Case (iii)} $k_1$, $k_2$, $k_5$ : arbitrary,
$ k_3=0,\,k_4=0,\,k_6=\frac{2(3+2 k_1)k_2^2}{(5+4 k_1)^2}$
The parametric restriction given above fix the equation of motion (\ref{eq6}) as
\begin{eqnarray}
\ddot{x}+k_1 \frac{\dot{x}^2}{x}+\rho_5(5+4k_1)\dot{x}+k_5
x^2+2(3+2 k_1)\rho_5^2
x=0,\nonumber\qquad\qquad\qquad\qquad\qquad\quad(\ref{td3eq})
\end{eqnarray}
where $\displaystyle \rho_5=\frac{k_2}{(5+4k_1)}$.
The first integral for this equation is
\begin{eqnarray}
&\,I_1=x^{2k_1}\bigg(\dot{x}^2+4 \rho_5^2 x^2
+\frac{2 k_5 x^3}{3+2k_1}+4 \rho_5x\dot{x} \bigg)e^{2(3+2k_1) \rho_5t}.\label{integral3}
\end{eqnarray}
This time independent integral can be transformed into a time independent integral
by introducing a transformation of the form
\begin{eqnarray}
w=\frac{1}{\sqrt{2}}x^{1+k_1}
e^{2(1+k_1)\rho_5t},\,
z=-\frac{e^{-\rho_5t}}{\rho_5}.
\end{eqnarray}
where $w$ and $z$ are new dependent and independent variables, respectively.
Rewriting the integral of motion in terms of the new variables, we obtain
\begin{eqnarray}
I_1=w'^{2}+2^{\left(\frac{2k_1+3}{2(1+k_1)}\right)}\frac{k_5(1+k_1)^2}
{(3+2k_1)}w^{\frac{3+2k_1}{(1+k_1)}}.
\label{eq90a}
\end{eqnarray}
Eq. (\ref{eq90a})
can be integrated further and one can obtain the general solution as
\begin{eqnarray}
z-z_0=w\sqrt{\frac{I_1}{I_1-\hat{k} w^{\frac{3+2 k_1}{1+k_1}}}}
F\bigg[\frac{1+k_1}{3+2
k_1},\frac{1}{2},\frac{4+3k_1}{3+2k_1},\frac{\hat{k}w^{\frac{3+2k_1}{1+k_1}}}{I_1}
\bigg]
\end{eqnarray}
where $F$ is the hypergeometric function{\footnotesize$^{24}$} and
$\hat{k} = 2^{\frac{(3+2k_1)}{2(1+k_1)}}\frac{k_5(1+k_1)^2} {(3+2k_1)}$.
\vskip 6pt
\noindent {\bf Case (iv) } $k_1$, $k_2$, $k_3$ : arbitrary,
$ k_5=\frac{k_2 k_3}{3 + 2k_1},\,k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2}.
$
The equation of motion in this case turns out to be
\begin{eqnarray}
&&\ddot{x}+k_1\frac{\dot{x}^2}{x}+(k_2+k_3x)\dot{x}+k_4x^3+\frac{
k_2k_3}{(3+2k_1)} x^2+\frac{k_2^2(2+k_1)}{(3+2k_1)^2}x=0.\nonumber\qquad\qquad\qquad
(\ref{eqiv})
\end{eqnarray}
The first integral reads
\begin{eqnarray}
&&\hspace{-1cm}\mbox{\bf(iva)}\,\,I_1=
\log\bigg[x^{2k_1}\bigg((2+k_1)\left(\dot{x}+\rho_6x\right)^2+k_3x^2(\dot{x}+\rho_6x)+k_4x^4
\bigg)\bigg]\nonumber\\
&&\qquad\qquad-\frac{2k_3
\tan^{-1}\left(\frac{2(2+k_1)(\dot{x}+\rho_6x)+k_3x^2}
{x^2\sqrt{4(2+k_1)k_4-k_3^2}}\right)}
{\sqrt{4(2+k_1)k_4-k_3^2}}+2(2+k_1)\rho_6t,\quad4k_4(2+k_1)>k_3^2\nonumber\\\vspace {-0.5 cm}\nonumber\\
&&\hspace{-1cm}\mbox{\bf(ivb)}\,\,I_1=\frac{x^{k_1(2-r)}\left(k_3x^2+(2+k_1)r(\dot{x}+x\rho_6)\right)e^{(2+k_1)(2-r)\rho_6 t}}
{\left(k_3(r-1)x^2+(2+k_1)r(\dot{x}+x\rho_6)\right)^{r-1}},\qquad
4k_4(2+k_1)<k_3^2\nonumber\\
&&\hspace{-1cm}\mbox{\bf(ivc)}\,\, I_1=(2+k_1)\rho_6t+\log\left[k_3x^{2+k_1}+
2(2+k_1)x^{k_1}(\dot{x}+\rho_6x)\right]
\nonumber\\
&&\hspace{4cm}-
\frac{2(2+k_1)(\dot{x}+\rho_6x)}{k_3x^2+2(2+k_1)(\dot{x}+x\rho_6)},\qquad\quad\,
\hspace{0.3cm}4k_4(2+k_1)=k_3^2,\nonumber
\end{eqnarray}
where $\rho_6=\frac{k_2}{(3+2k_1)}$, $r=\frac{k_3^2\pm k_3\sqrt{k_3^2-4k_4(2+k_1)}}{2k_4(2+k_1)}$.
In the above forms of $I_1$, we now
introduce the following transformation,
\begin{eqnarray}
w=xe^{\rho_6t},\quad\,z=-\frac{e^{-\rho_6t}}{\rho_6},
\end{eqnarray}
so that in the new variables the integrals of motion read,
\begin{eqnarray}
&&\hspace{-2cm}\mbox{\bf(iva)}\,\,
I_1=
\displaystyle\log\left[w^{2k_1}\{(2+k_1)w'^2+
[k_3w^2w'
+ k_4w^4]\}\right]-\frac{2k_3\tan^{-1}\left[\frac{ k_3w^2+2(2+k_1)
w'}{w^2\sqrt{4(2+k_1)k_4-k_3^2}}\right]}
{\sqrt{4k_4(2+k_1)-k_3^2}},
\\
&&\hspace{-2cm}\mbox{\bf(ivb)}\,\,I_1=
\displaystyle\frac{w^{k_1(2-r)}[k_3w^2+(2+k_1)rw']}
{\left\{k_3(r-1)w^2+(2+k_1)rw'\right\}^{r-1}},
\\
&&\hspace{-2cm}\mbox{\bf(ivc)}\,\,I_1=
\displaystyle\log\bigg[\bigg(k_3w^2+2(2+k_1)w'\bigg)w^{k_1}\bigg]-\frac{2(2+k_1)w'}
{k_3w^2+2(2+k_1)w'}.
\end{eqnarray}
In terms of the new variables $w$ and $z$ Eq. (\ref{eqiv}) reduces to the form
\begin{eqnarray}
w''+k_1\frac{w'^2}{w}+k_3ww'+k_4w^3=0\label{rho0}
\end{eqnarray}
which is equivalent to Eq. (\ref{tid5}) with $k_2=0$.
We find that it is very difficult to integrate the resultant integrals even after removing the time dependent factors. To
establish the integrability of Eq. (\ref{eqiv}) we transform these time independent integrals
into time independent Hamiltonian and thereby establish the Liouville integrability of Eq. (\ref{eqiv}).
By following the procedure given in Sec. \ref{hamdes} we obtain the following Hamiltonian for the above cases, namely
\begin{eqnarray}
&&\hspace{-1.1cm}\mbox{\bf(iva)}\,\,H
=\displaystyle\log\left[w^{2+k_1}\sec\left[\frac{w^2\sqrt{4(2+k_1)k_4-k_3^2} p}{2(2+k_1)}\right]\right]
-\frac{k_3pw^2}{2(2+k_1)},\,\,4k_4(2+k_1)>k_3^2\label{ivham1}\\
&&\hspace{-1.1cm}\mbox{\bf(ivb)}\,\,H
=\displaystyle\frac{(r-1)}{(r-2)}\left(pw^{-k_1}\right)^{\frac{(r-2)}{(r-1)}}-\frac{k_3p(r-1)}{r(2+k_1)}
w^2,\qquad\qquad\qquad\quad\,\,\,4k_4(2+k_1)<k_3^2\label{ivham2}\\
&&\hspace{-1.1cm}\mbox{\bf(ivc)}\,\,H
=\displaystyle\log\left[\frac{w^{k_1}}{p}\right]+\frac{k_3pw^2}{2(2+k_1)},\qquad\qquad\qquad\qquad\qquad\qquad\quad\,\,\,4k_4(2+k_1)=k_3^2\label{ivham3}
\end{eqnarray}
where $p$ is the canonical momenta defined by
\begin{eqnarray}
&&\hspace{-1.1cm}\mbox{\bf(iva)}\,\,p=
\frac{2(2+k_1)}{w^2\sqrt{4(2+k_1)k_4-k_3^2}}\tan^{-1}\bigg[\frac{k_3w^2+2(2+k_1)w'}
{w^2\sqrt{4(2+k_1)k_4-k_3^2}}\bigg],\qquad 4k_4(2+k_1)>k_3^2\\
&&\hspace{-1.1cm}\mbox{\bf(ivb,c)}\,\,p=
\frac{w^{(2-r)k_1}}
{\left(w'+\frac{k_3(r-1)w^2}{r(2+k_1)}\right)^{r-1}},\qquad\qquad\qquad\qquad\qquad
\qquad\qquad\quad 4k_4(2+k_1)\le k_3^2
\end{eqnarray}
The existence of the time independent Hamiltonian confirms that the system (\ref{eqiv}) is
an integrable one. However, in the following we briefly point out the method of integrating the underlying Hamilton equations of motion
associated with the Hamiltonians (\ref{ivham1}) -
(\ref{ivham3})
\subsection{\bf General solution }
\label{caseivgensol}
To derive the general solution, we use the following
canonical transformations.
{\bf{Case {\bf (iv a)}: Parametric choice $k_3^2<4k_4(2+k_1)$}}
By introducing the canonical transformation, $w=\frac{U}{P}$ and $p=\frac{P^2}{2}$, where
$U$ and $P$ are new canonical variables,
we transform the Hamiltonian (\ref{ivham1}) (for the choice $k_3^2<4k_4(2+k_1)$) to the form
\begin{eqnarray}
H\equiv E=4(2+k_1)\log\left[\left(\frac{U}{P}\right)^{2+k_1}
\sec\left[\frac{U^2\sqrt{4(2+k_1)k_4-k_3^2}}{4(2+k_1)}\right]\right]-k_3U^2.
\label{tidcanh2}
\end{eqnarray}
The underlying canonical equations of motion then become
\begin{subequations}
\label{eq39}
\begin{eqnarray}
&&\hspace{-0.5cm}U'=\frac{-4(2+k_1)^2}{P},\qquad\qquad \left('=\frac{d}{dz}\right)\label{tidcaneq2}\\
&&\hspace{-0.5cm} P'=2k_3U-\frac{4(2+k_1)^2}{U}-2U\left(\sqrt{4(2+k_1)k_4-k_3^2}\right)
\tan\left[\frac{U^2\sqrt{4(2+k_1)k_4-k_3^2}}{4(2+k_1)}\right].
\end{eqnarray}
\end{subequations}
Eq. (\ref{eq39}) can be solved in the following way.
Expressing $P$ in terms of $\dot{U}$, by using the Eq. (\ref{tidcaneq2}), and substituting it in
(\ref{tidcanh2}) we obtain
\begin{eqnarray}
E=4(2+k_1)\log\left[\mu_1(UU')^{2+k_1}\sec\left[\frac{U^2\sqrt{4(2+k_1)k_4-k_3^2}}
{4(2+k_1)}\right]\right]-k_3U^2,\label{eq41}
\end{eqnarray}
where $\mu_1=\frac{4^{-(2+k_1)}(-1)^{k_1}}{(2+k_1)^{2(2+k_1)}}$ and $E$ is an arbitrary constant.
By splitting $U'$ and $U$ in (\ref{eq41}),
\begin{eqnarray}
U'=\frac{\mbox{exp}\left[\frac{E+k_3U^2}{4(2+k_1)^2}\right]}
{U\left(\mu_1\sec\left[\frac{U^2\sqrt{4(2+k_1)k_4-k_3^2}}{4(2+k_1)}\right]\right)^{\frac{1}{(2+k_1)}}},
\end{eqnarray}
and integrating the above expression, we get
\begin{eqnarray}
z-z_0=&&F\left[\frac{1}{2+k_1},\frac{4a_1+ik_3}{8a_1(2+k_1)},
\frac{4a_1(5+2k_1)+ik_3}{8a_1(2+k_1)},-e^{2ia_1U^2}\right]\nonumber\\
&&\times\frac{(2+k_1)2^{\frac{3+k_1}{2+k_1}}e^{\frac{-(E+k_3U^2)+2ia_1U^2}{2(2+k_1)}}}{(k_3-4ia_1)
\mu_1^{\frac{1}{2+k_1}}}
\end{eqnarray}
where $a_1=\frac{\sqrt{4(2+k_1)k_4-k_3^2}}{4(2+k_1)}$ and $F$ is the hypergeometric function{\footnotesize$^{2}$}.\\
{\bf{Case {\bf (iv b)}: Parametric choice $k_3^2>4k_4(2+k_1)$}}
The Hamiltonian (\ref{ivham2}) (for the choice $k_3^2>4k_4(2+k_1)$) can be rewritten in terms of the new
canonical variables, $w=\frac{U}{P}$ and $p=\frac{P^2}{2}$, as
\begin{eqnarray}
H=2(2+k_1)rr_{12}\left(P^{2+k_1}U^{-k_1}\right)^{r_{12}}
-2^{r_{12}}k_3(r-1)U^2,\label{tidcanh1}
\end{eqnarray}
where we have defined $r_{12}=\frac{(r-1)}{(r-2)}$.
The Hamilton equations of motion corresponding to the above Hamiltonian are
\begin{subequations}
\label{eq44}
\begin{eqnarray}
&&U'=\frac{2(2+k_1)^2rr_{12}^2}{P}\left(\frac{P^{2+k_1}}
{U^{k_1}}\right)^{r_{12}},\label{tidcaneq1}\\
&&P'=2^{r_{12}}k_3(r-1)2U+\frac{2k_1(2+k_1)rr_{12}^2}{U}
\left(\frac{P^{(2+k_1)}}{U^{k_1}}\right)^{r_{12}}.
\end{eqnarray}
\end{subequations}
Now we integrate Eq. (\ref{tidcaneq1}) by
following the same analogy described in the previous subcase. First
we rewrite Eq. (\ref{tidcaneq1}) for $P$ and obtain
\begin{eqnarray}
P=\left(\frac{U^{k_1r_{12}}U'}{2(2+k_1)^2rr_{12}^2}\right)^{\frac{1}{(2+k_1)r_{12}-1}}.\label{tidcanp1}
\end{eqnarray}
\begin{eqnarray}
H=\mu_2U^{m_1}U'^{m_2}+\mu_3U^2\equiv E,\label{eq48}
\end{eqnarray}
where
\begin{eqnarray}
&&\mu_2=\frac{2(2+k_1)rr_{12}}{\left(2(2+k_1)^2rr_{12}^2\right)^{m_3(2+k_1)r_{12}}}
,\,\,\,\,\mu_3=-2^{r_{12}}k_3(r-1),\,
\nonumber\\
&&m_1=k_1\left((2+k_1)m_3-r_{12}\right),\qquad \quad m_2=r_{12}(2+k_1)m_3.\nonumber
\end{eqnarray}
From (\ref{eq48}) one can express
\begin{eqnarray}
U'=\left(\frac{E-\mu_3U^2}{\mu_2U^{m_1}}\right)^{\frac{1}{m_2}}.
\label{eq47a}
\end{eqnarray}
Integrating (\ref{eq47a}) we get
\begin{eqnarray}
z-z_0=\frac{m_2U}{m_1+m_2}
\left(\frac{\mu_2U^{m_1}}{E}\right)^{\frac{1}{m_2}}F\left[\frac{m_1+m_2}{2m_2},\frac{1}{m_2},\frac{m_1+3m_2}{2m_2},
\frac{\mu_3U^2}{E}\right],
\end{eqnarray}
where $F$ is the hypergeometric function{\footnotesize$^{2}$} and $t_0$ is an integration constant.
{\bf{Case {\bf (iv c)}: Parametric choice $k_3^2=4k_4(2+k_1)$}}
We use the same canonical transformation $w=\frac{U}{P}$, $p=\frac{P^2}{2}$
and rewrite the Hamiltonian (\ref{ivham3}) with the parametric choice $k_3^2=4(2+k_1)$ as
\begin{eqnarray}
H=\log\left[\frac{U^{k_1}}{P^{k_1}}\right]-\log[P^2]
+\frac{k_3}{4(2+k_1)}U^2.\label{tidcanh3}
\end{eqnarray}
The associated canonical equations of motion now become
\begin{subequations}
\begin{eqnarray}
&&U'=-\frac{1}{P}(2+k_1),\label{tidcaneq3}\\
&&P'=-\left(\frac{k_1}{U}+\frac{2k_3U}{4(2+k_1)}\right).
\end{eqnarray}
\end{subequations}
Rewriting (\ref{tidcaneq3}) for $P=\frac{(2+k_1)}{-U'}$ and
substituting the latter into (\ref{tidcanh3}) we get
\begin{eqnarray}
H=\log[U^{k_1}(-4U')^{2+k_1}]+\frac{k_3U^2}{4(2+k_1)}\equiv E
\end{eqnarray}
which in turn can be brought to the form
\begin{eqnarray}
U'=-\frac{1}{4}
\left(U^{-k_1}\mbox{exp}[E-\frac{k_3}{4(2+k_1)}U^2]\right)^{\frac{1}{(2+k_1)}}.
\label{53a}
\end{eqnarray}
Now integrating the above equation (\ref{53a}) we get
\begin{eqnarray}
z-z_0=\frac{\tilde{E} U^{\frac{2(1+k_1)}{2+k_1}}}
{\left(\frac{-k_3U^2}{(2+k_1)^2}\right)
^{\frac{(1+k_1)}{2+k_1}}}
\Gamma\left[\frac{1+k_1}{2+k_1},\frac{-k_3U^2}{4(2+k_1)^2}\right],
\end{eqnarray}
where $\tilde{E}=-2^{\frac{4+4k_1}{2+k_1}}\mbox{exp}\left(\frac{-E}{2+k_1}\right)$ and
$\Gamma$ is the gamma function{\footnotesize$^{24}$}.
We summarize the results obtained in this section. Solving the determining equations
(\ref{lin02})-(\ref{lin04}) we find that the system (\ref{eq6}) admits time dependent integrals for four parametric
choices and their respective equations are (\ref{caseia,b}),(\ref{td2-1}), (\ref{td3eq}) and (\ref{eqiv}). For (\ref{caseia,b}) explicit solution is deduced by
integrating the corresponding time dependent integral of motion. Using suitable
variable transformations, solutions of Eq. (\ref{td2-1}), Eq. (\ref{td3eq}) and Eq. (\ref{eqiv}) are
found in an implicit form.
\section{Connection with 2D Lotka-Volterra system (LV)}
\label{connection}
The detailed study made on the integrability of the second order
ODE (\ref{eq6}) in the previous sections helps one to identify the
dynamics of 2D-LV system which we find to be related to the second
order ODE (\ref{eq6}) under appropriate choice of $k_i$s. The
2D-LV system
\begin{subequations}
\begin{eqnarray}
&&\dot {x}=x(a_{1}+b_{11}x+b_{12}y),\nonumber\\
&&\dot {y}=y(a_{2}+b_{21}x+b_{22}y),\qquad\qquad\qquad (\ref{eq1})\nonumber
\end{eqnarray}
\end{subequations}
models the the population dynamics of two interacting
species{\footnotesize$^{3}$}. Interestingly Lotka-Volterra systems arises in
other branches of physics also such as the coupling of
waves in laser physics{\footnotesize$^{25}$} and the evolution of electrons, ions and
neutral species in plasma physics. In hydrodynamics they
model the convective instability in the Benard problem{\footnotesize$^{26}$}.
Similarly, they appear in the interaction of gases in a background
host medium{\footnotesize$^{27}$}. In the theory of partial differential equation
they can be obtained as a discretized form of the Korteweg-de
Vries equation{\footnotesize$^{28}$}.
Since the 2D-LV system is a planar dynamical system it is also
being thoroughly investigated from a mathematical point of view.
Due to the multi-faceted importance of the 2D-LV system, several
in-depth and independent studies have been made to classify the
integrable cases{\footnotesize$^{7-19}$}.
Integrals of motion of the 2D-LV system (\ref{eq1}) have been studied for several
parametric choices all of which reduces to any one of the following 3 parametric choices or subcases thereof :
\begin{eqnarray}
a_2&=&\frac{a_1b_{22}(b_{11}-b_{21})}{b_{11}(b_{12}-b_{22})},\\
b_{21}&=&\frac{b_{11}b_{22}}{b_{12}},\\
a_1&=&a_2.
\end{eqnarray}
for which integrals of motion have been explicitly deduced{\footnotesize$^{9}$}.
Interestingly, we find that all these three cases (see LV 1, LV 8, LV 15 below), in addition to
several subcases of the above parametric choices,
and their associated integrals of motion can be deduced from the
results of Eq. (\ref{eq6}) straightforwardly.
In the following we show that Eq. (\ref{eq1}) can be transformed to the form (\ref{eq6}) and thus the integrable cases
of (\ref{eq6}) can be correlated with the integrable cases of
(\ref{eq1}).
\subsection{Transformation}
\label{sec12}
\noindent To transform the system (\ref{eq1}) to the form of Eq. (\ref{eq6}), first we
rewrite Eq. (\ref {eq1a}) for the variable $y$ as
\begin{eqnarray}
y=\frac{1}{b_{12}}\left(\frac{\dot {x}}{x}-b_{11}x-a_{1}\right),\,\,b_{12}\ne0\label {eq5a}
\end{eqnarray}
Then we substitute the latter into Eq. (\ref {eq1b}) and obtain the following equation,
\begin{eqnarray}
&&\hspace{-1cm}\ddot{x}-\left(1+\frac{b_{22}}{b_{12}}\right)\frac{\dot{x}^{2}}{x}+
\left((2b_{11}\frac{b_{22}}{b_{12}}-b_{11}-b_{21})x+(2a_{1}\frac{b_{22}}
{b_{12}}-a_2)\right)\dot{x}+\left(b_{21}b_{11}-\frac{b_{22}}{b_{12}}b_{11}^{2}\right)x^{3}
\nonumber\\&&\quad
+\left(b_{11}a_2+b_{21}a_{1}-2a_{1}b_{11}\frac{b_{22}}{b_{12}}\right)x^{2}
+\left(a_1a_2-\frac{b_{22}}{b_{12}}a_{1}^{2}\right)x=0.
\label {eq5}
\end{eqnarray}
which is of the same form as (\ref{eq6}).
Now comparing (\ref{eq5}) with Eq. (\ref{eq6}) we find the parameters are connected in the following way
\begin{eqnarray}
&&\hspace{-0.5cm}k_1=-(1+\frac{b_{22}}{b_{12}}),
\quad
k_{2}=(2a_{1}\frac{b_{22}}{b_{12}}-a_2),
\quad
k_{3}=2b_{11}\frac{b_{22}}{b_{12}}-b_{11}-b_{21},
\nonumber\\
&&\hspace{-0.5cm}k_{4}=(b_{21}b_{11}-\frac{b_{22}}{b_{12}}b_{11}^{2}),\quad
k_{5}=(b_{11}a_2+b_{21}a_{1}-2a_{1}b_{11}\frac{b_{22}}{b_{12}}),
\quad k_{6}=(a_1a_2-\frac{b_{22}}{b_{12}}a_{1}^{2}).\label {eq7}
\end{eqnarray}
One may note that both the Eqs. (\ref{eq6}) and (\ref{eq5}) contain six parameters and thus one ends up with six
relations connecting them. Here we emphasize that one can obtain the results of the LV equation from the results of Eq. (\ref{eq6}) straightforwardly, while
it is difficult to deduce all the results pertaining to Eq. (\ref{eq6}) from the known results of LV equation. This is illustrated with an example.
One of the integrable parametric choices of the LV equation (\ref{eq1}) is
$a_2=\frac{a_1b_{22}(b_{11}-b_{21})}{b_{11}(b_{12}-b_{22})}$. Upon substituting the above
parametric choice in the relations (\ref{eq7}) and solving for $k_i$'s we get,
$k_5=\frac{k_2k_4(3+2k_1)}{k_3(1+k_1)}$, $k_6=\frac{k_4k_2^2(2+k_1)}{k_3^2(1+k_1)}$ and
$b_{11}=\frac{-k_3\pm\sqrt{k_3^2-4(2+k_1)k_4}}{2(2+k_1)}$ where $b_{11}$ is a real and arbitrary constant. We find that
this $b_{11}$ is equivalent to $r=\frac{-k_3\pm\sqrt{k_3^2-4(2+k_1)k_4}}{2(2+k_1)}$ of the parametric choice $k_3^2>4(2+k_1)$.
However, we have obtained integrals of motion
for all the three parametric choices $k_3^2\ge4(2+k_1)$ and $k_3^2<4(2+k_1)$ which cannot be obtained going back from the results of LV.
\section{Results in terms of Lotka-Volterra equation parameters}
\label{seclv}
\label{resultcon}
Now we rewrite the results obtained in Secs. \ref{sectid} - \ref{sectd} in terms
of the LV parameters using the relation (\ref{eq7}). In total we obtain 16 parametric choices in terms of the LV parameters
(designated as LV 1 - LV 16) and the results are
tabulated in Tables \ref{table3} and \ref{table4}, corresponding to time independent and time dependent integrals, respectively.
It may be noted that out of the 16 parametric choices, 15 of them reduce to any one of the
three parametric choices LV 1, LV 8 and LV 15 (given in the last columns of Tables \ref{table3} and \ref{table4}), while the remaining one (LV 6) is the uncoupled case.
Also we note that no new parametric choice is obtained other than the ones already reported in the literature {\footnotesize$^{9,12,13,15,16}$} as far as the Lotka-Volterra
system (\ref{eq1}) is concerned.
\vskip 6pt
\begin{sidewaystable}
\caption{Parametric cases of Eq. (\ref{eq6}) possessing time independent integrals of motion in terms of the LV
parameters}
\noindent\begin{tabular}{llll} \hline
Case\,\, & Parametric restrictions
\quad& \qquad Parametric restrictions in terms of LV parameters& \qquad Integrable choices of the LV equation
\\
&\qquad in Eq. (\ref{eq6})& &\\
\hline\\
{\bf (i)}&$k_5=\frac{k_2k_4(3+2k_1)}{k_3(1+k_1)}$& \qquad
(i) $(a_2b_{11}(b_{12}-b_{22})+a_1(b_{21}-b_{11})b_{22})
=0,$ & {\bf LV 1} $
a_2=\frac{a_1b_{22}(b_{11}-b_{21})}{b_{11}(b_{12}-b_{22})}$\\
&$k_6=\frac{k_4k_2^2(2+k_1)}{k_3^2(1+k_1)}$.&\qquad (ii) $a_1=a_2$, $b_{21}=0$ & {\bf LV 2} $a_1=a_2,\,b_{21}=0$\\
&&\\
{\bf (ii)} &$k_2=0,\,k_3=0$& \qquad $a_2-\frac {2a_1b_{22}}{b_{12}}=0$,\,
$ b_{11}+b_{21}-\frac {2b_{11}b_{22}}{b_{12}}=0$ &{\bf LV 3} $a_2=\frac{2a_1b_{22}}
{b_{12}},\, b_{21}=b_{11}\left (\frac{2b_{22}}{b_{12}}-1\right )$\\
&&\\
{\bf (iii)} &$k_1=0,\,k_2=0,$ &\qquad $\frac{b_{22}}{b_{12}}+1=0,\,2a_1\frac{b_{22}}{b_{12}}-a_2=0,$ &{\bf LV 4} $b_{12}=-b_{22},\,a_1=-\frac{a_2}{2},\,b_{21}=0$
\\
& $k_5=0$& \qquad $b_{11}a_2+b_{21}a_1-2a_1b_{11}\frac{b_{22}}{b_{12}}=0$ &{\bf LV 5} $b_{12}=-b_{22},\,a_1=a_2=0$\\
& &\\
\hline
\end{tabular}
\label{table3}
\end{sidewaystable}
\begin{sidewaystable}
\caption{Parametric cases of Eq. (\ref{eq6}) possessing time dependent integrals of motion in terms of the LV
parameters}
\noindent\begin{tabular}{llll}
\hline Case\,\, & Parametric restrictions &\qquad Parametric restrictions in terms of LV parameters & Integrable choices of the LV equation
\\
&\qquad in Eq. (\ref{eq6})&&\\
\hline
(i)&$k_4=\frac{(1+k_1) k_3^2}{(3+2 k_1)^2}$, & \qquad$\frac{b_{21}(b_{11}(b_{12}-2b_{22})+b_{21}b_{22})}{b_{12}-2b_{22}}=0,$& {\bf LV 6} $b_{21}=0,\,\,b_{12}\ne 0$\\
&$k_5=\frac{k_2k_3}{3+2k_1}$ & \qquad$\frac{b_{21}(a_1-a_2)}{b_{12}-2b_{22}}=0$ & {\bf LV 7} $a_1=a_2,\,b_{11}=-\frac{b_{21}b_{22}}{b_{12}-2b_{22}}$\\
&&&\\
(ii)& $k_4=0,$& \qquad$b_{11}b_{21}-\frac{b_{11}^2b_{22}}{b_{12}}=0,$ & {\bf LV 8} $b_{21}=\frac{b_{11}b_{22}}{b_{12}}$\\
&$k_5=\frac{k_3(k_2\pm\omega)}{2 (2+k_1)}$& \qquad$\frac{a_2b_{12}(b_{11}-b_{21})+2a_1(b_{12}b_{21}-b_{11}b_{22})\pm
a_2(b_{11}b_{12}+b_{12}b_{21}-2b_{11}b_{22})}{b_{12}-b_{22}}=0$ & {\bf LV 9} $a_1=a_2,\,b_{11}=0$,\,\,\mbox{\bf LV 10} $a_1=0,\,b_{11}=0$\\
&&&\\
(iii)& $k_3=0,\,k_4=0,$& \qquad$\frac{2 b_{11}b_{22}}{b_{12}}-b_{21}-b_{11}=0,\,b_{11}b_{21}-\frac{b_{11}^2 b_{22}}{b_{12}}=0,$& {\bf LV 11} $a_1=2a_2,\,b_{11}=0,\,b_{21}=0$\\
& $k_6=\frac{2(3+2 k_1)k_2^2}{(5+4 k_1)^2}$& \qquad$\frac{(a_1-2a_2)(a_1b_{22}+2a_2b_{22}-a_2b_{12})}{b_{12}-4b_{22}}
=0$& {\bf LV 12} $a_1=\frac{a_2(b_{12}-2b_{22})}{b_{22}},\,b_{11}=0,\,b_{21}=0$,\\
&&&\mbox{\bf LV 13} $a_1=2a_2,\,b_{11}=b_{21},\,b_{12}=b_{22}$\\
&&&\mbox{\bf LV 14} $a_1=-a_2,\,b_{11}=b_{21},\,b_{12}=b_{22}$\\
&&&\\
(iv)&$k_5=\frac{k_2k_3}{3+2k_1}$,& \qquad$\frac{b_{21}(a_1-a_2)}{b_{12}-2b_{22}}=0$,& {\bf LV 15} $a_1=a_2\label{LV15}$\\
&$k_6=\frac{(2+k_1)k_2^2}{(3+2k_1)^2}$ & \qquad$\frac{(a_1-a_2)
(a_1b_{22}+a_2(b_{22}-b_{12}))}{b_{12}-2b_{22}}=0$& {\bf LV 16} $a_1=\frac{a_2(b_{12}-b_{22})}{b_{22}},\,b_{21}=0$\\
& & &\\
\hline
\end{tabular}
\label{table4}
\end{sidewaystable}
To deduce the integrable choices LV 1 - LV 16 from the results of Eq. (\ref{eq6}) one can derive the corresponding
integrals of motion for each one of the cases from the results of the second order equation and the relation (\ref{eq7}). As we pointed out earlier, the integrability
of the parametric choices obtained in this analysis have already been established in Ref. [9,12,13,15,16].
In the following we illustrate the procedure to
deduce
the time independent integral from the results of the second order equation (\ref{eq6}) for the parametric choices LV 1 and LV 2. Integrals of motion for the remaining cases can be
deduced similarly and the procedure is straightforward. Therefore we do not present the details here.
Considering the integrals of motion of the second order system (\ref{eq6})
we note that case (i) have three types of
integrals (vide Eqs. (\ref{integral6})) depending on the values of the parameters. To
rewrite these integrals for the first order LV system (\ref{eq1}) first we
check whether the LV parametric relations LV 1 and LV 2 are consistent with
these conditions. While verifying this we find that both the LV
systems are subcases of the overdamped parametric choices $k_3^2>4k_4(2+k_1)$, see Eq. (\ref{integral6}).
Once the respective integral has been identified
then one
can replace the variable $\dot{x}$ in terms $x$ and $y$ (vide eq (\ref{eq1}))
in (\ref{integral6}). The integrals of motion
for the above two LV cases turn out to be, respectively,
\begin{eqnarray}
&&\hspace{-0.7cm}\mbox{\bf LV 1}\,\,I_1=yx^{\frac{b_{22}(b_{21}-b_{11})}
{b_{11}(b_{12}-b_{22})}}(a_1(b_{11}-b_{21})+b_{11}(b_{11}-b_{21})x
+b_{11}(b_{12}-b_{22})
y)^{\frac{b_{22}b_{11}-b_{12}b_{21}}{b_{11}(b_{12}-b_{22})}}\\
&&\hspace{-0.7cm}\mbox{\bf LV 2}\,\,I_1=(a_1+b_{22}y)(b_{11}x+
(b_{12}-b_{22})y)^{\frac{b_{22}}{b_{12}-b_{22}}}x^{\frac{-b_{22}}{b_{12}-b_{22}}}
\end{eqnarray}
We also note that the well known LV equation
\begin{subequations}
\begin{eqnarray}
&&\dot{x}=x(a_1-b_{21}y),\\
&&\dot{y}=y(a_2+b_{21}x),
\end{eqnarray}
\label{classical}
\end{subequations}
is a subcase of LV 1. The integral of motion corresponding to (\ref{classical}) is deduced from the integral (\ref{integral6}) (with $k_1=-1$) as
\begin{eqnarray}
I_1=b_{21}(x+y)+a_2\log[x]-a_1\log[y]).
\end{eqnarray}
For the remaining cases LV 3- LV 16, similar analysis can be performed straightforwardly.
\section{Conclusion}
\label{seccon}
In this paper, we have investigated the integrability properties of Eq. (\ref{eq6}) and shown
that it admits a set of integrable parametric choices. To identify them we have divided
our analysis into two categories, that is systems which admit time independent first integrals and equations which
possess time dependent integrals. After carrying out the detailed analysis we found that there exists a new equation
which admits time independent integral. To interpret this integral as a Hamiltonian we first
deduce the corresponding Lagrangian and then construct the Hamiltonian using the Legendre transformation. Since
we have identified a conservative Hamiltonian description for a dissipative system, we expect the study can be
extended to the quantum case as well in future. The other two systems which admit time independent integrals are
already known in the literature. However we have also given the Hamiltonian description for both of them.
We then moved on to identify the systems which admit time dependent integrals. Our results show that there exist
four integrable cases in (\ref{eq6}) that admit time dependent integrals. We have also reported the explicit forms of
these integrals. For the first three equations we have also found the general solution from these integrals. Since the integral
of the
fourth equation turned out to be a very complicated one it became difficult to integrate it straightforwardly. So first
we have transformed the
time dependent integral into a time independent one. Then from the latter we identified a Hamiltonian. We then
introduced a canonical transformation to this Hamiltonian and transformed the latter into a relatively simpler Hamiltonian. This Hamiltonian
has been integrated to obtain the general solution.
We have transformed the
identified integrable choices of the second order equation to the LV system. Out of the 16 LV parametric choices obtained,
15 reduces to any one of the three parametric choice LV 1, LV 8, LV 15. The 16$^{th}$ one, namely LV 6, is an uncoupled case.
Interestingly our results reproduce
all the known integrable cases of the LV system in the literature.
\section{Acknowledgements}
The work forms a part of a research project of MS and an IRPHA
project of ML sponsored by the Department of Science \& Technology
(DST), Government of India. ML is also supported by a DST Ramanna
Fellowship.
\vskip 14pt
\begin{tabular}{p{.15cm}p{14cm}}
{\footnotesize$^{1}$} &
E. Kamke: \emph{Differentialgleichungen Losungsmethoden und Losungen},
Stuggart: Teubner, 1983.\\
{\footnotesize$^{2}$} &
George M. Murphy \emph{Ordinary Differential Equations and Their Solutions},
An East-West Editon 1969, New Delhi.\\
{\footnotesize$^{3}$} &
J.D Murray \emph{Mathematical Biology} (Springer-Verlag, New York, 1989)\\
{\footnotesize$^{4}$} &
M. Lakshmanan and S. Rajasekar \emph{Nonlinear Dynamics: Integrability
Chaos and Patterns} (Springer-Verlag, New York, 2003)
\end{tabular}
\newpage
\begin{tabular}{p{.15cm}p{14cm}}
{\footnotesize$^{5}$} &
V. K.Chandrasekar, M. Senthilvelan and M. Lakshmanan,
Proc. R. Soc. London A {\bf 461}, 2451 (2005)\\
{\footnotesize$^{6}$} &
V. K.Chandrasekar, M. Senthilvelan and M. Lakshmanan, J. Nonlinear
Math. Phys. {\bf 12}, 184 (2005)\\
{\footnotesize$^{7}$} &
D.D Hua, L. Cairo, M.R. Feix, K.S. Govinder and P.G.L. Leach,
Proc. R. Soc. London A {\bf 452}, 859 (1996)\\
{\footnotesize$^{8}$} &
P.L. Sachdev and Sharadha Ramanan, J. Math. Phys. {\bf 34}
4025 (1992)\\
{\footnotesize$^{9}$} &
L. Cairo,M. R. Feix and J. Goedert, Phys. Lett. A {\bf 140}, 421 (1989)\\
{\footnotesize$^{10}$} &
L. Cairo, and M.R. Feix, J. Math. Phys. {\bf 33} 2440 (1992)\\
{\footnotesize$^{11}$} &
M.A. Almeida, M.E. Magalhaes and I.C Moreira, J. Math. Phys. {\bf 36} 1854 (1995)\\
{\footnotesize$^{12}$} &
L. Cairo, J. Llibre, J. Phys. A {\bf 33}, 2407 (2000)\\
{\footnotesize$^{13}$} &
L. Cairo, M.R. Feix and J. Llibre J. Math. Phys. {\bf 40} 2074 (1999)\\
{\footnotesize$^{14}$} &
L. Cairo and M.R. Feix J. Phys. A {\bf 25}, L1287 (1992)\\
{\footnotesize$^{15}$} &
J. Llibre and C. Valls, J. Math. Phys. {\bf 48}, 033507 (2007)\\
{\footnotesize$^{16}$} &
L. Cairo, H. Giacomini and J. Llibre, Rend. Circ. Mat. Palermo {\bf 52}, 389 (2003)\\
{\footnotesize$^{17}$} &
J. Moulin Ollagnier, Bull. Sci. Math. {\bf 121}, 463 (1997)\\
{\footnotesize$^{18}$} &
J. Moulin Ollagnier, Bull. Sci. Math. {\bf 123}, 437 (1999)\\
{\footnotesize$^{19}$}
& J. Moulin Ollagnier, Qualitative
Theory of Dynamical Systems, {\bf 2} 307 (2001)\\
{\footnotesize$^{20}$} &
V. K. Chandrasekar, S.N. Pandey, M. Senthilvelan and M. Lakshmanan,
J. Math. Gen. {\bf 47} 023508
(2006)\\
{\footnotesize$^{21}$} &
V K Chandrasekar, M Senthilvelan and M Lakshmanan Phys. Rev. E {\bf 72},
066203 (2005)\\
{\footnotesize$^{22}$} &
R Gladwin Pradeep, V K Chandrasekar, M Senthilvelan and M Lakshmanan J. Math. Phys. {\bf 50}, 052901 (2009)\\
{\footnotesize$^{23}$} &
R. Iacono, J. Phys. A : Math. Theor. {\bf 41}, 068001, (2008)\\
{\footnotesize$^{24}$} &
I.S Gradshteyn and I.M. Ryzhik, \emph {Table of Integrals, Series and Products}
(Academic Press, London, 1980)\\
{\footnotesize$^{25}$} &
W. E. Lamb, Phys. Rev. A 134, 1429 (1964)\\
{\footnotesize$^{26}$} &
F.H. Busse, \emph{Synergetics} (Springer, Berlin, 1978)\\
\footnotesize$^{27}$ &
R. Lupin and G. Spiga, Phys. Fluids 31, 2048 (1988)\\
\footnotesize$^{28}$ &
O.I. Bogoyavlensky, Phys. Lett. A 134, 34 (1988)
\end{tabular}
\end{document}
|
\section{Introduction}
As is well known, the study of polarization phenomena in hadron and hadron-nucleus
interactions gives more detailed information on dynamics of their interactions
and the structure of colliding particles. The quark structure and relativistic
effects of light nuclei, in particular, deuterons, is one of important problems in nuclear
physics at intermediate and high energies. The theoretical and experimental
study of reactions like the elastic
$e-d$
\cite{Arnold:1981}
and $p-d$
\cite{Arvieux:1984,Keister:1981}
scattering, deuteron break-up reactions induced
by electrons or protons
\cite{Rekalo,Rekalo1}
and the deuteron stripping processes on protons and
nuclei at intermediate and high energies
\cite{Perdrisat:1987,DL:1990},
can allow us to find out new information on the deuteron structure at short
distances. The elastic backward proton-deuteron scattering has been
experimentally and theoretically studied in Saclay
\cite{Arvieux:1984}, Dubna
and at JLab (USA)
\cite{Azhgirei,Azhgirei1,Azhgirei2,Punjabi}.
Usually these processes are analyzed within a simple
impulse approximation.
Up to now all these data have not been described within the one-nucleon exchange
model (ONE) including even the relativistic effects in the deuteron
\cite{KermKissl:1969,IL:2001}.
In this paper we analyze the elastic backward proton-deuteron scattering within
the relativistic approach including the ONE and the high order graphs
corresponding to the emission, rescattering and absorption of the virtual pion
by a nucleon of deuteron
\cite{ILV:2007}.
\section{Light cone dynamics for $dp\rightarrow pd$}
\subsection{The leading order diagrams}
Let us analyze the elastic $dp\rightarrow pd$ scattering within the Weinberg
diagrammatic technique in the infinite-momentum frame (IMF)
\cite{Weinberg:1966,Brodsky:1973}.
The four-momentum of the fast deuteron $P_d$ and its nucleons $k_1$ and $k_2$
have the following components in the IMF:
\begin{eqnarray}
P_d\left(P+\frac{m_d^2}{2P}, 0, P\right) \\
\nonumber
k_1\left(xP+\frac{m_t^2}{2xP}, {\bf k}_t, xP\right) \\
\nonumber
k_2\left((1-x)P+\frac{m_t^2}{2(1-x)P}, -{\bf k}_t, (1-x)P\right)~,
\label{def:IMF}
\end{eqnarray}
where $P$ is the magnitude of the three-momentum of the incident particle,
in particular, deuteron;
$x=(E(k)+k_z)/(E_d(p_d)+p_{dz})$ is the light cone variable,
$E(k)=\sqrt{{\bf k}^2+m^2_N}$ and
$E_d(p_d)=\sqrt{{\bf p}^2_d+m^2_N}$ are the total energies of the nucleon inside
deuteron and of the deuteron, respectively; ${\bf k},{\bf p}_d$ are three-momenta of
this nucleon and deuteron; $k_z,p_{dz}$ are their longitudinal components;
$m_N$ is the nucleon mass.
The first order diagrams or the one-nucleon exchange graphs are presented in Fig.1.
In the general relativistic case the deuteron vertex $dpn$ does not reduce merely
to dissociation of the deuteron into two nucleons; it may also include the
annihilation ${\bar N}d\rightarrow N$ and therefore, the deuteron decay vertex
cannot always be reduced to an ordinary deuteron wave function whose square is
a probability of finding a nucleon in the deuteron with a definite momentum. As
is well known, Feynman graph of $n$th order is equivalent to $n!$ time ordered
graphs of the old perturbative theory (OPT). If $dp$ processes are analyzed within
the IMF, then many graphs in the old perturbative theory make a contribution
of order $O(1/P)$
\cite{Weinberg:1966,Brodsky:1973}.
There remain only the diagrams that correspond
to the dissociation of the deuteron into two nucleons.
As an example, Fig.1 shows the Feynman diagram of the $dp\rightarrow pd$ process
(Fig.1a) and two equivalent diagrams of the OPT, ordered in the time $t$
(Figs.1(b,c)). The graph of Fig.1a corresponds to the deuteron dissociation and
the graph of Fig.1b is the so called $z$-diagram corresponding to the
${\bar N}d\rightarrow N$ annihilation.
Therefore, one can introduce the concept of a deuteron wave function (d.w.f.)
with the usual probability interpretation. In this case the d.w.f. $\Psi$ depends
on the following relativistic invariant variable
\cite{FS:1981,SF:1980}:
\begin{equation}
k^2~=~\frac{m_t^2}{4x(1-x)}~-~m_N^2.
\label{def:ksqure}
\end{equation}
In \cite{SF:1980} it was shown that the variable $k^2$ is proportional to the
difference of the initial and final energies in the $d\rightarrow pn$ dissociation
vertex of Fig.1b.
Note that in each vertex of the OPT graph (Figs.1(b,c)) the three-momentum is
conserved but the energy is not, although the energy and the three-momentum is
conserved for the complete reaction. All the particles, including those in the
intermediate state, are on the mass shell. In the Feynman-diagram technique the
four-momentum is conserved at each vertex of the diagram (Fig.1a), but the
intermediate particle with four-momentum $k_N$ is off the mass shell, e.g.,
$k_N^2\neq m^2$.
\begin{figure}[htb]
\includegraphics[width=0.48\textwidth]{3.eps}
\caption[Fig.1]{The Feynman graph corresponding to the one-nucleon exchange
graph (a) for the process $d p\rightarrow p d$ (a) and its equivalent to two
time-ordered diagrams within the OPT (b,c) .}
\label{Fig.1}
\end{figure}
In such approach
\cite{FS:1981,Kobushkin:1982},
the d.w.f. $\Psi_d$ is related to the nonrelativistic d.w.f. $\Phi^{n.r.}_d$,
that depends on the relativistic invariant variable $k^2$ given by Eq.
(\ref{def:ksqure}):
\begin{equation}
\Psi_d(x,k_t)~=~\left(\frac{m^2_t}{4x(1-x)}\right)^{1/4}\Phi^{n.r.}_d(k^2)~,
\label{def:psid}
\end{equation}
with the following normalization equation:
\begin{equation}
\frac{1}{2}\int_0^1\frac{dx}{x(1-x)}\int\mid\Psi_d(x,k_t)\mid^2d^2k_t~=~1~.
\label{def:normpsi}
\end{equation}
There are also several covariant approaches to construct the relativistic d.w.f.
For example, one of them \cite{BuckGross:1979,Tokarev:1991} is based on the
assumption that one nucleon inside deuteron is on its mass shell ($k_1^2=m_N^2$)
while the other one is off-mass-shell ($k_2^2-m_N^2<0$ or $k_2^2-m_N^2>0$).
In that case additionally to the $S$- and $D$-wave functions ($u,w$) in the d.w.f.
two components also appear: the triplet $P$-state wave function ($v_t$) and
the singlet $P$-state wave function ($v_s$). Actually, as is shown in
\cite{BuckGross:1979},
the total probability for the $P$-wave components in the d.w.f. ($W_p$) is too
small, it is less than $1\%$. Recently it has been shown
\cite{IL:2001}
that within the relativistic one-nucleon exchange model (RONE) proposed in
\cite{BuckGross:1979}
one can not describe all the polarization experimental data for the elastic
backward $p-d$ scattering even by increasing the parameter $W_p$ till several
percent.
Another interesting covariant approach within the light cone dynamics to
construct the relativistic d.w.f. \cite{Karmanov,Karmanov1} is based on the
three-dimensional formalism for the quantum field theory. Within this approach two
nucleons inside the deuteron are mass-shell, however, the four-vector $\omega$
determining the light cone surface is introduced to satisfy the three-momentum
and energy conservation by the deuteron break-up. The relativistic d.w.f.
constructed within this approach depends on the relativistic invariant variable
$k^2$ and direction vector ${\bf n}$ of the IMF. For example, choosing
${\bf n}$ in the opposite direction to the deuteron moving in the IMF, one gets
the same dependence of $\Psi_d$ on $k^2$ given by Eq.
(\ref{def:psid})
like in
\cite{FS:1981,Kobushkin:1982}.
Other approaches to get the relativistic d.w.f. can be found, for example, in
\cite{Keister:1981,Tjon:1982}
and
\cite{Lev}.
The amplitude for the elastic backward $dp$ scattering within the impulse
approximation of the OPT (Fig.1b) in the LCD has the following form
\cite{GL:1993,Yudin:2000}:
\begin{equation}
{\cal F}^{(1)}_{LCD}=\sqrt{3}\frac{M_d^2-m_N^2/(x(1-x))}{1-x}\Psi_d^2(k^2)~,
\label{def:FLCONE}
\end{equation}
On the other hand, the amplitude corresponding to the one-nucleon exchange Feynman
graph (Fig.1a) can be presented in the following form
\cite{Rekalo,Sitnik:1994,Sitnik96,Sitnik98,IL:2001}:
\begin{equation}
{\cal F}^{RONE}=8\sqrt{3} m_N(m_N^2-u)\Psi_d^2(k^2)~,
\label{de:FRONE}
\end{equation}
where
$u$ is the square of momentum transfer from initial deuteron to final proton;
$k^2$ can be also written in the following form:
$k^2=\frac{1}{4}s_{12}-m^2_N$; $s_{12}=(k_1+k_2)^2$; $k_1,k_2$ are the four-momenta
of neutron and proton in the deuteron. Unfortunately, the ONE and the RONE do not
allow a satisfactory description of all the observables at the kinetic energy of
backward scattered protons $T_p>0.6$ GeV
\cite{IL:2001}.
\subsection{Next to leading order diagrams}
As was shown in
\cite{Wilkin:1969,Nakamura:1985},
the contribution of the high-order graphs in the $p-d$ backward elastic scattering
corresponding to the emission, scattering and absorption of the virtual pion by a
deuteron nucleon, can be sizable at initial energies corresponding to possible
production of the $\Delta$-isobar at the $\pi-N$ vertex, see Fig.2(a,b). In
\cite{Kaptari:1998} this process was analyzed within the
Bethe-Saltpeter approach using the impulse approximation, however, the one-pion
exchange contribution in the intermediate state was also included. The
contribution of the $\Delta$-isobar exchange graph to the elastic $p-d$
scattering was studied in
\cite{Uzikov,Uzikov1}.
All these models reproduce the gross features of the backward cross section
and describe the experimental data rather well, however, there is a
difficulty to describe both the cross section and the polarization observables
like the tensor analyzing power $T_{20}$ of deuteron and the transfer
polarization $\kappa_0$ within all these approaches. In
\cite{Vasan:1973} it is stressed that the energy dependence of $T_{20}$ should
be sensitive to the microscopic structure of the model.
\begin{figure}[htb]
\includegraphics[width=0.45\textwidth]{1.eps}
\caption[Fig.2]{The triangle Feynman graph with one-pion
exchange for the process
$d p\rightarrow p d$, and its equivalent graph (b).}
\label{Fig.2}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=0.45\textwidth]{2.eps}
\caption[Fig.3]{The time ordered graphs with one-pion exchange
within the OPT (a,b) and their equivalent graph (c)
for the process $d p\rightarrow p d$.}
\label{Fig.3}
\end{figure}
For example, the inclusion of the
triangle Feynman graph of Fig.2a in addition to the one-nucleon exchange
diagram of Fig.1a allows to describe the experimental data on $T_{20}$
at the deuteron momentum $P_d$ less than $4$ GeV$/c$ only
\cite{Nakamura:1985},
and this calculation does not describe the tail of $T_{20}$ at$P_d\geq 4$ GeV$/c$.
The corrections to the ONE graph of Fig.1a were also analyzed in other papers,
see, for example,
\cite{Uzikov}
and references therein. As was shown in
\cite{DL:1990,DL1:1990,DN:1989},
the contribution of the one-pion exchange graphs to the
deuteron stripping reaction of the type $d+p\rightarrow p+X$, can be also sizable
at the initial energies close to a possible $\Delta$-isobar production in the
intermediate state.
Let us apply the Weinberg diagram formalism
\cite{Weinberg:1966}
within the LCD analyzed in
\cite{GL:1993,DL:1990} for the deuteron stripping reactions $dp\rightarrow pX$
to the elastic $d-p$ scattering. As is known, Feynman graph of the $n$th order
is equivalent to $n!$ time-order graphs of the old perturbative theory (OPT). In
\cite{Brodsky:1973}, it is shown that the time ordered diagrams of the order $n$
at $x>0$, are finite, whereas at $x<0$ they can be suppressed as $1/P^{n-1}$. One
Feynman diagram of the $3$ order presented in Fig.2a is equivalent to $6$ time-
ordered diagrams calculated within the OPT
\cite{Brodsky:1973,GL:1993},
however, only two diagrams presented in Fig.3(a,b) are finite, while the other $4$
graphs are suppressed as $1/P^2$ or as $1/P$ when the spin structure of the
vertices is included, therefore they can be neglected at high values of $P$.
Actually, these results were obtained in \cite{Brodsky:1973} for a $\phi^3$
interaction, nevertheless, it can be also applied for $d-p$ reactions, shown in
\cite{DL:1990,DL1:1990,GL:1993}.
The calculation of the graphs of Fig.3(a,b) is equivalent to
the calculation of the diagram in Fig.3c.
The four-momentum of the fast deuteron $P_d$ and its nucleons $k_1$ and $k_2$ are
represented within the IMF in the same forms, as in \cite{DL:1990}, see Eq.(\ref{def:IMF}).
The part of the $d-p$ elastic scattering amplitude corresponding to the graph of
Fig.(1c) within the OPT in the LCD, can be presented in the following form:
\cite{GL:1993,DL:1990}:
\begin{widetext}
\begin{eqnarray}
{\cal F}^{(3)}_{LCD}=-\left(\frac{g P}{(2\pi)^3}\right)^2\int\frac{dx dx^\prime
d^2k_t d^2k^\prime_t}{4\sqrt{E_N(k_1)E_N(k_2)}4\sqrt{E_N(k_1^f)E_N(k_2^f)}
E_\pi(q_1)E_\pi(q_2)} \\
\nonumber
\Psi^+_d(x^\prime, k^\prime_t)\Gamma_N^{(2)}
G(q_2)F_{\pi N}(q_2^2)f_{\pi N}^{el}(s_1,t_1)F_{\pi N}(q_1^2)G(q_1)\Gamma_N^{(1)}
\Psi_d(x,k_t)~,
\label{def:lctr}
\end{eqnarray}
\end{widetext}
where $x,x^\prime$ are the light cone variables for nucleons inside the initial and
final deuterons, respectively, while ${\bf k}_t,{\bf k}^\prime_t$ are the transverse
momenta of these nucleons; the energy Green functions $G(q_{1,2})$ within the OPT
have the following forms:
\begin{eqnarray}
G(q_1)=\left(E_N(p^f)-E_N(k_1)-E_\pi(q_1)+i\epsilon\right)^{-1}\\
\nonumber
G(q_2)=\left(E_N(k_1^f)-E_N(p )-E_\pi(q_2)-i\epsilon\right)^{-1}~,
\label{def:GOPT}
\end{eqnarray}
$$E_N(k_{1,2})=\sqrt{{\bf k}_{1,2}^2+m_N^2},$$
$$ E_N(k^f_{1,2})=\sqrt{{{\bf k}^f}^2_{1,2}+m_N^2},$$
$$E_\pi(q_{1,2})=\sqrt{{\bf q}_{1,2}^2+\mu_\pi^2},$$
where ${\bf k}_{1,2}$ and ${\bf k}^f_{1,2}$ are the three-momenta of nucleons inside
the initial and final deuterons respectively; ${\bf q}_{1,2}$ are three-momenta
of the intermediate pion in Fig.(1c); $\mu_\pi$ is the pion bar mass;
$F_\pi(q^2_{1.2})$ is the pion form factor taking into account the virtuality of
the intermediate pion depending on its four-momentum squared $q_{1,2}^2$. The
pion form factor was taken in the monopole form
$F_\pi=\Lambda_\pi^2/(\Lambda_\pi^2-q_{1,2}^2)$ , where the value of the cut-off
parameter $\Lambda_\pi$ was taken as $\Lambda_\pi=0.7-0.8$ GeV$/c$ also used in
\cite{DL:1990,DL1:1990}
and enabled us to make a rather satisfactory description of the experimental on the
$dp\rightarrow pX$ reactions. The d.w.f. $\Psi_d$ is related to the nonrelativistic
d.w.f $\Phi^{n.r.}_d$, see Eq.(\ref{def:psid}), that has the following form
\cite{Reid:1968}:
\begin{equation}
\Phi^{n.r.}_d(k^2)~=~\left(u(k^2)~-~\frac{1}{\sqrt{8}}w(k^2){\cal S}_{np}\right)
\chi_{1M}~,
\label{def:Phinr}
\end{equation}
where $u(k^2)$ and $w(k^2)$ are the $S$- and $D$-waves of the d.w.f., $\chi_{1M}$
is the spin triplet wave function, ${\cal S}_{np}=3({\vec\sigma}_n\cdot{\hat{
\vec k}})({\vec\sigma}_p\cdot{\hat{\vec k}})-({\vec\sigma}_n\cdot{\vec \sigma}_p)$;
$\hat{\vec k}$ is the unit vector of the relative momentum of nucleons in the
deuteron; $f_{\pi N}(s_1,t_1)$ is the amplitude of the elastic $\pi-N$ scattering,
see Fig.2b; the vertex $\Gamma_N^{(1)}={\bar u}(p^f)\gamma_5 u(k_1)=\xi^+
({\vec\sigma}_N\cdot{\vec\tau}_1)\xi$ corresponds to the absorption of the virtual
pion by the final nucleon (the bottom $\pi N$ vertex in Fig.2b) and the vertex
$\Gamma_N^{(2)}={\bar u}(k_1^f)\gamma_5 u(p)=\xi^+{\vec\sigma}_N\cdot{\vec\tau}_2\xi$
corresponds to the emission of the virtual pion by initial nucleon (the top $\pi-N$
vertex in Fig.2), here $u$ is the four-component spinor of the nucleon,
whereas ${\bar u}$ is the conjugated four-component spinor of the nucleon in the
deuteron; $\xi$ is the two-component spinor of the nucleon; the forms for the vectors
${\vec \tau}_1,{\vec \tau}_2$ are presented in the APPENDIX.
The amplitude of the elastic $\pi-N$ scattering $f_{\pi N}(s_1,t_1)$ depends on the
square of the energy in the $\pi-N$ c.m.s. $s_1=(q_2+k_2)^2$ and the
the four-momentum transferred square $t_1=(q_2-q_1)^2$, where $q_2,q_1$ are the
four-momenta of the virtual pion before and after the $\pi-N$ scattering, $k_1,k_2$
are the four-momenta of proton and neutron in the initial deuteron with the
four-momentum $p_d$, whereas $k_1^\prime,k_2\prime$ are the four-momenta of these
nucleons in the final deuteron with the four-momentum $p_d^\prime$. In the $pd$
c.m.s. the three-momenta of nucleons in the initial and final deuteron can be
presented in the following form:
\begin{eqnarray}
{\vec k}_1~=~\frac{1}{2}{\vec p}_d-{\vec k},~{\vec k}_2~=~\frac{1}{2}
{\vec p}_d+{\vec k};\\
{\vec k}_1^f~=~\frac{1}{2}{\vec p}_d^ f-{\vec k}^\prime,~{\vec k}_2^f~
=~\frac{1}{2}
{\vec p}_d^f+{\vec k}^ \prime~,
\label{def:vectorsk}
\end{eqnarray}
where ${\vec k}$ and ${\vec k}^\prime$ are the relative momenta of nucleons in the
initial and final deuterons respectively.
To calculate the amplitude ${\cal F}^{(3)}$ given by Eq.(\ref{def:GOPT}), we
removed the integral $f_{\pi N}$ at the mean value of the nucleon relative momentum
in deuteron ${\bar \mid{\vec k}\mid}\simeq 0.07-0.1$ GeV$/c$ because the d.w.f.
$\Psi_d(x,{\vec k}_t)$ is sharply decreasing as $x$ and $\mid{\vec k}_t\mid$ grow,
as it was done, for example, in
\cite{DL:1990,Kolybasov:1971}.
The amplitude $f_{\pi N}^{el}$ was presented in the following form
\cite{Ponomarev:1976}:
\begin{equation}
f_{\pi N}^{el}(s_1,t_1)~=~A(s_1,t_1)~+~iB(s_1,t_1)({\vec\sigma}\cdot{\vec n})~,
\label{def:fpiN}
\end{equation}
where $n=({\vec q}_2^*\times {\vec q}_1^*)/\mid({\vec q}_2^*\times {\vec q}_1^*)\mid$
is the unit vector, ${\vec q}_2^*$ and ${\vec q}_1^*$ are the three-momenta of the
intermediate pion before and after $\pi-N$ scattering in the $\pi-N$ c.m.s.
The details for the kinematics corresponding to the elastic $\pi-N$ scattering and the
backward $d p$ scattering are presented in the APPENDIX. The functions $A(s_1,t_1)$ and
$B(s_1,t_1)$ were found from the phase shift analysis for the elastic $\pi N$
scattering
\cite{Strakovsky:2003}.
\subsection{Observables for $dp\rightarrow pd$ reaction }
We calculated the differential cross section $d\sigma/d\Omega$, the transfer
polarization $\kappa$ and the tensor analyzing power $T_{20}$.
\begin{equation}
\frac{d\sigma}{d\Omega}~=~\frac{1}{64\pi^2 s}\mid{\cal F}^{tot}\mid^2~,
\label{def:diffcrs}
\end{equation}
where the total amplitude ${\cal F}^{tot}$ calculated, for example, within the
LCD has the following form:
\begin{equation}
{\cal F}^{tot}_{LCD}~=~{\cal F}^{(1)}_{LCD}~+~{\cal F}^{(3)}_{LCD}~.
\label{def:FtotLCD}
\end{equation}
Assuming that the calculation of the triangle graphs of Fig.3 within the LCD
can give the same results as the calculation of the
triangle Feynman graph of Fig.2 we also compute the sum of the Feynman graph
of Fig.2b and the diagram of Fig.3c. The total amplitude within this combined
relativistic calculation (RC) is presented in the following form:
\begin{equation}
{\cal F}^{tot}_{RC}~=~{\cal F}^{(RONE)}~+~{\cal F}^{(3)}_{LCD}~.
\label{def:FtotRC}
\end{equation}
Then we compare all the results obtained within the LCD using the total amplitude given
by Eq.(\ref{def:FtotLCD}) and the RC using Eq.(\ref{def:FtotRC}) for ${\cal F}^{tot}$.
The reason for this assumption is based on the results of \cite{Brodsky:1973} which show that
the triangle diagrams with $x<0$ can be more suppressed than the $Z$-diagrams within the impulse
approximation of Fig.1c.
The tensor analyzing power of the deuteron $T_{20}$ has the following form:
\begin{equation}
T_{20}~=~\frac{Tr\left(\rho_d({\cal F}^{tot})^+\Omega_{20}{\cal F}^{tot}\right)}
{Tr\left(\rho_d({\cal F}^{tot})^+{\cal F}^{tot}\right)}~,
\label{def:T20}
\end{equation}
where \cite{Vasan:1973}
\begin{equation}
\!\Omega_{20}~=~\frac{1}{\sqrt{2}}\left(3S_z^2-2\right)\equiv
\frac{1}{\sqrt{2}}\left(\frac{3}{2}(1+\sigma_{pz}\sigma_{nz})-2\right)\!
\label{def:Omega20}
\end{equation}
is the spin-tensor operator corresponding to the tensor component of the deuteron
polarization. Here $S_z$ is the projection of the deuteron spin operator $S$ on the
quantization axis $z$, which in our case is the direction of the initial deuteron,
whereas $\sigma_{pz}$ and $\sigma_{nz}$ are the $z$ components of the Pauli matrices
corresponding to the proton and the neutron, respectively.
The transfer polarization and the tensor analyzing power for the deuteron
were studied within the impulse approximations in
\cite{Ableev:1988,Strokovsky:1999}.
The transfer polarization is defined as
\begin{equation}
\kappa_0~=~\frac{{\vec{\cal P}}^\prime\cdot {\vec n}}{{\vec{\cal P}}\cdot {\vec n}
(1-\rho_{20}T_{20})}~,
\label{def:kappa}
\end{equation}
where ${\vec{\cal P}}^\prime$ is the vector polarization of the final proton,
${\vec n}$ is the unit vector transverse to the reaction plane, ${\vec{\cal P}}$
is the vector polarization of the initial deuteron and
\begin{equation}
{\vec{\cal P}}^\prime\cdot {\vec n}~=~\frac{Tr\left(\rho_d({\cal F}^{tot})^+
{\vec\sigma}\cdot{\vec n}{\cal F}^{tot}\right)}
{Tr\left(\rho_d({\cal F}^{tot})^+{\cal F}^{tot}\right)}~,
\label{def:Pprimen}
\end{equation}
$\rho_d$ is the density matrix of the deuteron, it has the following form:
\begin{equation}
\rho_d~=~\frac{1}{3}P_T\left(1+\frac{3}{2}{\vec{\cal P}}\cdot {\vec S}-
\frac{1}{2}\rho_{20}(3S_z^2-2)\right)~.
\label{def:rhod}
\end{equation}
Here $P_T=(3+{\vec\sigma}_p\cdot {\vec\sigma}_n)/4$ is the projection operator of
the triplet deuteron state, $\rho_{20}$ is its tensor polarization.
Calculating the traces in Eqs.
(\ref{def:Pprimen},\ref{def:T20})
we have the following general form for $\kappa$ and $T_{20}$:
\begin{equation}
\kappa_0~=~\frac{u^2(k^2)-\frac{1}{\sqrt{2}}u(k^2)w(k^2)-w^2(k^2)+
{\tilde\Delta}_\kappa}
{(u^2(k^2)+w^2(k^2)+\Delta)(1-\rho_{20}T_{20}^{sp})}~,
\label{def:kappatot}
\end{equation}
\begin{equation}
T_{20}~=~\frac{1}{\sqrt{2}}\frac{2\sqrt{2}u(k^2)w(k^2)-w^2(k^2)+
{\tilde\Delta}_{T_{20}}}{u^2(k^2)+w^2(k^2)+\Delta}~,
\label{def:T20tot}
\end{equation}
where ${\tilde\Delta}_\kappa,{\tilde\Delta}_{T_{20}}$ and $\Delta$ are the
corrections due to the contributions of the triangle graphs. Here $T_{20}^{sp}$ is
the tensor analyzing power of the deuteron calculated within the spectator model
\cite{Ableev:1988}
that has the form given by Eq.
(\ref{def:T20tot})
at ${\tilde\Delta}_{T_{20}}=\Delta=0$. In the spectator model, when
${\tilde\Delta}_\kappa=\Delta=0$, the transfer polarization was analyzed in
details in
\cite{Strokovsky:1990,Strokovsky:1999}.
The forms for the correction functions ${\tilde\Delta}_\kappa,\Delta$ and
${\tilde\Delta}_{T_{20}}$ are presented in the APPENDIX.
\section{Results and discussion}
We calculated the center-of-mass differential cross section $d\sigma/d\Omega$,
the tensor analyzing power of the deuteron $T_{20}$ and the transfer polarization
$\kappa_0$ in the elastic backward $p-d$ scattering.
This calculation was done within the RONE (the graph of Fig.1a)
and the impulse approximation of the LCD (the graph of Fig.1b).We also calculated these
observables including both simple graphs of Fig.1 and the triangle graphs of Fig.3.
These results obtained within the LCD and
the RC are presented in Figs.(4,5) as a function of the deuteron momentum $p_d^{l.s.}$
in the laboratory system (l.s.).
In Fig.4 curves 1 and 2 correspond to the total calculation within the RC,
see Eq.(\ref{def:FtotRC}) for ${\cal F}^{tot}_{RC}$ with the Reid soft core
d.w.f. \cite{Reid:1968} and the Argon-18 d.w.f. \cite{AV18} respectively;
curves 3 and 4 correspond to the LCD, see Eq.(\ref{def:FtotLCD}) for
${\cal F}^{tot}_{LCD}$ with the same kinds of the d.w.f.; curves 5 and 6 correspond
to the RONE (Fig.1a) and the LCD impulse approximation (Fig.1b) with the Reid soft
core d.w.f. \cite{Reid:1968} and curves 7,8 correspond to the same calculations as
for curves 5,6 but with the AV18 d.w.f. \cite{AV18}.
One can see from Fig.4 that the total calculation within the LCD and RC using
both the Reid soft cor{e d.w.f. and the AV18 d.w.f. give approximately the same
results for the differential cross section which are very close to the experimental data
that are taken from \cite{Azhgirei}. As is seen from Fig.4 both impulse
approximations corresponding to Fig.1a and Fig.1b do not describe the experimental
data on $d\sigma/d\Omega$ at $p_d^{l.s.}~>~1.5$ GeV$/c$.
\begin{center}
\begin{figure}[htb]
\includegraphics[width=0.45\textwidth]{4.eps}
\caption[Fig.4]{The center-of-mass differential cross section
$d\sigma/d\Omega_{c.m.s.}$ for the elastic backward $p-D$ scattering
as a function of the deuteron momentum $p_d^{l.s}$ in the laboratory system.}
\end{figure}
\end{center}
\begin{figure}[htb]
{\includegraphics[width=0.45\textwidth]{5.eps}}
{\includegraphics[width=0.45\textwidth]{6.eps}}
\caption[Fig.5]{The tensor analyzing power of the deuteron $T_{20}$ as a function
of $p_d^{l.s}$ (top)
and the transfer polarization $\kappa_0$ as a function of $p_d^{l.s}$ (bottom).}
\end{figure}
In Fig.5 the tensor analyzing power of the deuteron $T_{20}$ (top ) and the
transfer polarization $\kappa$ (bottom) are presented as a function of the
initial deuteron momentum $p_d$ in the l.s. using the Reid
soft core $N-N$ potential for the d.w.f.\cite{Reid:1968}.
Curves 1 and 2 in Fig.5 correspond to the total calculation within the RC and the LCD
respectively, whereas curves 4 and 5 correspond to the RONE calculation (Fig.1a) and
the LCD impulse approximation (Fig.1b). One can see from Fig.5 that the impulse
approximations (Fig.1a and Fig.1b) do not describe $T_{20}$ and $\kappa_0$ at
$p_d^{l.s.}~>~1$ GeV$/c$. As is seen from Fig.5 (bottom), the total calculations of
the transfer polarization $\kappa_0$ within both the RC and the LCD give the same
results at $p_d^{l.s.}~\leq~4$ GeV$/c$ which are very close to the experimental data.
The not so large difference between the RC and the LCD calculations of $\kappa_0$
appears at $p_d^{l.s.}~>~4$ GeV$/c$, where no experimental data are available now.
Therefore, analyzing the experimental data on the transfer polarization
one can not differentiate between the RC and the LCD calculations.
In contrast, Fig.5 (top) shows that the tensor analyzing power $T_{20}$ is very
sensitive to the total calculations within the RC and the LCD approximations.
As is seen from Fig.5 (top), the total RC calculation results in a better description
of the experimental data on $T_{20}$ in the whole region of the initial deuteron momenta, whereas
the LCD calculation gives a worse description of the data at $1.2~<~p_d^{l.s.}~<~1.8$ GeV$/c$
and especially at $p_d^{l.s.}~>~5$ GeV$/c$. It can be due to a sizable contribution from the
$Z$-diagram of Fig.1c to $T_{20}$ that is included by the Feynman graph of Fig.1a corresponding
to the relativistic one nucleon exchange (RONE).
On the other hand, as is mentioned above, the inclusion of the relativistic triangle Feynman
graph of Fig.2a in \cite{Nakamura:1985} did not allow a description of $T_{20}$ at
$p_d^{l.s.}~>~4$ GeV$/c$
that corresponds to the intradeuteron nucleon momenta $k~>~0.5$ GeV$/c$ or the light cone
variables $x~>~0.4$ \cite{GL:1993}. It can probably be caused by the following. In the calculation of
the Feynman graph of Fig.2a the relativistic invariant $d-N$ vertex is related in \cite{Nakamura:1985}
to the nonrelativistic d.w.f., while within the LCD we relate the $d-N$ vertex to $\Phi^{n.r.}_d(k^2)$
for the time-ordered graphs corresponding only to the deuteron dissociation (Fig.3(a,b)) and neglect
the $Z$-diagrams in the triangle graphs corresponding to the annihilation ${\bar N}d\rightarrow N$.
This is the difference between our calculation of the triangle diagram within the LCD (Fig.3) and
the calculation of the Feynman triangle graph (Fig.2a) in \cite{Nakamura:1985}.
Note that we also calculated all the observables presented in Figs.(4,5) using the CD Bonn d.w.f.
\cite{CDBonn} and the N3L0 d.w.f. \cite{N3LO}; however, the description of $T_{20}$ and
$\kappa_0$ was worse, especially at $p_d^{l.s.}>2$ GeV$/c$. Therefore, we do not present these
results because the figures will be rather cumbersome. Nevertheless, in the APPENDIX we present
the approximations of these d.w.f. by the simple Gauss forms that could be useful for other
calculations.
\section{ Conclusion}
The theoretical analysis of the elastic backward $p-d$ or the forward $d-p$ scattering
within the light cone dynamics allows us to draw the following conclusions. The calculation
of the differential cross section and the polarization observables, the tensor analyzing
power $T_{20}$ and the transfer polarization $\kappa_0$ within the impulse approximation
(diagrams of Fig.1) is not able to describe the experimental data at the initial deuteron
momenta $p_d\geq 1.2 $ GeV$/c$. In this kinematic region the contribution
of the triangle graphs (Fig.3) is very sizable because it is mainly due to the
possible creation of the $\Delta$ isobar in the intermediate state. The inclusion
of these graphs results in a rather satisfactory description of the experimental data
on the differential cross section $d\sigma/d\Omega$ in a wide region of the
initial momenta. We show that the contribution of the RONE graph (Fig.1a)
and the LCD impulse approximation (Fig.1b) give approximately similar results for the differential cross
section; therefore, the contribution of the $Z$-diagram
(Fig.1c) to $d\sigma/d\Omega$ is not large. However, its contribution to the tensor analyzing
power is sizable because the total RC calculation including graphs of Fig.1a and Fig.3
describes the experimental data on $T_{20}$ better than the total LCD calculations
(graphs of Fig.1b and Fig.3). The experimental data on the transfer polarization $\kappa_0$
are described rather satisfactorily by both the total RC and total LCD calculations.
One can conclude that the calculation of all the observables for the elastic
backward $p-d$ scattering within the light cone dynamics including the triangle graphs
of Fig.3 and the $Z$-diagrams of Fig.1c results in a rather satisfactory description
of the experimental data at initial deuteron momenta up to $7 GeV/c$. Note that we
do not include the six-quark admixture in the deuteron wave function.
This effect can probably be important at larger initial momenta because the
contribution of the graphs of Figs.(1-3) decreases with increasing $p_d^{l.s}$, as is shown
in Fig.4.
{\bf Acknowledgment.}
We are very grateful to E.A.Strokovsky for extremely useful discussions and
help in the preparation of this paper. We also thank
F.Gross, V.A.Karmanov, A.P.Kobushkin, I.M.Sitnik and Yu.N.Uzikov
for very useful discussions. This work was supported in part by the RFBR grant
No. 08-02-01003.
\begin{widetext}
\section{\bf Appendix}
\subsection{\bf Corrections ${\tilde\Delta}_{T_{20}},{\tilde\Delta}_{\kappa_0},\Delta $}
\vspace{1.0cm}
Let us present the general forms for the corrections ${\tilde\Delta}_{\kappa_0}$,
${\tilde\Delta}_{T_{20}}$ and $\Delta$ entering into Eq.(\ref{def:T20}) for $T_{20}$
and Eq.(\ref{def:kappa}) for $\kappa_0$.
\begin{equation}
{\tilde\Delta}_{T_{20}}~=~Tr\left(\rho_d[({\cal F}^{(3)})^+
\Omega_{20}{\cal F}^{(1)}+({\cal F}^{(1)})^+\Omega_{20}{\cal F}^{(3)}+
({\cal F}^{(3)})^+\Omega_{20}{\cal F}^{(3)}]
\right)
\label{def:DeltaT20}
\end{equation}
\vspace{0.3cm}
\begin{equation}
{\tilde\Delta}_{\kappa_0}~=~Tr\left(\rho_d[({\cal F}^{(3)})^+
{\vec\sigma}\cdot{\vec n}{\cal F}^{(1)}+
({\cal F}^{(1)})^+{\vec\sigma}\cdot{\vec n}{\cal F}^{(3)}+
({\cal F}^{(3)})^+{\vec\sigma}\cdot{\vec n}{\cal F}^{(3)}]
\right)
\label{def:Deltakappa}
\end{equation}
\vspace{0.3cm}
\begin{equation}
\Delta~=~Tr\left(\rho_d[2Re(({\cal F}^{(3)})^+
{\cal F}^{(1)})+({\cal F}^{(3)})^+{\cal F}^{(3)}]\right)
\label{def:Delta}
\end{equation}
\end{widetext}
\subsection{\bf Vertices $\Gamma_N^{(1)},\Gamma_N^{(2)}$}
The vector $ {\vec\tau}_1$ entering into the $\pi$-absorption vertex
\begin{equation}
\Gamma_N^{(1)}={\bar u}(p^f)\gamma_5 u(k_1)=\xi^+({\vec\sigma}_N\cdot{\bf \tau}_1)\xi
\label{def:GamNf}
\end{equation}
has the following form:
\begin{equation}
{\vec\tau}_1=a_1\frac{{\vec p}^f-{\vec k}_1}{2m}-b_1\frac{E_N({\vec p}^f)-E_N({\vec k}_1)}
{2m}\frac{{\vec p}^f+{\vec k}_1}{2m}~,
\label{def:tauf}
\end{equation}
where
\begin{equation}
a_1=\frac{(E_N({\vec p}^f)+E_N({\vec k}_1))/2+m}
{\sqrt{(E_N({\vec p}^f)+m)(E_N({\vec k}_1)+m)}}~,
\label{def:afirst}
\end{equation}
\begin{equation}
b_1=\frac{1}
{\sqrt{(E_N({\vec p}^f)+m)(E_N({\vec k}_1)+m)}}~.
\label{def:bfirst}
\end{equation}
The vector ${\vec\tau}_2$ entering into the $\pi$-emission vertex
\begin{equation}
\Gamma_N^{(2)}={\bar u}(k_1^f)\gamma_5 u(p)=\xi^+{\vec\sigma}_N\cdot{\vec \tau}_2\xi
\label{def:GamNs}
\end{equation}
has the form
\begin{equation}
{\vec\tau}_2=a_2\frac{{\vec k}^f_1-{\vec p}}{2m}-b_2\frac{E_N({\vec k}^f_1)-E_N({\vec p})}
{2m}\frac{{\vec k}^f_1+{\vec p}}{2m}~,
\label{def:taus}
\end{equation}
where
\begin{equation}
a_2=\frac{(E_N({\vec p})+E_N({\vec k}^f_1))/2+m}
{\sqrt{(E_N({\vec p})+m)(E_N({\vec k}^f_1)+m)}}
\label{def:asecond}
\end{equation}
and
\begin{equation}
b_2=\frac{1}
{\sqrt{(E_N({\vec p})+m)(E_N({\vec k}^f_1)+m)}}~.
\label{def:bsecond}
\end{equation}
\subsection{\bf Kinematics for elastic $\pi-N$ and backward $p-d$ scattering }
The square of the initial energy in the $\pi-N$ c.m.s. reads
\begin{equation}
s_1=(q_2+k_2)^2~,
\end{equation}
where $q_2$ and $k_2$ are the four-momenta of the colliding intermediate pion and a nucleon
in the initial deuteron. Introducing the variable ${\vec\Delta}={\vec p}_d^f/2-{\vec p}$
and using Eqs.(\ref{def:vectorsk}) one can get the following form for $s_1$:
\begin{equation}
s_1\simeq\left(\mid{{\vec k}^\prime-{\vec\Delta}}\mid+E_N(p_d/2)\right)^2-
({\vec k}^\prime-{\vec\Delta})\cdot ({\vec k}+{\vec p}_d/2)~,
\label{def:sfirst}
\end{equation}
The transfer in the $\pi-N$ elastic scattering is
\begin{equation}
t_1=(q_2-q_1)^2~,
\end{equation}
where $q_1$ is the four-momentum of the rescattered pion. Introducing the variable
${\vec\Delta}^\prime={\vec p}_d/2-{\vec p}^f$ and taking into account that for the
backward $p-d$ scattering ${\vec\Delta}^\prime=-{\vec\Delta}$ we have the following form for
$t_1$:
\begin{equation}
t_1\simeq\left(\mid{{\vec k}^\prime-{\vec\Delta}}\mid-
\mid{{\vec k}+{\vec\Delta}}\mid\right)^2-({\vec k}^\prime-{\vec k}-2{\vec\Delta})^2~.
\label{def:tfirst}
\end{equation}
Note that getting Eqs.(\ref{def:sfirst},\ref{def:tfirst}) we neglected the pion mass squared
$\mu_\pi^2$.
\subsection{\bf Deuteron wave functions}
\vspace{0.2cm}
We presented the d.w.f of the type of Reid soft core \cite{Reid:1968}, AV18 \cite{AV18},
N3LO \cite{N3LO} and CD Bonn \cite{CDBonn} in the following forms of the Gauss functions
and found all the parameters from their fits.
\begin{equation}
u(p)~=~\sum_{n=1}^{n_{max}}A_n\exp(-\alpha_n p^2)
\label{def:up}
\end{equation}
\vspace{-0.5cm}
\begin{equation}
w(p)~=~p^2\sum_{n=1}^{n_{max}}B_n\exp(-\beta_n p^2)
\label{def:wp}
\end{equation}
\vspace{0.2cm}
\begin{table}[h]
\caption{\bf The Reid soft core d.w.f.~\cite{Reid:1968} ($n_{max}=5$):}
\vspace{0.2cm}
\label{Tb1}
\begin{center}
\begin{tabular}[t]{|c|c|c|c|c|}
\hline
$n$ & $A_n$ & $\alpha_n$ & $B_n$ & $\beta_n$ \\
\hline
1 & 9.007 & 1277.26 & 1.358 & 5.165 \\
\hline
2 & 20.035 & 370.595 & 11.289 & 15.774 \\
\hline
3 & 9.724 & 88.625 & 15.376 & 50.065 \\
\hline
4 & 2.142 & 18.904 & 43.963 & 52.592 \\
\hline
5 & -0.184 & 2.494 & 227.617 & 205.697 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{0.5cm}
\begin{table}[h]
\caption{\bf AV18 d.w.f.~\cite{AV18} ($n_{max}=7$):}
\vspace{0.1cm}
\label{Tb2}
\begin{center}
\begin{tabular}[t]{|c|c|c|c|c|}
\hline
$n$ & $A_n$ & $\alpha_n$ & $B_n$ & $\beta_n$ \\
\hline
1 & 5.33818388 & 1277.26 & 1.26183415 & 5.165 \\
\hline
2 & 17.7506951 & 370.595 & 10.7790333 & 15.774 \\
\hline
3 & 10.1156672 & 88.625 & -30.4158329 & 50.065 \\
\hline
4 & 2.00231994 & 18.904 & 91.7607541 & 52.592 \\
\hline
5 & -0.129987968 & 2.494 & 193.350066 & 205.697 \\
\hline
6 & 1.85353863 & 15000.0 & 51.1855721 & 600.0 \\
\hline
7 & 4.94736493 & 650.0 & 221.427665 & 1000.0 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{0.5cm}
\begin{table}[h]
\caption{\bf CD Bonn d.w.f.~\cite{CDBonn} ($n_{max}=7$):}
\vspace{0.5cm}
\label{Tb3}
\begin{center}
\begin{tabular}[t]{|c|c|c|c|c|}
\hline
$n$ & $A_n$ & $\alpha_n$ & $B_n$ & $\beta_n$ \\
\hline
1 & 6.25524152 & 1277.26 & 0.836559861 & 5.165 \\
\hline
2 & 18.0668014 & 370.595 & 11.225581 & 15.774 \\
\hline
3 & 10.0949093 & 88.625 & -54.150829 & 50.065 \\
\hline
4 & 1.96622794 & 18.904 & 117.382885 & 52.592 \\
\hline
5 & -0.0681124745 & 2.494 & 192.905072 & 205.697 \\
\hline
6 & 1.02009756 & 15000.0 & 83.961456 & 600.0 \\
\hline
7 & 4.06853301 & 650.0 & 161.799081 & 1000.0 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{0.5cm}
\begin{table}[h]
\caption{\bf N3LO d.w.f.~\cite{N3LO} ($n_{max}=7$):}
\vspace{-0.1cm}
\label{Tb4}
\begin{center}
\begin{tabular}[t]{|c|c|c|c|c|}
\hline
$n$ & $A_n$ & $\alpha_n$ & $B_n$ & $\beta_n$ \\
\hline
1 & 5.75328843 & 1277.26 & -0.411245062 & 5.165 \\
\hline
2 & 17.6314 & 370.595 & 17.5832912 & 15.774 \\
\hline
3 & 10.125005 & 88.625& -251.958128 & 50.065 \\
\hline
4 & 2.1933269 & 18.904 & 318.379763 & 52.592 \\
\hline
5 & -0.228481595 & 2.494 & 166.840281 & 205.697 \\
\hline
6 & 0.787045975 & 15000.0 & 194.835805 & 600.0 \\
\hline
7 & 4.82444602 & 650.0 & 3.08550887 & 1000.0 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{10.5cm}
\newpage
|
\section{Introduction}
\label{sec:Intoro}
\subsection{}
In \cite{LR} Lunts and Rosenberg constructed the quantized flag manifold for a quantized enveloping algebra as a non-commutative projective scheme.
They also defined a category of $D$-modules on it, and conjectured a Beilinson-Bernstein type equivalence of categories.
In \cite{T2} we proposed a modification of the definition of the ring of differential operators on the quantized flag manifold, and established a Beilinson-Bernstein type equivalence for the modified ring of differential operators
(see also Backelin-Kremnizer \cite{BK1}).
The above mentioned results are for a quantized enveloping algebra when the parameter $q$ is transcendental.
The aim of this paper is to investigate the ring of differential operators on the quantized flag manifold when the parameter is a root of unity.
It is a general phenomenon that quantized objects at roots of unity resembles ordinary objects in positive characteristics.
Hence it is natural to pursue analogue of the theory of $D$-modules on flag manifolds in positive characteristics due to Bezrukavnikov-Mirkovi\'{c}-Rumynin \cite{BMR}.
In \cite{BMR} an analogue of the Beilinson-Bernstein equivalence was established on the level of derived categories.
Moreover, it was also shown there that the ring of differential operators satisfies certain Azumaya properties.
In this paper we will be concerned with the Azumaya properties in the quantized situation.
\subsection{}
Let $G$ be a connected simply-connected simple algebraic group over ${\mathbb{C}}$, and let ${\mathfrak{g}}$ be its Lie algebra.
We fix Borel subgroups $B^+$ and $B^-$ of $G$ such that $H=B^+\cap B^-$ is a maximal torus of $G$.
We denote by $N^\pm$ the unipotent radical of $B^\pm$.
We denote by $Q$ and $\Lambda$ the root lattice and the weight lattice respectively.
We also denote by $\Lambda^+$ the set of dominant weights.
Set ${\mathbb{F}}={\mathbb{Q}}(q^{1/|\Lambda/Q|})$, where $q^{1/|\Lambda/Q|}$ is an indeterminate.
We denote by $U_{\mathbb{F}}$ the quantized enveloping algebra of ${\mathfrak{g}}$ over ${\mathbb{F}}$.
It is a Hopf algebra over ${\mathbb{F}}$, and is generated as an ${\mathbb{F}}$-algebra by the elements
$k_\lambda, e_i, f_i\,\,(\lambda\in\Lambda, i\in I)$, where $I$ is the index set for simple roots for ${\mathfrak{g}}$.
We can define a $q$-analogue $C_{\mathbb{F}}$ of the coordinate algebra of $G$ as a Hopf algebra dual of $U_{\mathbb{F}}$.
More precisely, we define $C_{\mathbb{F}}$ to be the subspace of $\mathop{\rm Hom}\nolimits_{\mathbb{F}}(U_{\mathbb{F}},{\mathbb{F}})$ spanned by the matrix coefficients of type 1 representations of $U_{\mathbb{F}}$.
Then we have a $U_{\mathbb{F}}$-bimodule structure of $C_{\mathbb{F}}$ given by
\[
\langle u_1\cdot \varphi\cdot u_2,u\rangle
=\langle \varphi,u_2uu_1\rangle
\qquad
(u, u_1, u_2\in U_{\mathbb{F}}, \varphi\in C_{\mathbb{F}}).
\]
Set
\[
A_{\mathbb{F}}=\bigoplus_{\lambda\in\Lambda^+}A_{\mathbb{F}}(\lambda)\subset C_{\mathbb{F}}
\]
with
\[
A_{\mathbb{F}}(\lambda)
=\{\varphi\in C_{\mathbb{F}}\mid
\varphi\cdot v=\chi_\lambda(v)\varphi\quad(v\in U_{\mathbb{F}}^{\leqq0})\},
\]
where $U^{\leqq0}_{\mathbb{F}}$ is the subalgebra of $U_{\mathbb{F}}$ generated by $k_\lambda, f_i\,\,(\lambda\in\Lambda, i\in I)$ and $\chi_\lambda$ is the character of $U^{\leqq0}_{\mathbb{F}}$ corresponding to $\lambda$.
Note that $A_{\mathbb{F}}$ is a non-commutative $\Lambda$-graded ${\mathbb{F}}$-algebra.
The quantized flag manifold ${\mathcal B}_q$ is defined as a non-commutative projective scheme by
\[
{\mathcal B}_q=\mathop{\rm Proj}\nolimits_\Lambda(A_{\mathbb{F}}).
\]
This actually means that we are given an abelian category $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})$ of ``quasi-coherent sheaves on ${\mathcal B}_q$'' defined by
\[
\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})=\mathop{\rm Mod}\nolimits_\Lambda(A_{\mathbb{F}})/{\rm{Tor}}_{\Lambda^+}(A_{\mathbb{F}}),
\]
where $\mathop{\rm Mod}\nolimits_\Lambda(A_{\mathbb{F}})$ is the category of $\Lambda$-graded left $A_{\mathbb{F}}$-modules, and ${\rm{Tor}}_{\Lambda^+}(A_{\mathbb{F}})$ denotes its full subcategory consisting of $M\in\mathop{\rm Mod}\nolimits_\Lambda(A_{\mathbb{F}})$ such that for each $m\in M$ there exists some $\lambda\in\Lambda^+$ such that $A_{\mathbb{F}}(\lambda+\mu)m=0$ for any $\mu\in\Lambda^+$.
The natural functor
$\omega^*:\mathop{\rm Mod}\nolimits_\Lambda(A_{\mathbb{F}})\to\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})$ admits a right adjoint
$\omega_*:
\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})\to\mathop{\rm Mod}\nolimits_\Lambda(A_{\mathbb{F}})$.
Taking the degree zero part in $\omega_*$ we obtain a left exact functor $\Gamma:\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})\to\mathop{\rm Mod}\nolimits({\mathbb{F}})$, called the global section functor.
Here $\mathop{\rm Mod}\nolimits({\mathbb{F}})$ denotes the category of ${\mathbb{F}}$-modules.
Define an ${\mathbb{F}}$-subalgebra $D_{\mathbb{F}}$ of $\mathop{\rm{End}}\nolimits_{\mathbb{F}}(A_{\mathbb{F}})$ by
\[
D_{\mathbb{F}}=
\langle \ell_\varphi, r_\varphi,\partial_u, \sigma_\lambda\mid\varphi\in A_{\mathbb{F}}, u\in U_{\mathbb{F}}, \lambda\in\Lambda\rangle,
\]
where $\ell_\varphi$ and $r_\varphi$ are the left and the right multiplications of $\varphi$ respectively, $\partial_u$ denotes the natural left action of $u\in U_{\mathbb{F}}$ on $A_{\mathbb{F}}$ induced by that on $C_{\mathbb{F}}$, and $\sigma_\lambda$ is the grading operator given by
$\sigma_\lambda(\varphi)=q^{(\lambda,\mu)}\varphi$ for $\varphi\in A_{\mathbb{F}}(\mu)$.
Then $D_{\mathbb{F}}$ is a $\Lambda$-graded ring by
$D_{\mathbb{F}}(\lambda)=\{\Phi\in D_{\mathbb{F}}\mid\Phi(A_{\mathbb{F}}(\mu))\subset A_{\mathbb{F}}(\mu+\lambda)\,\,(\mu\in\Lambda)\}$ for $\lambda\in\Lambda$.
By the aid of the universal $R$-matrix we can show that $r_\varphi$ can be expressed using other type of generators.
Hence we have
\[
D_{\mathbb{F}}=
\langle \ell_\varphi, \partial_u, \sigma_\lambda\mid\varphi\in A_{\mathbb{F}}, u\in U_{\mathbb{F}}, \lambda\in\Lambda\rangle.
\]
A $D_{\mathbb{F}}$-module is regarded as an $A_{\mathbb{F}}$-module by the algebra homomorphism $A_{\mathbb{F}}\ni\varphi\mapsto\ell_\varphi\in D_{\mathbb{F}}$ in the following.
We define an abelian category $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q})$ of ``quasi-coherent ${\mathcal D}_{{\mathcal B}_q}$-modules'' by
\[
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q})
=
\mathop{\rm Mod}\nolimits_\Lambda(D_{\mathbb{F}})/\mathop{\rm Mod}\nolimits_\Lambda(D_{\mathbb{F}})\cap{\rm{Tor}}_{\Lambda^+}(A_{\mathbb{F}}).
\]
Denote by $H({\mathbb{F}})$ the set of ${\mathbb{F}}$-rational points of the maximal torus $H$ of $G$.
For $t\in H({\mathbb{F}})$ we define an abelian category $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q,t})$ of ``quasi-coherent ${\mathcal D}_{{\mathcal B}_q,t}$-modules'' by
\[
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q,t})
=
\mathop{\rm Mod}\nolimits_{\Lambda,t}(D_{\mathbb{F}})/\mathop{\rm Mod}\nolimits_{\Lambda,t}(D_{\mathbb{F}})\cap{\rm{Tor}}_{\Lambda^+}(A_{\mathbb{F}}),
\]
where $\mathop{\rm Mod}\nolimits_{\Lambda,t}(D_{\mathbb{F}})$ is the full subcategory of $\mathop{\rm Mod}\nolimits_{\Lambda}(D_{\mathbb{F}})$ consisting of $M\in\mathop{\rm Mod}\nolimits_{\Lambda}(D_{\mathbb{F}})$ such that $\sigma_\mu|_{M(\lambda)}=
\mu(t)q^{(\lambda,\mu)}\mathop{\rm id}\nolimits$ for any $\lambda\in\Lambda$.
Then ${\mathcal D}_{{\mathcal B}_q,1}$ is ``the sheaf of differential operators on ${\mathcal B}_q$'', and other ${\mathcal D}_{{\mathcal B}_q,t}$'s are its twisted analogues (although they have only symbolical meanings).
Let us consider the specialization of the parameter $q$ to roots of unity.
We take an odd integer $\ell>1$ which is prime to $|\Lambda/Q|$, and prime to 3 if ${\mathfrak{g}}$ is of type $G_2$.
We fix a primitive $\ell$-th root of unity $\zeta'\in{\mathbb{C}}$ and consider the specialization
\begin{equation}
\label{eq:specialization}
q^{1/|\Lambda/Q|}\mapsto \zeta',\qquad
q\mapsto \zeta=(\zeta')^{|\Lambda/Q|}.
\end{equation}
Note that $\zeta$ is also a primitive $\ell$-th root of unity by our assumption.
Set
\[
{\mathbb{A}}=
\{
f(q^{1/|\Lambda/Q|})\in{\mathbb{F}}\mid
f(x) \mbox{ is regular at } x=\zeta'
\},
\]
and
\begin{align*}
U_{\mathbb{A}}^L&=\langle
k_\lambda, e_i^{(n)}, f_i^{(n)}\mid
\lambda\in\Lambda, i\in I, n\in{\mathbb{Z}}_{\geqq0}
\rangle_{{\mathbb{A}}-alg}\subset U_{\mathbb{F}},
\\
U_{\mathbb{A}}&=\langle
k_\lambda, e_i, f_i\mid
\lambda\in\Lambda, i\in I
\rangle_{{\mathbb{A}}-alg}\subset U_{\mathbb{F}},
\end{align*}
where $e_i^{(n)}, f_i^{(n)}$ denote standard divided powers.
$U_{\mathbb{A}}^L$ and $U_{\mathbb{A}}$ are Hopf algebras over ${\mathbb{A}}$ called the Lusztig form and the De Concini-Kac form of $U_{\mathbb{F}}$ respectively.
Taking the Hopf algebra dual of $U^L_{\mathbb{A}}$ we obtain an ${\mathbb{A}}$-form $C_{\mathbb{A}}$ of $C_{\mathbb{F}}$.
We set
\begin{align*}
A_{\mathbb{A}}&=A_{\mathbb{F}}\cap C_{\mathbb{A}}=
\bigoplus_{\lambda\in\Lambda^+}
A_{\mathbb{A}}(\lambda),\\
D_{\mathbb{A}}&=
\langle \ell_\varphi, r_\varphi,\partial_u, \sigma_\lambda\mid\varphi\in A_{\mathbb{A}}, u\in U_{\mathbb{A}}, \lambda\in\Lambda\rangle\\
&=
\langle \ell_\varphi, \partial_u, \sigma_\lambda\mid\varphi\in A_{\mathbb{A}}, u\in U_{\mathbb{A}}, \lambda\in\Lambda\rangle\subset\mathop{\rm{End}}\nolimits_{\mathbb{A}}(A_{\mathbb{A}})
\subset\mathop{\rm{End}}\nolimits_{\mathbb{F}}(A_{\mathbb{F}}).
\end{align*}
Now we consider the specializations
\begin{align*}
A_\zeta&={\mathbb{C}}\otimes_{\mathbb{A}} A_{\mathbb{A}},
\qquad
U_\zeta^L={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}}^L,\qquad
U_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}},\\
D_\zeta&={\mathbb{C}}\otimes_{\mathbb{A}} D_{\mathbb{A}},
\end{align*}
where ${\mathbb{A}}\to{\mathbb{C}}$ is given by \eqref{eq:specialization}.
Note that the natural homomorphism $D_\zeta\to\mathop{\rm{End}}\nolimits_{\mathbb{C}}(A_\zeta)$ is not injective.
Similarly to ${\mathcal B}_q$ we obtain a non-commutative projective scheme ${\mathcal B}_\zeta=\mathop{\rm Proj}\nolimits_\Lambda(A_\zeta)$, which actually means we are given an abelian category $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})$ defined similarly to $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_q})$.
We also have abelian categories
$\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})$, $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})$ for $t\in H({\mathbb{C}})$ defined similarly to $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q})$ and $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_q,t})$ respectively.
Let $U_\zeta^L\to U({\mathfrak{g}})$ be Lusztig's Frobenius morphism, where $U({\mathfrak{g}})$ is the enveloping algebra of ${\mathfrak{g}}$.
By taking the dual Hopf algebras we obtain a central embedding
${\mathbb{C}}[G]\to C_\zeta$ of the coordinate algebra ${\mathbb{C}}[G]$ of $G$ into $C_\zeta$.
Let $A_1$ be the subalgebra of ${\mathbb{C}}[G]$ defined similarly to $A_{\mathbb{F}}$.
Then $A_1$ is a commutative $\Lambda$-graded ${\mathbb{C}}$-algebra such that $\mathop{\rm Proj}\nolimits_\Lambda(A_1)$ is naturally isomorphic to the flag manifold ${\mathcal B}=B^-\backslash G$ of $G$.
Under the identification ${\mathbb{C}}[G]\subset C_\zeta$ we have $A_1\subset A_\zeta$ and $A_1(\lambda)\subset A_\zeta(\ell\lambda)$ for $\lambda\in\Lambda^+$.
We denote by $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$ the ${\mathcal O}_{\mathcal B}$-module corresponding to the $\Lambda$-graded $A_1$-module
$A_\zeta^{(\ell)}=\bigoplus_{\lambda\in\Lambda^+}A_\zeta(\ell\lambda)$.
The $A_1$-algebra structure of $A_\zeta^{(\ell)}$ endows with $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$ a canonical ${\mathcal O}_{\mathcal B}$-algebra structure.
Then we have an equivalence
\[
\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})\cong\mathop{\rm Mod}\nolimits(Fr_*{\mathcal O}_{{\mathcal B}_\zeta})
\]
of abelian categories, where $\mathop{\rm Mod}\nolimits(Fr_*{\mathcal O}_{{\mathcal B}_\zeta})$ denotes the category of quasi-coherent $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$-modules.
Similarly, we have
\begin{align*}
&\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})\cong\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta}),\\&\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta, t})\cong\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t}) \quad(t\in H({\mathbb{C}})),
\end{align*}
where $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$ and $Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t}$ are ${\mathcal O}_{\mathcal B}$-algebras corresponding to $A_1$-algebras
$D_\zeta^{(\ell)}$ and $D_\zeta^{(\ell)}\otimes_{{\mathbb{C}}[\Lambda]}{\mathbb{C}}$ respectively.
Here $D_\zeta^{(\ell)}=\bigoplus_{\lambda\in\Lambda^+}D_\zeta(\ell\lambda)$, and ${\mathbb{C}}[\Lambda]=\bigoplus_{\lambda\in\Lambda}{\mathbb{C}} e(\lambda)$ denotes the group algebra of $\Lambda$.
Moreover, ${\mathbb{C}}[\Lambda]\to D_\zeta^{(\ell)}$ and ${\mathbb{C}}[\Lambda]\to{\mathbb{C}}$ are given by
$e(\lambda)\mapsto\sigma_\lambda$ and $e(\lambda)\mapsto \lambda(t)$ respectively.
Denote by $ZD_\zeta^{(\ell)}$ the central subalgebra of $D_\zeta^{(\ell)}$ generated by elements $\ell_\varphi, \partial_u, \sigma_\lambda\,\,(\varphi\in A_1, u\in Z_{Fr}(U_\zeta), \lambda\in\Lambda)$, where $Z_{Fr}(U_\zeta)$ denotes the Frobenius center of $U_\zeta$.
Let ${\mathcal Z}_\zeta$ be the central ${\mathcal O}_{\mathcal B}$-subalgebra of $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$ corresponding to $ZD_\zeta^{(\ell)}$.
By \cite{DP} $Z_{Fr}(U_\zeta)$ is a Hopf subalgebra of $U_\zeta$ isomorphic to the coordinate algebra ${\mathbb{C}}[K]$ of the algebraic group
\[
K=
\{(hg_+,h^{-1}g_-)\mid
h\in H, g_\pm\in N^\pm\}\subset B^+\times B^-.
\]
We note also that the group algebra ${\mathbb{C}}[\Lambda]$ is naturally isomorphic to the coordinate algebra ${\mathbb{C}}[H]$ of $H$.
Hence we have a natural surjective algebra homomorphism $A_1\otimes {\mathbb{C}}[K]\otimes{\mathbb{C}}[H]\to ZD_\zeta^{(\ell)}$.
Correspondingly, we have a natural surjective ${\mathcal O}_{\mathcal B}$-algebra homomorphism
$p_{*}{\mathcal O}_{{\mathcal B}\times K\times H}\to{\mathcal Z}_\zeta$, where $p:{\mathcal B}\times K\times H\to{\mathcal B}$ is the projection.
Define $\kappa:K\to G$ by
$\kappa(k_1,k_2)=k_1k_2^{-1}$.
\begin{theorem}
\label{theorem01}
We have
${\mathcal Z}_\zeta\cong p_{*}{\mathcal O}_{\mathcal V}$, where
\[
{\mathcal V}=\{(B^-g,k,t)\in{\mathcal B}\times K\times H\mid
g\kappa(k)g^{-1}\in t^{2\ell}N^-\}.
\]
\end{theorem}
The proof of this theorem is accomplished as follows.
In order to show that the kernel of $p_{*}{\mathcal O}_{{\mathcal B}\times K\times H}\to{\mathcal Z}_\zeta$
contains defining equations of ${\mathcal V}$ one needs to establish certain relations among elements of $ZD_\zeta^{(\ell)}\left(\subset D_\zeta^{(\ell)}\right)$.
As mentioned earlier, $r_\varphi$ for $\varphi\in A_\zeta$ can be expressed using other generators by the aid of the universal $R$-matrix ${\mathcal R}$.
In fact we have two universal $R$-matrices ${\mathcal R}$ and ${}^t{\mathcal R}^{-1}$ by which we obtain two different expressions of the same element $r_\varphi$.
This gives our desired relations.
Hence ${\mathcal Z}_\zeta$ is a quotient of $p_{*}{\mathcal O}_{\mathcal V}$.
To show that ${\mathcal Z}_\zeta$ is isomorphic to $p_{*}{\mathcal O}_{\mathcal V}$ we use Poisson geometry.
We have a natural Poisson structure of $Y=(N^-\backslash G)\times K\times H$, and the support of the pullback of ${\mathcal Z}_\zeta$ to $Y$ is a Poisson subvariety of $Y$ by \cite{DP}.
On the other hand we can show that the pullback of ${\mathcal V}$ to $Y$ is a connected symplectic leaf of the Poisson manifold $Y$.
Hence the assertion follows from the fact that ${\mathcal Z}_\zeta\ne0$ which is easy to check.
We denote by $\tilde{{\mathcal D}}$ the localization of $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$ on ${\mathcal V}$.
\begin{theorem}
\label{theorem02}
$\tilde{{\mathcal D}}$ is locally free over ${\mathcal O}_{\mathcal V}$ of finite rank.
Moreover, for any $v\in {\mathcal V}$ the fiber $\tilde{{\mathcal D}}(v)$ of $\tilde{{\mathcal D}}$ at $v$ is isomorphic to the matrix algebra $M_{\ell^{N}}({\mathbb{C}})$, where $N$ is the number of the positive roots.
In particular, $\tilde{{\mathcal D}}$ is an Azumaya algebra.
\end{theorem}
This result follows from a result of
Brown-Gordon \cite{BG} and the fact that $\tilde{{\mathcal D}}$ is locally generated by $\ell^{2N}$ sections.
Using the action of the braid group on $\tilde{{\mathcal D}}$ the proof of the latter fact is reduced to a calculation on a standard open subset of ${\mathcal B}$.
Let $W$ denote the Weyl group.
Consider the fiber product $K\times _{H/W}H$,
where $K\to H/W$ is the composite of $\kappa:K\to G$ and the map $G\to H/W$ associating $g\in G$ with its semisimple part, and $H\to H/W$ is given by associating $t\in H$ with the $W$-orbit of $t^{2\ell}$.
We define $\delta:{\mathcal V}\to K\times _{H/W}H$ by
$\delta(B^-g,k,t)=(k,t)$.
\begin{theorem}
\label{theorem03}
For any $(k,t)\in K\times _{H/W}H$ there exists a locally free ${\mathcal O}_{\delta^{-1}(k,t)}$-module ${\mathcal M}$ such that $\tilde{{\mathcal D}}|_{\delta^{-1}(k,t)}\cong\mathop{{\mathcal{E}}nd}\nolimits_{{\mathcal O}_{\delta^{-1}(k,t)}}({\mathcal M})$.
Hence $\tilde{{\mathcal D}}|_{\delta^{-1}(k,t)}$ is a split Azumaya algebra.
\end{theorem}
The proof of Theorem \ref{theorem03} is similar to that for the corresponding fact in positive characteristics due to Bezrukavnikov-Mirkovi\'{c}-Rumynin \cite{BMR}.
By Brown-Gordon \cite{BG} the result is already known when $t\in H$ belongs to certain open dense subset $H_{ur}$ of $H$.
The proof for the general case is reduced to this special case by using certain isomorphisms of Azumaya algebras.
The content of this paper is as follows.
In Section 1 and Section 2 we recall basic facts on a quantized enveloping algebra and its representations respectively.
In Section 3 the quantized flag manifold is introduced and some of its properties are investigated.
In Section 4 we define the ring of differential operators on the quantized flag manifold and establish some properties.
In particular, we show that it acquires an action of the braid group.
Theorem \ref{theorem01} is proved in Section 5.
Theorem \ref{theorem02} and Theorem \ref{theorem03} are proved in Section 6.
We note that a closely related result is given in Backelin-Kremnizer \cite{BK2}.
\subsection{}
In this paper we shall use the following notation for a Hopf algebra $H$ over a field ${\mathbb{K}}$.
The comultiplication, the counit, and the antipode of $H$ are denoted by
\begin{align}
&\Delta_H:H\to H\otimes_{\mathbb{K}} H,\\
&\varepsilon_H:H\to{\mathbb{K}},\\
&S_H:H\to H
\end{align}
respectively.
The subscript $H$ will often be omitted.
For $n\in{\mathbb{Z}}_{>0}$ we denote by
\[
\Delta_n:H\to H^{\otimes n+1}
\]
the algebra homomorphism given by
\[
\Delta_1=\Delta,\qquad
\Delta_n=(\Delta\otimes \mathop{\rm id}\nolimits_{H^{\otimes n-1}})\circ\Delta_{n-1},
\]
and write
\[
\Delta(h)=\sum_{(h)}h_{(0)}\otimes h_{(1)},
\qquad
\Delta_n(h)=\sum_{(h)_n}h_{(0)}\otimes\cdots\otimes h_{(n)}
\quad(n\geqq2).
\]
Moreover, for a ${\mathbb{K}}$-algebra $A$ we denote by
\[
m:A\otimes A\to A
\]
the ${\mathbb{K}}$-linear map induced by the multiplication of $A$.
\section{Quantized enveloping algebras}
\label{sec:QE}
\subsection{}
Let $G$ be a connected simply-connected simple algebraic group over the complex number field ${\mathbb{C}}$.
We fix Borel subgroups $B^+$ and $B^-$ such that $H=B^+\cap B^-$ is a maximal torus of $G$.
Set $N^+=[B^+,B^+]$ and $N^-=[B^-,B^-]$.
We denote the Lie algebras of $G$, $B^+$, $B^-$, $H$, $N^+$, $N^-$ by ${\mathfrak{g}}$, ${\mathfrak{b}}^+$, ${\mathfrak{b}}^-$, ${\mathfrak{h}}$, ${\mathfrak{n}}^+$, ${\mathfrak{n}}^-$ respectively.
Let $\Delta\subset{\mathfrak{h}}^*$ be the root system of $({\mathfrak{g}},{\mathfrak{h}})$.
For $\alpha\in\Delta$ we denote by ${\mathfrak{g}}_\alpha$ the corresponding root space.
We denote by $\Lambda\subset{\mathfrak{h}}^*$ and $Q\subset{\mathfrak{h}}^*$ the weight lattice and the root lattice respectively.
For $\lambda\in\Lambda$ we denote the corresponding character of $H$ by
$\theta_\lambda:H\to{\mathbb{C}}^\times$.
We take a system of positive roots $\Delta^+$ such that ${\mathfrak{b}}^+$ is the sum of weight spaces with weights in $\Delta^+\cup\{0\}$.
Let $\{\alpha_i\}_{i\in I}$ be the set of simple roots, and
$\{\varpi_i\}_{i\in I}$ the corresponding set of fundamental weights.
We denote by
$\Lambda^+$ be the set of dominant integral weights.
We set $Q^+=\bigoplus_{i\in I}{\mathbb{Z}}_{\geqq0}\alpha_i$.
Let $W\subset GL({\mathfrak{h}}^*)$ be the Weyl group.
For $i\in I$ we denote by $s_i\in W$ the corresponding simple reflection.
We take a $W$-invariant symmetric bilinear form
\begin{equation}
\label{eq:killing}
(\,,\,):{\mathfrak{h}}^*\times{\mathfrak{h}}^*\to{\mathbb{C}}
\end{equation}
such that $(\alpha,\alpha)=2$ for short roots $\alpha$.
\begin{lemma}
We have
\[
(\Lambda,Q)\subset{\mathbb{Z}},\qquad
(\Lambda,\Lambda)\subset\frac1{|\Lambda/Q|}{\mathbb{Z}}.
\]
\end{lemma}
\begin{proof}
For $\alpha\in\Delta$ and $\lambda\in\Lambda$ we have
\[
(\lambda,\alpha)=\frac{2(\lambda,\alpha)}{(\alpha,\alpha)}
\frac{(\alpha,\alpha)}2\in{\mathbb{Z}}.
\]
The second formula follows from the first one.
\end{proof}
For $\alpha\in\Delta$ we set $\alpha^\vee=2\alpha/(\alpha,\alpha)$.
For $i\in I$ we fix $\overline{e}_i\in{\mathfrak{g}}_{\alpha_i}$, $\overline{f}_i\in{\mathfrak{g}}_{-\alpha_i}$ such that $[\overline{e}_i,\overline{f}_i]=\alpha_i^\vee$ under the identification ${\mathfrak{h}}={\mathfrak{h}}^*$ induced by $(\,\,,\,\,)$.
For $\lambda=\sum_{i\in I}c_i\alpha_i\in{\mathfrak{h}}^*$ we set $\mathop{\rm ht}\nolimits(\lambda)=\sum_{i\in I}c_i$.
We define $\rho\in\Lambda$ by $(\rho,\alpha_i^\vee)=1$ for any $i\in I$.
We set $N=|\Delta^+|$.
\subsection{}
For $n\in{\mathbb{Z}}_{\geqq0}$ we set
\[
[n]_t=\frac{t^n-t^{-n}}{t-t^{-1}}\in{\mathbb{Z}}[t,t^{-1}],
\qquad
[n]_t!=[n]_t[n-1]_t\cdots[2]_t[1]_t\in{\mathbb{Z}}[t,t^{-1}].
\]
We denote by
$U_{\mathbb{F}}$
the quantized enveloping algebra over ${\mathbb{F}}={\mathbb{Q}}(q^{1/{|\Lambda/Q|}})$ associated to ${\mathfrak{g}}$.
Namely, $U_{\mathbb{F}}$ is the associative algebra over ${\mathbb{F}}$ generated by elements
\[
k_\lambda\quad(\lambda\in\Lambda),\qquad
e_i, f_i\quad( i\in I)
\]
satisfying the relations
\begin{align}
&k_0=1,\quad
k_\lambda k_\mu=k_{\lambda+\mu}
\qquad(\lambda,\mu\in\Lambda),
\label{eq:def1}\\
&k_\lambda e_ik_\lambda^{-1}=q^{(\lambda,\alpha_i)}e_i,\qquad(\lambda\in\Lambda, i\in I),
\label{eq:def2a}\\
&k_\lambda f_ik_\lambda^{-1}=q^{-(\lambda,\alpha_i)}f_i
\qquad(\lambda\in\Lambda, i\in I),
\label{eq:def2b}\\
&e_if_j-f_je_i=\delta_{ij}\frac{k_i-k_i^{-1}}{q_i-q_i^{-1}}
\qquad(i, j\in I),
\label{eq:def3}\\
&\sum_{n=0}^{1-a_{ij}}(-1)^ne_i^{(1-a_{ij}-n)}e_je_i^{(n)}=0
\qquad(i,j\in I,\,i\ne j),
\label{eq:def4}\\
&\sum_{n=0}^{1-a_{ij}}(-1)^nf_i^{(1-a_{ij}-n)}f_jf_i^{(n)}=0
\qquad(i,j\in I,\,i\ne j),
\label{eq:def5}
\end{align}
where $q_i=q^{(\alpha_i,\alpha_i)/2}, k_i=k_{\alpha_i}, a_{ij}=2(\alpha_i,\alpha_j)/(\alpha_i,\alpha_i)$ for $i, j\in I$, and
\[
e_i^{(n)}=
e_i^n/[n]_{q_i}!,
\qquad
f_i^{(n)}=
f_i^n/[n]_{q_i}!
\]
for $i\in I$ and $n\in{\mathbb{Z}}_{\geqq0}$.
We will use the Hopf algebra structure of $U_{\mathbb{F}}$ given by
\begin{align}
&\Delta(k_\lambda)=k_\lambda\otimes k_\lambda,\\
&\Delta(e_i)=e_i\otimes 1+k_i\otimes e_i,\quad
\Delta(f_i)=f_i\otimes k_i^{-1}+1\otimes f_i,
\nonumber\\
&\varepsilon(k_\lambda)=1,\quad
\varepsilon(e_i)=\varepsilon(f_i)=0,\\
&S(k_\lambda)=k_\lambda^{-1},\quad
S(e_i)=-k_i^{-1}e_i, \quad S(f_i)=-f_ik_i.
\end{align}
Define subalgebras $U_{\mathbb{F}}^{0}$, $U_{\mathbb{F}}^{+}$, $U_{\mathbb{F}}^{-}$, $U_{\mathbb{F}}^{\geqq0}$, $U_{\mathbb{F}}^{\leqq0}$ of $U_{\mathbb{F}}$ by
\begin{align*}
&U_{\mathbb{F}}^{0}=\langle k_\lambda\mid\lambda\in\Lambda\rangle,\qquad
U_{\mathbb{F}}^+=\langle e_i\mid i\in I\rangle, \qquad
U_{\mathbb{F}}^-=\langle f_i\mid i\in I\rangle, \\
&U_{\mathbb{F}}^{\geqq0}=\langle k_\lambda, e_i\mid\lambda\in\Lambda, i\in I\rangle
,\qquad
U_{\mathbb{F}}^{\leqq0}=\langle k_\lambda, f_i\mid\lambda\in\Lambda, i\in I\rangle.
\end{align*}
Then the multiplication of $U_{\mathbb{F}}$ induces isomorphisms
\begin{align}
\label{eq:tri:F1}
&U_{\mathbb{F}}\cong U_{\mathbb{F}}^{-}\otimes U_{\mathbb{F}}^{0}\otimes U_{\mathbb{F}}^{+},\\
\label{eq:tri:F2}
&
U_{\mathbb{F}}^{\geqq0}\cong U_{\mathbb{F}}^{0}\otimes U_{\mathbb{F}}^{+}
\cong U_{\mathbb{F}}^{+}\otimes U_{\mathbb{F}}^{0},\qquad
U_{\mathbb{F}}^{\leqq0}\cong U_{\mathbb{F}}^{0}\otimes U_{\mathbb{F}}^{-}
\cong U_{\mathbb{F}}^{-}\otimes U_{\mathbb{F}}^{0}
\end{align}
of vector spaces.
Moreover, $\{k_\lambda\mid\lambda\in\Lambda\}$ is an ${\mathbb{F}}$-basis of $U_{\mathbb{F}}^0$.
We have
\begin{align}
\label{eq:Upm-decomp1}
&U_{\mathbb{F}}^{\pm}=\bigoplus_{\gamma\in Q^+}U_{{\mathbb{F}},\pm\gamma}^{\pm},\quad\\
\label{eq:Upm-decomp2}
&U_{{\mathbb{F}},\pm\gamma}^{\pm}=
\{u\in U_{{\mathbb{F}}}^{\pm}\mid
k_\lambda u k_\lambda^{-1}=q^{\pm(\lambda,\gamma)}u\,\,(\lambda\in\Lambda)\}.
\end{align}
We denote by
${\mathbb{B}}$
the braid group corresponding to $W$.
Namely, ${\mathbb{B}}$ is a group generated by elements $T_i\,\,(i\in I)$ satisfying relations
\[
\underbrace{T_iT_j\cdots\cdots}_{\mbox{\footnotesize{${\mathop{\rm ord}\nolimits}(s_is_j)$-times}}}=
\underbrace{T_jT_i\cdots\cdots}_{\mbox{\footnotesize{${\mathop{\rm ord}\nolimits}(s_is_j)$-times}}}
\qquad(i, j\in I, i\ne j),
\]
where ${\mathop{\rm ord}\nolimits}(s_is_j)$ denotes the order of $s_is_j\in W$.
For $w\in W$ we set $T_w=T_{i_1}\cdots T_{i_r}$ where $w=s_{i_1}\cdots s_{i_r}$ is a reduced expression of $w$.
It does not depend on the choice of a reduced expression.
We have a group homomorphism
\[
{\mathbb{B}}\to
{\rm Aut}_{alg}(U_{{\mathbb{F}}})
\]
given by
\begin{align*}
&T_i(k_\mu)=k_{s_i\mu}\qquad(\mu\in\Lambda),\\
&T_i(e_j)=
\begin{cases}
\sum_{k=0}^{-a_{ij}}(-1)^kq_i^{-k}e_i^{(-a_{ij}-k)}e_je_i^{(k)}\qquad
&(j\in I,\,\,j\ne i),\\
-f_ik_i\qquad
&(j=i),
\end{cases}\\
&T_i(f_j)=
\begin{cases}
\sum_{k=0}^{-a_{ij}}(-1)^kq_i^{k}f_i^{(k)}f_jf_i^{(-a_{ij}-k)}\qquad
&(j\in I,\,\,j\ne i),\\
-k_i^{-1}e_i\qquad
&(j=i)
\end{cases}
\end{align*}
(see Lusztig \cite{Lbook}).
Let $w_0$ be the longest element of $W$.
We fix a reduced expression
\[
w_0=s_{i_1}\cdots s_{i_N}
\]
of $w_0$, and set
\[
\beta_k=s_{i_1}\cdots s_{i_{k-1}}(\alpha_{i_k})\qquad
(1\leqq k\leqq N).
\]
Then we have $\Delta^+=\{\beta_k\mid1\leqq k\leqq N\}$.
For $1\leqq k\leqq N$ set
\begin{equation}
\label{eq:root vector}
e_{\beta_k}=T_{i_1}\cdots T_{i_{k-1}}(e_{i_k}),\quad
f_{\beta_k}=T_{i_1}\cdots T_{i_{k-1}}(f_{i_k}).
\end{equation}
Then $\{e_{\beta_{N}}^{m_N}\cdots e_{\beta_{1}}^{m_1}\mid
m_1,\dots, m_N\geqq0\}$ (resp.
\newline
$\{f_{\beta_{N}}^{m_N}\cdots f_{\beta_{1}}^{m_1}\mid
m_1,\dots, m_N\geqq0\}$)
is an ${\mathbb{F}}$-basis of $U_{\mathbb{F}}^+$ (resp. $U_{\mathbb{F}}^-$), called the PBW-basis (see Lusztig \cite{L2}).
We have $e_{\alpha_i}=e_i$ and $f_{\alpha_i}=f_i$ for any $i\in I$.
For $1\leqq k\leqq N,\, m\geqq0$ we also set
\begin{equation}
e^{(m)}_{\beta_k}=e^{m}_{\beta_k}/[m]_{q_{\beta_k}}!,\quad
f^{(m)}_{\beta_k}=f^{m}_{\beta_k}/[m]_{q_{\beta_k}}!,
\end{equation}
where $q_\beta=q^{(\beta,\beta)/2}$ for $\beta\in\Delta^+$.
Denote by
\begin{equation}
\label{eq:Drinfeld-paring}
\tau: U_{\mathbb{F}}^{\geqq0}\times U_{\mathbb{F}}^{\leqq0}\to{\mathbb{F}}
\end{equation}
the Drinfeld paring.
It is characterized as a bilinear form satisfying
\begin{align}
&\tau(x,y_1y_2)=(\tau\otimes\tau)(\Delta(x),y_1\otimes y_2)
&(x\in U_{\mathbb{F}}^{\geqq0},\,y_1,y_2\in U_{\mathbb{F}}^{\leqq0}),\\
&\tau(x_1x_2,y)=(\tau\otimes\tau)(x_2\otimes x_1,\Delta(y))
&(x_1, x_2\in U_{\mathbb{F}}^{\geqq0},\,y\in U_{\mathbb{F}}^{\leqq0}),\\
&\tau(k_\lambda,k_\mu)=q^{-(\lambda,\mu)}
&(\lambda,\mu\in\Lambda),\\
&\tau(k_\lambda, f_i)=\tau(e_i,k_\lambda)=0
&(\lambda\in\Lambda,\,i\in I),\\
&\tau(e_i,f_j)=\delta_{ij}/(q_i^{-1}-q_i)
&(i,j\in I)
\end{align}
(see \cite{T1}, \cite{Lbook}).
It satisfies the following (see \cite{T1}, \cite{Lbook}).
\begin{lemma}
\label{lem:Drinfeld paring}
\begin{itemize}
\item[\rm(i)]
$\tau(S(x),S(y))=\tau(x,y)$ for $x\in U_{\mathbb{F}}^{\geqq0}, y\in U_{\mathbb{F}}^{\leqq0}$.
\item[\rm(ii)]
For $x\in U_{\mathbb{F}}^{\geqq0}, y\in U_{\mathbb{F}}^{\leqq0}$ we have
\begin{align*}
yx=\sum_{(x)_2,(y)_2}
\tau(x_{(0)},S(y_{(0)}))\tau(x_{(2)},y_{(2)})x_{(1)}y_{(1)},\\
xy=\sum_{(x)_2,(y)_2}
\tau(x_{(0)},y_{(0)})\tau(x_{(2)},S(y_{(2)}))y_{(1)}x_{(1)}.
\end{align*}
\item[\rm(iii)]
$\tau(xk_\lambda, yk_\mu)=q^{-(\lambda,\mu)}\tau(x,y)$ for $\lambda, \mu\in\Lambda, x\in U_{\mathbb{F}}^+, y\in U_{\mathbb{F}}^-$.
\item[\rm(iv)]
$\tau(U^+_{{\mathbb{F}},\beta}, U^-_{{\mathbb{F}},-\gamma})=\{0\}$ for $\beta, \gamma\in Q^+$ with $\beta\ne\gamma$.
\item[\rm(v)]
For any $\beta\in Q^+$ the restriction $\tau|_{U^+_{{\mathbb{F}},\beta}\times U^-_{{\mathbb{F}},-\beta}}$ is non-degenerate.
\end{itemize}
\end{lemma}
Denote by $Z(U_{{\mathbb{F}}})$ the center of $U_{{\mathbb{F}}}$.
Let
\[
{\mathbb{F}}[\Lambda]=\bigoplus_{\lambda\in\Lambda}{\mathbb{F}}
e(\lambda)
\]
be the group algebra of $\Lambda$.
Define a linear map
\[
\iota:Z(U_{\mathbb{F}})\to{\mathbb{F}}[\Lambda]
\]
as the composite of
\[
Z(U_{\mathbb{F}})
\subset
U_{\mathbb{F}}
\simeq
U^-_{\mathbb{F}}\otimes U^0_{\mathbb{F}}\otimes U^+_{\mathbb{F}}
\xrightarrow{\varepsilon\otimes 1\otimes\varepsilon}
U_{\mathbb{F}}^0
\cong
{\mathbb{F}}[\Lambda],
\]
where $U_{\mathbb{F}}^0\cong {\mathbb{F}}[\Lambda]$ is given by $k_\lambda\leftrightarrow e(\lambda)$ for $\lambda\in\Lambda$.
Then $\iota$ is an injective algebra homomorphism and its image is described as follows.
Note that the Weyl group $W$ naturally acts on ${\mathbb{F}}[\Lambda]$ by
\[
we(\lambda)=e(w\lambda)\qquad
(w\in W,\,\lambda\in\Lambda).
\]
We also consider a twisted action of $W$ on ${\mathbb{F}}[\Lambda]$ given by
\[
w\circ e(\lambda)=q^{(w\lambda-\lambda,\rho)}e(w\lambda)\qquad
(w\in W,\,\lambda\in\Lambda).
\]
Then the image of $\iota$ coincides with
\[
{\mathbb{F}}[2\Lambda]^{W\circ}
=\{f\in{\mathbb{F}}[2\Lambda]\mid
w\circ f=f\quad(w\in W)\}
\]
(note that the twisted action of $W$ on ${\mathbb{F}}[\Lambda]$ preserves ${\mathbb{F}}[2\Lambda]$).
In particular, we have an isomorphism
\begin{equation}
\label{eq:HC-center-F}
Z(U_{\mathbb{F}})\simeq{\mathbb{F}}[2\Lambda]^{W\circ}
\end{equation}
of ${\mathbb{F}}$-algebras (see, e.g.\ \cite{T1}).
For $\lambda\in\Lambda^+$ we denote by $m(\lambda)$ the element of $Z(U_{\mathbb{F}})$ which corresponds to
\[
\sum_{w\in W/W_\lambda}w\circ e(-2\lambda)\in {\mathbb{F}}[2\Lambda]^{W\circ}
\]
under the identification \eqref{eq:HC-center-F}, where
$W_\lambda=\{w\in W\mid w\lambda=\lambda\}$.
Then we have
\begin{equation}
\label{eq:mlambda:F}
Z(U_{\mathbb{F}})=\bigoplus_{\lambda\in\Lambda^+}{\mathbb{F}} m(\lambda).
\end{equation}
\subsection{}
We fix an integer $\ell>1$ satisfying
\begin{itemize}
\item[(a)]
$\ell$ is odd,
\item[(b)]
$\ell$ is prime to 3 if $G$ is of type $G_2$,
\item[(c)]
$\ell$ is prime to $|\Lambda/Q|$,
\end{itemize}
and a primitive $\ell$-th root $\zeta'\in{\mathbb{C}}$ of 1.
Define a subring ${\mathbb{A}}$ of ${\mathbb{F}}$ by
\[
{\mathbb{A}}=\{f(q^{1/|\Lambda/Q|})\mid
f(x)\in{\mathbb{Q}}(x),\,\mbox{$f$ is regular at $x=\zeta'$}\}.
\]
We set $\zeta=(\zeta')^{|\Lambda/Q|}$.
We note that $\zeta$ is also a primitive $\ell$-th root of 1 by the condition (c).
We denote by $U_{\mathbb{A}}^L$, $U_{\mathbb{A}}$ the ${\mathbb{A}}$-forms of $U_{\mathbb{F}}$ called the Lusztig form and the De Concini-Kac form respectively.
Namely, we have
\begin{align*}
U_{\mathbb{A}}^L
&=\langle
e_i^{(m)},\, f_i^{(m)},\,k_\lambda\mid
i\in I,\,m\in{\mathbb{Z}}_{\geqq0},\,\lambda\in\Lambda
\rangle_{{\mathbb{A}}-{\rm alg}}
\subset U_{\mathbb{F}}
,\\
U_{\mathbb{A}}
&=\langle
e_i,\, f_i,\,k_\lambda\mid
i\in I,\,\lambda\in\Lambda
\rangle_{{\mathbb{A}}-{\rm alg}}
\subset U_{\mathbb{F}}.
\end{align*}
We have obviously
$U_{\mathbb{A}}\subset U^L_{\mathbb{A}}$.
The Hopf algebra structure of $U_{\mathbb{F}}$ induces Hopf algebra structures over ${\mathbb{A}}$ of $U_{\mathbb{A}}^{L}$ and $U_{\mathbb{A}}$.
Setting
\[
U_{\mathbb{A}}^{L,\flat}=U_{\mathbb{A}}^{L}\cap U_{\mathbb{F}}^{\flat},\qquad
U_{\mathbb{A}}^{\flat}=U_{\mathbb{A}}\cap U_{\mathbb{F}}^{\flat}
\qquad(\flat=+, -, \geqq0, \leqq0),
\]
we have
\begin{align}
\label{eq:tri:L1}
&
U_{\mathbb{A}}^{L}\cong U_{\mathbb{A}}^{L,-}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,+},\\
\label{eq:tri:L2}
&
U_{\mathbb{A}}^{L,\geqq0}\cong U_{\mathbb{A}}^{L,0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,+}
\cong U_{\mathbb{A}}^{L,+}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,0},\\
\label{eq:tri:L3}
&
U_{\mathbb{A}}^{L,\leqq0}\cong U_{\mathbb{A}}^{L,0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,-}
\cong U_{\mathbb{A}}^{L,-}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,0},
\end{align}
and
\begin{align}
\label{eq:tri:DK1}
&
U_{\mathbb{A}}\cong U_{\mathbb{A}}^{-}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{+},\\
\label{eq:tri:DK2}
&
U_{\mathbb{A}}^{\geqq0}\cong U_{\mathbb{A}}^{0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{+}
\cong U_{\mathbb{A}}^{+}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{0},\\
\label{eq:tri:DK3}
&
U_{\mathbb{A}}^{\leqq0}\cong U_{\mathbb{A}}^{0}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{-}
\cong U_{\mathbb{A}}^{-}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{0}.
\end{align}
Set
\begin{align*}
\begin{bmatrix}{k_i;c}\\{m}\end{bmatrix}
&=\prod_{s=0}^{m-1}
\frac{q_i^{c-s}k_i-q_i^{-c+s}k_i^{-1}}
{q_i^{s+1}-q_i^{-s-1}}
\qquad(i\in I, m\in{\mathbb{Z}}_{\geqq0}, c\in{\mathbb{Z}}),\\
\begin{bmatrix}{k_i}\\{m}\end{bmatrix}
&=\begin{bmatrix}{k_i;0}\\{m}\end{bmatrix}
\qquad(i\in I, m\in{\mathbb{Z}}_{\geqq0}).
\end{align*}
Then we have
\[
\begin{bmatrix}{k_i;c}\\{m}\end{bmatrix}
\in U_{\mathbb{A}}^{L,0}
\qquad(i\in I, m\in{\mathbb{Z}}_{\geqq0}, c\in{\mathbb{Z}})
\]
and
\begin{align*}
U_{\mathbb{A}}^{L,0}
&=\bigoplus_{\lambda\in \Lambda',
(\varepsilon_i)\in\{0,1\}^I,
(n_i)\in{\mathbb{Z}}_{\geqq0}^I}
{\mathbb{A}}
k_\lambda\prod_{i\in I}
k_i^{\varepsilon_i}
\begin{bmatrix}{k_i}\\{n_i}
\end{bmatrix}
,\\
U_{\mathbb{A}}^{0}
&
=\bigoplus_{\lambda\in\Lambda}{\mathbb{A}} k_\lambda,
\end{align*}
where
$\Lambda'\subset \Lambda$ is a representative of $\Lambda/Q$.
By Lusztig \cite{L2} we have the following.
\begin{lemma}
\label{lem:PBW-L}
\begin{itemize}
\item[\rm(i)]
$\{{e}_{\beta_{N}}^{(m_N)}\cdots {e}_{\beta_{1}}^{(m_1)}\mid
m_1,\dots, m_N\geqq0\}$ $($resp.
\newline
$\{{f}_{\beta_{N}}^{(m_N)}\cdots {f}_{\beta_{1}}^{(m_1)}\mid
m_1,\dots, m_N\geqq0\}$$)$
is an ${{\mathbb{A}}}$-basis of $U_{{\mathbb{A}}}^{L,+}$ $($resp. $U_{{\mathbb{A}}}^{L,-}$$)$.
\item[\rm(ii)]
$U^L_{\mathbb{A}}$ is ${\mathbb{B}}$-stable.
\end{itemize}
\end{lemma}
By De Concini-Kac \cite{DK} we have also the following.
\begin{lemma}
\label{lem:PBW-DK}
\begin{itemize}
\item[\rm(i)]
$\{{e}_{\beta_{N}}^{m_N}\cdots {e}_{\beta_{1}}^{m_1}\mid
m_1,\dots, m_N\geqq0\}$ $($resp.
\newline
$\{{f}_{\beta_{N}}^{m_N}\cdots {f}_{\beta_{1}}^{m_1}\mid
m_1,\dots, m_N\geqq0\}$$)$
is an ${{\mathbb{A}}}$-basis of $U_{{\mathbb{A}}}^+$ $($resp. $U_{{\mathbb{A}}}^-$$)$.
\item[\rm(ii)]
$U_{\mathbb{A}}$ is ${\mathbb{B}}$-stable.
\end{itemize}
\end{lemma}
By Lemma \ref{lem:PBW-DK} (i), Lemma \ref{lem:PBW-L} (i) and Jantzen \cite[8.28]{Jan} we have
\begin{lemma}
\label{lem:pm-duality}
\[
U_{\mathbb{A}}^+=\{u\in U_{\mathbb{F}}^+\mid
\tau(u,U_{\mathbb{A}}^{L,-})\subset{\mathbb{A}})\},
\quad
U_{\mathbb{A}}^-=\{u\in U_{\mathbb{F}}^-\mid
\tau(U_{\mathbb{A}}^{L,+},u)\subset{\mathbb{A}})\}.
\]
\end{lemma}
\subsection{}
Now we consider the specialization
\[
{\mathbb{A}}\to{\mathbb{C}}\qquad
(q^{1/|\Lambda/Q|}\mapsto\zeta').
\]
Note that $q$ is mapped to $\zeta=(\zeta')^{|\Lambda/Q|}\in{\mathbb{C}}$, which is also a primitive $\ell$-th root of 1.
We set
\begin{align*}
&U_\zeta^L={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}}^L,\qquad
U_\zeta^{L,\flat}={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{L,\flat}\quad(\flat=+,-,\geqq0,\leqq0),\\
&U_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}},\qquad
U_\zeta^{\flat}={\mathbb{C}}\otimes_{\mathbb{A}} U_{\mathbb{A}}^{\flat}\quad(\flat=+,-,\geqq0,\leqq0).
\end{align*}
Then $U^L_\zeta$ and $U_\zeta$ are Hopf algebras over ${\mathbb{C}}$, and we have
\begin{align*}
&
U_\zeta\cong U_\zeta^{-}\otimes_{\mathbb{C}} U_\zeta^{0}\otimes_{\mathbb{C}} U_\zeta^{+},\\
&
U_\zeta^{\geqq0}\cong U_\zeta^{0}\otimes_{\mathbb{C}} U_\zeta^{+}
\cong U_\zeta^{+}\otimes_{\mathbb{C}} U_\zeta^{0},\\
&
U_\zeta^{\leqq0}\cong U_\zeta^{0}\otimes_{\mathbb{C}} U_\zeta^{-}
\cong U_\zeta^{-}\otimes_{\mathbb{C}} U_\zeta^{0}.
\end{align*}
We denote by
\[
^L\tau:U^{L,\geqq0}_\zeta\times U^{\leqq0}_\zeta\to{\mathbb{C}},\qquad
\tau^L:U^{\geqq0}_\zeta\times U^{L,\leqq0}_\zeta\to{\mathbb{C}}
\]
the bilinear forms induced by the Drinfeld paring $\tau$.
In general for a Lie algebra ${\mathfrak{s}}$ we denote its enveloping algebra by $U({\mathfrak{s}})$.
We denote by
\begin{equation}
\label{eq:LFrobenius}
\pi:U_\zeta^L\to U({\mathfrak{g}})
\end{equation}
Lusztig's Frobenius homomorphism (\cite{L2}).
Namely $\pi$ is the ${\mathbb{C}}$-algebra homomorphism given by
\[
\pi(e_i^{(m)})=
\begin{cases}
\overline{e}_i^{(m/\ell)}\,\,&(\ell|m)\\
0&(\ell\not|m),
\end{cases}
\quad
\pi(f_i^{(m)})=
\begin{cases}
\overline{f}_i^{(m/\ell)}\,\,&(\ell|m)\\
0&(\ell\not|m),
\end{cases}
\quad
\pi(k_\lambda)=1
\]
for $i\in I$, $m\in{\mathbb{Z}}_{\geqq0}$, $\lambda\in\Lambda$.
Here, $\overline{e}_i^{(n)}=\overline{e}_i^n/n!$, $\overline{f}_i^{(n)}=\overline{f}_i^n/n!$ for $i\in I$ and $n\in{\mathbb{Z}}_{\geqq0}$.
Then $\pi$ is a homomorphism of Hopf algebras.
\subsection{}
We recall the description of the center $Z(U_\zeta)$ of the algebra $U_\zeta$ due to De Concini-Kac \cite{DK} and De Concini-Procesi \cite{DP}.
Denote by $Z(U_{{\mathbb{A}}})$ the center of $U_{{\mathbb{A}}}$.
Then by \cite{DK} we have
\begin{equation*}
Z(U_{\mathbb{A}})=\bigoplus_{\lambda\in\Lambda^+}{\mathbb{A}} m(\lambda).
\end{equation*}
Define a subalgebra $Z_{Har}(U_\zeta)$ of $Z(U_\zeta)$ by
\[
Z_{Har}(U_\zeta)=\mathop{\rm Im}\nolimits(Z(U_{{\mathbb{A}}})\to U_\zeta).
\]
We define a twisted action of $W$ on the group algebra
${\mathbb{C}}[\Lambda]=\bigoplus_{\lambda\in\Lambda}{\mathbb{C}}
e(\lambda)$
of $\Lambda$ by
\[
w\circ e(\lambda)=\zeta^{(w\lambda-\lambda,\rho)}e(w\lambda)\qquad
(w\in W,\,\lambda\in\Lambda).
\]
By \cite{DK} \eqref{eq:HC-center-F} induces an isomorphism
\begin{equation}
\label{eq:HC-center}
Z_{Har}(U_\zeta)\simeq{\mathbb{C}}[2\Lambda]^{W\circ}
\end{equation}
of ${\mathbb{C}}$-algebras.
Namely,
the linear map
\[
\iota:Z_{Har}(U_\zeta)\to{\mathbb{C}}[\Lambda]
\]
defined as the composite of
\[
Z_{Har}(U_\zeta)
\subset
U_\zeta
\simeq
U^-_\zeta\otimes U^0_\zeta\otimes U^+_\zeta
\xrightarrow{\varepsilon\otimes 1\otimes\varepsilon}
U_\zeta^0
\cong
{\mathbb{C}}[\Lambda]
\]
is an injective algebra homomorphism whose image coincides with
${\mathbb{C}}[2\Lambda]^{W\circ}$.
Moreover, we have
\begin{equation}
\label{eq:mlambda:zeta}
Z_{Har}(U_\zeta)=\bigoplus_{\lambda\in\Lambda^+}{\mathbb{C}} m(\lambda),
\end{equation}
where we also denote by $m(\lambda)$ its image in $U_\zeta$ by abuse of notation.
By \cite{DK} the elements
\[
{e_\beta}^\ell,\quad
{f_\beta}^\ell,\quad
k_{\ell\lambda}\qquad(\beta\in\Delta^+,\,\lambda\in\Lambda)
\]
are central in $U_\zeta$.
Let $Z_{Fr}(U_\zeta)$ be the subalgebra of $U_\zeta$ generated by them.
$Z_{Fr}(U_\zeta)$ turns out to be a ${\mathbb{B}}$-stable Hopf subalgebra of $U_\zeta$.
Define an algebraic subgroup $K$ of $B^+\times B^-$ by
\[
K=\{(gh, g'h^{-1})\mid
h\in H,\,g\in N^+,\,g'\in N^-\}.
\]
By \cite{DP} we have an isomorphism
\begin{equation}
\label{eq:Fr-center}
Z_{Fr}(U_\zeta)\cong {\mathbb{C}}[K]
\end{equation}
of Hopf algebras.
The following description of the isomorphism \eqref{eq:Fr-center} is due to Gavarini \cite{Gav}.
Let us identify ${\mathbb{C}}[K]$ with ${\mathbb{C}}[N^-]\otimes{\mathbb{C}}[N^+]\otimes{\mathbb{C}}[H]$ via the isomorphism
\[
N^-\times N^+\times H\cong
K\qquad
((g,g',h)\leftrightarrow(gh,g'h^{-1}))
\]
of algebraic varieties.
For $f\in{\mathbb{C}}[N^+]$, $f'\in{\mathbb{C}}[N^-]$, $\lambda\in\Lambda$ the element of $Z_{Fr}(U_\zeta)$ corresponding to $f'\otimes f\otimes \theta_\lambda$ is given by $uk_{\ell\lambda}(Su')$ where $u\in U_\zeta^+$, $u'\in U_\zeta^-$ are given by
\begin{align*}
&\tau^L(u,y)=\langle f,\pi(y)\rangle\qquad(y\in U^{L,-}_\zeta),\\
&{}^L\tau(x,u')=\langle f',\pi(x)\rangle\qquad(x\in U^{L,+}_\zeta).
\end{align*}
Here we identify ${\mathbb{C}}[N^\pm]$ with a subspace of $U({\mathfrak{n}}^\pm)^*$ via the canonical Hopf paring.
Define
\begin{equation}
\label{eq:kappa}
\kappa:K\to G
\end{equation}
by $\kappa(g_1, g_2)=g_1g_2^{-1}$.
Define $\eta:G\to H/W$ as follows.
For $g\in G$ let $g_s\in G$ be the semisimple part of $g$ with respect to the Jordan decomposition.
Then $\mathop{\rm Ad}\nolimits(G)(g_s)\cap H$ coincides with a single $W$-orbit in $H$.
We define $\eta(g)\in H/W$ to be this $W$-orbit.
The morphism $\eta\circ\kappa:K\to H/W$ of algebraic varieties induces an injective algebra homomorphism $(\eta\circ\kappa)^*:{\mathbb{C}}[H/W]\to{\mathbb{C}}[K]$.
We identify ${\mathbb{C}}[H/W]$ with
\[
{\mathbb{C}}[2\ell \Lambda]^W
=\{f\in{\mathbb{C}}[2\ell\Lambda]\mid
wf=f\quad(w\in W)\}
\]
using the identification
\[
{\mathbb{C}}[2\ell\Lambda]\cong{\mathbb{C}}[H]
\qquad
(e(2\ell\lambda)\leftrightarrow\theta_\lambda).
\]
\begin{proposition}
[De Concini-Procesi \cite{DP}]
There exists an isomorphism
\begin{equation}
\label{eq:HCFr-center}
Z_{Har}(U_\zeta)\cap Z_{Fr}(U_\zeta)
\cong {\mathbb{C}}[2\ell\Lambda]
\end{equation}
of algebras such that the diagram
\[
\begin{CD}
Z_{Har}(U_\zeta)
@<<<
Z_{Har}(U_\zeta)\cap Z_{Fr}(U_\zeta)
@>>>
Z_{Fr}(U_\zeta)
\\
@VVV @VVV @VVV
\\
{\mathbb{C}}[2\Lambda]^{W\circ}
@<<<
{\mathbb{C}}[2\ell \Lambda]^W
@>>>
{\mathbb{C}}[K]
\end{CD}
\]
commutes.
Here the vertical arrows are the isomorphisms
\eqref{eq:HC-center}, \eqref{eq:HCFr-center}, \eqref{eq:Fr-center},
the upper horizontal arrows are the inclusions, and the lower horizontal arrows are the inclusion ${\mathbb{C}}[2\ell \Lambda]^W
\subset{\mathbb{C}}[2\Lambda]^{W\circ}$ and $(\eta\circ\kappa)^*$.
Moreover, we have an isomorphism
\[
Z(U_\zeta)\cong
Z_{Har}(U_\zeta)\otimes_{Z_{Har}(U_\zeta)\cap Z_{Fr}(U_\zeta)}
Z_{Fr}(U_\zeta)
\qquad(z_1z_2\leftrightarrow z_1\otimes z_2)
\]
of algebras.
In particular, we have
\begin{equation}
\label{eq:full-center}
Z(U_\zeta)\cong
{\mathbb{C}}[2\Lambda]^{W\circ}
\otimes_{{\mathbb{C}}[2\ell \Lambda]^W}
{\mathbb{C}}[K].
\end{equation}
\end{proposition}
\begin{corollary}
\label{cor:center}
We have
\[
{\rm{Spec}}\, Z(U_\zeta)
\cong K\times_{H/W}H/{W\circ},
\]
where $H/{W\circ}\to{H/W}$ is given by $[t]\mapsto[t^\ell]$.
\end{corollary}
\section{Representation}
\label{sec:rep}
\subsection{}
If $R$ is a ring, we denote by $\mathop{\rm Mod}\nolimits(R)$ the category of left $R$-modules.
\begin{remark}
{\rm
We will also use the notation like $\mathop{\rm Mod}\nolimits({\mathcal R})$ even when ${\mathcal R}$ is not a ring (e.g.\ $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})$).
The meaning of this type of notation will be explained separately when they appear.
}
\end{remark}
For $M_1, M_2\in\mathop{\rm Mod}\nolimits(U_{\mathbb{F}})$ the tensor product $M_1\otimes M_2$ has a natural $U_{\mathbb{F}}$-module structure by
\[
u\cdot(m_1\otimes m_2)=\Delta(u)(m_1\otimes m_2)
\qquad
(u\in U_{\mathbb{F}},\,m_1\in M_1,\, m_2\in M_2).
\]
For $\lambda\in\Lambda$ we define an algebra homomorphism $\chi_\lambda:U_{\mathbb{F}}^0\to{\mathbb{F}}$ by $\chi_\lambda(k_\mu)=q^{(\lambda,\mu)}\,(\mu\in\Lambda)$.
For $M\in\mathop{\rm Mod}\nolimits(U_{\mathbb{F}})$ and $\lambda\in\Lambda$ we set
\[
M_\lambda=\{m\in M\mid
hm=\chi_\lambda(h)m\quad(h\in U_{\mathbb{F}}^0)\}.
\]
We denote by $\mathop{\rm Mod}\nolimits_f(U_{\mathbb{F}})$ the category of finite dimensional $U_{\mathbb{F}}$-modules $M$ such that $M=\bigoplus_{\lambda\in\Lambda}M_\lambda$.
We also denote by $\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ the category of $U_{\mathbb{F}}$-modules $M$ which is a sum of modules in $\mathop{\rm Mod}\nolimits_f(U_{\mathbb{F}})$.
It is well-known that a $U_{\mathbb{F}}$-module $M$ belongs to $\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ if and only if $M=\bigoplus_{\lambda\in\Lambda}M_\lambda$ and
for any $m\in M$ there exists $r\in{\mathbb{Z}}_{>0}$ such that
$e_i^{(r)}m=f_i^{(r)}m=0$ for any $i\in I$.
For $M_1, M_2\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ we have $M_1\otimes M_2\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
For $\lambda\in\Lambda$ we define
$M_{+,{\mathbb{F}}}(\lambda), M_{-,{\mathbb{F}}}(\lambda)\in\mathop{\rm Mod}\nolimits(U_{\mathbb{F}})$
by
\begin{align*}
M_{+,{\mathbb{F}}}(\lambda)
=&U_{\mathbb{F}}/
\sum_{y\in U_{\mathbb{F}}^-}U_{\mathbb{F}}(y-\varepsilon(y))+
\sum_{h\in U_{\mathbb{F}}^0}U_{\mathbb{F}}(h-\chi_\lambda(h)),\\
M_{-,{\mathbb{F}}}(\lambda)
=&U_{\mathbb{F}}/
\sum_{x\in U_{\mathbb{F}}^+}U_{\mathbb{F}}(x-\varepsilon(x))+
\sum_{h\in U_{\mathbb{F}}^0}U_{\mathbb{F}}(h-\chi_\lambda(h)).
\end{align*}
$M_{+,{\mathbb{F}}}(\lambda)$ is a lowest weight module with lowest weight $\lambda$, and $M_{-,{\mathbb{F}}}(\lambda)$ is a highest weight module with highest weight $\lambda$.
By \eqref{eq:tri:F1} we have isomorphisms
\[
M_{+,{\mathbb{F}}}(\lambda)
\cong
U_{\mathbb{F}}^{+}\quad
(\overline{u}\leftrightarrow u),\qquad
M_{-,{\mathbb{F}}}(\lambda)
\cong
U_{\mathbb{F}}^{-}\quad
(\overline{u}\leftrightarrow u)
\]
of ${\mathbb{F}}$-modules.
Moreover, we have weight space decompositions
\[
M_{+,{\mathbb{F}}}(\lambda)
=\bigoplus_{\mu\in\lambda+Q^+}M_{+,{\mathbb{F}}}(\lambda)_\mu,\qquad
M_{-,{\mathbb{F}}}(\lambda)
=\bigoplus_{\mu\in\lambda-Q^+}M_{-,{\mathbb{F}}}(\lambda)_\mu.
\]
For $\lambda\in\Lambda^+$ we define
$L_{+,{\mathbb{F}}}(-\lambda), L_{-,{\mathbb{F}}}(\lambda)\in\mathop{\rm Mod}\nolimits_f(U_{\mathbb{F}})$
by
\begin{align*}
&L_{+,{\mathbb{F}}}(-\lambda)\\
=&U_{\mathbb{F}}/
\sum_{y\in U_{\mathbb{F}}^-}U_{\mathbb{F}}(y-\varepsilon(y))+
\sum_{h\in U_{\mathbb{F}}^0}U_{\mathbb{F}}(h-\chi_{-\lambda}(h))
+\sum_{i\in I}U_{\mathbb{F}} e_i^{((\lambda,\alpha_i^\vee)+1)},\\
&L_{-,{\mathbb{F}}}(\lambda)\\
=&U_{\mathbb{F}}/
\sum_{x\in U_{\mathbb{F}}^+}U_{\mathbb{F}}(x-\varepsilon(x))+
\sum_{h\in U_{\mathbb{F}}^0}U_{\mathbb{F}}(h-\chi_\lambda(h))
+\sum_{i\in I}U_{\mathbb{F}} f_i^{((\lambda,\alpha_i^\vee)+1)}.
\end{align*}
$L_{+,{\mathbb{F}}}(-\lambda)$ is a finite-dimensional irreducible lowest weight module with lowest weight $-\lambda$, and $L_{-,{\mathbb{F}}}(\lambda)$ is a finite-dimensional irreducible highest weight module with highest weight $\lambda$.
We have weight space decompositions
\[
L_{+,{\mathbb{F}}}(-\lambda)
=\bigoplus_{\mu\in-\lambda+Q^+}L_{+,{\mathbb{F}}}(-\lambda)_\mu,\qquad
L_{-,{\mathbb{F}}}(\lambda)
=\bigoplus_{\mu\in\lambda-Q^+}L_{-,{\mathbb{F}}}(\lambda)_\mu.
\]
We have also $L_{-,{\mathbb{F}}}(\lambda)\cong L_{+,{\mathbb{F}}}(w_0\lambda)$.
Moreover, the category $\mathop{\rm Mod}\nolimits_f(U_{\mathbb{F}})$ is semisimple, and its simple objects are $L_{-,{\mathbb{F}}}(\lambda)$ for $\lambda\in\Lambda^+$ (see Lusztig \cite{L1}).
Let $M$ be a $U_{\mathbb{F}}$-module with weight space decomposition
$M=\bigoplus_{\mu\in\Lambda}M_\mu$ such that $\dim M_\mu<\infty$ for any $\mu\in\Lambda$.
We define a $U_{\mathbb{F}}$-module $M^\bigstar$ by
\[
M^\bigstar=\bigoplus_{\mu\in\Lambda}
M_\mu^*\subset M^*=\mathop{\rm Hom}\nolimits_{\mathbb{F}}(M,{\mathbb{F}}),
\]
where the action of $U_{\mathbb{F}}$ is given by
\[
\langle um^*,m\rangle=\langle m^*,(Su)m\rangle
\qquad(u\in U_{\mathbb{F}}, m^*\in M^\bigstar, m\in M).
\]
Here $\langle\,,\,\rangle:M^\bigstar\times M\to{\mathbb{F}}$ is the natural paring.
We set
\begin{align*}
M^*_{\pm,{\mathbb{F}}}(\lambda)&=(M_{\mp,{\mathbb{F}}}(-\lambda))^\bigstar\qquad
(\lambda\in\Lambda),\\
L^*_{\pm,{\mathbb{F}}}(\mp\lambda)&=(L_{\mp,{\mathbb{F}}}(\pm\lambda))^\bigstar\qquad
(\lambda\in\Lambda^+).
\end{align*}
Since $L_{\mp,{\mathbb{F}}}(\pm\lambda)$ is irreducible, we have
\[
L^*_{\pm,{\mathbb{F}}}(\mp\lambda)\cong L_{\pm,{\mathbb{F}}}(\mp\lambda)\qquad
(\lambda\in\Lambda^+).
\]
Note that we have an injective $U_{\mathbb{F}}$-homomorphism
\begin{equation}
\label{eq:LASTARtoMASTAR:F}
L^*_{\pm,{\mathbb{F}}}(\mp\lambda)\to M^*_{\pm,{\mathbb{F}}}(\mp\lambda)
\qquad(\lambda\in\Lambda^+).
\end{equation}
induced by the natural homomorphism
$M_{\mp,{\mathbb{F}}}(\pm\lambda)\to L_{\mp,{\mathbb{F}}}(\pm\lambda)$.
For $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ we have a group homomorphism
\[
{\mathbb{B}}\to \mathop{\rm{End}}\nolimits(M)^\times
\]
given by
\begin{align}
\label{eq:Ti}
T_i
&=
\exp_{q_i^{-1}}(q_ik_if_i)
\exp_{q_i^{-1}}(-e_i)
\exp_{q_i^{-1}}(q_i^{-1}k_i^{-1}f_i)H_i\\
\nonumber
&=
\exp_{q_i^{-1}}(-q_ik_i^{-1}e_i)
\exp_{q_i^{-1}}(f_i)
\exp_{q_i^{-1}}(-q_i^{-1}k_ie_i)H_i,
\end{align}
where
\[
\exp_t(x)=\sum_{n=0}^\infty\frac{t^{n(n-1)/2}}{[n]_t!}x^n
\in{\mathbb{Q}}(t)[[x]],
\]
and $H_i$ is the operator on $M$ which acts by $q^{(\lambda,\alpha_i)((\lambda,\alpha_i^\vee)+1)/2}\mathop{\rm id}\nolimits$ on $M_\lambda$ for each $\lambda\in\Lambda$.
This operator $T_i$ coincides with Lusztig's operator $T''_{i,1}$ in \cite[5.2]{Lbook}.
We have
\[
T_w(M_\lambda)=M_{w\lambda}\qquad
(M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}}),\,\,\lambda\in\Lambda),
\]
and
\begin{equation*}
T_w(um)=T_w(u)T_w(m)
\qquad(w\in W, u\in U_{\mathbb{F}}, m\in M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})).
\end{equation*}
For $M_1, M_2\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$
we sometimes write the action of $T\in{\mathbb{B}}$ on $M_1\otimes M_2\in \mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ by
$\Delta T:M_1\otimes M_2\to M_1\otimes M_2$.
We will need the following (see Lusztig \cite[5.3]{Lbook}).
\begin{lemma}
\label{lem:DeltaT}
For $M_1, M_2\in \mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ and $i\in I$ we have
\begin{align*}
\Delta T_i&=
\exp_{q_i}(q_i^{-2}(q_i-q_i^{-1})e_ik_i^{-1}\otimes f_ik_i)(T_i\otimes T_i)\\
&=(T_i\otimes T_i)\exp_{q_i}((q_i-q_i^{-1})f_i\otimes e_i)
\end{align*}
in $\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M_1\otimes M_2)^\times$.
\end{lemma}
\subsection{}
For $M\in\mathop{\rm Mod}\nolimits(U^L_{\mathbb{A}})$ and $\lambda\in\Lambda$ we set
\[
M_\lambda=\{m\in M\mid
hm=\chi_\lambda(h)m\quad(h\in U^{L,0}_{\mathbb{A}})\}.
\]
\begin{lemma}
\label{lemma:chi-indep0}
Let $\lambda, \mu\in\Lambda$ such that $\lambda\ne\mu$.
Then there exists $h\in U^{L,0}_{\mathbb{A}}$ such that
$\chi_\lambda(h)=1$ and $\chi_\mu(h)=0$.
In particular, we have
$\chi_\lambda\ne\chi_\mu$
in $\mathop{\rm Hom}\nolimits_{\mathbb{A}}(U_{\mathbb{A}}^{L,0},{\mathbb{A}})$.
\end{lemma}
\begin{proof}
Take $i\in I$ such that $(\lambda,\alpha_i^\vee)\ne(\mu,\alpha_i^\vee)$.
We may assume $(\lambda,\alpha_i^\vee)>(\mu,\alpha_i^\vee)$.
Then the assertion holds for
$h=\begin{bmatrix}
{k_i;-(\mu,\alpha_i^\vee)}\\
{(\lambda-\mu,\alpha_i^\vee)}
\end{bmatrix}
$.
\end{proof}
For $\lambda\in\Lambda$ we define
$M_{+,{\mathbb{A}}}(\lambda), M_{-,{\mathbb{A}}}(\lambda)\in\mathop{\rm Mod}\nolimits(U^L_{\mathbb{A}})$
by
\begin{align*}
M_{+,{\mathbb{A}}}(\lambda)
=&U^L_{\mathbb{A}}/
\sum_{y\in U_{\mathbb{A}}^{L,-}}U^L_{\mathbb{A}}(y-\varepsilon(y))+
\sum_{h\in U_{\mathbb{A}}^{L,0}}U^L_{\mathbb{A}}(h-\chi_\lambda(h)),\\
M_{-,{\mathbb{A}}}(\lambda)
=&U^L_{\mathbb{A}}/
\sum_{x\in U_{\mathbb{A}}^{L,+}}U^L_{\mathbb{A}}(x-\varepsilon(x))+
\sum_{h\in U_{\mathbb{A}}^{L,0}}U^L_{\mathbb{A}}(h-\chi_\lambda(h)).
\end{align*}
By \eqref{eq:tri:L1} we have isomorphisms
\[
M_{+,{\mathbb{A}}}(\lambda)
\cong
U_{\mathbb{A}}^{L,+}\quad
(\overline{u}\leftrightarrow u),\qquad
M_{-,{\mathbb{A}}}(\lambda)
\cong
U_{\mathbb{A}}^{L,-}\quad
(\overline{u}\leftrightarrow u)
\]
of ${\mathbb{A}}$-modules.
In particular, $M_{\pm,{\mathbb{A}}}(\lambda)$ is a free ${\mathbb{A}}$-module and we have ${\mathbb{F}}\otimes_{\mathbb{A}} M_{\pm,{\mathbb{A}}}(\lambda)\cong M_{\pm,{\mathbb{F}}}(\lambda)$.
Moreover, we have weight space decompositions
\[
M_{+,{\mathbb{A}}}(\lambda)
=\bigoplus_{\mu\in\lambda+Q^+}M_{+,{\mathbb{A}}}(\lambda)_\mu,\qquad
M_{-,{\mathbb{A}}}(\lambda)
=\bigoplus_{\mu\in\lambda-Q^+}M_{-,{\mathbb{A}}}(\lambda)_\mu.
\]
For $\lambda\in\Lambda^+$ we define
$L_{+,{\mathbb{A}}}(-\lambda)\in\mathop{\rm Mod}\nolimits(U^L_{\mathbb{A}})$ (resp. $L_{-,{\mathbb{A}}}(\lambda)\in\mathop{\rm Mod}\nolimits(U^L_{\mathbb{A}})$)
to be the $U^L_{\mathbb{A}}$-submodule of
$L_{+,{\mathbb{F}}}(-\lambda)$ (resp. $L_{-,{\mathbb{F}}}(\lambda)$)
generated by $\overline{1}\in L_{+,{\mathbb{F}}}(-\lambda)$ (resp. $\overline{1}\in L_{-,{\mathbb{F}}}(\lambda)$).
By definition $L_{\pm,{\mathbb{A}}}(\mp\lambda)$ is a free ${\mathbb{A}}$-module and we have ${\mathbb{F}}\otimes_{\mathbb{A}} L_{\pm,{\mathbb{A}}}(\mp\lambda)\cong L_{\pm,{\mathbb{F}}}(\mp\lambda)$.
Moreover, we have weight space decompositions
\[
L_{+,{\mathbb{A}}}(-\lambda)
=\bigoplus_{\mu\in-\lambda+Q^+}L_{+,{\mathbb{A}}}(-\lambda)_\mu,\qquad
L_{-,{\mathbb{A}}}(\lambda)
=\bigoplus_{\mu\in\lambda-Q^+}L_{-,{\mathbb{A}}}(\lambda)_\mu.
\]
The canonical surjective $U_{\mathbb{F}}$-homomorphism $M_{\pm,{\mathbb{F}}}(\mp\lambda)\to L_{\pm,{\mathbb{F}}}(\mp\lambda)$ induces a surjective $U_{\mathbb{A}}^L$-homomorphism
\begin{equation}
\label{eq:MAtoLA}
M_{\pm,{\mathbb{A}}}(\mp\lambda)\to L_{\pm,{\mathbb{A}}}(\mp\lambda)
\qquad(\lambda\in\Lambda^+).
\end{equation}
Note that \eqref{eq:MAtoLA} is a split epimorphism of ${\mathbb{A}}$-modules since ${\mathbb{A}}$ is PID and $L_{\pm,{\mathbb{A}}}(\mp\lambda)_\mu$ is a torsion free finitely generated ${\mathbb{A}}$-module.
Let $M$ be a $U^L_{\mathbb{A}}$-module with weight space decomposition
$M=\bigoplus_{\mu\in\Lambda}M_\mu$ such that $M_\mu$ is a free ${\mathbb{A}}$-module of finite rank for any $\mu\in\Lambda$.
We define a $U^L_{\mathbb{A}}$-module $M^\bigstar$ by
\[
M^\bigstar=\bigoplus_{\mu\in\Lambda}
\mathop{\rm Hom}\nolimits_{\mathbb{A}}(M_\mu,{\mathbb{A}})\subset \mathop{\rm Hom}\nolimits_{\mathbb{A}}(M,{\mathbb{A}}),
\]
where the action of $U^L_{\mathbb{A}}$ is given by
\[
\langle um^*,m\rangle=\langle m^*,(Su)m\rangle
\qquad(u\in U^L_{\mathbb{A}}, m^*\in M^\bigstar, m\in M).
\]
Here $\langle\,,\,\rangle:M^\bigstar\times M\to{\mathbb{A}}$ is the natural paring.
We set
\begin{align*}
M^*_{\pm,{\mathbb{A}}}(\lambda)&=(M_{\mp,{\mathbb{A}}}(-\lambda))^\bigstar\qquad
(\lambda\in\Lambda),\\
L^*_{\pm,{\mathbb{A}}}(\mp\lambda)&=(L_{\mp,{\mathbb{A}}}(\pm\lambda))^\bigstar\qquad
(\lambda\in\Lambda^+).
\end{align*}
Then $M^*_{\pm,{\mathbb{A}}}(\lambda)$ for $\lambda\in\Lambda$ and $L^*_{\pm,{\mathbb{A}}}(\mp\lambda)$ for $\lambda\in\Lambda^+$ are free ${\mathbb{A}}$-modules satisfying
\[
{\mathbb{F}}\otimes_{\mathbb{A}} M^*_{\pm,{\mathbb{A}}}(\lambda)
\cong
M^*_{\pm,{\mathbb{F}}}(\lambda),\qquad
{\mathbb{F}}\otimes_{\mathbb{A}} L^*_{\pm,{\mathbb{A}}}(\mp\lambda)
\cong
L^*_{\pm,{\mathbb{F}}}(\mp\lambda).
\]
Moreover, we can identify $M^*_{\pm,{\mathbb{A}}}(\lambda)$ and $L^*_{\pm,{\mathbb{A}}}(\mp\lambda)$ with ${\mathbb{A}}$-submodules of
$M^*_{\pm,{\mathbb{F}}}(\lambda)$ and $L^*_{\pm,{\mathbb{F}}}(\mp\lambda)$ respectively.
Under this identification we have
\begin{equation}
\label{eq:MLAF}
L^*_{\pm,{\mathbb{A}}}(\mp\lambda)=
L^*_{\pm,{\mathbb{F}}}(\mp\lambda)\cap M^*_{\pm,{\mathbb{A}}}(\mp\lambda)\qquad(\lambda\in\Lambda^+).
\end{equation}
In particular, the $U_{\mathbb{A}}^L$-homomorphism
\begin{equation}
\label{eq:LASTARtoMASTAR}
L^*_{\pm,{\mathbb{A}}}(\mp\lambda)\to M^*_{\pm,{\mathbb{A}}}(\mp\lambda)
\qquad(\lambda\in\Lambda^+)
\end{equation}
is a split monomorphism of ${\mathbb{A}}$-modules.
\subsection{}
Let $\lambda\in\Lambda$.
By abuse of notation
we also denote by $\chi_\lambda:U^{L,0}_\zeta\to{\mathbb{C}}$ the ${\mathbb{C}}$-algebra homomorphism induced by $\chi_\lambda:U^{L,0}_{\mathbb{A}}\to{\mathbb{A}}$.
\begin{lemma}
\label{lemma:chi-indep}
\begin{itemize}
\item[\rm(i)]
Let $\lambda, \mu\in\Lambda$.
If we have $\chi_\lambda=\chi_\mu$ in $\mathop{\rm Hom}\nolimits_{\mathbb{C}}(U^{L,0}_\zeta,{\mathbb{C}})$, then we have $\lambda=\mu$.
\item[\rm(ii)]
$\{\chi_\lambda\}_{\lambda\in\Lambda}$ is a linearly independent subset of $\mathop{\rm Hom}\nolimits_{\mathbb{C}}(U^{L,0}_\zeta,{\mathbb{C}})$.
\end{itemize}
\end{lemma}
\begin{proof}
(i) is a consequence of Lemma \ref{lemma:chi-indep0}, and
(ii) follows from (i) easily.
\end{proof}
For $M\in\mathop{\rm Mod}\nolimits(U^L_\zeta)$ and $\lambda\in\Lambda$ we set
\[
M_\lambda=\{m\in M\mid hm=\chi_\lambda(h)m\,\,(h\in U^{L,0}_\zeta)\}.
\]
We denote by $\mathop{\rm Mod}\nolimits_f(U^L_\zeta)$ the category of finite dimensional $U_\zeta^L$-module $M$ such that $M=\bigoplus_{\lambda\in\Lambda}M_\lambda$.
We also denote by $\mathop{\rm Mod}\nolimits_{int}(U^L_\zeta)$ the category of $U^L_\zeta$-modules $M$ which is a sum of modules in $\mathop{\rm Mod}\nolimits_f(U^L_\zeta)$.
It is known that a $U^L_\zeta$-module $M$ belongs to $\mathop{\rm Mod}\nolimits_{int}(U^L_\zeta)$ if and only if $M=\bigoplus_{\lambda\in\Lambda}M_\lambda$ and
for any $m\in M$ there exists $r\in{\mathbb{Z}}_{>0}$ such that
$e_i^{(n)}m=f_i^{(n)}m=0$ for any $i\in I$ and $n\geqq r$ (see, for example, Andersen-Polo-Wen \cite{APW1}).
For $M\in\mathop{\rm Mod}\nolimits_f(U^L_\zeta)$ we have a group homomorphism
\[
{\mathbb{B}}\to GL(M)
\]
given by the formula similar to \eqref{eq:Ti} ($q$ is replaced by $\zeta$).
\begin{lemma}
\label{lem:braid-Fr}
Let $M$ be a finite-dimensional $U({\mathfrak{g}})$-module.
If we regard $M$ as a $U_\zeta^L$-module via $\pi:U_\zeta^L\to U({\mathfrak{g}})$ $($see \eqref{eq:LFrobenius}$)$,
then the action of $T_i$ on the $U_\zeta^L$-module $M$ is given by
\[
T_i=\exp(\overline{f}_i)\exp(-\overline{e}_i)\exp(\overline{f}_i).
\]
\end{lemma}
\begin{proof}
Note that for $\lambda\in\Lambda$ and $m\in M$ satisfying $hm=\lambda(h)m$ for any $h\in{\mathfrak{h}}$ we have $tm=\chi_{\ell\lambda}(t)m$ for any $t\in U^{L,0}_\zeta$.
In particular, $k_i$ and $H_i$ in \eqref{eq:Ti} act trivially on $M$. From this we see easily that the assertion holds.
\end{proof}
For $\lambda\in\Lambda$ we set
\[
M_{\pm,\zeta}(\lambda)={\mathbb{C}}\otimes_{\mathbb{A}} M_{\pm,{\mathbb{A}}}(\lambda),
\qquad
M^*_{\pm,\zeta}(\lambda)={\mathbb{C}}\otimes_{\mathbb{A}} M^*_{\pm,{\mathbb{A}}}(\lambda).
\]
For $\lambda\in\Lambda^+$ we set
\[
L_{\pm,\zeta}(\mp\lambda)={\mathbb{C}}\otimes_{\mathbb{A}} L_{\pm,{\mathbb{A}}}(\mp\lambda),
\qquad
L^*_{\pm,\zeta}(\mp\lambda)={\mathbb{C}}\otimes_{\mathbb{A}} L^*_{\pm,{\mathbb{A}}}(\mp\lambda).
\]
We have canonical $U^L_\zeta$-homomorphisms
\begin{align}
\label{eq:MtoL}
M_{\pm,\zeta}(\mp\lambda)\to L_{\pm,\zeta}(\mp\lambda)
\qquad(\lambda\in\Lambda^+),
\\
\label{eq:LSTARtoMSTAR}
L^*_{\pm,\zeta}(\mp\lambda)\to M^*_{\pm,\zeta}(\mp\lambda)
\qquad(\lambda\in\Lambda^+).
\end{align}
Note that \eqref{eq:MtoL} is surjective, and \eqref{eq:LSTARtoMSTAR} is injective.
\section{Quantized flag manifold}
\label{qflag}
\subsection{}
We denote by $C_{\mathbb{F}}$ the subspace of $U_{\mathbb{F}}^*=\mathop{\rm Hom}\nolimits_{\mathbb{F}}(U_{\mathbb{F}},{\mathbb{F}})$ spanned by the matrix coefficients of $U_{\mathbb{F}}$-modules belonging to $\mathop{\rm Mod}\nolimits_f(U_{\mathbb{F}})$, and denote by
\begin{equation}
\label{eq:Hopf-paring}
\langle\,,\,\rangle:C_{\mathbb{F}}\times U_{\mathbb{F}}\to{\mathbb{F}}
\end{equation}
the canonical paring.
Then $C_{\mathbb{F}}$ is endowed with a Hopf algebra structure dual to $U_{\mathbb{F}}$ via \eqref{eq:Hopf-paring}.
Moreover, the bilinear paring \eqref{eq:Hopf-paring} is a Hopf paring in the sense that we have
\begin{itemize}
\item[(a)]
$u\in U_{\mathbb{F}},\,\langle C_{\mathbb{F}}, u\rangle=0
\,\Longrightarrow
u=0$,
\item[(b)]
$\varphi\in C_{\mathbb{F}},\,\langle \varphi, U_{\mathbb{F}}\rangle=0
\,\Longrightarrow
\varphi=0$,
\item[(c)]
$\langle \varphi\psi,u\rangle=\langle \varphi\otimes\psi,\Delta u\rangle
\qquad(\varphi, \psi\in C_{\mathbb{F}}, u\in U_{\mathbb{F}})$,
\item[(d)]
$\langle \varphi,uv\rangle=\langle \Delta\varphi,u\otimes v\rangle
\qquad(\varphi\in C_{\mathbb{F}}, u, v\in U_{\mathbb{F}})$,
\item[(e)]
$\langle 1,u\rangle=\varepsilon(u)
\qquad(u\in U_{\mathbb{F}})$,
\item[(f)]
$\langle \varphi,1\rangle=\varepsilon(\varphi)
\qquad(\varphi\in C_{\mathbb{F}})$,
\item[(g)]
$\langle S\varphi,Su\rangle=\langle \varphi,u\rangle
\qquad(\varphi\in C_{\mathbb{F}}, u\in U_{\mathbb{F}})$
\end{itemize}
(see, for example, \cite[5.11]{Jan} for (a) and \cite{T1} for (b),\dots, (g)).
Note also that we have a $U_{\mathbb{F}}$-bimodule structure of $C_{\mathbb{F}}$ by
\[
\langle u_1\cdot\varphi\cdot u_2,u\rangle
=\langle \varphi,u_2uu_1\rangle
\qquad(\varphi\in C_{\mathbb{F}}, u, u_1, u_2\in U_{\mathbb{F}}).
\]
Define a $\Lambda$-graded ring $A_{\mathbb{F}}=\bigoplus_{\lambda\in\Lambda^+}A_{\mathbb{F}}(\lambda)$ by
\begin{align*}
A_{\mathbb{F}}&=\{\varphi\in C_{\mathbb{F}}\mid\varphi\cdot f_i=0\quad(i\in I)\},\\
A_{\mathbb{F}}(\lambda)&=
\{\varphi\in A_{\mathbb{F}}\mid \varphi\cdot k_\mu=q^{(\mu,\lambda)}\varphi\quad(\mu\in\Lambda)\}.
\end{align*}
Note that $A_{\mathbb{F}}$ is a left $U_{\mathbb{F}}$-submodule of $C_{\mathbb{F}}$.
Moreover, for $\lambda\in\Lambda^+$ we have $A_{\mathbb{F}}(\lambda)\cong L_{\mathbb{F}}(\lambda)$ as a $U_{\mathbb{F}}$-module
(see \cite{T2}).
We set
\[
(U_{\mathbb{F}}^\pm)^\bigstar=\bigoplus_{\beta\in Q^+}\mathop{\rm Hom}\nolimits_{\mathbb{F}}(U_{{\mathbb{F}},\pm\beta}^\pm,{\mathbb{F}})\subset
\mathop{\rm Hom}\nolimits_{\mathbb{F}}(U_{{\mathbb{F}}}^\pm,{\mathbb{F}})=(U_{\mathbb{F}}^\pm)^*.
\]
We identify $(U^-_{{\mathbb{F}}})^*\otimes (U^0_{{\mathbb{F}}})^*\otimes (U^+_{{\mathbb{F}}})^*$ with a subspace of $U_{\mathbb{F}}^*$ by the embedding
\begin{align*}
&(U^-_{{\mathbb{F}}})^*\otimes (U^0_{{\mathbb{F}}})^*\otimes (U^+_{{\mathbb{F}}})^*\to U_{\mathbb{F}}^*\\
&\quad
(f\otimes\chi\otimes g\mapsto[uhv\mapsto
f(u)\chi(h)g(v)] \mbox{ for }u\in
U^-_{{\mathbb{F}}}, h\in U^0_{{\mathbb{F}}}, v\in U^+_{{\mathbb{F}}}).
\end{align*}
Then we have
\begin{align}
\label{eq:Cdecomp}
C_{\mathbb{F}}&\subset
(U^-_{{\mathbb{F}}})^\bigstar\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{F}}\chi_\lambda)\otimes (U^+_{{\mathbb{F}}})^\bigstar,\\
\label{eq:Adecomp1}
A_{\mathbb{F}}&=
(\varepsilon\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{F}}\chi_\lambda)\otimes (U^+_{{\mathbb{F}}})^\bigstar)
\cap\, C_{\mathbb{F}},\\
\label{eq:Adecomp2}
A_{\mathbb{F}}(\lambda)&=
(\varepsilon\otimes \chi_\lambda\otimes (U^+_{{\mathbb{F}}})^\bigstar)
\cap\, C_{\mathbb{F}}.
\end{align}
\subsection{}
We define ${\mathbb{A}}$-forms $C_{\mathbb{A}}$, $A_{\mathbb{A}}$, $A_{\mathbb{A}}(\lambda)$ ($\lambda\in\Lambda^+$) of $C_{\mathbb{F}}$, $A_{\mathbb{F}}$, $A_{\mathbb{F}}(\lambda)$ respectively by
\begin{align*}
&C_{\mathbb{A}}=\{\varphi\in C_{\mathbb{F}}\mid
\langle \varphi,U_{\mathbb{A}}^L\rangle\subset{\mathbb{A}}\},\quad
A_{\mathbb{A}}=A_{\mathbb{F}}\cap C_{\mathbb{A}},\quad
A_{\mathbb{A}}(\lambda)=A_{\mathbb{F}}(\lambda)\cap C_{\mathbb{A}}.
\end{align*}
Then $C_{\mathbb{A}}$ is an ${\mathbb{A}}$-algebra and $A_{\mathbb{A}}$ is its subalgebra.
Moreover, $C_{\mathbb{A}}$ is a $U^L_{\mathbb{A}}$-bimodule and $A_{\mathbb{A}}$ is its left $U^L_{\mathbb{A}}$-submodule.
We set
\[
(U_{\mathbb{A}}^{L,\pm})^\bigstar=\bigoplus_{\beta\in Q^+}\mathop{\rm Hom}\nolimits_{\mathbb{A}}(U_{{\mathbb{A}},\pm\beta}^{L,\pm},{\mathbb{A}})\subset
\mathop{\rm Hom}\nolimits_{\mathbb{A}}(U_{{\mathbb{A}}}^{L,\pm},{\mathbb{A}}).
\]
By Lemma \ref{lemma:chi-indep0} we can easily show
\[
(\bigoplus_{\lambda\in \Lambda}{\mathbb{F}}\chi_\lambda)
\cap \mathop{\rm Hom}\nolimits_{\mathbb{A}}(U^{L,0}_{\mathbb{A}},{\mathbb{A}})=
\bigoplus_{\lambda\in \Lambda}{\mathbb{A}}\chi_\lambda.
\]
Hence by \eqref{eq:Cdecomp}, \eqref{eq:Adecomp1}, \eqref{eq:Adecomp2}
we have
\begin{align}
\label{eq:CdecompA}
C_{\mathbb{A}}&=
(
(U^{L,-}_{{\mathbb{A}}})^\bigstar\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{A}}\chi_\lambda)\otimes (U^{L,+}_{{\mathbb{A}}})^\bigstar)
\cap C_{\mathbb{F}},\\
\label{eq:Adecomp1A}
A_{\mathbb{A}}&=
(\varepsilon\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{A}}\chi_\lambda)\otimes (U^{L,+}_{{\mathbb{A}}})^\bigstar)
\cap\, C_{\mathbb{F}},\\
\label{eq:Adecomp2A}
A_{\mathbb{A}}(\lambda)&=
(\varepsilon\otimes \chi_\lambda\otimes (U^{L,+}_{{\mathbb{A}}})^\bigstar)
\cap\, C_{\mathbb{F}}.
\end{align}
In particular, we have
\begin{equation}
A_{\mathbb{A}}=\bigoplus_{\lambda\in\Lambda^+}
A_{\mathbb{A}}(\lambda).
\end{equation}
By \eqref{eq:CdecompA} we can easily show that $C_{\mathbb{A}}$ is naturally a Hopf algebra over ${\mathbb{A}}$.
\begin{lemma}
\label{lem:AA}
We have an isomorphism
\[
A_{\mathbb{A}}(\lambda)\cong L^*_{-,{\mathbb{A}}}(\lambda)
\]
of $U_{\mathbb{A}}^L$-modules.
\end{lemma}
\begin{proof}
Note that we have an isomorphism
\[
g_\lambda:L^*_{-,{\mathbb{F}}}(\lambda)\to A_{\mathbb{F}}(\lambda)
\]
of $U_{\mathbb{F}}$-modules given by
\[
\langle g_\lambda(\ell^*),u\rangle
=\langle \ell^*, \overline{Su}\rangle
\qquad(\ell^*\in L^*_{-,{\mathbb{F}}}(\lambda), u\in U_{\mathbb{F}}).
\]
Here, the paring in the right side is the canonical one
$L^*_{-,{\mathbb{F}}}(\lambda)\times
L_{+,{\mathbb{F}}}(-\lambda)\to{\mathbb{F}}$.
We have $g_\lambda(\ell^*)\in A_{\mathbb{A}}(\lambda)$ if and only if $\langle g_\lambda(\ell^*),U^{L}_{\mathbb{A}}\rangle\subset{\mathbb{A}}$.
By
$\langle g_\lambda(\ell^*),U^{L}_{\mathbb{A}}\rangle
=\langle \ell^*, \overline{U^L_{\mathbb{A}}}\rangle
=\langle \ell^*, L_{+,{\mathbb{A}}}(-\lambda)\rangle
$
we obtain
$g_\lambda^{-1}(A_{\mathbb{A}}(\lambda))=L^*_{-,{\mathbb{A}}}(\lambda)$.
\end{proof}
By setting
\[
A_{\mathbb{A}}(\lambda)_\xi
=A_{\mathbb{F}}(\lambda)_\xi\cap A_{\mathbb{A}}
\qquad(\lambda\in\Lambda^+, \xi\in\Lambda)
\]
we have
\begin{equation}
A_{\mathbb{A}}(\lambda)=\bigoplus_{\gamma\in Q^+}
A_{\mathbb{A}}(\lambda)_{\lambda-\gamma}.
\end{equation}
\subsection{}
We set
\[
C_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} C_{\mathbb{A}},\quad
A_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} A_{\mathbb{A}},\quad
A_\zeta(\lambda)={\mathbb{C}}\otimes_{\mathbb{A}} A_{\mathbb{A}}(\lambda)\qquad
(\lambda\in\Lambda^+).
\]
Then $C_\zeta$ is a Hopf algebra over ${\mathbb{C}}$.
Moreover, the $U_{\mathbb{F}}$-bimodule structure of $C_{\mathbb{F}}$ induces a $U_\zeta^L$-bimodule structure of
$C_\zeta$.
Let
\begin{equation}
\label{eq:Hopf-paring2}
\langle\,,\,\rangle:C_\zeta\times U^L_\zeta\to{\mathbb{C}}.
\end{equation}
be the paring induced by \eqref{eq:Hopf-paring}.
We set
\[
(U_\zeta^{L,\pm})^\bigstar=\bigoplus_{\beta\in Q^+}(U_{\zeta,\pm\beta}^{L,\pm})^*\subset
(U_{\zeta}^{L,\pm})^*.
\]
By Lemma \ref{lemma:chi-indep} (ii) we have
\[
(U^{L,-}_{\zeta})^\bigstar\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{C}}\chi_\lambda)\otimes (U^{L,+}_{\zeta})^\bigstar
\subset
(U^{L}_{\zeta})^*.
\]
Moreover, by \eqref{eq:CdecompA}, \eqref{eq:Adecomp1A}, \eqref{eq:Adecomp2A}
we have
\begin{align}
\label{eq:CdecompZ}
C_\zeta&\subset
(U^{L,-}_{\zeta})^\bigstar\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{C}}\chi_\lambda)\otimes (U^{L,+}_{\zeta})^\bigstar,\\
\label{eq:Adecomp1Z}
A_\zeta&\subset
\varepsilon\otimes (\bigoplus_{\lambda\in \Lambda}{\mathbb{C}}\chi_\lambda)\otimes (U^{L,+}_{\zeta})^\bigstar,\\
\label{eq:Adecomp2Z}
A_\zeta(\lambda)&\subset
\varepsilon\otimes \chi_\lambda\otimes (U^{L,+}_{\zeta})^\bigstar.
\end{align}
In particular, we have
\begin{equation}
A_\zeta=\bigoplus_{\lambda\in\Lambda^+}
A_\zeta(\lambda)
\subset C_\zeta\subset(U_\zeta^L)^*.
\end{equation}
Hence we have
\begin{align*}
A_\zeta
&=\{\varphi\in C_\zeta\mid
\varphi\cdot u=\varepsilon(u)\varphi\quad(u\in U_\zeta^{L,-})\},\\
A_\zeta(\lambda)
&=\{\varphi\in A_\zeta\mid
\varphi\cdot h=\chi_\lambda(h)\varphi\quad(h\in U_\zeta^{L,0})\}
\qquad(
\lambda\in\Lambda^+).
\end{align*}
By Andersen-Wen \cite[4.2(2)]{AW}
(see also Andersen-Polo-Wen \cite[1.31 Theorem(iii)]{APW1}) we have the following.
\begin{proposition}
$C_\zeta$ coincides with the subspace of $(U^L_\zeta)^*$ spanned by the matrix coefficients of $U^L_\zeta$-modules belonging to $\mathop{\rm Mod}\nolimits_f(U^L_\zeta)$.
\end{proposition}
By Lemma \ref{lem:AA} we have the following.
\begin{lemma}
\label{lem:Azeta}
For any $\lambda\in\Lambda^+$ we have an isomorphism
\[
A_\zeta(\lambda)\cong L^*_{-,\zeta}(\lambda)
\]
of $U_\zeta^L$-modules.
\end{lemma}
\begin{lemma}
\label{lem:AAtoA}
For $\lambda,\mu\in\Lambda^+$ the canonical map
\begin{equation}
\label{eq:AAtoA}
A_\zeta(\lambda)\otimes A_\zeta(\mu)\to A_\zeta(\lambda+\mu)
\end{equation}
induced by the multiplication of $A_\zeta$ is surjective.
\end{lemma}
\begin{proof}
Note that \eqref{eq:AAtoA} is a homomorphism of $U^L_\zeta$-modules.
Hence by Lemma \ref{lem:Azeta} we have only to show that the unique (up to scalar) homomorphism $L_{-,\zeta}(\lambda+\mu)\to
L_{-,\zeta}(\lambda)\otimes
L_{-,\zeta}(\mu)$ of $U^L_\zeta$-modules is injective.
For that it it sufficient to show that any non-trivial homomorphism
$L_{-,{\mathbb{A}}}(\lambda+\mu)\to
L_{-,{\mathbb{A}}}(\lambda)\otimes
L_{-,{\mathbb{A}}}(\mu)$
of
$U_{\mathbb{A}}^L$-modules which maps $L_{-,{\mathbb{A}}}(\lambda+\mu)_{\lambda+\mu}$ onto $L_{-,{\mathbb{A}}}(\lambda)_\lambda\otimes
L_{-,{\mathbb{A}}}(\mu)_\mu$
is a split monomorphism of ${\mathbb{A}}$-modules.
This follows from \cite[Chapter 27]{Lbook}.
\end{proof}
\begin{lemma}
\eqref{eq:Hopf-paring2} is a Hopf paring.
\end{lemma}
\begin{proof}
It is sufficient to show that the canonical map
$U_\zeta^L\to(C_\zeta)^*$ is injective.
Hence we have only to show that if $u\in U^L_\zeta$
satisfies
$u\mapsto0$ under $U^L_\zeta\to
\mathop{\rm{End}}\nolimits_{\mathbb{C}}(L_{-,\zeta}(\lambda)\otimes L_{+,\zeta}(-\mu))$ for any $\lambda, \mu\in\Lambda^+$, then $u=0$.
This can be proved as in \cite[5.11]{Jan}.
Details are omitted.
\end{proof}
\subsection{}
Assume that we are given a homomorphism $\iota:A\to B$ of $\Lambda$-graded rings satisfying
\begin{equation}
\label{eq:cond-AK}
\iota(A(\lambda))B(\mu)=B(\mu)\iota(A(\lambda))\qquad
(\lambda, \mu\in\Lambda).
\end{equation}
For $M\in\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ let ${\rm{Tor}}(M)$ be the subset of $M$ consisting of $m\in M$
such that for any $m\in M$ there exists $\lambda\in\Lambda^+$ satisfying $\iota(A(\lambda+\mu))m=\{0\}$ for any $\mu\in\Lambda^+$.
Then ${\rm{Tor}}(M)$ is a subobject of $M$ in $\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ by \eqref{eq:cond-AK}.
We denote by ${\rm{Tor}}_{\Lambda^+}(A,B)$ the full subcategory of $\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ consisting of $M\in\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ such that ${\rm{Tor}}(M)=M$.
Note that ${\rm{Tor}}_{\Lambda^+}(A,B)$ is closed under taking subquotients and extensions in $\mathop{\rm Mod}\nolimits_{\Lambda}(B)$.
Let $\Sigma(A,B)$ denote the collection of morphisms $f$ of $\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ such that its kernel $\mathop{\rm Ker\hskip.5pt}\nolimits(f)$ and its cokernel $\mathop{\rm Coker}\nolimits(f)$ belong to ${\rm{Tor}}_{\Lambda^+}(A,B)$.
Then we define an abelian category ${\mathcal C}(A,B)=
\mathop{\rm Mod}\nolimits_\Lambda(B)/{\rm{Tor}}_{\Lambda^+}(A,B)$ as
the localization
\[
{\mathcal C}(A,B)
=\Sigma(A,B)^{-1}\mathop{\rm Mod}\nolimits_{\Lambda}(B)
\]
of $\mathop{\rm Mod}\nolimits_{\Lambda}(B)$ with respect to the multiplicative system $\Sigma(A,B)$
(see \cite{GZ}, \cite{Popescu} for the notion of localization of categories).
We denote by
\begin{equation}
\label{eq:omega1}
\omega(A,B)^*:\mathop{\rm Mod}\nolimits_{\Lambda}(B)\to
{\mathcal C}(A,B)
\end{equation}
the canonical exact functor.
It
admits a right adjoint
\begin{equation}
\label{eq:omega2}
\omega(A,B)_*:{\mathcal C}(A,B)
\to
\mathop{\rm Mod}\nolimits_{\Lambda}(B),
\end{equation}
which is left exact.
It is known that $\omega(A,B)^*\circ\omega(A,B)_*\cong \mathop{\rm Id}\nolimits$.
By taking the degree zero part of \eqref{eq:omega2} we obtain a left exact functor
\begin{equation}
\label{eq:GammaAK}
\Gamma_{(A,B)}:{\mathcal C}(A,B)
\to
\mathop{\rm Mod}\nolimits(B(0)).
\end{equation}
The abelian category ${\mathcal C}(A,B)$ has enough injectives, and we have the right derived functors
\begin{equation}
\label{eq:RGamma}
R^i\Gamma_{(A,B)}:{\mathcal C}(A,B)
\to
\mathop{\rm Mod}\nolimits(B(0))
\qquad(i\in{\mathbb{Z}})
\end{equation}
of \eqref{eq:GammaAK}.
We apply the above arguments to the case $A=B=A_\zeta$.
By Lemma \ref{lem:AAtoA}
${\rm{Tor}}(M)$ for $M\in\mathop{\rm Mod}\nolimits_{\Lambda}(A_\zeta)$ consists of
$m\in M$ such that there exists $\lambda\in\Lambda^+$ satisfying $A_\zeta(\lambda)m=\{0\}$.
We set
\begin{equation}
\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})=
{\mathcal C}(A_\zeta,A_\zeta).
\end{equation}
In this case the natural functors \eqref{eq:omega1}, \eqref{eq:omega2}, \eqref{eq:GammaAK} are simply denoted as
\begin{align}
\label{eq:omega1A}
\omega^*&:\mathop{\rm Mod}\nolimits_{\Lambda}(A_\zeta)\to
\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta}),\\
\omega_*&:\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})
\to
\mathop{\rm Mod}\nolimits_{\Lambda}(A_\zeta),\\
\Gamma&:\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})
\to
\mathop{\rm Mod}\nolimits({\mathbb{C}}).
\end{align}
\begin{remark}
{\rm
In the terminology of non-commutative algebraic geometry
$\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})$ is the category of ``quasi-coherent sheaves'' on the quantized flag manifold ${\mathcal B}_\zeta$, which is a ``non-commutative projective scheme''.
The notations ${{\mathcal B}_\zeta}$, ${\mathcal O}_{{\mathcal B}_\zeta}$ have only symbolical meaning.
}
\end{remark}
\subsection{}
Using Lusztig's Frobenius homomorphism \eqref{eq:LFrobenius} we will relate the quantized flag manifold ${\mathcal B}_\zeta$ with the ordinary flag manifold ${\mathcal B}=B^-\backslash G$.
Taking the dual Hopf algebras in \eqref{eq:LFrobenius} we obtain an injective homomorphism
${\mathbb{C}}[G]\to C_\zeta$
of Hopf algebras.
Moreover, its image is contained in the center of $C_\zeta$ (see Lusztig \cite{L2}).
We will regard ${\mathbb{C}}[G]$ as a central Hopf subalgebra of $C_\zeta$ in the following.
Setting
\begin{align*}
A_1=&\{\varphi\in{\mathbb{C}}[G]\mid
\varphi(ng)=\varphi(g)\,\,(n\in N^-,\,g\in G)\},\\
A_1(\lambda)
=&\{\varphi\in A_1\mid
\varphi(tg)=\theta_\lambda(t)\varphi(g)\,\,(t\in H,\,g\in G)\}
\qquad(\lambda\in\Lambda^+)
\end{align*}
we have a $\Lambda$-graded algebra
\[
A_1=\bigoplus_{\lambda\in\Lambda^+}A_1(\lambda).
\]
We have a left $G$-module structure of $A_1$ given by
\[
(x\varphi)(g)=\varphi(gx)
\qquad(\varphi\in A_1, x, g\in G).
\]
In particular, $A_1$ is a $U({\mathfrak{g}})$-module.
Moreover, for each $\lambda\in\Lambda^+$, $A_1(\lambda)$ is a $U({\mathfrak{g}})$-submodule of $A_1$ which is an irreducible highest weight module with highest weight $\lambda$.
For $\lambda\in\Lambda^+$ and $\xi\in\Lambda$ we set
\[
A_1(\lambda)_\xi
=
\{\varphi\in A_1(\lambda)\mid h\varphi=\xi(h)\varphi\,\,(h\in{\mathfrak{h}})\}.
\]
For $\lambda,\mu\in\Lambda^+$ the canonical map
\[
A_1(\lambda)\otimes A_1(\mu)\to A_1(\lambda+\mu)
\]
induced by the multiplication of $A_1$ is surjective
since it is a non-trivial homomorphism of $U({\mathfrak{g}})$-modules into an irreducible module.
Regarding ${\mathbb{C}}[G]$ as a subalgebra of $C_\zeta$ we have
\[
A_1=A_\zeta\cap{\mathbb{C}}[G],\qquad
A_1(\lambda)=A_\zeta(\ell\lambda)\cap{\mathbb{C}}[G],
\]
\begin{proposition}
\label{prop:AzetaoverA1}
$A_\zeta$ is finitely generated as an $A_1$-module.
\end{proposition}
\begin{proof}
For $i\in I$ set
\[
A_{\zeta, i}
=\bigoplus_{n\geqq0}A_\zeta(n\varpi_i)\subset
A_\zeta,\qquad
A_{1, i}
=\bigoplus_{n\geqq0}A_1(n\varpi_i)\subset A_1.
\]
Then the natural maps
\[
\bigotimes_{i\in I}A_{\zeta, i}\to A_\zeta,
\qquad
\bigotimes_{i\in I}A_{1, i}\to A_1
\]
induced by the multiplications of $A_\zeta$ and $A_1$ respectively are surjective.
Here the tensor product is defined with respect to some fixed ordering of $I$.
Since $A_1$ is a central subalgebra of $A_\zeta$, it is sufficient to show that $A_{\zeta,i}$ is a finitely generated $A_{1,i}$-module for any $i\in I$.
Set $A_{\zeta,i}^{(\ell)}=
\bigoplus_{n\geqq0}A_\zeta(\ell n\varpi_i)$.
Since $A_{\zeta,i}$ is generated by
$\bigoplus_{n<\ell}A_\zeta(n\varpi_i)$
as an $A_{\zeta,i}^{(\ell)}$-module,
it is sufficient to show that $A_{\zeta,i}^{(\ell)}$ is a finitely generated $A_{1,i}$-module.
Assume we could show
\begin{equation}
\label{eq:formula-i}
A_1(\varpi_i)A_\zeta(\ell m\varpi_i)=A_\zeta(\ell(m+1)\varpi_i)
\end{equation}
for some $m>0$.
Then for any $n\geqq m$ we have
\begin{align*}
&A_1(\varpi_i)A_\zeta(\ell n\varpi_i)
=A_1(\varpi_i)A_\zeta(\ell m\varpi_i)A_\zeta(\ell (n-m)\varpi_i)\\
=&A_\zeta(\ell(m+1)\varpi_i)A_\zeta(\ell (n-m)\varpi_i)
=A_\zeta(\ell(n+1)\varpi_i),
\end{align*}
and hence
\[
A_1(k\varpi_i)A_\zeta(m\ell\varpi_i)
=A_\zeta(\ell(m+k)\varpi_i).
\]
This implies that $A_{\zeta,i}^{(\ell)}$ is generated by
$\bigoplus_{n=0}^mA_\zeta(\ell n\varpi_i)$.
Hence it is sufficient to show \eqref{eq:formula-i}.
Set $\Delta^+_\sharp=\Delta^+\cap \sum_{j\ne i}{\mathbb{Z}}\alpha_j$ and
${\mathfrak{p}}_\sharp={\mathfrak{b}}^-\oplus\bigoplus_{\alpha\in\Delta^+_\sharp}{\mathfrak{g}}_\alpha$.
We denote by $P_\sharp$ the parabolic subgroup of $G$ with Lie algebra ${\mathfrak{p}}_\sharp$.
We also denote by $U_\zeta^{L,\sharp}$ the subalgebra of $U_\zeta^L$ generated by $U_\zeta^{L,\leqq0}$ and $e_j^{(n)}$ for $j\ne i, n\geqq0$.
We define a Hopf algebra $C_\zeta^\sharp$ as the image of the composite of $C_\zeta\to \mathop{\rm Hom}\nolimits_{\mathbb{C}}(U_\zeta^{L},{\mathbb{C}})\to\mathop{\rm Hom}\nolimits_{\mathbb{C}}(U_\zeta^{L,\sharp},{\mathbb{C}})$.
For a Hopf algebra $H$ we denote by $\mathop{\rm Comod}\nolimits(H)$ the category of right $H$-comodules.
We have functors
\begin{align*}
&r:\mathop{\rm Comod}\nolimits(C_\zeta^\sharp)\to\mathop{\rm Comod}\nolimits(C_\zeta^{\leqq0}),\\
&r':\mathop{\rm Comod}\nolimits({\mathbb{C}}[P_\sharp])\to\mathop{\rm Comod}\nolimits(C_\zeta^{\leqq0})
\end{align*}
such that $r(N)=N$ and $r'(N')=N'$ as vector spaces and the left
$U_\zeta^{L,\leqq0}$-actions on $r(N)$ and $r'(N')$ are given by the algebra homomorphisms
\[
U_\zeta^{L,\leqq0}
\to U_\zeta^{L,\sharp},\qquad
U_\zeta^{L,\leqq0}
\to U_\zeta^{L,\sharp}\to U({\mathfrak{p}}^-)
\]
respectively.
Let $m>0$.
Denote by $M$ the kernel of the homomorphism
$A_1(\varpi_i)\to{\mathbb{C}}_{\varpi_i}$ of $U({\mathfrak{p}}_\sharp)$-modules.
Then we have an exact sequence
\[
0\to r'(M\otimes{\mathbb{C}}_{m\varpi_i})
\to r'(A_1(\varpi_i))\otimes{\mathbb{C}}_{\ell m\varpi_i}
\to{\mathbb{C}}_{\ell(m+1)\varpi_i}\to0
\]
of right $C_\zeta^{\leqq0}$-comodules.
Applying the induction functor
\[
\mathop{\rm Ind}\nolimits:\mathop{\rm Comod}\nolimits(C_\zeta^{\leqq0})\to\mathop{\rm Comod}\nolimits(C_\zeta)
\]
(see \cite{APW1}) to this exact sequence
we obtain an exact sequence
\[
A_1(\varpi_i)\otimes A_\zeta(\ell m\varpi_i)
\to
A_\zeta(\ell(m+1)\varpi_i)
\to
R^1\mathop{\rm Ind}\nolimits(r'(M\otimes{\mathbb{C}}_{m\varpi_i}))
\]
of right $C_\zeta$-comodules.
By
\[
\dim\mathop{\rm Hom}\nolimits_{U^L_\zeta}(A_1(\varpi_i)\otimes A_\zeta(\ell m\varpi_i),
A_\zeta(\ell(m+1)\varpi_i))=1
\]
the map $A_1(\varpi_i)\otimes A_\zeta(\ell m\varpi_i)
\to
A_\zeta(\ell(m+1)\varpi_i)$
in the above exact sequence coincides with the one given by the multiplication in $A_\zeta$ up to a non-zero constant multiple.
Hence it is sufficient to show that for any finite-dimensional right ${\mathbb{C}}[P_\sharp]$-comodule $N$ there exists some $m>0$ such that $R^1\mathop{\rm Ind}\nolimits(r'(N\otimes{\mathbb{C}}_{m\varpi_i}))=0$.
We may assume that $N$ is irreducible.
Hence it is sufficient to show $R^1\mathop{\rm Ind}\nolimits(r'(N_1))=0$ for the irreducible $U({\mathfrak{p}}_\sharp)$-module $N_1$ with highest weight $\lambda\in\Lambda^+$.
Note that we have natural induction functors
\[
\mathop{\rm Ind}\nolimits_1:\mathop{\rm Comod}\nolimits(C_\zeta^{\leqq0})\to\mathop{\rm Comod}\nolimits(C_\zeta^\sharp),\quad
\mathop{\rm Ind}\nolimits_2:\mathop{\rm Comod}\nolimits(C_\zeta^{\sharp})\to\mathop{\rm Comod}\nolimits(C_\zeta)
\]
such that $\mathop{\rm Ind}\nolimits=\mathop{\rm Ind}\nolimits_2\circ\mathop{\rm Ind}\nolimits_1$ (see \cite{APW1}).
By the Frobenius splitting theorem of Kumar-Littelmann \cite{KR}
$r'(N_1)$ is a direct summand of $r(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda}))$.
Hence it is sufficient to show
$R^1\mathop{\rm Ind}\nolimits(r(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda}))=0$.
By a standard fact on induction functors
we have $\mathop{\rm Ind}\nolimits_1(r(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda})))=\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda})$ and $R^i\mathop{\rm Ind}\nolimits_1(r(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda})))=0$ for $i>0$.
Moreover, $R^i\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda})=0$ for $i>0$ by (a relative version of) the Kempf type vanishing theorem
(\cite{R-H}, \cite{W}, see also \cite{APW1}, \cite{AW}, \cite{Ka1}, \cite{Ka2}).
Hence we obtain
\[
R^1\mathop{\rm Ind}\nolimits(r(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda})))
=R^1\mathop{\rm Ind}\nolimits_2(\mathop{\rm Ind}\nolimits_1({\mathbb{C}}_{\ell\lambda}))
=R^1\mathop{\rm Ind}\nolimits({\mathbb{C}}_{\ell\lambda})=0
\]
again by the Kempf type vanishing theorem.
\end{proof}
Since $A_1$ is a noetherian ring, we obtain
from Proposition \ref{prop:AzetaoverA1} the following.
\begin{proposition}
\label{prop:A-noether}
$A_\zeta$ is a left and right noetherian ring.
\end{proposition}
Note that the $\Lambda$-graded algebra $A_1$ is the
homogeneous coordinate algebra of the projective variety ${\mathcal B}=B^-\backslash G$.
Hence we have an identification
\begin{equation}
\label{eq:A1A1}
\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})={\mathcal C}(A_1,A_1)
\end{equation}
of abelian categories, where $\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})$ denotes the category of quasi-coherent ${\mathcal O}_{\mathcal B}$-modules on the ordinary flag manifold ${\mathcal B}$.
We set
\begin{equation}
\omega_{{\mathcal B}*}=\omega(A_1,A_1)_*:\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})\to\mathop{\rm Mod}\nolimits_\Lambda(A_1).
\end{equation}
For $\lambda\in\Lambda$ we denote by ${\mathcal O}_{\mathcal B}(\lambda)\in \mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})$ the invertible $G$-equivariant ${\mathcal O}_{\mathcal B}$-module corresponding to $\lambda$.
Then under the identification \eqref{eq:A1A1} we have
\[
\omega_{{\mathcal B}*}M=
\bigoplus_{\lambda\in\Lambda}\Gamma({\mathcal B},M\otimes_{{\mathcal O}_{\mathcal B}}{\mathcal O}_{\mathcal B}(\lambda))
\qquad(M\in\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})),
\]
where $\Gamma({\mathcal B},\,\,):\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}})\to{\mathbb{C}}$ is the global section functor for the algebraic variety ${\mathcal B}$.
In particular, the functor $\Gamma_{(A_1,A_1)}:\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})\to\mathop{\rm Mod}\nolimits({\mathbb{C}})$ is identified with
$\Gamma({\mathcal B},\,\,)$.
For $w\in W$
we set
\[
\Theta_{w}=\bigcup_{\lambda\in\Lambda^+}
(A_1(\lambda)_{w^{-1}\lambda}\setminus\{0\})
\subset A_1.
\]
Then $\Theta_{w}$ is a multiplicative subset of the commutative ring $A_1$, and
the localization $\Theta_{w}^{-1}A_1$ turns out to be a $\Lambda$-graded ${\mathbb{C}}$-algebra.
Moreover,
the ${\mathbb{C}}$-algebra $(\Theta_{w}^{-1}A_1)(0)$ is naturally regarded as the coordinate algebra of the affine open subset ${\mathcal B}_w:=B^-\backslash B^-N^+w$ of ${\mathcal B}$.
We denote by $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_w})$ the category of quasi-coherent ${\mathcal O}_{{\mathcal B}_w}$-modules.
We have $\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_w})=\mathop{\rm Mod}\nolimits((\Theta_{w}^{-1}A_1)(0))$.
The functor
\[
j_{w}^*:\mathop{\rm Mod}\nolimits({\mathcal O}_{\mathcal B})\to\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_w})
\]
induced by
\[
\mathop{\rm Mod}\nolimits_\Lambda(A_1)\to\mathop{\rm Mod}\nolimits((\Theta_{w}^{-1}A_1)(0))
\qquad
(M\mapsto (\Theta_{w}^{-1}A_1\otimes_{A_1}M)(0))
\]
is nothing but the inverse image functor with respect to the embedding $j_{w}:{\mathcal B}_w\to{\mathcal B}$.
\subsection{}
For a $\Lambda$-graded ${\mathbb{C}}$-algebra $B$ we define a new $\Lambda$-graded ${\mathbb{C}}$-algebra $B^{(\ell)}$ by
\[
B^{(\ell)}(\lambda)=B(\ell\lambda)
\qquad(\lambda\in\Lambda).
\]
Let
\begin{equation}
\label{eq:(ell)}
(\,\,)^{(\ell)}:\mathop{\rm Mod}\nolimits_\Lambda(B)\to\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})
\end{equation}
be the exact functor given by
\[
M^{(\ell)}(\lambda)=M(\ell\lambda)\qquad
(\lambda\in\Lambda)
\]
for $M\in\mathop{\rm Mod}\nolimits_\Lambda(B)$.
\begin{lemma}
\label{lem:Fr1}
Let $B$ be a $\Lambda$-graded ${\mathbb{C}}$-algebra.
Assume that we are given a homomorphism $\iota:A_\zeta\to B$ of $\Lambda$-graded ${\mathbb{C}}$-algebras.
We denote by $\iota':A_1\to B^{(\ell)}$ the induced homomorphism of $\Lambda$-graded ${\mathbb{C}}$-algebras.
Assume
\begin{align*}
\iota(A_\zeta(\lambda)) B(\mu)&=B(\mu)\iota(A_\zeta(\lambda))
\qquad(\lambda,\mu\in\Lambda),\\
\iota'(A_1(\lambda))B^{(\ell)}(\mu)
&=B^{(\ell)}(\mu)\iota'(A_1(\lambda))
\qquad(\lambda,\mu\in\Lambda).
\end{align*}
Then the exact functor
\begin{equation*}
(\,\,)^{(\ell)}:\mathop{\rm Mod}\nolimits_\Lambda(B)\to\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})
\end{equation*} induces an equivalence
\begin{equation}
Fr_*:{\mathcal C}(A_\zeta,B)
\to
{\mathcal C}(A_1,B^{(\ell)})
\end{equation}
of abelian categories.
Moreover, we have
\begin{equation}
\label{eq:Frobenius}
\omega(A_1,B^{(\ell)})_{*}\circ Fr_*=(\,\,)^{(\ell)}\circ\omega(A_\zeta,B)_{*}.
\end{equation}
\end{lemma}
\begin{proof}
For simplicity we write
$\omega(A_\zeta,B)^{*}$,
$\omega(A_1,B^{(\ell)})^{*}$,
$\omega(A_\zeta,B)_{*}$,
$\omega(A_1,B^{(\ell)})_{*}$
as
$\omega_1^{*}$,
$\omega_2^{*}$,
$\omega_{1*}$,
$\omega_{2*}$
respectively.
By Proposition \ref{prop:AzetaoverA1} we see easily that
for any $\lambda\in\Lambda^+$ there exists some $\mu\in\Lambda^+$ such that $A_\zeta(\nu)A_1(\lambda)=A_\zeta(\ell\lambda+\nu)$ for $\nu\in\mu+\Lambda^+$.
From this we obtain
\begin{equation}
\label{eq:Fr1}
({\rm{Tor}}(M))^{(\ell)}={\rm{Tor}}(M^{(\ell)})\qquad(M\in\mathop{\rm Mod}\nolimits_\Lambda(B)).
\end{equation}
Hence $M\in {\rm{Tor}}_{\Lambda^+}(A_\zeta,B)$ implies $M^{(\ell)}\in{\rm{Tor}}_{\Lambda^+}(A_1,B^{(\ell)})$.
It follows that we have a well-defined functor
$Fr_*:{\mathcal C}(A_\zeta,B)
\to
{\mathcal C}(A_1,B^{(\ell)})$
satisfying $Fr_*\circ\omega_1^*=\omega_2^*\circ (\,\,)^{(\ell)}$.
We see easily that
\begin{equation*}
\label{eq:Fr2}
(B\otimes_{B^{(\ell)}}N)^{(\ell)}\cong N
\qquad(N\in\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})).
\end{equation*}
Hence we have
\[
(Fr_*\circ\omega_1^*)(B\otimes_{B^{(\ell)}}N)
=
\omega_2^*((B\otimes_{B^{(\ell)}}N)^{(\ell)})
=
\omega_2^*(N)
\]
for any $N\in\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})$.
It follows that
$Fr_*$ is a dense functor.
Let us show that
\[
\mathop{\rm Hom}\nolimits(\omega_1^*M,\omega_1^*N)
\to
\mathop{\rm Hom}\nolimits(\omega_2^*(M^{(\ell)}), \omega_2^*(N^{(\ell)}))
\]
is bijective for any $M, N\in\mathop{\rm Mod}\nolimits_\Lambda(B)$.
By $(B\otimes_{B^{(\ell)}}M^{(\ell)})^{(\ell)}\cong M^{(\ell)}$
we see easily that the canonical morphism
$B\otimes_{B^{(\ell)}}M^{(\ell)}\to M$
belongs to $\Sigma(A_\zeta,B)$, that is,
$\omega_1^*(B\otimes_{B^{(\ell)}}M^{(\ell)})\cong\omega_1^* M$.
Hence we have
\begin{align*}
&\mathop{\rm Hom}\nolimits(\omega_1^*M,\omega_1^*N)\cong
\mathop{\rm Hom}\nolimits(\omega_1^*(B\otimes_{B^{(\ell)}}M^{(\ell)}),\omega_1^*N)\\
\cong&
\mathop{\rm Hom}\nolimits(B\otimes_{B^{(\ell)}}M^{(\ell)},\omega_{1*}\omega_1^*N)
\cong
\mathop{\rm Hom}\nolimits(M^{(\ell)},(\omega_{1*}\omega_1^*N)^{(\ell)}).
\end{align*}
On the other hand we have
\begin{align*}
&\mathop{\rm Hom}\nolimits(\omega_2^*(M^{(\ell)}), \omega_2^*(N^{(\ell)}))
\cong
\mathop{\rm Hom}\nolimits(M^{(\ell)}, \omega_{2*}\omega_2^*(N^{(\ell)})).
\end{align*}
Therefore, it is sufficient to show
\begin{equation}
\label{eq:lem:Frr}
(\omega_{1*}\omega_1^*N)^{(\ell)}
\cong
\omega_{2*}\omega_2^*(N^{(\ell)})
\end{equation}
(note that \eqref{eq:lem:Frr} is equivalent to \eqref{eq:Frobenius}).
We may assume that $N=B\otimes_{B^{(\ell)}}P$ for some $P\in\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})$.
We may further assume that $\omega_{2*}\omega_2^*P\cong P$.
Hence it is sufficient to show for $P\in\mathop{\rm Mod}\nolimits_\Lambda(B^{(\ell)})$ satisfying $\omega_{2*}\omega_2^*P\cong P$ that
$P\cong (\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P))^{(\ell)}$.
Since the canonical morphism
$B\otimes_{B^{(\ell)}}P\to
\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P)$
belongs to $\Sigma(A_\zeta,B)$,
the corresponding morphism
$f:P\to
(\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P))^{(\ell)}$
belongs to $\Sigma(A_1,B^{(\ell)})$.
By $\omega_{2*}\omega_2^*P\cong P$ we see that $f$ is injective and its cokernel is isomorphic to a submodule of
$(\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P))^{(\ell)}$.
By
\[
{\rm{Tor}}((\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P))^{(\ell)})
=({\rm{Tor}}(\omega_{1*}\omega_1^*(B\otimes_{B^{(\ell)}}P)))^{(\ell)}=0
\]
we obtain $\mathop{\rm Coker}\nolimits(f)=0$.
It follows that $f$ is an isomorphism.
\end{proof}
The following fact is concerned with ordinary (commutative) projective algebraic geometry and its proof is straightforward. Details are omitted.
\begin{lemma}
\label{lem:Fr2}
Let $F$ be a $\Lambda$-graded ${\mathbb{C}}$-algebra, and let $A_1\to F$ be a homomorphism of $\Lambda$-graded ${\mathbb{C}}$-algebras.
Assume that $\mathop{\rm Im}\nolimits(A_1\to F)$ is central in $F$.
Regard $F$ as an object of $\mathop{\rm Mod}\nolimits_\Lambda(A_1)$ and consider
$\omega_{\mathcal B}^*F\in\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}})$.
Then the multiplication of $F$ induces an ${\mathcal O}_{\mathcal B}$-algebra structure of $\omega_{\mathcal B}^*F$, and we have an identification
\begin{equation}
\label{eq:lem:Fr2}
{\mathcal C}(A_1,F)
=
\mathop{\rm Mod}\nolimits(\omega_{\mathcal B}^*F),
\end{equation}
of abelian categories,
where $\mathop{\rm Mod}\nolimits(\omega_{\mathcal B}^*F)$ denotes the category of quasi-coherent $\omega_{\mathcal B}^*F$-modules.
Moreover, under the identification \eqref{eq:lem:Fr2}
we have
\[
\Gamma_{(A_1,F)}(M)=\Gamma({\mathcal B},M)
\in\mathop{\rm Mod}\nolimits(F(0))
\qquad(M\in\mathop{\rm Mod}\nolimits(\omega_{\mathcal B}^*F)).
\]
\end{lemma}
We define an ${\mathcal O}_{\mathcal B}$-algebra $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$ by
\[
Fr_*{\mathcal O}_{{\mathcal B}_\zeta}=\omega_{\mathcal B}^*(A_\zeta^{(\ell)}).
\]
We denote by $\mathop{\rm Mod}\nolimits(Fr_*{\mathcal O}_{{\mathcal B}_\zeta})$ the category of quasi-coherent $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$-modules.
By Lemma \ref{lem:Fr1} and Lemma \ref{lem:Fr2} we have the following.
\begin{lemma}
We have an equivalence
\[
Fr_*:\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})\to\mathop{\rm Mod}\nolimits(Fr_*{\mathcal O}_{{\mathcal B}_\zeta})
\]
of abelian categories.
Moreover,
for $M\in\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}_\zeta})$ we have
\[
R^i\Gamma(M)\simeq R^i\Gamma({\mathcal B},Fr_*(M)),
\]
where $\Gamma({\mathcal B},\,\,):\mathop{\rm Mod}\nolimits({\mathcal O}_{{\mathcal B}})\to\mathop{\rm Mod}\nolimits({\mathbb{C}})$ in the right side is the global section functor for ${\mathcal B}$.
\end{lemma}
\begin{proposition}
\label{prop:AzetaA1}
$(\Theta_e^{-1}A_\zeta)(0)$ is a free $(\Theta_e^{-1}A_1)(0)$-module of rank $\ell^{|\Delta^+|}$. Here $e$ is the identity element of $W$.
Hence the restriction $j_e^*Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$ of $Fr_*{\mathcal O}_{{\mathcal B}_\zeta}$ to ${\mathcal B}_e=B^-\backslash B^-N^+$ is a free ${\mathcal O}_{{\mathcal B}_e}$-module of rank $\ell^{|\Delta^+|}$.
\end{proposition}
\begin{proof}
Denote by
\[
g:A_\zeta\to(U_\zeta^{L,\geqq0})^*
\]
the composite of
\[
A_\zeta\subset C_\zeta\subset (U_\zeta^L)^*\to (U_\zeta^{L,\geqq0})^*.
\]
Then $g$ is an algebra homomorphism with respect to the multiplication of $(U_\zeta^{L,\geqq0})^*$ induced by the comultiplication of $U_\zeta^{L,\geqq0}$.
Set
\[
(U_\zeta^{L,+})^\bigstar=\bigoplus_{\gamma\in Q^+}
(U_{\zeta,\gamma}^{L,+})^*\subset
(U_{\zeta}^{L,+})^*,
\]
and identify $(U_\zeta^{L,+})^\bigstar$ with a subspace of $(U_\zeta^{L,\geqq0})^*$ by the embedding
$(U_\zeta^{L,+})^\bigstar\ni\varphi\mapsto\tilde{\varphi}\in(U_\zeta^{L,\geqq0})^*$
given by
\[
\tilde{\varphi}(hx)=\varepsilon(h)\varphi(x)\qquad
(h\in U^{L,0}_\zeta, x\in U^{L,+}_\zeta).
\]
For $\lambda\in\Lambda$ define algebra homomorphism $\chi_\lambda:U^{L,\geqq0}_\zeta\to{\mathbb{C}}$ by
\[
\chi_\lambda(hx)=\chi_\lambda(h)\varepsilon(x)\qquad
(h\in U^{L,0}_\zeta, x\in U^{L,+}_\zeta).
\]
Then for $\varphi\in(U_\zeta^{L,+})^\bigstar$ and $\lambda\in\Lambda$ we have
\[
(\tilde{\varphi}\chi_\lambda)(hx)=\chi_\lambda(h)\tilde{\varphi}(x)\qquad
(h\in U^{L,0}_\zeta, x\in U^{L,+}_\zeta).
\]
Moreover, $(U_\zeta^{L,+})^\bigstar$ is a subalgebra of
$(U_\zeta^{L,\geqq0})^*$, and
\begin{align*}
&\chi_\lambda\chi_\mu=\chi_{\lambda+\mu}\qquad
(\lambda, \mu\in\Lambda),\\
&\chi_\lambda\tilde{\varphi}=\zeta^{(\lambda,\gamma)}\tilde{\varphi}\chi_\lambda
\qquad
(\lambda\in\Lambda, \varphi\in (U_{\zeta,\gamma}^{L,+})^*)
\end{align*}
in the algebra $(U_\zeta^{L,\geqq0})^*$.
In particular,
\[
(U_\zeta^{L,\geqq0})^\bigstar
:=\bigoplus_{\lambda\in\Lambda}(U_\zeta^{L,+})^\bigstar
\chi_\lambda
\]
is a subalgebra of $(U_\zeta^{L,\geqq0})^*$.
By \eqref{eq:Adecomp1Z} $g$ induces an injective algebra homomorphism
\[
g':A_\zeta\to(U_\zeta^{L,\geqq0})^\bigstar.
\]
For $\varphi\in A_\zeta(\lambda)_\lambda\setminus\{0\}$ we have $g'(\varphi)\in{\mathbb{C}}\chi_\lambda\setminus\{0\}$, and hence $g'$ induces an injective algebra homomorphism
\[
g'':\Theta_e^{-1}A_\zeta\to(U_\zeta^{L,\geqq0})^\bigstar.
\]
Let us show that $g''$ is surjective.
It is sufficient to show that for any $\gamma\in Q^+$ there exists $\lambda\in\Lambda^+$ such that
$g'(A_\zeta(\lambda)_{\lambda-\gamma})=
(U_{\zeta,\gamma}^{L,+})^*\chi_\lambda$.
This is a consequence of the injectivity of $U_{\zeta,\gamma}^{L,+}\ni u\mapsto \overline{u}\in L_{+,\zeta}(-\lambda)_{-\lambda+\gamma}$ for sufficiently large $\lambda$ (see for example \cite[Lemma 2.1]{T2}).
Hence $g''$ is an isomorphism.
Similarly to the above argument the natural algebra homomorphism
\[
g_1:A_1\to(U({\mathfrak{b}}^+))^*
\]
induces an algebra isomorphism
\[
g_1'':\Theta_e^{-1}A_1\to (U({\mathfrak{b}}^+))^\bigstar,
\]
where
\begin{align*}
&(U({\mathfrak{b}}^+))^\bigstar=\bigoplus_{\lambda\in\Lambda}
(U({\mathfrak{n}}^+))^\bigstar\chi_{1,\lambda},\\
&(U({\mathfrak{n}}^+))^\bigstar
=\bigoplus_{\gamma\in Q^+}(U({\mathfrak{n}}^+)_\gamma)^*,\\
&\chi_{1,\lambda}(hx)=\langle\lambda, h\rangle\varepsilon(x)\qquad
(\lambda\in\Lambda, h\in U({\mathfrak{h}}), x\in U({\mathfrak{n}}^+)).
\end{align*}
Moreover, we have the following commutative diagram
\[
\begin{CD}
\Theta_e^{-1}A_1@>{g_1''}>> (U({\mathfrak{b}}^+))^\bigstar\\
@VVV @VVV\\
\Theta_e^{-1}A_\zeta @>>{g''}> (U_\zeta^{L,\geqq0})^\bigstar,
\end{CD}
\]
where
$\Theta_e^{-1}A_1\to\Theta_e^{-1}A_\zeta$ is the embedding induced from the inclusion $A_1\subset A_\zeta$, and $(U({\mathfrak{b}}^+))^\bigstar\to(U_\zeta^{L,\geqq0})^\bigstar$ is the injective algebra homomorphism induced by the restriction of \eqref{eq:LFrobenius}.
Restricting to the degree zero part we obtain algebra isomorphisms
\[
(\Theta_e^{-1}A_\zeta)(0) \to(U_\zeta^{L,+})^\bigstar,\qquad
(\Theta_e^{-1}A_1)(0)\to (U({\mathfrak{n}}^+))^\bigstar
\]
and the commutative diagram
\[
\begin{CD}
(\Theta_e^{-1}A_1)(0)@>{g_1''}>> (U({\mathfrak{n}}^+))^\bigstar\\
@VVV @VVV\\
(\Theta_e^{-1}A_\zeta)(0) @>>{g''}> (U_\zeta^{L,+})^\bigstar.
\end{CD}
\]
Define a linear map $F:S({U}^-_\zeta)\to(U_\zeta^{L,\geqq0})^\bigstar$ by $(F(y))(z)={}^L\tau(z,y)$ for
$y\in S({U}^-_\zeta)$ and $z\in U_\zeta^{L,\geqq0}$.
Then we see easily that $F$ is an injective algebra homomorphism whose image is $(U_\zeta^{L,+})^\bigstar$.
Hence we can identify the algebra $(U_\zeta^{L,+})^\bigstar$ with $S({U}^-_\zeta)$.
Under this identification the image of $(U({\mathfrak{n}}^+))^\bigstar\to(U_\zeta^{L,+})^\bigstar$ coincides with the subalgebra of $S({U}^-_\zeta)$ generated by the central elements $S(f_{\beta_j}^\ell)\,\,(j=1,\dots, N)$.
Hence our assertion is clear from Lemma \ref{lem:PBW-DK}.
\end{proof}
\section{Ring of differential operators}
\subsection{}
We define a subalgebra $D_{\mathbb{F}}$ of $\mathop{\rm{End}}\nolimits_{\mathbb{F}}(A_{\mathbb{F}})$ by
\[
D_{\mathbb{F}}=
\langle
\ell_\varphi, r_\varphi, \partial_u, \sigma_\lambda\mid
\varphi\in A_{\mathbb{F}}, u\in U_{\mathbb{F}}, \lambda\in\Lambda\rangle,
\]
where
\[
\ell_\varphi(\psi)=\varphi\psi,\quad
r_\varphi(\psi)=\psi\varphi,\quad
\partial_u(\psi)=u\cdot\psi,\quad
\sigma_\lambda(\psi)=q^{(\lambda,\mu)}\psi
\]
for $\psi\in A_{\mathbb{F}}(\mu)$.
We have a natural grading
\begin{align*}
&D_{\mathbb{F}}=\bigoplus_{\lambda\in\Lambda^+}D_{\mathbb{F}}(\lambda),\\
&D_{\mathbb{F}}(\lambda)=\{\Phi\in D_{\mathbb{F}}\mid
\Phi(A_{\mathbb{F}}(\mu))\subset A_{\mathbb{F}}(\lambda+\mu)\quad(\mu\in\Lambda)\}\qquad(\lambda\in\Lambda)
\end{align*}
of $D_{\mathbb{F}}$.
We have
\begin{align*}
\partial_u\ell_\varphi=&\sum_{(u)}\ell_{u_{(0)}\cdot\varphi}\partial_{u_{(1)}}
\qquad(u\in U_{\mathbb{F}}, \varphi\in A_{\mathbb{F}}),\\
\partial_u\sigma_\lambda=&\sigma_\lambda\partial_u
\qquad(u\in U_{\mathbb{F}}, \lambda\in\Lambda),\\
\sigma_\lambda\ell_\varphi=&
q^{(\lambda,\mu)}\ell_\varphi\sigma_\lambda
\qquad(\lambda\in\Lambda, \varphi\in A_{\mathbb{F}}(\mu)).
\end{align*}
We have also
\begin{equation}
\label{eq:Har-center-in-D0}
z\in Z(U_{\mathbb{F}}),\,\,
\iota(z)=\sum_{\lambda\in\Lambda}a_\lambda e({2\lambda})\quad\Longrightarrow\quad
\partial_z
=\sum_{\lambda\in\Lambda}a_\lambda\sigma_{2\lambda}.
\end{equation}
We take bases $\{x_p\}_p$ and $\{y_p\}_p$ of $U_{\mathbb{F}}^+$ and $U_{\mathbb{F}}^-$ respectively and elements $\beta_p\in Q^+$ for each $p$ such that
\begin{align}
\label{eq:xy1}
&x_p\in U_{{\mathbb{F}},\beta_p}^+,\qquad
y_p\in U_{{\mathbb{F}},-\beta_p}^-,\\
\label{eq:xy2}
&\tau(x_{p_1},y_{p_2})=\delta_{p_1,p_2}.
\end{align}
\begin{lemma}
\label{lem:rvarphi}
Let $\lambda\in\Lambda^+$ and $\xi\in \Lambda$.
For $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$ we have
\begin{equation}
\label{eq:rvarphi}
r_\varphi=
\sum_p\ell_{{y_p}\cdot\varphi}\partial_{x_pk_{-\xi}}\sigma_\lambda
=\sum_p\ell_{(Sx_p)\cdot\varphi}\partial_{y_pk_{\beta_p}k_\xi}\sigma_{-\lambda}.
\end{equation}
\end{lemma}
\begin{proof}
The first equality is shown in \cite[Lemma 5.1]{T2} by the following argument.
Let ${\mathcal R}\in U_{\mathbb{F}}\hat{\otimes}U_{\mathbb{F}}$ be the universal $R$-matrix.
Then we have
\begin{align*}
&\langle r_\varphi(\psi),u\rangle=
\langle \psi\varphi,u\rangle=
\langle \psi\otimes\varphi,\Delta(u)\rangle=
\langle \varphi\otimes\psi,\tau(\Delta(u))\rangle\\
=&
\langle \varphi\otimes\psi,{\mathcal R}\Delta(u){\mathcal R}^{-1}\rangle=
\langle {\mathcal R}^{-1}\cdot(\varphi\otimes\psi)\cdot{\mathcal R},\Delta(u)\rangle
\\
=&
\langle m({\mathcal R}^{-1}\cdot(\varphi\otimes\psi)\cdot{\mathcal R}),u\rangle
\end{align*}
for $\varphi, \psi\in A_{\mathbb{F}}$ and $u\in U_{\mathbb{F}}$.
Here $\tau:U_{\mathbb{F}}\otimes U_{\mathbb{F}}\to U_{\mathbb{F}}\otimes U_{\mathbb{F}}$ is the linear map sending $a\otimes b$ to $b\otimes a$.
Hence we obtain $r_\varphi(\psi)=m({\mathcal R}^{-1}\cdot(\varphi\otimes\psi)\cdot{\mathcal R})$.
By rewriting it using an explicit form of ${\mathcal R}$ we obtain the first equality in \eqref{eq:rvarphi}.
Applying the same argument to another $R$-matrix $\tau({\mathcal R}^{-1})$ we also obtain the second equality in \eqref{eq:rvarphi}.
Details are omitted.
\end{proof}
Set
\begin{equation}
E_{\mathbb{F}}=A_{\mathbb{F}}\otimes U_{\mathbb{F}}\otimes{\mathbb{F}}[\Lambda],
\end{equation}
and regard $A_{\mathbb{F}}$, $U_{\mathbb{F}}$, ${\mathbb{F}}[\Lambda]$ as subspaces of $E_{\mathbb{F}}$ by the natural embeddings
$A_{\mathbb{F}}\ni\varphi\mapsto\varphi\otimes1\otimes1\in E_{\mathbb{F}}$ e.t.c.
Then we have an ${\mathbb{F}}$-algebra structure of $E_{\mathbb{F}}$ such that
the natural embeddings $A_{\mathbb{F}}\to E_{\mathbb{F}}$, $U_{\mathbb{F}}\to E_{\mathbb{F}}$, ${\mathbb{F}}[\Lambda]\to E_{\mathbb{F}}$ are algebra homomorphisms, and
\begin{align*}
u\varphi=
&\sum_{(u)}
({u_{(0)}\cdot\varphi}){u_{(1)}}
\qquad(u\in U_{\mathbb{F}}, \varphi\in A_{\mathbb{F}}),\\
ue(\lambda)=&e(\lambda) u
\qquad(u\in U_{\mathbb{F}}, \lambda\in\Lambda),\\
e(\lambda)\varphi=&
q^{(\lambda,\mu)}\varphi e(\lambda)
\qquad(\lambda\in\Lambda, \varphi\in A_{\mathbb{F}}(\mu))
\end{align*}
in ${E_{\mathbb{F}}}$.
Moreover, we have a surjective algebra homomorphism $E_{\mathbb{F}}\to D_{\mathbb{F}}$ given by $\varphi\mapsto\ell_\varphi\,(\varphi\in A_{\mathbb{F}})$,
$u\mapsto\partial_u\,(u\in U_{\mathbb{F}})$, $e(\lambda)\mapsto\sigma_\lambda\,(\lambda\in\Lambda)$.
For $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$ with $\lambda\in\Lambda^+$, $\xi\in\Lambda$ we set
\begin{align}
\Omega_1(\varphi)&=
\sum_p({y_p}\cdot\varphi){x_pk_{-\xi}}e(\lambda)
\in E_{\mathbb{F}},\\
\Omega_2(\varphi)&=
\sum_p({(Sx_p)\cdot\varphi}){y_pk_{\beta_p}k_\xi}e({-\lambda})
\in E_{\mathbb{F}},\\
\Omega(\varphi)&=\Omega_1(\varphi)-\Omega_2(\varphi)
\in E_{\mathbb{F}}.
\end{align}
We extend $\Omega$, $\Omega_1$, $\Omega_2$ to whole $A_{\mathbb{F}}$ by linearity.
By Lemma \ref{lem:rvarphi} we have $\Omega(\varphi)\in\mathop{\rm Ker\hskip.5pt}\nolimits(E_{\mathbb{F}}\to D_{\mathbb{F}})$.
We set
\[
D'_{\mathbb{F}}=E_{\mathbb{F}}/
\sum_{\varphi\in A_{\mathbb{F}}}E_{\mathbb{F}}\Omega(\varphi)E_{\mathbb{F}}.
\]
Then we have a sequence of surjective algebra homomorphisms
\begin{equation}
\label{EDD}
E_{\mathbb{F}}\to{D}'_{\mathbb{F}}\to{D}_{\mathbb{F}}.
\end{equation}
\begin{lemma}
\label{lem:Omega}
For $\varphi\in A_{\mathbb{F}}(\lambda)$ with $\lambda\in\Lambda^+$ and $i=1, 2$ we have
\begin{align}
\label{eq:lem:Omega1}
e(\mu)\Omega_i(\varphi)&=
q^{(\lambda,\mu)}\Omega_i(\varphi)e(\mu)
&(\mu\in\Lambda),\\
\label{eq:lem:Omega2}
\psi\Omega_i(\varphi)&=
\Omega_i(\varphi)\psi
&(\psi\in A_{\mathbb{F}}),\\
\label{eq:lem:Omega3}
u\Omega_i(\varphi)&=
\sum_{(u)}\Omega_i(u_{(1)}\cdot\varphi)u_{(0)}
&(u\in U_{\mathbb{F}}),\\
\label{eq:lem:Omega4}
\Omega_i(\varphi\psi)&=
\Omega_i(\psi)\Omega_i(\varphi)
&(\varphi, \psi\in A_{\mathbb{F}})
\end{align}
in $E_{\mathbb{F}}$.
\end{lemma}
\begin{proof}
We will only give a proof for the case $i=1$.
The proof for the case $i=2$ is similar.
The proof of \eqref{eq:lem:Omega1} is easy and omitted.
Let us show \eqref{eq:lem:Omega2}.
Let $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$, $\psi\in A_{\mathbb{F}}(\mu)_\eta$.
By the formula
\begin{equation}
\label{eq:lem:Omega:pf1}
\sum_p \Delta x_p\otimes y_p=
\sum_{p,r}x_pk_{\beta_r}\otimes x_r\otimes y_py_r
\end{equation}
(see \cite[(4.3.16)]{T1}) we have
\begin{align*}
\Omega_1(\varphi)\psi&=
\sum_p({y_p}\cdot\varphi){x_pk_{-\xi}}e(\lambda)\psi\\
&=q^{(\lambda,\mu)-(\xi,\eta)}
\sum_p({y_p}\cdot\varphi)x_p\psi k_{-\xi}e(\lambda)\\
&=q^{(\lambda,\mu)-(\xi,\eta)}
\sum_{p, r}({y_py_r}\cdot\varphi)(x_pk_{\beta_r}\cdot\psi)
x_r k_{-\xi}e(\lambda)\\
&=q^{(\lambda,\mu)-(\xi,\eta)}
\sum_{r}
q^{(\beta_r,\eta)}
\left(
\sum_p({y_py_r}\cdot\varphi)(x_p\cdot\psi)
\right)
x_r k_{-\xi}e(\lambda).
\end{align*}
By Lemma \ref{lem:rvarphi} we have
\begin{align*}
&\sum_p({y_py_r}\cdot\varphi)(x_p\cdot\psi)\\
=&
\sum_p
r_{x_p\cdot\psi}({y_py_r}\cdot\varphi)\\
=&
\sum_{p,s}
\ell_{S(x_s)x_p\cdot\psi}
\partial_{y_sk_{\beta_s+\beta_p+\eta}}\sigma_{-\mu}
({y_py_r}\cdot\varphi)\\
=&\sum_{p,s}
q^{-(\lambda,\mu)+(\beta_s+\beta_p+\eta,\xi-\beta_p-\beta_r)}
(S(x_s)x_p\cdot\psi)
({y_sy_py_r}\cdot\varphi)\\
=&\sum_{p,s}
q^{-(\lambda,\mu)+(\beta_s+\beta_p+\eta,\xi-\beta_r)}
(S(x_sk_{\beta_p})x_p\cdot\psi)
({y_sy_py_r}\cdot\varphi).
\end{align*}
By the formula
\begin{equation}
\label{eq:lem:Omega:pf2}
\sum_{\beta_p+\beta_r=\gamma}
S(x_pk_{\beta_r})x_r\otimes y_py_r
=\begin{cases}
1\otimes1\quad&(\gamma=0)\\
0&(\gamma\ne0),
\end{cases}
\end{equation}
which is a consequence of
\eqref{eq:lem:Omega:pf1} and $m\circ(S\otimes1)\circ\Delta=\varepsilon$,
we obtain
\[
\sum_p({y_py_r}\cdot\varphi)(x_p\cdot\psi)
=
q^{-(\lambda,\mu)+(\eta,\xi-\beta_r)}
\psi
({y_r}\cdot\varphi).
\]
It follows that
\[
\Omega_1(\varphi)\psi
=\psi
\sum_{r}
({y_r}\cdot\varphi)
x_r k_{-\xi}e(\lambda)
=\psi\Omega_1(\varphi).
\]
The formula \eqref{eq:lem:Omega2} is verified.
Let us show \eqref{eq:lem:Omega3}.
It is sufficient to consider the three cases; $u\in U_{\mathbb{F}}^0$, $u\in U_{\mathbb{F}}^-$, $u\in U_{\mathbb{F}}^+$.
Let $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$.
For $u\in U_{\mathbb{F}}$ we have
\[
u\Omega_1(\varphi)
=\sum_{p,(u)}(u_{(0)}y_p\cdot\varphi)u_{(1)}x_pk_{-\xi}e(\lambda)
\]
and
\begin{align*}
&\sum_{(u)}
\Omega_1(u_{(1)}\cdot\varphi)u_{(0)}\\
=&
\sum_{p,(u)}(y_pu_{(1)}\cdot\varphi)x_pk_{-\xi-\mathop{\rm wt}\nolimits(u_{(1)})}e(\lambda)
u_{(0)}\\
=&
\sum_{p,(u)}
q^{-(\xi+\mathop{\rm wt}\nolimits(u_{(1)}),\mathop{\rm wt}\nolimits(u_{(0)}))}
(y_pu_{(1)}\cdot\varphi)x_pu_{(0)}k_{-\xi-\mathop{\rm wt}\nolimits(u_{(1)})}e(\lambda).
\end{align*}
Hence our assertion is equivalent to
\begin{equation}
\label{eq:lem:Omega:pf3}
\begin{split}
&\sum_{p,(u)}(u_{(0)}y_p\cdot\varphi)u_{(1)}x_p\\
=&\sum_{p,(u)}
q^{-(\xi+\mathop{\rm wt}\nolimits(u_{(1)}),\mathop{\rm wt}\nolimits(u_{(0)}))}
(y_pu_{(1)}\cdot\varphi)x_pu_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})}.
\end{split}
\end{equation}
Here, we have used the following notation.
For $u\in U_{\mathbb{F}}$ such that $k_\nu uk_\nu^{-1}=q^{(\nu,\mu)}u$ for any $\nu\in\Lambda$ we write $\mu=\mathop{\rm wt}\nolimits(u)$.
Moreover, in the expansion $\Delta u=\sum_{(u)}u_{(0)}\otimes u_{(1)}$ the elements $u_{(0)}$ and $u_{(1)}$ are taken to be weight vectors.
The proof of \eqref{eq:lem:Omega:pf3} in the case $u\in U_{\mathbb{F}}^0$ is easy and omitted.
Let us consider the case $u\in U_{\mathbb{F}}^-$.
By Lemma \ref{lem:Drinfeld paring} and the formula
\begin{equation}
\label{eq:lem:Omega:pf4}
\sum_p\Delta_2 x_p\otimes y_p
=\sum_{p,r,s}
x_pk_{\beta_r+\beta_s}\otimes
x_rk_{\beta_s}\otimes
x_s\otimes
y_py_ry_s,
\end{equation}
which is a consequence of \eqref{eq:lem:Omega:pf1}
we have
\begin{align*}
&\sum_{p,(u)}(u_{(0)}y_p\cdot\varphi)u_{(1)}x_p\\
=&
\sum_{p,(u)_3}(u_{(0)}y_p\cdot\varphi)
\left(
\sum_{(x_p)_2}
\tau(x_{p(0)},Su_{(1)})
\tau(x_{p(2)},u_{(3)})
x_{p(1)}u_{(2)}
\right)\\
=&
\sum_{p,r,s,(u)_3}
\tau(x_pk_{\beta_r+\beta_s},Su_{(1)})
\tau(x_s,u_{(3)})
(u_{(0)}y_py_ry_s\cdot\varphi)
x_rk_{\beta_s}u_{(2)}\\
=&
\sum_{p,r,s,(u)_3}
q^{(\beta_s+\beta_r,\mathop{\rm wt}\nolimits(u_{(0)})+\mathop{\rm wt}\nolimits(u_{(1)}))}
\tau(x_p,(Su_{(1)})k_{\mathop{\rm wt}\nolimits(u_{(0)})+\mathop{\rm wt}\nolimits(u_{(1)})})\\
&\qquad\qquad\qquad\qquad\qquad\qquad
\tau(x_s,u_{(3)})(u_{(0)}y_py_ry_s\cdot\varphi)
x_rk_{\beta_s}u_{(2)}
\\
=&
\sum_{r,s,(u)_3}
q^{(\beta_s+\beta_r,\mathop{\rm wt}\nolimits(u_{(0)})+\mathop{\rm wt}\nolimits(u_{(1)}))}
\tau(x_s,u_{(3)})\\
&\qquad\qquad\qquad\qquad\qquad
(u_{(0)}(Su_{(1)})k_{\mathop{\rm wt}\nolimits(u_{(0)})+\mathop{\rm wt}\nolimits(u_{(1)})}y_ry_s\cdot\varphi)
x_rk_{\beta_s}u_{(2)}
\\
=&
\sum_{r,s,(u)}
\tau(x_s,u_{(1)})
(y_ry_s\cdot\varphi)
x_rk_{\beta_s}u_{(0)}
\\
=&
\sum_{r,s,(u)}
\tau(x_s,u_{(1)}k_{-\mathop{\rm wt}\nolimits(u_{(0)})})
(y_ry_s\cdot\varphi)
x_rk_{\beta_s}u_{(0)}
\\
=&
\sum_{r,(u)}
(y_ru_{(1)}k_{-\mathop{\rm wt}\nolimits(u_{(0)})}\cdot\varphi)
x_rk_{-\mathop{\rm wt}\nolimits(u_{(1)})}u_{(0)}
\\
=&\sum_{r,(u)}
q^{-(\xi+\mathop{\rm wt}\nolimits(u_{(1)}),\mathop{\rm wt}\nolimits(u_{(0)}))}
(y_ru_{(1)}\cdot\varphi)x_ru_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})}.
\end{align*}
The formula \eqref{eq:lem:Omega:pf3} for $u\in U_{\mathbb{F}}^-$ is shown.
Let us consider the case $u\in U_{\mathbb{F}}^+$.
By Lemma \ref{lem:Drinfeld paring} and the formula
\[
\sum_px_p\otimes \Delta_2 y_p
=\sum_{p,r,s}
x_sx_rx_p\otimes
y_p\otimes
y_rk_{-\beta_p}\otimes
y_sk_{-\beta_p-\beta_r},
\]
which is shown similarly to \eqref{eq:lem:Omega:pf4}
we have
\begin{align*}
&\sum_{p,(u)}(u_{(0)}y_p\cdot\varphi)u_{(1)}x_p\\
=&
\sum_{p,(u)_3}
\sum_{(y_p)_2}
\tau(u_{(0)},y_{p(0)})
\tau(u_{(2)},Sy_{p(2)})
(y_{p(1)}u_{(1)}\cdot\varphi)u_{(3)}x_p\\
=&
\sum_{p,r,s,(u)_3}
\tau(u_{(0)},y_{p})
\tau(u_{(2)},S(y_sk_{-\beta_p-\beta_r}))
(y_rk_{-\beta_p}u_{(1)}\cdot\varphi)u_{(3)}x_sx_rx_p\\
=&
\sum_{p,r,s,(u)_3}
q^{-(\beta_p+\beta_r,\mathop{\rm wt}\nolimits(u_{(2)})+\mathop{\rm wt}\nolimits(u_{(3)}))}
\tau(u_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})-\mathop{\rm wt}\nolimits(u_{(2)})-\mathop{\rm wt}\nolimits(u_{(3)})},y_{p})
\\
&\qquad\qquad
\tau((S^{-1}u_{(2)})k_{\mathop{\rm wt}\nolimits(u_{(2)})+\mathop{\rm wt}\nolimits(u_{(3)})},y_s)
(y_rk_{-\beta_p}u_{(1)}\cdot\varphi)u_{(3)}x_sx_rx_p\\
=&
\sum_{r,(u)_3}
q^{-(\mathop{\rm wt}\nolimits(u_{(0)})+\beta_r,\mathop{\rm wt}\nolimits(u_{(2)})+\mathop{\rm wt}\nolimits(u_{(3)}))}
(y_rk_{-\mathop{\rm wt}\nolimits(u_{(0)})}u_{(1)}\cdot\varphi)
\\
&\qquad\qquad
u_{(3)}(S^{-1}u_{(2)})k_{\mathop{\rm wt}\nolimits(u_{(2)})+\mathop{\rm wt}\nolimits(u_{(3)})}
x_ru_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})-\mathop{\rm wt}\nolimits(u_{(2)})-\mathop{\rm wt}\nolimits(u_{(3)})}\\
=&
\sum_{r,(u)}
(y_rk_{-\mathop{\rm wt}\nolimits(u_{(0)})}u_{(1)}\cdot\varphi)
x_ru_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})}\\
=&
\sum_{r,(u)}
q^{-(\xi+\mathop{\rm wt}\nolimits(u_{(1)}),\mathop{\rm wt}\nolimits(u_{(0)}))}
(y_ru_{(1)}\cdot\varphi)
x_ru_{(0)}k_{-\mathop{\rm wt}\nolimits(u_{(1)})}.
\end{align*}
The formula \eqref{eq:lem:Omega:pf3} for $u\in U_{\mathbb{F}}^+$ is also shown.
Let us finally show \eqref{eq:lem:Omega4}.
We may assume $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$ and $\psi\in A_{\mathbb{F}}(\mu)_\eta$ for $\lambda, \mu\in \Lambda^+, \xi,\eta\in\Lambda$.
Then we have
\begin{align*}
&\Omega_1(\varphi\psi)\\
=&
\sum_p
\sum_{(y_p)}
(y_{p(0)}\cdot\varphi)(y_{p(1)}\cdot\psi)x_pk_{-(\xi+\eta)}
e(\lambda+\mu)\\
=&
\sum_{p,r}
(y_{p}\cdot\varphi)(y_{r}k_{-\beta_p}\cdot\psi)x_rx_pk_{-(\xi+\eta)}
e(\lambda+\mu)\\
=&
\sum_{p,r}
q^{-(\beta_p,\eta)}
(y_{p}\cdot\varphi)(y_{r}\cdot\psi)x_rx_pk_{-(\xi+\eta)}
e(\lambda+\mu).
\end{align*}
On the other hand we have
\begin{align*}
\Omega_1(\psi)\Omega_1(\varphi)
&=\Omega_1(\psi)
\sum_p
(y_{p}\cdot\varphi)x_pk_{-\xi}
e(\lambda)\\
&=
\sum_p
(y_{p}\cdot\varphi)
\Omega_1(\psi)
x_pk_{-\xi}
e(\lambda)\\
&=
\sum_{p,r}
(y_{p}\cdot\varphi)
(y_{r}\cdot\psi)x_rk_{-\eta}
e(\mu)
x_pk_{-\xi}
e(\lambda)\\
&=
\sum_{p,r}
q^{-(\beta_p,\eta)}
(y_{p}\cdot\varphi)(y_{r}\cdot\psi)x_rx_pk_{-(\xi+\eta)}
e(\lambda+\mu).
\end{align*}
We are done.
\end{proof}
By Lemma \ref{lem:Omega} we have
\[
D'_{\mathbb{F}}=
E_{\mathbb{F}}/
\sum_{\varphi\in A_{\mathbb{F}}}
A_{\mathbb{F}}\Omega(\varphi)U_{\mathbb{F}}\BF[\Lambda].
\]
We have a $\Lambda$-graded ${\mathbb{F}}$-algebra structure of $E_{\mathbb{F}}$ given by $E_{\mathbb{F}}(\lambda)=A_{\mathbb{F}}(\lambda)U_{\mathbb{F}}\BF[\Lambda]$ for $\lambda\in\Lambda$.
This also induces $\Lambda$-graded ${\mathbb{F}}$-algebra structure of
${D}'_{\mathbb{F}}$ by
${D}'_{\mathbb{F}}(\lambda)=\mathop{\rm Im}\nolimits(E_{\mathbb{F}}(\lambda)\to{D}'_{\mathbb{F}})$.
Then \eqref{EDD} is a sequence of homomorphisms of $\Lambda$-graded algebras.
\subsection{}
Since $A_{\mathbb{F}}$ belongs to $\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ as a $U_{\mathbb{F}}$-module, we have a natural group homomorphism
\begin{equation}
\label{eq:braid-D0}
{\mathbb{B}}\to \mathop{\rm{End}}\nolimits_{\mathbb{F}}(A_{\mathbb{F}})^\times.
\end{equation}
It induces a group homomorphism
\begin{equation}
\label{eq:braid-D}
{\mathbb{B}}\to{\rm Aut}_{alg}(D_{\mathbb{F}})
\qquad(T\mapsto[\Phi\mapsto T\star \Phi:=T\circ \Phi\circ T^{-1}])
\end{equation}
(see \cite[Proposition 5.2]{T2}).
We will show that this naturally lifts to
group homomorphisms
\[
{\mathbb{B}}\to{\rm Aut}_{alg}(E_{\mathbb{F}}),\qquad
{\mathbb{B}}\to{\rm Aut}_{alg}(D'_{\mathbb{F}}).
\]
\begin{lemma}
\label{lem:braid-D}
\begin{itemize}
\item[\rm(i)]
For $T\in{\mathbb{B}}$ we have
\begin{align*}
&T\star\partial_u=\partial_{T(u)}\qquad
(u\in U_{\mathbb{F}}),\\
&T\star\sigma_\lambda=\sigma_\lambda\qquad
(\lambda\in\Lambda).
\end{align*}
\item[\rm(ii)]
For $i\in I$ write
\begin{align*}
\exp_{q_i}((q_i-q_i^{-1})k_i^{-1}e_i\otimes f_ik_i)
&=\sum_{n=0}^\infty a_{i,n}\otimes b_{i, n},\\
\exp_{q_i^{-1}}(-(q_i-q_i^{-1})f_i\otimes e_i)
&=\sum_{n=0}^\infty a'_{i,n}\otimes b'_{i, n}.
\end{align*}
Then for $\varphi\in A_{\mathbb{F}}$ we have
\begin{align*}
T_i\star\ell_\varphi
&=\sum_{n=0}^\infty \ell_{a_{i,n}\cdot(T_i(\varphi))}\partial_{b_{i, n}},\\
T_i^{-1}\star\ell_\varphi
&=\sum_{n=0}^\infty
\ell_{a'_{i,n}\cdot(T_i^{-1}(\varphi))}\partial_{b'_{i, n}}.
\end{align*}
\end{itemize}
\end{lemma}
\begin{proof}
The proof of (i) is easy and omitted.
Let us show (ii).
In general for $T\in{\mathbb{B}}$ and $\varphi, \psi\in A_{\mathbb{F}}$ we have
\begin{align*}
&(T\star\ell_\varphi)(\psi)
=T(\varphi\cdot T^{-1}(\psi))
=Tm(\varphi\otimes T^{-1}(\psi))\\
=&m(\Delta T)(\varphi\otimes T^{-1}(\psi))
=m(\Delta T)(T^{-1}\otimes T^{-1})
(T(\varphi)\otimes \psi).
\end{align*}
Hence the assertion follows from Lemma \ref{lem:DeltaT}.
\end{proof}
In particular, the action of ${\mathbb{B}}$ on $D_{\mathbb{F}}$ preserves the subalgebra
\[
D^\dagger_{\mathbb{F}}=\langle\partial_u,\ell_\varphi\mid u\in U_{\mathbb{F}}, \varphi\in A_{\mathbb{F}}\rangle\subset D_{\mathbb{F}}.
\]
We first define an action of ${\mathbb{B}}$ on the subalgebra $E^\dagger_{\mathbb{F}}=A_{\mathbb{F}}\otimes U_{\mathbb{F}}$ of
$E_{\mathbb{F}}$.
For $\Phi=\sum_i\varphi_i\otimes u_i\in E^\dagger_{\mathbb{F}}$ and $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ we define
\[
\Phi_M:M\to A_{\mathbb{F}}\otimes M
\]
by $\Phi_M(m)=\sum_i\varphi_i\otimes u_im\,\,(m\in M)$.
By \cite[5.11]{Jan} we have the following.
\begin{lemma}
\label{lem:unique-PhiM}
Let $\Phi\in E^\dagger_{\mathbb{F}}$.
If $\Phi_M=0$ for any $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$, then we have $\Phi=0$.
\end{lemma}
\begin{proposition}
\label{prop:braid-action-Edagger}
There exists a group homomorphism
\[
{\mathbb{B}}\to{\rm Aut}_{alg}(E^\dagger_{\mathbb{F}})
\qquad(T\mapsto[\Phi\mapsto T\star \Phi])
\]
such that
$(T\star \Phi)_M=(\Delta T)\Phi_M T^{-1}$ for $T\in{\mathbb{B}}$, $\Phi\in E^\dagger_{\mathbb{F}}$, $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
Here, $\Delta T:A_{\mathbb{F}}\otimes M\to A_{\mathbb{F}}\otimes M$ is the action of $T\in{\mathbb{B}}$ on $A_{\mathbb{F}}\otimes M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
\end{proposition}
\begin{proof}
We first note the following formula whose proof is easy and omitted;
\begin{equation}
(\Phi\Psi)_M=(m\otimes 1)\Phi_{A_{\mathbb{F}}\otimes M}\Psi_M\qquad
(\Phi, \Psi\in E_{\mathbb{F}}^\dagger, M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})).
\end{equation}
Let $T\in{\mathbb{B}}$.
For $\Phi\in E_{\mathbb{F}}^\dagger$ there exists at most one $T\star\Phi\in E_{\mathbb{F}}^\dagger$ satisfying $(T\star \Phi)_M=(\Delta T)\Phi_M T^{-1}$ for any $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ (see Lemma \ref{lem:unique-PhiM}).
We claim that if $T\star\Phi$ exists for any $\Phi\in E_{\mathbb{F}}^\dagger$, then
$E_{\mathbb{F}}^\dagger\ni\Phi\mapsto T\star\Phi\in E_{\mathbb{F}}^\dagger$ is an algebra homomorphism.
We have
\[
(T\star 1)_M(v)
=(\Delta T)1_M T^{-1}(v)
=(\Delta T)(1\otimes T^{-1}(v))=1\otimes v
\]
for any $v\in M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
Here the last equality is a consequence of ${\mathbb{F}}\otimes M\cong M$ as a $U_{\mathbb{F}}$-module.
By $1_M(v)=1\otimes v$ we obtain $T\star 1=1$ by Lemma \ref{lem:unique-PhiM}.
For $\Phi, \Psi\in E_{\mathbb{F}}^\dagger$
we have
\begin{align*}
&((T\star\Phi)(T\star\Psi))_M
=(m\otimes 1)(T\star\Phi)_{A\otimes M}(T\star\Psi)_M\\
=&(m\otimes 1)(\Delta_2T)\Phi_{A\otimes M}
(\Delta T)^{-1}(\Delta T)\Psi_M T^{-1}\\
=&(\Delta T)(m\otimes 1)\Phi_{A\otimes M}\Psi_M T^{-1}
=(\Delta T)(\Phi\Psi)_M T^{-1}=(T\star(\Phi\Psi))_M
\end{align*}
for any $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
Here, we used the fact that the multiplication $m:A_{\mathbb{F}}\otimes A_{\mathbb{F}}\to A_{\mathbb{F}}$ is a homomorphism of $U_{\mathbb{F}}$-modules.
Hence we have $(T\star\Phi)(T\star\Psi)=T\star(\Phi\Psi)$.
Our claim is verified.
Let $T, T'\in{\mathbb{B}}$, and assume that
$T\star\Phi$, $T'\star\Phi$ exist for any $\Phi\in E_{\mathbb{F}}^\dagger$.
Then we have
\[
(T\star(T'\star\Phi))_M
=(\Delta T)(\Delta T')\Phi_M(T')^{-1}T^{-1}
=\Delta(TT')\Phi_M(TT')^{-1}
\]
for $\Phi\in E_{\mathbb{F}}^\dagger$, $M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
Hence
$(TT')\star\Phi$ exists and we have $(TT')\star\Phi=T\star(T'\star\Phi)$ for any $\Phi\in E_{\mathbb{F}}^\dagger$.
It remains to show the existence of $T\star\Phi$ for $T=T_i^{\pm1}$, $\Phi\in E_{\mathbb{F}}^\dagger$.
For $\Phi=\varphi\otimes u\in E_{\mathbb{F}}^\dagger$ \,\,($\varphi\in A_{\mathbb{F}}$, $u\in U_{\mathbb{F}}$) we have
\begin{align*}
&((\Delta T_i) \Phi_M T_i^{-1})(v)
=(\Delta T_i) (\varphi\otimes uT_i^{-1}v)\\
=&(\Delta T_i)(T_i^{-1}\otimes T_i^{-1})
(T_i(\varphi)\otimes T_i(u)v)\\
=&\sum_n
a_{i,n}\cdot T_i(\varphi)\otimes b_{i,n}T_i(u)v\\
=&
\left(\sum_n
a_{i,n}\cdot T_i(\varphi)\otimes b_{i,n}T_i(u)\right)_M(v)
\end{align*}
for $v\in M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
Hence we obtain
\[
T_i\star(\varphi\otimes u)=\sum_n
a_{i,n}\cdot T_i(\varphi)\otimes b_{i,n}T_i(u)
\qquad(\varphi\in A_{\mathbb{F}}, u\in U_{\mathbb{F}}).
\]
Similarly, we have
\[
T_i^{-1}\star(\varphi\otimes u)=\sum_n
a'_{i,n}\cdot T_i^{-1}(\varphi)\otimes b'_{i,n}T_i^{-1}(u)
\qquad(\varphi\in A_{\mathbb{F}}, u\in U_{\mathbb{F}}).
\]
We are done.
\end{proof}
\begin{lemma}
\label{lem:TstarFormula}
\begin{itemize}
\item[\rm(i)]
$T\star u=T(u)$
\qquad$(T\in {\mathbb{B}}, \,\,u\in U_{\mathbb{F}})$.
\item[\rm(ii)]
$T_i\star\varphi=\sum_n
a_{i,n}\cdot T_i(\varphi)\otimes b_{i,n}$
\qquad$(i\in I,\,\,\varphi\in A_{\mathbb{F}})$.
\item[\rm(iii)]
$T_i^{-1}\star\varphi=\sum_n
a'_{i,n}\cdot T_i^{-1}(\varphi)\otimes b'_{i,n}$
\qquad$(i\in I,\,\,\varphi\in A_{\mathbb{F}})$.
\end{itemize}
\end{lemma}
\begin{proof}
For $T\in{\mathbb{B}}$, $u\in U_{\mathbb{F}}$, $v\in M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ we have
\begin{align*}
&(T\star u)_M(v)
=(\Delta T)u_MT^{-1}(v)
=(\Delta T)(1\otimes uT^{-1}(v))\\
=&1\otimes TuT^{-1}v
=1\otimes T(u)v=(T(u))_M(v).
\end{align*}
Hence (i) holds.
The statements (ii) and (iii) are already shown in the proof of Proposition \ref{prop:braid-action-Edagger}.
\end{proof}
By Lemma \ref{lem:TstarFormula}
we have
\[
T\star(A_{\mathbb{F}}(\lambda)\otimes U_{\mathbb{F}})=A_{\mathbb{F}}(\lambda)\otimes U_{\mathbb{F}}
\qquad(\lambda\in\Lambda^+).
\]
Hence, for $T\in{\mathbb{B}}$ the algebra automorphism $T\star(\cdot):E^\dagger_{\mathbb{F}}\to E^\dagger_{\mathbb{F}}$ is naturally extended to that of $E_{\mathbb{F}}$ by setting
\[
T\star e(\mu)=e(\mu)\qquad(\mu\in \Lambda).
\]
By this we obtain
a group homomorphism
\begin{equation}
\label{eq:braid-action-E}
{\mathbb{B}}\to{\rm Aut}_{alg}(E_{\mathbb{F}})
\qquad(T\mapsto[\Phi\mapsto T\star \Phi]).
\end{equation}
\begin{proposition}
\label{prop:braid-action-Dprime}
For $T\in{\mathbb{B}}$ we have
\[
T\star\mathop{\rm Ker\hskip.5pt}\nolimits(E_{\mathbb{F}}\to D_{\mathbb{F}}')\subset
\mathop{\rm Ker\hskip.5pt}\nolimits(E_{\mathbb{F}}\to D_{\mathbb{F}}').
\]
\end{proposition}
\begin{proof}
It is sufficient to show
\[
T\star\Omega(\varphi)\in\sum_{\psi\in A_{\mathbb{F}}}
E_{\mathbb{F}}\Omega(\psi)E_{\mathbb{F}}
\]
for $T=T_i^{\pm1}$, $\varphi\in A_{\mathbb{F}}$.
In general, for $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$, we write
\begin{align*}
&
\Omega(\varphi)=\Omega_1'(\varphi)e(\lambda)-\Omega_2'(\varphi)e(-\lambda),
\\
&
\Omega_1'(\varphi)=
\sum_p(y_p\cdot\varphi)x_pk_{-\xi}
\in E_{\mathbb{F}}^\dagger,
\\
&
\Omega_2'(\varphi)=
\sum_p((Sx_p)\cdot\varphi)y_pk_{\beta_p}k_\xi\in E_{\mathbb{F}}^\dagger.
\end{align*}
Let $M, M'\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
We define a linear automorphism $\kappa_{M,M'}$ of $M\otimes M'$ by
$\kappa_{M,M'}|_{M_\lambda\otimes M'_\mu}=q^{(\lambda,\mu)}\mathop{\rm id}\nolimits$ for $\lambda, \mu\in \Lambda$.
We also define a linear isomorphism $\tau_{M,M'}:M\otimes M'\to M'\otimes M$ by $\tau_{M,M'}(v\otimes v')=v'\otimes v$.
Set
\[
{\mathcal R}_{M,M'}
=\kappa_{M,M'}^{-1}\circ
\left(
\sum_pq^{(\beta_p,\beta_p)}
k_{\beta_p}^{-1}x_p\otimes k_{\beta_p}y_p
\right)
\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M').
\]
Then ${\mathcal R}_{M,M'}$ is invertible and we have
\[
{\mathcal R}_{M,M'}^{-1}
=
\left(
\sum_pq^{(\beta_p,\beta_p)}
Sx_p\otimes k_{\beta_p}y_p
\right)
\circ \kappa_{M,M'}
\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M').
\]
Moreover, we have
\begin{equation}
\label{eq:univR}
(\Delta' u){\mathcal R}_{M,M'}={\mathcal R}_{M,M'}(\Delta u)\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M')
\qquad(u\in U_{\mathbb{F}}),
\end{equation}
where $\Delta':U_{\mathbb{F}}\to U_{\mathbb{F}}\otimes U_{\mathbb{F}}$ is the opposite comultiplication given by $\Delta'=\tau\circ\Delta$ with
$\tau(a\otimes b)=b\otimes a$
(see, for example, \cite[2.2]{T2}).
By \eqref{eq:univR}
we have
\begin{equation}
\label{eq:univR-braid}
(\Delta' T){\mathcal R}_{M,M'}={\mathcal R}_{M,M'}(\Delta T)\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M')
\qquad(T\in{\mathbb{B}}),
\end{equation}
where $\Delta' T=\tau_{M,M'}^{-1}\circ\Delta T\circ\tau_{M,M'}$.
Let $\varphi\in A_{\mathbb{F}}(\lambda)_\xi$
and $v\in M\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$.
By the definition we see easily that
\begin{equation}
{\mathcal R}_{A_{\mathbb{F}},M}^{-1}
(\varphi\otimes v)=\Omega_2'(\varphi)_M(v).
\end{equation}
Hence we have
\begin{align*}
&(T\star\Omega_2'(\varphi))_M(v)
=(\Delta T)\Omega_2'(\varphi)_MT^{-1}(v)
=(\Delta T)
{\mathcal R}_{A_{\mathbb{F}},M}^{-1}(\varphi\otimes T^{-1}v)\\
=&
{\mathcal R}_{A_{\mathbb{F}},M}^{-1}(\Delta' T)(\varphi\otimes T^{-1}v)
=
{\mathcal R}_{A_{\mathbb{F}},M}^{-1}(\Delta' T)(T^{-1}\otimes T^{-1})(T\varphi\otimes v)
\end{align*}
for any $T\in{\mathbb{B}}$.
For $T=T_i^{\pm1}$ we can write
$(\Delta T)(T^{-1}\otimes T^{-1})=\sum_nc_n\otimes d_n$ for $c_n, d_n\in U_{\mathbb{F}}$ (see Lemma \ref{lem:DeltaT}),
and hence
\begin{align*}
&(T\star\Omega_2'(\varphi))_M(v)
=
{\mathcal R}_{A_{\mathbb{F}},M}^{-1}(\sum_nd_n\cdot T(\varphi)\otimes c_n v)
\\
=&
\sum_n
\Omega_2'(d_n\cdot T(\varphi))_M(c_nv)
=
\left(
\sum_n
\Omega_2'(d_n\cdot T(\varphi))c_n
\right)_M
(v).
\end{align*}
It follows that
\begin{equation}
\label{eq:TOmega1}
T\star\Omega_2'(\varphi)
=\sum_n\Omega_2'(d_n\cdot T(\varphi))c_n
\qquad
(T=T_i^{\pm1}).
\end{equation}
For $M, M'\in\mathop{\rm Mod}\nolimits_{int}(U_{\mathbb{F}})$ define
${\mathcal R}'_{M, M'}\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M')$
by ${\mathcal R}'_{M, M'}=\tau_{M, M'}^{-1}\circ{\mathcal R}_{M', M}\circ \tau_{M, M'}$.
Then by a similar argument as above using
\begin{align}
\label{eq:univR-braid2}
&(\Delta T){\mathcal R}'_{M,M'}={\mathcal R}'_{M,M'}(\Delta' T)\in\mathop{\rm{End}}\nolimits_{\mathbb{F}}(M\otimes M')
\qquad(T\in{\mathbb{B}}),\\
&({\mathcal R}'_{A_{\mathbb{F}},M})^{-1}
(\varphi\otimes v)=\Omega_1'(\varphi)_M(v)
\quad
(\varphi\in A_{\mathbb{F}}(\lambda)_\xi,\,\,
v\in M)
\end{align}
we obtain
\begin{equation}
\label{eq:TOmega2}
T\star\Omega_1'(\varphi)
=\sum_n\Omega_1'(d_n\cdot T(\varphi))c_n
\qquad
(T=T_i^{\pm1}).
\end{equation}
By \eqref{eq:TOmega1}, \eqref{eq:TOmega2} we finally obtain
\begin{equation}
\label{eq:TOmega}
T\star\Omega(\varphi)
=\sum_n\Omega(d_n\cdot T(\varphi))c_n
\qquad
(T=T_i^{\pm1},\,\,\varphi\in A_{\mathbb{F}}).
\end{equation}
We are done.
\end{proof}
We see from Proposition \ref{prop:braid-action-Dprime} that \eqref{eq:braid-action-E}
induces a group homomorphism
\begin{equation}
\label{eq:braid-action-Dprime}
{\mathbb{B}}\to{\rm Aut}_{alg}(D'_{\mathbb{F}})
\qquad(T\mapsto[\Phi\mapsto T\star \Phi]).
\end{equation}
Note that \eqref{eq:braid-action-E} and \eqref{eq:braid-action-Dprime} are natural lifts of \eqref{eq:braid-D} by Lemma \ref{lem:braid-D}.
\subsection{}
Set
\[
D_{\mathbb{A}}=
\langle
\ell_\varphi, r_\varphi, \partial_u,\sigma_\lambda\mid
\varphi\in A_{\mathbb{A}}, u\in U_{\mathbb{A}}, \lambda\in\Lambda\rangle_{{\mathbb{A}}-{\rm{alg}}}\subset D_{\mathbb{F}}.
\]
We have a canonical embedding
\[
D_{\mathbb{A}}\to\mathop{\rm{End}}\nolimits_{\mathbb{A}}(A_{\mathbb{A}}).
\]
Recall that we have fixed ${\mathbb{F}}$-bases $\{x_p\}_p$ and $\{y_p\}_p$ of $U_{\mathbb{F}}^+$ and $U_{\mathbb{F}}^-$ respectively and $\beta_p\in Q^+$ for each $p$ satisfying \eqref{eq:xy1}, \eqref{eq:xy2}.
By lemma \ref{lem:pm-duality} we can renormalize them in the following two manners;
\begin{itemize}
\item[(a)]
$\{x_p\}_p$ and $\{y_p\}_p$ are ${\mathbb{A}}$-bases of $U_{\mathbb{A}}^{L,+}$ and $U_{\mathbb{A}}^-$ respectively,
\item[(b)]
$\{x_p\}_p$ and $\{y_p\}_p$ are ${\mathbb{A}}$-bases of $U_{\mathbb{A}}^{+}$ and $U_{\mathbb{A}}^{L,-}$ respectively.
\end{itemize}
In particular, we have
\[
D_{\mathbb{A}}=
\langle
\ell_\varphi, \partial_u,\sigma_\lambda\mid
\varphi\in A_{\mathbb{A}}, u\in U_{\mathbb{A}}, \lambda\in\Lambda\rangle_{{\mathbb{A}}-{\rm{alg}}}\subset D_{\mathbb{F}}
\]
by Lemma \ref{lem:rvarphi}.
In the case (a) (resp. (b)) we write
$\{x_p\}_p$ and $\{y_p\}_p$ as above as
$\{x^L_p\}_p$ and $\{y_p\}_p$
(resp. $\{x_p\}_p$ and $\{y^L_p\}_p$).
Define an ${\mathbb{A}}$-subalgebra $E_{\mathbb{A}}$ of $E_{\mathbb{F}}$ by
\[
E_{\mathbb{A}}=A_{\mathbb{A}} U_{\mathbb{A}} {\mathbb{A}}[\Lambda]\,
(\cong A_{\mathbb{A}}\otimes_{\mathbb{A}} U_{\mathbb{A}}\otimes_{\mathbb{A}} {\mathbb{A}}[\Lambda]),
\]
and set
\[
D'_{\mathbb{A}}
=\mathop{\rm Im}\nolimits(E_{\mathbb{A}}\to D'_{\mathbb{F}})\subset D'_{\mathbb{F}}.
\]
For $\varphi\in A_{\mathbb{A}}(\lambda)_\xi$ with $\lambda\in\Lambda^+$, $\xi\in\Lambda$ we have
\begin{align*}
\Omega_1(\varphi)&=
\sum_p({y^L_p}\cdot\varphi){x_pk_{-\xi}}e(\lambda)
\in E_{\mathbb{A}},\\
\Omega_2(\varphi)&=
\sum_p({(Sx^L_p)\cdot\varphi}){y_pk_{\beta_p}k_\xi}e({-\lambda})
\in E_{\mathbb{A}},
\end{align*}
and hence
$\Omega(\varphi)\in E_{\mathbb{A}}$.
It follows that
\begin{equation*}
D'_{\mathbb{A}}=
E_{\mathbb{A}}/
\sum_{\varphi\in A_{\mathbb{A}}}E_{\mathbb{A}}\Omega(\varphi)E_{\mathbb{A}}
=
E_{\mathbb{A}}/
\sum_{\varphi\in A_{\mathbb{A}}}A_{\mathbb{A}}\Omega(\varphi)U_{\mathbb{A}} {\mathbb{A}}[\Lambda].
\end{equation*}
Note that
$E_{\mathbb{A}}$, $D'_{\mathbb{A}}$, $D_{\mathbb{A}}$ are $\Lambda$-graded ${\mathbb{A}}$-algebras
by
\[
E_{\mathbb{A}}(\lambda)=E_{\mathbb{F}}(\lambda)\cap E_{\mathbb{A}},\quad
D'_{\mathbb{A}}(\lambda)=D'_{\mathbb{F}}(\lambda)\cap D'_{\mathbb{A}},\quad
D_{\mathbb{A}}(\lambda)=D_{\mathbb{F}}(\lambda)\cap D_{\mathbb{A}}\quad
\]
for $\lambda\in\Lambda^+$.
We have a sequence
\[
E_{\mathbb{A}}\to D'_{\mathbb{A}}\to D_{\mathbb{A}}
\]
of surjective homomorphisms of $\Lambda$-graded ${\mathbb{A}}$-algebras.
The braid group actions on $E_{\mathbb{F}}$, $D'_{\mathbb{F}}$, $D_{\mathbb{F}}$ induces those on
$E_{\mathbb{A}}$, $D'_{\mathbb{A}}$, $D_{\mathbb{A}}$.
Namely we have group homomorphisms
\begin{equation}
\label{eq:braidA}
{\mathbb{B}}\to{\rm Aut}_{alg}(E_{\mathbb{A}}),\qquad
{\mathbb{B}}\to{\rm Aut}_{alg}(D'_{\mathbb{A}}),\qquad
{\mathbb{B}}\to{\rm Aut}_{alg}(D_{\mathbb{A}})
\end{equation}
denoted by $T\mapsto[\Phi\mapsto T\star \Phi]$.
\subsection{}
We set
\[
E_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} E_{\mathbb{A}},\qquad
D'_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} D'_{\mathbb{A}},\qquad
D_\zeta={\mathbb{C}}\otimes_{\mathbb{A}} D_{\mathbb{A}}.
\]
Then we have
\begin{align}
\label{eq:Ezeta1}
&E_\zeta=A_\zeta\otimes U_\zeta\otimes {\mathbb{C}}[\Lambda]\\
\label{eq:Ezeta2}
&D'_\zeta=
E_\zeta/
\sum_{\varphi\in A_\zeta}E_\zeta\Omega(\varphi)E_\zeta
=
E_\zeta/
\sum_{\varphi\in A_\zeta}A_\zeta\Omega(\varphi)U_\zeta {\mathbb{C}}[\Lambda].
\end{align}
We have a sequence
\begin{equation}
\label{eq:ED'D:zeta}
E_\zeta\to D'_\zeta\to D_\zeta
\end{equation}
of surjective homomorphisms of $\Lambda$-graded ${\mathbb{C}}$-algebras.
By \eqref{eq:HC-center-F}, \eqref{eq:HC-center}, \eqref{eq:Har-center-in-D0}
we have
\begin{equation}
\label{eq:Har-center-in-D}
z\in Z_{Har}(U_\zeta),\,\,
\iota(z)=\sum_{\lambda\in\Lambda}a_\lambda e({2\lambda})\quad\Longrightarrow\quad
\partial_z
=\sum_{\lambda\in\Lambda}a_\lambda\sigma_{2\lambda}
\end{equation}
in $D_\zeta$.
\begin{remark}
{\rm
The natural algebra homomorphism
$D_\zeta\to\mathop{\rm{End}}\nolimits_{\mathbb{C}}(A_\zeta)$
is not injective.
}
\end{remark}
The actions of ${\mathbb{B}}$ on $E_{\mathbb{A}}$, $D'_{\mathbb{A}}$, $D_{\mathbb{A}}$ induce the group homomorphisms
\begin{equation}
\label{eq:braid-zeta}
{\mathbb{B}}\to{\rm Aut}_{alg}(E_\zeta),\qquad
{\mathbb{B}}\to{\rm Aut}_{alg}(D'_\zeta),\qquad
{\mathbb{B}}\to{\rm Aut}_{alg}(D_\zeta)
\end{equation}
denoted by $T\mapsto[\Phi\mapsto T\star \Phi]$.
We have natural algebra homomorphisms
\begin{equation}
\label{eq:A1toED}
A_1\to E_\zeta,\qquad
A_1\to D'_\zeta,\qquad
A_1\to D_\zeta
\end{equation}
induced by \eqref{eq:ED'D:zeta} and the embedding
\[
A_1\subset A_\zeta=A_\zeta\otimes 1\otimes 1\subset A_\zeta\otimes U_\zeta\otimes{\mathbb{C}}[\Lambda]
=E_\zeta.
\]
Note that the images of \eqref{eq:A1toED} are contained in the center.
\begin{lemma}
\label{lem:braid-action-on-D}
Identify $\Theta_x$ for $x\in W$ as a subset of $E_\zeta$ via \eqref{eq:A1toED}.
Then we have
$T_i^{-1}\star\Theta_w=\Theta_{ws_i}$
for $w\in W$ and $i\in I$ such that $ws_i>w$ with respect to the standard partial order.
\end{lemma}
\begin{proof}
Note that for $\lambda\in\Lambda^+$
we have
$T_i^{-1}(A_\zeta(\lambda)_{w^{-1}\lambda})
=A_\zeta(\lambda)_{s_iw^{-1}\lambda}$.
By our assumption on $w$ and $i$ we have
$(s_iw^{-1}\lambda,\alpha_i^\vee)\leqq0$, and hence
we obtain from Lemma \ref{lem:TstarFormula} that
$T_i^{-1}\star\varphi=T_i^{-1}(\varphi)$ for any $\varphi\in A_\zeta(\lambda)_{w^{-1}\lambda}$.
\end{proof}
Let $w\in W$.
By Lemma \ref{lem:braid-action-on-D}
we have $T_{w^{-1}}^{-1}\star\Theta_e=\Theta_w$.
Hence
the algebra automorphisms
\[
T_{w^{-1}}^{-1}\star(\bullet):E_\zeta\to E_\zeta,\qquad
T_{w^{-1}}^{-1}\star(\bullet):D'_\zeta\to D'_\zeta,\qquad
T_{w^{-1}}^{-1}\star(\bullet):D_\zeta\to D_\zeta
\]
induce isomorphisms
\[
\Theta_e^{-1}E_\zeta\to\Theta_w^{-1}E_\zeta,
\qquad
\Theta_e^{-1}D'_\zeta\to\Theta_w^{-1}D'_\zeta,
\qquad
\Theta_e^{-1}D_\zeta\to\Theta_w^{-1}D_\zeta
\]
of $\Lambda$-graded algebras.
For $w\in W$ set
\[
\tilde{\Theta}_w=\bigcup_{\lambda\in\Lambda^+}(A_\zeta(\lambda)_{w^{-1}\lambda}\setminus\{0\}).
\]
It is a multiplicative subset of $A_\zeta$.
Moreover, for any $s\in \tilde{\Theta}_w$ we have $s^{\ell}\in\Theta_w$.
Hence if we are given a ring homomorphism $A_\zeta\to R$ such that the image of $A_1$ is contained in the center of $R$, then the image of $\tilde{\Theta}_w$ in $R$ satisfies the left and right Ore conditions.
Moreover, in this situation we have $\tilde{\Theta}_w^{-1}R\cong {\Theta}_w^{-1}R$.
In particular, we have
\[
\Theta_w^{-1}A_\zeta\cong\tilde{\Theta}_w^{-1}A_\zeta,
\qquad
\Theta_w^{-1}E_\zeta\cong\tilde{\Theta}_w^{-1}E_\zeta,
\qquad
\Theta_w^{-1}D'_\zeta\cong\tilde{\Theta}_w^{-1}D'_\zeta.
\]
\begin{proposition}
\label{prop:local-verma}
We have a natural $U_\zeta^L$-module structure of $\tilde{\Theta}_e^{-1}A_\zeta$ such that $A_\zeta\to\tilde{\Theta}_e^{-1}A_\zeta$ is a homomorphism of $U_\zeta^L$-modules and
\[
u\cdot(\varphi\psi)=\sum_{(u)}(u_{(0)}\cdot\varphi)(u_{(1)}\cdot\psi)
\qquad
(u\in U_\zeta^L, \varphi, \psi\in \tilde{\Theta}_e^{-1}A_\zeta).
\]
Moreover, for any $\lambda\in\Lambda$ we have
\[
(\tilde{\Theta}_e^{-1}A_\zeta)(\lambda)\cong
M^*_{-,\zeta}(\lambda).
\]
as a $U^L_\zeta$-module.
\end{proposition}
\begin{proof}
It is not difficult to deduce our assertion from
the corresponding fact over ${\mathbb{F}}$, which is shown in
\cite[Proposition 4.3, Proposition 4.6]{T2}.
Details are omitted.
\end{proof}
Denote by
\[
\jmath:\tilde{\Theta}_e^{-1}E_\zeta\to\tilde{\Theta}_e^{-1}D'_\zeta
\]
the canonical algebra homomorphism.
\begin{proposition}
\label{prop:local-formula1}
Let $\lambda\in\Lambda^+$ and $\gamma\in Q^+$.
For $\varphi\in A_\zeta(\lambda)_{\lambda-\gamma}$ and $s\in A_\zeta(\lambda)_\lambda\setminus\{0\}$ we have
\[
\jmath(
\sum_{p}((Sx^L_p)\cdot(\varphi s^{-1}))
y_pk_{\beta_p}
)
=
\zeta^{(\lambda,\gamma)}
\jmath(
\sum_ps^{-1}(y^L_p\cdot\varphi)x_pk_{-2(\lambda-\gamma)}e(2\lambda)
)
\]
in $(\tilde{\Theta}_e^{-1}D'_\zeta)(0)$.
\end{proposition}
\begin{proof}
By Lemma \ref{lem:Omega} we have algebra anti-homomorphisms
\begin{equation}
\label{eq:Omega-bar}
\overline{\Omega}_i:A_\zeta\to D'_\zeta
\qquad(i=1, 2)
\end{equation}
as the composite of
\[
A_\zeta \xrightarrow{\Omega_i} E_\zeta\to D'_\zeta.
\]
By the definition of $D_\zeta'$ we have $\overline{\Omega}_1=\overline{\Omega}_2$.
For $\psi\in A_\zeta(\lambda)_\lambda$ with $\lambda\in\Lambda^+$ we have
$\overline{\Omega}_2(\psi)=
\overline{k_{\lambda}e(-\lambda)\psi}$
by definition.
Hence \eqref{eq:Omega-bar} induces an anti-homomorphism
\begin{equation}
\label{eq:Omega-bar2}
\tilde{\Theta}_e^{-1}\overline{\Omega}_1=
\tilde{\Theta}_e^{-1}\overline{\Omega}_2:\tilde{\Theta}_e^{-1}A_\zeta\to \tilde{\Theta}_e^{-1}D'_\zeta
\end{equation}
of $\Lambda$-graded algebras.
For $x\in U_\zeta^{L,+}$ we have
\begin{align*}
&\varepsilon(x)1=x\cdot 1=x\cdot(s^{-1}s)
=\sum_{(x)}(x_{(0)}\cdot s^{-1})(x_{(1)}\cdot s)\\
=&
\sum_{(x)}\varepsilon(x_{(1)})(x_{(0)}\cdot s^{-1})s
=(x\cdot s^{-1})s,
\end{align*}
and hence $x\cdot s^{-1}=\varepsilon(x)s^{-1}$.
Therefore
\begin{align*}
&(Sx^L_p)\cdot(\varphi s^{-1})=
\sum_{(x^L_p)}(S(x^L_{p(1)})\cdot \varphi)
(S(x^L_{p(0)})\cdot s^{-1})\\
=&((Sx^L_{p})\cdot \varphi)
(k_{-\beta_p}\cdot s^{-1})
=\zeta^{(\lambda,\beta_p)}
((Sx^L_{p})\cdot \varphi)s^{-1}.
\end{align*}
By $\overline{\Omega}_2(s)=\jmath(k_\lambda e(-\lambda)s)$ we have
$\jmath(s^{-1})=(\overline{\Omega}_2(s))^{-1}\jmath(k_\lambda e(-\lambda))$, and hence
\begin{align*}
\jmath
&((Sx^L_p)\cdot(\varphi s^{-1}))
=\zeta^{(\lambda,\beta_p)}
\jmath((Sx^L_{p})\cdot \varphi)
(\overline{\Omega}_2(s))^{-1}\jmath(k_\lambda e(-\lambda))\\
=&
\zeta^{(\lambda,\beta_p)}
(\overline{\Omega}_2(s))^{-1}
\jmath(((Sx^L_{p})\cdot \varphi) k_\lambda e(-\lambda))
\end{align*}
by Lemma \ref{lem:Omega}.
Therefore, we have
\begin{align*}
&\jmath(
\sum_{p}((Sx^L_p)\cdot(\varphi s^{-1}))
y_pk_{\beta_p}
)\\
=&
(\overline{\Omega}_2(s))^{-1}
\jmath(
\sum_{p}
\zeta^{(\lambda,\beta_p)}
((Sx^L_{p})\cdot \varphi) k_\lambda e(-\lambda)
y_pk_{\beta_p})\\
=&
(\overline{\Omega}_2(s))^{-1}
\jmath(
\sum_{p}
((Sx^L_{p})\cdot \varphi)y_p k_{\lambda+\beta_p} e(-\lambda))\\
=&
(\overline{\Omega}_2(s))^{-1}
(\overline{\Omega}_2(\varphi))
\jmath(k_\gamma)
\\
=&(\tilde{\Theta}_e^{-1}\overline{\Omega}_2)(\varphi s^{-1})\jmath(k_\gamma).
\end{align*}
On the other hand
we have
\begin{align*}
&(\tilde{\Theta}_e^{-1}\overline{\Omega}_1)(\varphi s^{-1})
=
\overline{\Omega}_2(s)^{-1}
\overline{\Omega}_1(\varphi)
\\
=&
\jmath(
\sum_p
s^{-1}e(\lambda)k_{-\lambda}(y^L_p\cdot\varphi)x_pk_{-(\lambda-\gamma)}e(\lambda))
\\
=&
\zeta^{(\lambda,\gamma)}
\jmath(
\sum_p
s^{-1}(y^L_p\cdot\varphi)x_pk_{-(2\lambda-\gamma)}e(2\lambda)).
\end{align*}
We obtain the desired result by $\tilde{\Theta}_e^{-1}\overline{\Omega}_1=\tilde{\Theta}_e^{-1}\overline{\Omega}_2$
\end{proof}
Considering the case $\varphi=s\in A_1(\lambda)_\lambda\setminus\{0\}$ in Proposition \ref{prop:local-formula1} we obtain the following.
\begin{proposition}
\label{prop:local-formula2}
Let $\lambda\in\Lambda^+$.
For $s\in A_\zeta(\lambda)_\lambda\setminus\{0\}$ we have
\[
\jmath(
k_{2\lambda})
=
\jmath(
\sum_ps^{-1}(y^L_p\cdot s)x_pe(2\lambda))
\]
in $(\Theta_e^{-1}D'_\zeta)(0)$.
\end{proposition}
\subsection{}
The natural embedding $A_\zeta\to E_\zeta$ induces
homomorphisms
\[
A_\zeta\to D'_\zeta\quad
(\varphi\mapsto \overline{\varphi}),
\qquad
A_\zeta\to D_\zeta\quad
(\varphi\mapsto \ell_\varphi)
\]
of graded ${\mathbb{C}}$-algebras.
We define abelian categories $\mathop{\rm Mod}\nolimits({\mathcal D}'_{{\mathcal B}_\zeta})$, $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})$ by
\[
\mathop{\rm Mod}\nolimits({\mathcal D}'_{{\mathcal B}_\zeta})=
{\mathcal C}(A_\zeta,D'_\zeta),\qquad
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})=
{\mathcal C}(A_\zeta,D_\zeta).
\]
We also define ${\mathcal O}_{\mathcal B}$-algebras $Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}$, $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$ by
\[
Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}=\omega_{\mathcal B}^*D_\zeta^{'(\ell)},
\qquad
Fr_*{\mathcal D}_{{\mathcal B}_\zeta}=\omega_{\mathcal B}^*D_\zeta^{(\ell)}.
\]
By Lemma \ref{lem:Fr1} and Lemma \ref{lem:Fr2}
we have
equivalences
\begin{align}
\label{eq:equiv-D1}
&Fr_*:\mathop{\rm Mod}\nolimits({\mathcal D}'_{{\mathcal B}_\zeta})\to
\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}),\\
\label{eq:equiv-D2}
&Fr_*:
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})\to
\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta})
\end{align}
of abelian categories,
where $\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}'_{{\mathcal B}_\zeta})$ (resp.\
$\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta})$) denote the category of quasi-coherent $Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}$-modules (resp.\ quasi-coherent $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$-modules).
Moreover, we have the following.
\begin{lemma}
For $M\in\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})$ we have
\[
R^i\Gamma_{(A_\zeta,D_\zeta)}(M)= R^i\Gamma({\mathcal B},Fr_*(M))
\in\mathop{\rm Mod}\nolimits(D_\zeta(0)),
\]
where $\Gamma({\mathcal B},\,\,)$ in the right side is the global section functor for ${\mathcal B}$.
\end{lemma}
For $t\in H$ we define an abelian category $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})$ by
\[
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})=
\mathop{\rm Mod}\nolimits_{\Lambda, t}(D_\zeta)/(\mathop{\rm Mod}\nolimits_{\Lambda,t}(D_\zeta)\cap{\rm{Tor}}_{\Lambda^+}(A_\zeta,D_\zeta)),
\]
where $\mathop{\rm Mod}\nolimits_{\Lambda,t}(D_\zeta)$ is the full subcategory of
$\mathop{\rm Mod}\nolimits_{\Lambda}(D_\zeta)$ consisting of $M\in\mathop{\rm Mod}\nolimits_{\Lambda}(D_\zeta)$ so that $\sigma_\lambda|_{M(\mu)}=\theta_\lambda(t) \zeta^{(\lambda,\mu)}\mathop{\rm id}\nolimits$ for any $\lambda, \mu\in\Lambda$.
Then we can regard $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})$ as a full subcategory of $\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta})$
(see \cite[Lemma 4.6]{T2}).
Set
\[
Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t}=Fr_*{\mathcal D}_{{\mathcal B}_\zeta}\otimes_{{\mathbb{C}}[\Lambda]}{\mathbb{C}}_t,
\]
where ${\mathbb{C}}_t$ denotes the one-dimensional ${\mathbb{C}}[\Lambda]$-module given by $e(\lambda)\mapsto\theta_\lambda(t)$ for $\lambda\in\Lambda$.
The equivalence \eqref{eq:equiv-D2} induces the equivalence
\begin{equation}
\label{eq:equiv-D3}
Fr_*:
\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})\to
\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t}),
\end{equation}
where $\mathop{\rm Mod}\nolimits(Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t})$ denotes the category of quasi-coherent $Fr_*{\mathcal D}_{{\mathcal B}_\zeta,t}$-modules.
In particular, for $M\in\mathop{\rm Mod}\nolimits({\mathcal D}_{{\mathcal B}_\zeta,t})$ we have
\[
R^i\Gamma_{(A_\zeta,D_\zeta)}(M)
=R^i\Gamma({\mathcal B},Fr_*M)
\in\mathop{\rm Mod}\nolimits(D_{\zeta,t}(0)),
\]
where
$D_{\zeta,t}(0)=
D_{\zeta}(0)/\sum_{\lambda\in\Lambda}
D_{\zeta}(0)(\sigma_\lambda-\theta_\lambda(t))$.
\section{Center}
\subsection{}
Note
\[
E_\zeta^{(\ell)}=A_\zeta^{(\ell)}\otimes_{\mathbb{C}} U_\zeta\otimes_{\mathbb{C}} {\mathbb{C}}[\Lambda].
\]
We set
\begin{equation}
ZE_\zeta^{(\ell)}=
A_1\otimes_{\mathbb{C}} Z_{Fr}(U_\zeta)\otimes_{\mathbb{C}} {\mathbb{C}}[\Lambda]
\subset E_\zeta^{(\ell)}.
\end{equation}
\begin{lemma}
\label{lem:center}
$ZE_\zeta^{(\ell)}$ is a central subalgebra of $E_\zeta^{(\ell)}$.
\end{lemma}
\begin{proof}
It is easily seen that ${\mathbb{C}}[\Lambda]$ is contained in the center of $E_\zeta^{(\ell)}$.
Hence it is sufficient to show $[A_1,U_\zeta]=[A_\zeta^{(\ell)},Z_{Fr}(U_\zeta)]=\{0\}$.
Let $j:U_\zeta\to U_\zeta^L$ be the homomorphism induced by the inclusion $U_{\mathbb{A}}\subset U^L_{\mathbb{A}}$.
Then we have
\[
u\varphi=\sum_{(u)}(j(u_{(0)})\cdot\varphi)u_{(1)}
\qquad
(u\in U_\zeta, \varphi\in A_\zeta)
\]
in $E_\zeta$.
Moreover,
$v\cdot\psi=\pi(v)\cdot\psi$ for $v\in U_\zeta^L$, $\psi\in A_1$.
Here, $\pi:U_\zeta^L\to U({\mathfrak{g}})$ is Lusztig's Frobenius morphism.
Hence it is sufficient to show $\pi(j(u))=\varepsilon(u)1$ for $u\in U_\zeta$ and $j(z)\cdot\varphi=\varepsilon(z)\varphi$ for $z\in Z_{Fr}(U_\zeta)$, $\varphi\in A_\zeta$.
The first statement is easily shown using the generators $e_i$, $f_i$, $k_\lambda$ ($i\in I$, $\lambda\in\Lambda$) of $U_\zeta$.
Since $j$ preserves the action of the braid group ${\mathbb{B}}$,
the second statement is a consequence of the fact that
$Z_{Fr}(U_\zeta)$ is generated by $e_i^\ell$, $f_i^\ell$, $k_{\ell\lambda}$ ($i\in I$, $\lambda\in\Lambda$) and their ${\mathbb{B}}$-conjugates.
\end{proof}
We denote by ${ZD'}_\zeta^{(\ell)}$, $ZD_\zeta^{(\ell)}$ the images of $ZE_\zeta^{(\ell)}$ in ${D'_\zeta}^{(\ell)}$, $D_\zeta^{(\ell)}$ respectively.
We have
\begin{align*}
\omega_{\mathcal B}^*ZE_\zeta^{(\ell)}&=
{\mathcal O}_{\mathcal B}\otimes_{\mathbb{C}} Z_{Fr}(U_\zeta)\otimes_{\mathbb{C}} {\mathbb{C}}[\Lambda]\\
&\cong
{\mathcal O}_{\mathcal B}\otimes_{\mathbb{C}} {\mathbb{C}}[K]\otimes_{\mathbb{C}} {\mathbb{C}}[H]\\
&\cong p_*{\mathcal O}_{{\mathcal B}\times K\times H},
\end{align*}
where $p:{\mathcal B}\times K\times H\to{\mathcal B}$ is the projection.
Note that we identify $Z_{Fr}(U_\zeta)$ and ${\mathbb{C}}[\Lambda]$ with ${\mathbb{C}}[K]$ and ${\mathbb{C}}[H]$ respectively (see \eqref{eq:Fr-center}).
Set
\[
{\mathcal Z}'_\zeta=\omega_{\mathcal B}^*{ZD'}_\zeta^{(\ell)},
\qquad
{\mathcal Z}_\zeta=\omega_{\mathcal B}^*{ZD}_\zeta^{(\ell)}.
\]
Then ${\mathcal Z}'_\zeta$ and ${\mathcal Z}_\zeta$ are central ${\mathcal O}_{\mathcal B}$-subalgebras of $Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}$ and $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$ respectively.
Moreover,
we have a sequence
\[
p_*{\mathcal O}_{{\mathcal B}\times K\times H}
\to
{\mathcal Z}'_\zeta
\to
{\mathcal Z}_\zeta
\]
of surjective ${\mathcal O}_{\mathcal B}$-algebra homomorphisms.
We define a subvariety ${\mathcal V}$ of ${\mathcal B}\times K\times H$ by
\[
{\mathcal V}=
\{
(B^-g,k,t)\in{\mathcal B}\times K\times H\mid
g\kappa(k)g^{-1}\in t^{2\ell}N^-\}.
\]
We denote by
\[
p_{\mathcal V}:{\mathcal V}\to{\mathcal B}
\]
the projection.
The aim of this section is to prove the following.
\begin{theorem}
\label{thm:center}
We have
\[
{\mathcal Z}'_\zeta\cong
{\mathcal Z}_\zeta
\cong
p_{{\mathcal V}*}{\mathcal O}_{\mathcal V}.
\]
\end{theorem}
\subsection{}
Set
\begin{align*}
\tilde{{\mathcal B}}=&N^-\backslash G,\\
\tilde{\mathcal V}=&
\{
(N^-g,k,t)\in\tilde{\mathcal B}\times K\times H\mid
g\kappa(k)g^{-1}\in t^{2\ell}N^-\}.
\end{align*}
\begin{lemma}
\label{lem:tilde-V}
$\tilde{{\mathcal V}}$ is a connected smooth variety with $\dim\tilde{{\mathcal V}}=2\dim\tilde{{\mathcal B}}$.
\end{lemma}
\begin{proof}
Set
\begin{align*}
{\mathcal W}&=
\{
(N^-g,x,t)\in\tilde{\mathcal B}\times G\times H\mid
gxg^{-1}\in t^{2\ell}N^-\},\\
{\mathcal W}^0&=\{
(N^-g,x,t)\in{\mathcal W}\mid
x\in N^+HN^-\}.
\end{align*}
Then ${\mathcal W}$ is a fiber bundle over $\tilde{{\mathcal B}}$ whose fiber at the origin $N^-\in\tilde{{\mathcal B}}$ is isomorphic to $H\times N^-$.
Hence ${\mathcal W}$ is a smooth connected variety with with $\dim{{\mathcal W}}=2\dim\tilde{{\mathcal B}}$.
It follows that its Zariski open subvariety ${\mathcal W}^0$ is also a smooth connected varietyn of the same dimension.
Note that $\kappa:K\to G$ is a composite of the Galois covering $\overline{\kappa}:K\to N^+HN^-$
with Galois group $\Gamma=\{\gamma\in H\mid\gamma^2=1\}$
and the open embedding $N^+HN^-\to G$.
By $\tilde{{\mathcal V}}\cong{\mathcal W}^0\times_{N^+HN^-}K$ we see that $\tilde{{\mathcal V}}$ is is a Galois covering of ${\mathcal W}^0$ with Galois group $\Gamma$.
Hence $\tilde{{\mathcal V}}$ is a smooth variety.
It remains to show that $\tilde{{\mathcal V}}$ is connected.
Since $\Gamma$ acts transitively on the set ${\mathcal X}$ of connected components of $\tilde{{\mathcal V}}$, it is sufficient to show that $\gamma(X)=X$ for any $\gamma\in\Gamma$ and $X\in{\mathcal X}$.
Since $G$ was chosen to be simply-connected, the group $\Gamma$ is generated by elements $\gamma_i\in\Gamma\,\,(i\in I)$ with $\theta_{\varpi_j}(\gamma_i)=1$
for $i\ne j$ and $\theta_{\varpi_i}(\gamma_i)=-1$.
Hence it is sufficient to show that $\gamma_i(X)=X$ for any $i\in I$ and $X\in{\mathcal X}$.
By the commutativity of $\Gamma$ we have only to show that for any $i\in I$ there exists some $X\in{\mathcal X}$ such that $\gamma_i(X)=X$.
Hence the proof is reduced to showing that for any $i\in I$ there exists some $v\in\tilde{{\mathcal V}}$ such that $v$ and $\gamma_i v$ are contained in the same connected component of $\tilde{{\mathcal V}}$.
We may assume that $G=SL_2({\mathbb{C}})$.
Then we can check the assertion by a direct computation.
Details are omitted.
\end{proof}
We regard $A_1$ as the ring of functions on the quasi-affine variety $\tilde{{\mathcal B}}$.
We also regard
$ZE_\zeta^{(\ell)}$ as the ring of functions on $\tilde{{\mathcal B}}\times K\times H$.
We denote by $\tilde{{\mathcal Z}}'_\zeta$, $\tilde{{\mathcal Z}}_\zeta$ the sheaf of ${\mathcal O}_{\tilde{{\mathcal V}}}$-algebras corresponding to
${ZD'}_\zeta^{(\ell)}$, ${ZD}_\zeta^{(\ell)}$ respectively.
In order to prove Theorem \ref{thm:center} it is sufficient to show
\[
\tilde{{\mathcal Z}}'_\zeta\cong\tilde{{\mathcal Z}}_\zeta\cong
{\mathcal O}_{\tilde{{\mathcal V}}},
\]
where $\tilde{{\mathcal V}}$ is regarded as a reduced scheme.
\begin{lemma}
\label{lem:defining-V}
For $\varphi\in A_1(\lambda)\subset A_\zeta(\ell\lambda)$ with $\lambda\in\Lambda^+$ we have
\[
\Omega_i(\varphi)\in ZE_\zeta^{(\ell)}\qquad(i=1, 2),
\]
and we have
\begin{align}
\label{eq:lem:defining-V1}
\Omega_1(\varphi)(N^-g,(n_1h,n_2h^{-1}),t)&=
\varphi(N^-t^\ell gn_2h^{-1}),
\\
\label{eq:lem:defining-V2}
\Omega_2(\varphi)(N^-g,(n_1h,n_2h^{-1}),t)&=
\varphi(N^-t^{-\ell} gn_1h)
\end{align}
for $g\in G$, $n_1\in N^+$, $n_2\in N^-$, $t, h\in H$.
\end{lemma}
\begin{proof}
We may assume that $\varphi\in A_1(\lambda)_\xi\subset
A_\zeta(\ell\lambda)_{\ell\xi}$.
Take bases $\{f_r\}_r$, $\{v_r\}_r$ of ${\mathbb{C}}[N^-]$ and $U({\mathfrak{n}}^-)$ respectively such that $\langle f_r,v_{r'}\rangle=\delta_{r, r'}$, where $\langle\,\,,\,\,\rangle:{\mathbb{C}}[N^-]\times U({\mathfrak{n}}^-)\to{\mathbb{C}}$ is the canonical Hopf paring.
Define $\{x_r\}_r\subset U_\zeta^+\cap Z_{Fr}(U_\zeta)$
by $\tau^L(x_r,y)=\langle f_r,\pi(y)\rangle$ for any $y\in U_\zeta^{L,-}$.
We can take
$\{y_r\}_r\subset U^{L,-}_\zeta$,
$\{y'_s\}_s\subset \mathop{\rm Ker\hskip.5pt}\nolimits(\pi|U^{L,-}_\zeta)$,
$\{x'_s\}_s\subset U^{+}_\zeta$,
such that $\pi(y_r)=v_r$ for any $r$,
$\{y_r\}_r\sqcup\{y'_s\}_s$ is a basis of $U_\zeta^{L,-}$,
$\{x_r\}_r\sqcup\{x'_s\}_s$ is a basis of $U_\zeta^{+}$,
$\tau^L(x'_s,y_r)=0$ for any $r, s$, and $\tau^L(x'_s,y'_{s'})=\delta_{s,s'}$.
Then we have
\begin{align*}
\Omega_1(\varphi)&=\sum_r(\pi(y_r)\cdot\varphi)x_rk_{-\ell\xi}e(\ell\lambda)
+
\sum_s(\pi(y'_s)\cdot\varphi)x'_sk_{-\ell\xi}e(\ell\lambda)\\
&=
\sum_r(v_r\cdot\varphi)x_rk_{-\ell\xi}e(\ell\lambda).
\end{align*}
By Lemma \ref{lem:unipotent} below we obtain
\begin{align*}
&\Omega_1(\varphi)(N^-g,(n_1h,n_2h^{-1}),t)
=\sum_r
((v_r\cdot\varphi)(N^-g))
f_r(n_2)
\theta_{-\xi}(h)
\theta_{\ell\lambda}(t)\\
=&((n_2\cdot\varphi)(N^-g))\theta_{\xi}(h^{-1})
\theta_{\lambda}(t^\ell)
=(n_2h^{-1}\cdot\varphi\cdot t^{\ell})(N^-g)\\
=&\varphi(N^-t^{\ell}gn_2h^{-1}).
\end{align*}
\eqref{eq:lem:defining-V1} is proved.
The proof of \eqref{eq:lem:defining-V2} is similar and omitted.
\end{proof}
The proof of the following result is standard and left to the readers.
\begin{lemma}
\label{lem:unipotent}
Let $N$ be a unipotent algebraic group over ${\mathbb{C}}$ with Lie algebra ${\mathfrak{n}}$.
Denote by $\sum_rv_r\otimes f_r$ the canonical element of (a completion of) ${\mathbb{C}}[N]\otimes U({\mathfrak{n}}^-)$ with respect to the canonical Hopf paring ${\mathbb{C}}[N]\times U({\mathfrak{n}})\to{\mathbb{C}}$.
Then for any finite dimensional $N$-module $M$ we have
\[
gm=\sum_rf_r(g)v_rm\qquad(g\in N, m\in M).
\]
\end{lemma}
We define $F_i:\tilde{{\mathcal B}}\times K\times H\to\tilde{{\mathcal B}}\,\,(i=1,2)$ by
\begin{align*}
&F_1(N^-g, (n_1h, n_2h^{-1}),t)=N^-t^{\ell} gn_2h^{-1},
\\
&F_2(N^-g, (n_1h, n_2h^{-1}),t)=N^-t^{-\ell} gn_1h
\end{align*}
for $g\in G, n_1\in N^+, n_2\in N^-, h, t\in H$.
\begin{lemma}
\label{lem:defining-V2}
The equations
\[
\varphi\circ F_1(x)=\varphi\circ F_2(x)
\qquad(\varphi\in A_1)
\]
for $x\in \tilde{{\mathcal B}}\times K\times H$ give defining equations of $\tilde{{\mathcal V}}$, which is reduced at any point of $\tilde{{\mathcal V}}$.
\end{lemma}
\begin{proof}
Since $\tilde{{\mathcal B}}$ is quasi-affine, the equations
\[
\varphi(x)=\varphi(y)
\qquad(\varphi\in A_1)
\]
for $(x,y)\in \tilde{{\mathcal B}}\times \tilde{{\mathcal B}}$ give defining equations of the diagonal subvariety
\[
(\tilde{{\mathcal B}}\times \tilde{{\mathcal B}})_{\rm diag}
=\{(x,x)\mid x\in\tilde{{\mathcal B}}\}
\]
of $\tilde{{\mathcal B}}\times \tilde{{\mathcal B}}$.
Hence by the cartesian diagram
\[
\begin{CD}
\tilde{{\mathcal V}}@>>>(\tilde{{\mathcal B}}\times \tilde{{\mathcal B}})_{\rm diag}\\
@VVV@VVV\\
\tilde{{\mathcal B}}\times K\times H
@>>{(F_1,F_2)}>
\tilde{{\mathcal B}}\times \tilde{{\mathcal B}}
\end{CD}
\]
it is sufficient to show that $(F_1,F_2):
\tilde{{\mathcal B}}\times K\times H
\to
\tilde{{\mathcal B}}\times \tilde{{\mathcal B}}$
is a smooth morphism.
Define
$\alpha:\tilde{{\mathcal B}}\times K\times H\to \tilde{{\mathcal B}}\times K\times H$,
$\beta:\tilde{{\mathcal B}}\times K\times H\to \tilde{{\mathcal B}}\times G\times H$ and
$\gamma:\tilde{{\mathcal B}}\times G\times H\to \tilde{{\mathcal B}}\times \tilde{{\mathcal B}}$ by
\begin{align*}
&\alpha(N^-g, (n_1h,n_2h^{-1}),t)
=(N^-t^{\ell}gn_2h^{-1},(n_1h,n_2h^{-1}),t),\\
&\beta(N^-g, (n_1h,n_2h^{-1}),t)
=(N^-g,hn_2^{-1}n_1h,t^2),\\
&\gamma(N^-g, x,t)
=(N^-g,N^-t^{-\ell}gx)
\end{align*}
for $g, x\in G$, $n_1\in N^+$, $n_2\in N^-$, $h, t\in H$.
Let us show that $\beta$ is smooth.
For that it is sufficient to show that $N^+\times N^-\times H
\ni(n_1,n_2,h)\mapsto hn_2^{-1}n_1h\in G$ is smooth.
This morphism is a composite of an isomorphism
$N^+\times N^-\times H
\ni(n_1,n_2,h)
\mapsto
(h^{-1}n_1h,hn_2h^{-1},h)\in
N^+\times N^-\times H$ and
a smooth morphism
$N^+\times N^-\times H
\ni(n_1,n_2,h)
\mapsto
n_2^{-1}h^2n_1\in
G$.
Hence $\beta$ is smooth.
Then by the cartesian diagram
\[
\begin{CD}
\tilde{{\mathcal B}}\times K\times H@>
{(F_1,F_2)}>>\tilde{{\mathcal B}}\times \tilde{{\mathcal B}}\\
@V{\alpha}VV@VV{\mathop{\rm id}\nolimits}V\\
\tilde{{\mathcal B}}\times K\times H
@>>{\gamma\circ\beta}>
\tilde{{\mathcal B}}\times \tilde{{\mathcal B}}
\end{CD}
\]
and the smoothness of $\beta$ it is sufficient to show that $\gamma$ is smooth.
Since the group $\tilde{G}=G\times H$ acts on $\tilde{{\mathcal B}}$ from the right by $N^-g\cdot(x,t)=N^-t^{-\ell}gx$, we can identify $\tilde{{\mathcal B}}$ with $\tilde{N}^-\backslash\tilde{G}$, where $\tilde{N}^-=\{(x,t)\in G\times H\mid t^{-\ell}x\in N^-\}$.
Under this identification $\gamma:(\tilde{N}^-\backslash\tilde{G})\times\tilde{G}\to
(\tilde{N}^-\backslash\tilde{G})\times(\tilde{N}^-\backslash\tilde{G})$ is given by
$\gamma(\tilde{N}^-\tilde{x},\tilde{g})=(\tilde{N}^-\tilde{x},\tilde{N}^-\tilde{x}\tilde{g})$, and hence the assertion is clear.
\end{proof}
By Lemma \ref{lem:defining-V}, Lemma \ref{lem:defining-V2} we have a sequence
\[
{\mathcal O}_{\tilde{\mathcal V}}\to\tilde{{\mathcal Z}}'_\zeta\to\tilde{{\mathcal Z}}_\zeta
\]
of surjective homomorphisms of ${\mathcal O}_{\tilde{{\mathcal B}}\times K\times H}$-algebras.
Hence in order to prove Theorem \ref{thm:center} it is sufficient to show that ${\mathcal O}_{\tilde{\mathcal V}}\to\tilde{{\mathcal Z}}_\zeta$ is an isomorphism.
\subsection{}
By De Concini-Procesi \cite[11.7]{DP} the center $Z$ of $E_\zeta^{(\ell)}$ is endowed with a Poisson algebra structure by
\begin{equation}
\label{eq:Poisson:E}
\{\overline{a},\overline{b}\}
=\overline{[a,b]/\ell(q^{\ell}-q^{-\ell})}
\qquad(a, b\in E_{\mathbb{A}}^{(\ell)}, \overline{a}, \overline{b}\in Z).
\end{equation}
Here
\[
E_{\mathbb{A}}^{(\ell)}=\bigoplus_{\lambda\in\Lambda^+}
A_{\mathbb{A}}(\ell\lambda)\otimes U_{\mathbb{A}}\otimes{\mathbb{A}}[\Lambda]
\subset E_{\mathbb{A}}.
\]
Note that
$ZE_\zeta^{(\ell)}$ is a subalgebra of $Z$.
We will show that $ZE_\zeta^{(\ell)}$ is a Poisson subalgebra of $Z$.
We endow $C_{\mathbb{A}}\otimes_{\mathbb{A}} U_{\mathbb{A}}$ with an ${\mathbb{A}}$-algebra structure such that $C_{\mathbb{A}}\otimes 1$, $1\otimes U_{\mathbb{A}}$ are subalgebras naturally isomorphic to $C_{\mathbb{A}}$, $U_{\mathbb{A}}$ respectively, and
\[
(1\otimes u)(\varphi\otimes 1)=
\sum_{(u)}u_{(0)}\cdot\varphi\otimes u_{(1)}\qquad
(u\in U_{\mathbb{A}}, \varphi\in C_{\mathbb{A}}).
\]
Then the center $Z(C_\zeta\otimes U_\zeta)$ of $C_\zeta\otimes U_\zeta$ is endowed with a Poisson algebra structure by
\begin{equation}
\label{eq:Poisson:CU}
\{\overline{a},\overline{b}\}
=\overline{[a,b]/\ell(q^{\ell}-q^{-\ell})}
\quad(a, b\in C_{\mathbb{A}}\otimes_{\mathbb{A}} U_{\mathbb{A}}, \overline{a}, \overline{b}\in Z(C_\zeta\otimes U_\zeta)).
\end{equation}
Similarly to Lemma \ref{lem:center}
we see that ${\mathbb{C}}[G]\otimes Z_{Fr}(U_\zeta)$ is a subalgebra of $Z(C_\zeta\otimes U_\zeta)$.
We first give a description of the Poisson bracket of ${\mathbb{C}}[G]\otimes Z_{Fr}(U_\zeta)$.
Denote by ${\mathfrak{k}}$ the Lie algebra of $K$.
We identify $K$ with a subgroup of $G\times G$ and regard ${\mathfrak{k}}$ as a subalgebra of ${\mathfrak{g}}\oplus {\mathfrak{g}}$.
Namely,
\[
{\mathfrak{k}}=\{(h+a,-h+b)\mid h\in{\mathfrak{h}}, a\in{\mathfrak{n}}^+, b\in{\mathfrak{n}}^-\}.
\]
Denote by $S$ the diagonal subgroup $\{(g,g)\mid g\in G\}$ of $G\times G$.
Then its Lie algebra ${\mathfrak{s}}$ is given by
\[
{\mathfrak{s}}=\{(x,x)\mid x\in {\mathfrak{g}}\}
\subset{\mathfrak{g}}\oplus{\mathfrak{g}}.
\]
In particular, we have ${\mathfrak{g}}\oplus{\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{s}}$.
We sometimes identify $S$ with $G$ by
$(g,g)\leftrightarrow g$.
Define a symmetric bilinear form
$\tilde{\epsilon}$ on ${\mathfrak{g}}\oplus{\mathfrak{g}}$ by
\[
\tilde{\epsilon}((x_1,x_2),(y_1,y_2))
=\epsilon(x_1,y_1)-\epsilon(x_2,y_2)
\qquad(x_1, x_2, y_1, y_2\in{\mathfrak{g}}),
\]
where $\epsilon:{\mathfrak{g}}\times{\mathfrak{g}}\to{\mathbb{C}}$ is the invariant symmetric bilinear form on ${\mathfrak{g}}$ whose restriction to ${\mathfrak{h}}\times{\mathfrak{h}}$ induces the bilinear form \eqref{eq:killing} on ${\mathfrak{h}}^*$.
Then $\tilde{\epsilon}|{\mathfrak{k}}\times{\mathfrak{k}}$ and $\tilde{\epsilon}|{\mathfrak{s}}\times{\mathfrak{s}}$ are identically zero, and $\tilde{\epsilon}|{\mathfrak{k}}\times{\mathfrak{s}}$ is non-degenerate.
In other words $({\mathfrak{g}}\oplus{\mathfrak{g}},{\mathfrak{k}},{\mathfrak{s}})$ is a Manin triple with respect to $\tilde{\epsilon}$.
In particular, ${\mathbb{C}}[K]$ and ${\mathbb{C}}[S]$ are Poisson Hopf algebras
(see Drinfeld \cite{D}, De Concini-Procesi \cite{DP}).
We will sometimes identify ${\mathfrak{g}}^*$ and ${\mathfrak{k}}^*$ with ${\mathfrak{k}}$ and ${\mathfrak{g}}$ respectively via the non-degenerate paring $\tilde{\epsilon}|{\mathfrak{k}}\times{\mathfrak{s}}$ and the identification ${\mathfrak{g}}\cong{\mathfrak{s}}$.
In general let $A$ be an algebraic group over ${\mathbb{C}}$ with Lie algebra ${\mathfrak{a}}$.
For $a\in{\mathfrak{a}}$ we define vector fields $L_a$, $R_a$ on $A$ by
\begin{align*}
&(L_af)(g)=\frac{d}{dt}f(g\exp(ta))|_{t=0}\qquad
(g\in A, f\in{\mathcal O}_{A,g}),\\
&(R_af)(g)=\frac{d}{dt}f(\exp(-ta)g)|_{t=0}\qquad
(g\in A, f\in{\mathcal O}_{A,g}).
\end{align*}
For $\xi\in{\mathfrak{a}}^*$ we define 1-forms
$L^*_\xi$, $R^*_\xi$ on $A$ by
\[
\langle L_a, L^*_\xi\rangle=
\langle R_a, R^*_\xi\rangle=\langle a,\xi\rangle\qquad
(a\in{\mathfrak{a}}).
\]
By De Concini-Lyubashenko \cite{DL}, De Concini-Procesi \cite{DP}, Gavarini \cite{Gav}, and Tanisaki \cite{T3} we have the following.
\begin{proposition}
\label{prop:Poisson1}
${\mathbb{C}}[G]\otimes Z_{Fr}(U_\zeta)$ is closed under the Poisson bracket \eqref{eq:Poisson:CU}.
More precisely we have the following.
\begin{itemize}
\item[\rm(i)]
${\mathbb{C}}[G]$ is closed under the Poisson bracket \eqref{eq:Poisson:CU}.
Moreover, the isomorphism ${\mathbb{C}}[G]\cong{\mathbb{C}}[S]$ induced by the identification $G\cong S$ preserves the Poisson structures,
where ${\mathbb{C}}[S]$ is regarded as a Poisson algebra via the Manin triple $({\mathfrak{g}}\oplus{\mathfrak{g}},{\mathfrak{k}},{\mathfrak{s}})$.
Namely, the Poisson tensor
$\delta\in\bigwedge^2TG$
is given by
\begin{align*}
&\delta_g((L^*_\eta)_g,(L^*_{\eta'})_g)
=\tilde{\epsilon}(p_{\mathfrak{s}}(\widetilde{\mathop{\rm Ad}\nolimits}(g)(\eta)),\widetilde{\mathop{\rm Ad}\nolimits}(g)(\eta')),\\
&\delta_g((R^*_\eta)_g,(R^*_{\eta'})_g)
=-\tilde{\epsilon}(p_{\mathfrak{s}}(\widetilde{\mathop{\rm Ad}\nolimits}(g^{-1})(\eta)),\widetilde{\mathop{\rm Ad}\nolimits}(g^{-1})(\eta'))
\end{align*}
for $g\in G$, $\eta, \eta'\in{\mathfrak{k}}\cong{\mathfrak{g}}^*$.
Here, $\widetilde{\mathop{\rm Ad}\nolimits}:G\to GL({\mathfrak{g}}\oplus{\mathfrak{g}})$ is the restriction of the adjoint action of $G\times G$ on ${\mathfrak{g}}\oplus{\mathfrak{g}}$ to $G\cong S$, and $p_{\mathfrak{s}}:{\mathfrak{g}}\oplus{\mathfrak{g}}\to{\mathfrak{s}}$ is the projection with respect to the direct sum decomposition ${\mathfrak{g}}\oplus{\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{s}}$.
\item[\rm(ii)]
$Z_{Fr}(U_\zeta)$ is closed under the Poisson bracket \eqref{eq:Poisson:CU}.
Moreover, the isomorphism $Z_{Fr}(U_\zeta)\cong{\mathbb{C}}[K]$ $($see \eqref{eq:Fr-center}$)$ preserves the Poisson structures,
where ${\mathbb{C}}[K]$ is regarded as a Poisson algebra via the Manin triple $({\mathfrak{g}}\oplus{\mathfrak{g}},{\mathfrak{k}},{\mathfrak{s}})$.
Namely, the Poisson tensor
$\delta\in\bigwedge^2TK$
is given by
\begin{align*}
&\delta_k((L^*_\xi)_k,(L^*_{\xi'})_k)
=\tilde{\epsilon}(p_{\mathfrak{k}}(\widetilde{\mathop{\rm Ad}\nolimits}(k)(\xi)),\widetilde{\mathop{\rm Ad}\nolimits}(k)(\xi')),\\
&\delta_k((R^*_\xi)_k,(R^*_{\xi'})_k)
=-\tilde{\epsilon}(p_{\mathfrak{k}}(\widetilde{\mathop{\rm Ad}\nolimits}(k^{-1})(\xi)),\widetilde{\mathop{\rm Ad}\nolimits}(k^{-1})(\xi'))
\end{align*}
for $k\in K$, $\xi, \xi'\in{\mathfrak{g}}\cong{\mathfrak{k}}^*$.
Here, $\widetilde{\mathop{\rm Ad}\nolimits}:K\to GL({\mathfrak{g}}\oplus{\mathfrak{g}})$ is the restriction of the adjoint action of $G\times G$ on ${\mathfrak{g}}\oplus{\mathfrak{g}}$ to $K$, and $p_{\mathfrak{k}}:{\mathfrak{g}}\oplus{\mathfrak{g}}\to{\mathfrak{k}}$ is the projection with respect to the direct sum decomposition ${\mathfrak{g}}\oplus{\mathfrak{g}}={\mathfrak{k}}\oplus{\mathfrak{s}}$.
\item[\rm(iii)]
The restriction of the Poisson tensor $\delta\in\bigwedge^2T(G\times K)$ with respect to \eqref{eq:Poisson:CU} to $TG\otimes TK$ is given by
\[
\delta_{(g,k)}((L^*_\eta)_g,(R^*_\xi)_k)=\tilde{\epsilon}(\xi,\eta)
\]
for
$g\in G, k\in K, \eta\in{\mathfrak{k}}\cong{\mathfrak{g}}^*, \xi\in{\mathfrak{g}}\cong{\mathfrak{k}}^*$.
\end{itemize}
\end{proposition}
\begin{lemma}
\label{lem:Poisson3}
The Poisson structure of ${\mathbb{C}}[G]$ induces that of ${\mathbb{C}}[N^-\backslash G]$.
\end{lemma}
\begin{proof}
It is sufficient to show that $A_1$ is closed under the Poisson bracket of ${\mathbb{C}}[G]$ given by
\[
\{\overline{a},\overline{b}\}
=\overline{[a,b]/\ell(q^{\ell}-q^{-\ell})}
\qquad(a, b\in C_{\mathbb{A}}, \overline{a},\overline{b}\in{\mathbb{C}}[G]).
\]
Let $\varphi, \psi\in A_1$.
Then we have
$\{\varphi,\psi\}=\overline{[a,b]/\ell(q^{\ell}-q^{-\ell})}$
for $a, b\in C_{\mathbb{A}}$ such that $\overline{a}=\varphi$ and $\overline{b}=\psi$.
We may assume $a, b\in A_{\mathbb{A}}$.
Then we have $\{\varphi,\psi\}\in{\mathbb{C}}[G]\cap A_\zeta=A_1$.
\end{proof}
For $a\in{\mathfrak{g}}$ we denote by $\overline{L}_a$ the vector field on $N^-\backslash G$ induced by $L_a$.
Namely,
\[
(\overline{L}_af)(N^-g)
=
\frac{d}{dt}f(N^-g\exp(ta))|_{t=0}\qquad
(g\in G, f\in{\mathcal O}_{N^-\backslash G,N^-g}).
\]
Then we have a surjective linear map
\[
{\mathfrak{g}}\to T(N^-\backslash G)_{N^-g}
\qquad(a\mapsto(\overline{L}_a)_{N^-g})
\]
with kernel $\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{n}}^-)$.
Hence we have
\begin{align*}
&T^*(N^-\backslash G)_{N^-g}\\
\cong&\{\eta\in{\mathfrak{g}}^*\mid\langle\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{n}}^-),\eta
\rangle=\{0\}\}\\
\cong&\{\eta\in{\mathfrak{k}}\mid
\tilde{\epsilon}(
\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{n}}^-),\eta
)=\{0\}\}\\
=&\{(y_1,y_2)\in{\mathfrak{k}}\mid
\epsilon(
\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{n}}^-),y_1-y_2
)=\{0\}\}\\
=&\{(y_1,y_2)\in{\mathfrak{k}}\mid
y_1-y_2\in\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{b}}^-)
\}.
\end{align*}
For $N^-g\in N^-\backslash G$ we set
\begin{align*}
{\mathfrak{k}}_{N^-g}=&
\{\eta\in{\mathfrak{g}}^*\mid\langle\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{n}}^-),\eta
\rangle=\{0\}\}\\
=&
\{(y_1,y_2)\in{\mathfrak{k}}\mid
y_1-y_2\in\mathop{\rm Ad}\nolimits(g^{-1})({\mathfrak{b}}^-)
\}.
\end{align*}
For $\eta\in{\mathfrak{k}}_{N^-g}$ we define $\overline{L}^*_\eta\in T^*(N^-\backslash G)_{N^-g}$ by
\[
\langle \overline{L}_a, \overline{L}^*_\eta\rangle=\langle a,\eta\rangle\qquad
(a\in{\mathfrak{g}}).
\]
We easily obtain the following from Proposition \ref{prop:Poisson1}.
\begin{proposition}
\label{prop:Poisson2}
$ZE_\zeta^{(\ell)}$ is closed under the Poisson bracket \eqref{eq:Poisson:E}.
Moreover, under the identification
$ZE_\zeta^{(\ell)}\cong{\mathbb{C}}[N^-\backslash G]\otimes {\mathbb{C}}[K]\otimes{\mathbb{C}}[H]$
the corresponding Poisson tensor of the manifold
$(N^-\backslash G)\times K\times H$
is given by
\begin{align*}
&\delta_{(N^-g,k,t)}(
\overline{L}^*_\eta,\overline{L}^*_{\eta'})
=\tilde{\epsilon}(p_{\mathfrak{s}}(\widetilde{\mathop{\rm Ad}\nolimits}(g)(\eta)),\widetilde{\mathop{\rm Ad}\nolimits}(g)(\eta'))
&(\eta, \eta'\in{\mathfrak{k}}_{N^-g}),
\\
&\delta_{(N^-g,k,t)}(
R^*_\xi,R^*_{\xi'})
=-\tilde{\epsilon}(p_{\mathfrak{k}}(\widetilde{\mathop{\rm Ad}\nolimits}(k^{-1})(\xi)),\widetilde{\mathop{\rm Ad}\nolimits}(k^{-1})(\xi'))
&(\xi, \xi'\in{\mathfrak{g}}),
\\
&\delta_{(N^-g,k,t)}(
\overline{L}^*_\eta,R^*_\xi)
=\tilde{\epsilon}(\xi,\eta)
&(\eta\in{\mathfrak{k}}_{N^-g}, \xi\in{\mathfrak{g}}),\\
&\delta_{(N^-g,k,t)}(\overline{L}^*_\eta,
L^*_\lambda)=
-\tilde{\epsilon}(\mathop{\rm Ad}\nolimits(g^{-1})(c_\lambda),\eta)/2\ell
&(\eta\in{\mathfrak{k}}_{N^-g}, \lambda\in{\mathfrak{h}}^*),
\\
&\delta_{(N^-g,k,t)}
((T^*H)_t,(T^*K)_{k}\oplus(T^*H)_t)=\{0\}.
\end{align*}
Here, $c_\lambda$ for $\lambda\in{\mathfrak{h}}^*$ denotes the element of ${\mathfrak{h}}$ such that $\mu(c_\lambda)=(\lambda,\mu)$ for any $\mu\in{\mathfrak{h}}^*$.
\end{proposition}
\begin{proposition}
\label{prop:symplectic}
$\tilde{{\mathcal V}}$ is a Poisson submanifold of $(N^-\backslash G)\times K\times H$ with non-degenerate Poisson tensor.
In particular $\tilde{{\mathcal V}}$ is a symplectic manifold.
\end{proposition}
\begin{proof}
Denote by
\[
\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})
\subset
(T^*(N^-\backslash G))_{N^-g}\oplus
(T^*K)_k\oplus (T^*H)_t
\]
the radical of the Poisson tensor $\delta$ at $(N^-g,k,t)
\in(N^-\backslash G)\times K\times H$.
Then it is sufficient to show $((T\tilde{{\mathcal V}})_{(N^-g,k,t)})^\perp
=\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})$
for any $(N^-g,k,t)\in\tilde{{\mathcal V}}$.
Here, $(T\tilde{{\mathcal V}})_{(N^-g,k,t)}$ is identified with a subspace of $(T(N^-\backslash G))_{N^-g}\oplus
(TK)_k\oplus (TH)_t$, and $((T\tilde{{\mathcal V}})_{(N^-g,k,t)})^\perp$ denotes the subspace of $(T^*(N^-\backslash G))_{N^-g}\oplus
(T^*K)_k\oplus (T^*H)_t$ which is orthogonal to $(T\tilde{{\mathcal V}})_{(N^-g,k,t)}$ with respect to the canonical paring between the tangent and the cotangent spaces.
Let us first compute $\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})$ using Proposition \ref{prop:Poisson2}.
Assume $y=\overline{L}^*_\eta+R^*_\xi+L^*_\lambda\in
\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})$ for
$\eta=(\eta_1,\eta_2)\in{\mathfrak{k}}_{N^-g}$, $\xi\in {\mathfrak{g}}$, $\lambda\in{\mathfrak{h}}$.
Note that the condition for $(\eta_1,\eta_2)\in{\mathfrak{k}}$ to be contained in ${\mathfrak{k}}_{N^-g}$ is equivalent to
\begin{equation}
\label{eq:prop:symplectic:1}
\epsilon(\mathop{\rm Ad}\nolimits(g)(\eta_1-\eta_2),{\mathfrak{n}}^-)=\{0\}.
\end{equation}
By $\delta_{(N^-g,k,t)}(y,R^*_{\xi'})=0$ for any $\xi'\in{\mathfrak{g}}$
we have
\begin{equation}
\label{eq:prop:symplectic:2}
\mathop{\rm Ad}\nolimits(k^{-1})(\eta)-\tilde{\mathop{\rm Ad}\nolimits}(k^{-1})(\xi)\in{\mathfrak{s}}.
\end{equation}
By $\delta_{(N^-g,k,t)}(y,L^*_\mu)=0$ for any $\mu\in{\mathfrak{h}}^*$
we have
\begin{equation}
\label{eq:prop:symplectic:3}
\epsilon(\mathop{\rm Ad}\nolimits(g)(\eta_1-\eta_2),{\mathfrak{h}})=\{0\}.
\end{equation}
By $\delta_{(N^-g,k,t)}(y,\overline{L}^*_{\eta'})=0$ for any $\eta'\in{\mathfrak{k}}_{N^-g}$
we have
\begin{equation}
\label{eq:prop:symplectic:4}
p_{\mathfrak{s}}(\tilde{\mathop{\rm Ad}\nolimits}(g)(\eta))-\mathop{\rm Ad}\nolimits(g)(\xi)+\frac{1}{2\ell}c_\lambda
\in{\mathfrak{n}}^-.
\end{equation}
By \eqref{eq:prop:symplectic:1} and \eqref{eq:prop:symplectic:3} we have
\begin{equation}
\label{eq:prop:symplectic:5}
\mathop{\rm Ad}\nolimits(g)(\eta_1-\eta_2)\in{\mathfrak{n}}^-.
\end{equation}
By
\[
\tilde{\mathop{\rm Ad}\nolimits}(g)(\eta)
=(\mathop{\rm Ad}\nolimits(g)(\eta_1),\mathop{\rm Ad}\nolimits(g)(\eta_1))+(0,-\mathop{\rm Ad}\nolimits(g)(\eta_1-\eta_2))
\]
and \eqref{eq:prop:symplectic:5} we have
$p_{\mathfrak{s}}(\tilde{\mathop{\rm Ad}\nolimits}(g)(\eta))=\mathop{\rm Ad}\nolimits(g)(\eta_1)$.
Hence \eqref{eq:prop:symplectic:4} is equivalent to
\begin{equation}
\label{eq:prop:symplectic:6}
\xi=\eta_1+\mathop{\rm Ad}\nolimits(g^{-1})(c_\lambda/2\ell+z)\qquad
(z\in{\mathfrak{n}}^-).
\end{equation}
Substituting \eqref{eq:prop:symplectic:6} to \eqref{eq:prop:symplectic:2} we obtain
\begin{equation}
\label{eq:prop:symplectic:7}
\mathop{\rm Ad}\nolimits(g\kappa(k)^{-1}g^{-1})(c_\lambda/2\ell+z)=
c_\lambda/2\ell+z+\mathop{\rm Ad}\nolimits(g)(\eta_1-\eta_2).
\end{equation}
In the case $(N^-g,k,t)\in\tilde{{\mathcal V}}$ we have $g\kappa(k)^{-1}g^{-1}\in t^{-2\ell}N^-$ and hence
\[
\mathop{\rm Ad}\nolimits(g\kappa(k)^{-1}g^{-1})(c_\lambda/2\ell+z)
-(c_\lambda/2\ell+z)\in{\mathfrak{n}}^-.
\]
Therefore, for each $\lambda\in{\mathfrak{h}}^*$ and $z\in{\mathfrak{n}}^-$ there exists unique $\eta=(\eta_1,\eta_2)\in{\mathfrak{k}}$ satisfying \eqref{eq:prop:symplectic:5}, \eqref{eq:prop:symplectic:7}.
We conclude that
$\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})$
for $(N^-g,k,t)\in\tilde{{\mathcal V}}$ consists of
\begin{equation}
\label{eq:prop:symplectic:7a}
y(\lambda,z)=\overline{L}^*_\eta+R^*_\xi+L^*_\lambda
\qquad(\lambda\in{\mathfrak{h}}^*, z\in{\mathfrak{n}}^-),
\end{equation}
where $\eta=(\eta_1,\eta_2)\in{\mathfrak{k}}_{N^-g}$ and $\xi\in{\mathfrak{g}}$ are uniquely determined by
\eqref{eq:prop:symplectic:7} and \eqref{eq:prop:symplectic:6}.
In particular we have $\dim \mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)})=\dim\tilde{{\mathcal B}}$.
Since the codimension of $\tilde{{\mathcal V}}$ in $\tilde{{\mathcal B}}\times K\times H$ is also $\dim\tilde{{\mathcal B}}$, we have only to show
\[
\langle
\mathop{\rm rad}\nolimits(\delta_{(N^-g,k,t)}),
(T\tilde{{\mathcal V}})_{(N^-g,k,t)}\rangle=\{0\}.
\]
By the description of $\tilde{{\mathcal V}}$
as a covering of an open subset of ${\mathcal W}$
(see proof of Lemma \ref{lem:tilde-V} for the notation)
we see easily that $(T\tilde{{\mathcal V}})_{(N^-g,k,t)}$ is spanned by the tangent vectors
$\overline{L}_a+R_b\,\,
(a\in{\mathfrak{g}}, b=(b_1,b_2)\in{\mathfrak{k}})$
with
\begin{equation}
\label{eq:prop:symplectic:8}
\mathop{\rm Ad}\nolimits(\kappa(k))(a)-a-\mathop{\rm Ad}\nolimits(\kappa(k))(b_2)+b_1=0,
\end{equation}
and
$R_{b'}+L_c\,\,(b'=(b'_1,b'_2)\in{\mathfrak{k}}, c\in{\mathfrak{h}})$ with
\begin{equation}
\label{eq:prop:symplectic:9}
\mathop{\rm Ad}\nolimits(g)(b'_1-\mathop{\rm Ad}\nolimits(\kappa(k))(b'_2))-2\ell c\in{\mathfrak{n}}^-.
\end{equation}
Take $y(\lambda,z)$ as in \eqref{eq:prop:symplectic:7a},
and set $u=c_\lambda/2\ell+z$.
For $\overline{L}_a+R_b$ satisfying \eqref{eq:prop:symplectic:8} we have
\begin{align*}
&\langle y(\lambda,z),\overline{L}_a+R_b\rangle\\
=&\epsilon(a,\eta_1-\eta_2)+\epsilon(b_1-b_2,\xi)\\
=&\epsilon(a,\mathop{\rm Ad}\nolimits(\kappa(k)^{-1}g^{-1})(u)-\mathop{\rm Ad}\nolimits(g^{-1})(u))
+\epsilon(b_1-b_2,\eta_1+\mathop{\rm Ad}\nolimits(g^{-1})(u))\\
=&\epsilon(\mathop{\rm Ad}\nolimits(\kappa(k))(a)-a,\mathop{\rm Ad}\nolimits(g^{-1})(u))
+\epsilon(b_1-b_2,\eta_1+\mathop{\rm Ad}\nolimits(g^{-1})(u))\\
=&\epsilon(-b_1+\mathop{\rm Ad}\nolimits(\kappa(k))(b_2),\mathop{\rm Ad}\nolimits(g^{-1})(u))
+\epsilon(b_1-b_2,\eta_1+\mathop{\rm Ad}\nolimits(g^{-1})(u))\\
=&\epsilon(\mathop{\rm Ad}\nolimits(\kappa(k))(b_2)-b_2,\mathop{\rm Ad}\nolimits(g^{-1})(u))
+\epsilon(b_1-b_2,\eta_1)\\
=&\epsilon(b_2,\mathop{\rm Ad}\nolimits(\kappa(k)^{-1}g^{-1})(u)-\mathop{\rm Ad}\nolimits(g^{-1})(u))
+\epsilon(b_1-b_2,\eta_1)\\
=&\epsilon(b_2,\eta_1-\eta_2)
+\epsilon(b_1-b_2,\eta_1)\\
=&-\epsilon(b_2,\eta_2)
+\epsilon(b_1,\eta_1)\\
=&0.
\end{align*}
For $R_{b'}+L_c$ satisfying \eqref{eq:prop:symplectic:9} we have
\begin{align*}
&\langle y(\lambda,z),R_{b'}+L_c\rangle\\
=&\epsilon(\xi,b_1'-b_2')+\lambda(c)\\
=&\epsilon(\eta_1+\mathop{\rm Ad}\nolimits(g^{-1})(u),b_1'-b_2')+\lambda(c)\\
=&\epsilon(u,\mathop{\rm Ad}\nolimits(g)(b_1'-b_2'))+
\epsilon(\eta_1,b_1'-b_2')+\lambda(c)\\
=&\epsilon(u,\mathop{\rm Ad}\nolimits(g\kappa(k))(b_2')-\mathop{\rm Ad}\nolimits(g)(b_2'))
-\epsilon(c_\lambda/2\ell,2\ell c)+
\epsilon(\eta_1,b_1'-b_2')+\lambda(c)\\
=&\epsilon(\mathop{\rm Ad}\nolimits(\kappa(k^{-1})g^{-1})(u)-\mathop{\rm Ad}\nolimits(g^{-1})(u),b_2')+
\epsilon(\eta_1,b_1'-b_2')\\
=&\epsilon(\eta_1-\eta_2,b_2')+
\epsilon(\eta_1,b_1'-b_2')\\
=&-\epsilon(\eta_2,b_2')+
\epsilon(\eta_1,b_1')\\
=&0.
\end{align*}
The proof is complete.
\end{proof}
\subsection{}
Let us finish the proof of Theorem \ref{thm:center}.
Set $J_1=\mathop{\rm Ker\hskip.5pt}\nolimits(ZE_\zeta^{(\ell)}\to {ZD}_\zeta^{(\ell)})$.
Since $ZE_\zeta^{(\ell)}\to {ZD}_\zeta^{(\ell)}$ is a morphism of Poisson algebras,
$J_1$ is a Poisson ideal of $ZE_\zeta^{(\ell)}$.
Hence $\sqrt{J_1}$ is also a Poisson ideal.
It follows that the support of $\tilde{{\mathcal Z}}_\zeta$ is a Poisson subvariety of $\tilde{{\mathcal B}}\times K\times H$ contained in $\tilde{\mathcal V}$.
By Lemma \ref{lem:tilde-V} and Proposition \ref{prop:symplectic}
$\tilde{\mathcal V}$ contains no non-empty Poisson subvariety except for $\tilde{\mathcal V}$ itself.
Therefore, we have only to show $\tilde{{\mathcal Z}}_\zeta\ne0$.
Since $A_1$ contains no non-trivial zero divisors,
the composit of
$A_1\to D_\zeta\to\mathop{\rm{End}}\nolimits_{\mathbb{C}}(A_\zeta)$ is injective, where $A_1\to D_\zeta$ is given by $\varphi\mapsto\ell_\varphi$.
Hence $A_1\to{ZD}_\zeta^{(\ell)}$ is also injective.
It follows that $\tilde{{\mathcal Z}}_\zeta\supset{\mathcal O}_{\tilde{{\mathcal B}}}\ne0$.
The proof of Theorem \ref{thm:center} is now complete.
\section{Azumaya properties}
\label{sec:Azumaya}
\subsection{}
By Lemma \ref{lem:center} and Theorem \ref{thm:center}
$Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}$ and $Fr_*{\mathcal D}_{{\mathcal B}_\zeta}$
are sheaves of ${\mathcal O}_{{\mathcal B}}$-algebras containing
$p_{{\mathcal V}*}{\mathcal O}_{\mathcal V}$ as a central subalgebra.
Hence we can consider their localizations
\[
\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}
=p_{{\mathcal V}}^{-1}Fr_*{\mathcal D}'_{{\mathcal B}_\zeta}
\otimes_{p_{{\mathcal V}}^{-1}p_{{\mathcal V}*}{\mathcal O}_{\mathcal V}}
{\mathcal O}_{\mathcal V},\qquad
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}
=p_{{\mathcal V}}^{-1}Fr_*{\mathcal D}_{{\mathcal B}_\zeta}
\otimes_{p_{{\mathcal V}}^{-1}p_{{\mathcal V}*}{\mathcal O}_{\mathcal V}}
{\mathcal O}_{\mathcal V}.
\]
on ${\mathcal V}$.
They are ${\mathcal O}_{\mathcal V}$-algebras, and we have a natural ${\mathcal O}_{{\mathcal B}}$-algebra homomorphism
$
\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}
\to
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}.
$
The first purpose of this section is to prove the following.
\begin{theorem}
\label{thm:Azumaya}
We have
\[
\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}
\cong
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}.
\]
Moreover, $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is an Azumaya algebra of
rank $\ell^{2|\Delta^+|}$ on ${\mathcal V}$.
Namely, $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is locally free as an ${\mathcal O}_{\mathcal V}$-module, and for any $v\in{\mathcal V}$ the fiber $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}(v)$ is isomorphic to the matrix algebra $M_{\ell^{|\Delta^+|}}({\mathbb{C}})$ as a ${\mathbb{C}}$-algebra.
\end{theorem}
We need some preliminaries.
\begin{lemma}
\label{lem:T-on-ZE}
For $w\in W$ regard $\Theta_w\subset A_1$ as a subset of $E_\zeta$.
Then with respect to the action of ${\mathbb{B}}$ on $E_\zeta$ we have $T_{w^{-1}}^{-1}\star \Theta_{e}=\Theta_{w}$.
Moreover, we have $T\star ZE_\zeta^{(\ell)}=ZE_\zeta^{(\ell)}$ for any $T\in{\mathbb{B}}$.
Hence we have also $T\star Z{D'_\zeta}^{(\ell)}=Z{D'_\zeta}^{(\ell)}$ and $T\star ZD_\zeta^{(\ell)}=ZD_\zeta^{(\ell)}$ for any $T\in{\mathbb{B}}$.
\end{lemma}
\begin{proof}
The first half is already shown in Lemma \ref{lem:braid-action-on-D}.
In view of Lemma \ref{lem:braid-D}
the only non-trivial part is to show $T_i^{\pm1}\star A_1\subset ZE_\zeta^{(\ell)}$ for $i\in I$.
Let $\varphi\in A_1$.
By Lemma \ref{lem:braid-Fr} we have $T_i^{\pm1}(\varphi)\in A_1$.
Then we see easily that $T_i^{\pm1}\star\varphi\in A_1\otimes Z_{Fr}(U_\zeta)$ by Lemma \ref{lem:braid-D} (ii).
\end{proof}
\begin{lemma}
$\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}$ is locally generated by $\ell^{2|\Delta^+|}$ sections.
\end{lemma}
\begin{proof}
It is sufficient to show that for any $w\in W$ the $(\Theta_{w}^{-1}Z{D'_\zeta}^{(\ell)})(0)$-module $(\Theta_{w}^{-1}{D'_\zeta}^{(\ell)})(0)$ is generated by $\ell^{2|\Delta^+|}$ elements.
By Lemma \ref{lem:T-on-ZE} we have $T_{w^{-1}}^{-1}\star \Theta_{e}=\Theta_{w}$ and $T_{w^{-1}}^{-1}\star Z{D'_\zeta}^{(\ell)}=Z{D'_\zeta}^{(\ell)}$.
Hence we may assume $w=e$ from the beginning.
Note that $(\Theta_{e}^{-1}{D'_\zeta}^{(\ell)})(0)$ is generated by the elements
\[
\jmath(u), \,\,\jmath(\Phi), \,\,\jmath(e(\lambda))\qquad
(u\in U_\zeta, \Phi\in(\Theta_{e}^{-1}A_\zeta)(0), \lambda\in\Lambda),
\]
while $(\Theta_{e}^{-1}Z{D'_\zeta}^{(\ell)})(0)$ is generated by the elements
\[
\jmath(u), \,\,\jmath(\Phi), \,\,\jmath(e(\lambda))\qquad
(u\in Z_{Fr}(U_\zeta), \Phi\in(\Theta_{e}^{-1}A_1)(0), \lambda\in\Lambda).
\]
We first show
\[
\jmath(y_pk_{\beta_p})\in
\jmath((\Theta_e^{-1}A_\zeta)(0)
U_\zeta^{\geqq0}{\mathbb{C}}[\Lambda])
\]
for any $p$ by induction on $\mathop{\rm ht}\nolimits(\beta_p)$.
By Proposition \ref{prop:local-verma} we can take $\lambda$, $\gamma$, $\varphi$, $s$ as in Proposition \ref{prop:local-formula1} satisfying $\gamma=\beta_p$, $(Sx^L_p)(\varphi s^{-1})=1$ and $(Sx^L_{p'})(\varphi s^{-1})=0$ for $p'\ne p$ with $\beta_{p'}=\beta_p$.
Then the assertion follows from Proposition \ref{prop:local-formula1}.
We next show
\[
\jmath(k_\mu)\in
\jmath((\Theta_e^{-1}A_\zeta)(0)
U_\zeta^{+})
(\Theta_{e}^{-1}Z{D'_\zeta}^{(\ell)})(0)
\]
for any $\mu\in\Lambda$.
We see easily that there exists some $\lambda\in\Lambda^+$ such that $\mu-2\lambda\in\ell\Lambda$.
Write
$\jmath(k_\mu)
=\jmath(k_{2\lambda})\jmath(k_{\mu-2\lambda})$.
Then we have $\jmath(k_{\mu-2\lambda})\in
(\Theta_{e}^{-1}Z{D'_\zeta}^{(\ell)})(0)$ by $k_{\mu-2\lambda}\in Z_{Fr}(U_\zeta)$.
Hence the assertion follows from Proposition \ref{prop:local-formula2}.
It follows that
\[
(\Theta_{e}^{-1}{D'_\zeta}^{(\ell)})(0)
=
\jmath((\Theta_e^{-1}A_\zeta)(0))
\jmath(U_\zeta^{+})
(\Theta_{e}^{-1}Z{D'_\zeta}^{(\ell)})(0).
\]
By definition $U_\zeta^+$ is a free $U_\zeta^+\cap Z_{Fr}(U_\zeta)$-module of rank $\ell^{|\Delta^+|}$.
Moreover, $(\Theta_e^{-1}A_\zeta)(0)$ is a free $(\Theta_e^{-1}A_1)(0)$-module of rank $\ell^{|\Delta^+|}$ by Proposition \ref{prop:AzetaA1}.
We are done.
\end{proof}
By \eqref{eq:Har-center-in-D}
we have an ${\mathcal O}_{\mathcal V}$-algebra homomorphism
\[
U_\zeta\otimes_{Z(U_\zeta)}{\mathcal O}_{\mathcal V}
\to
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta},
\]
where $Z(U_\zeta)\to{\mathcal O}_{\mathcal V}$ is given by
\[
{\mathcal V}\to K\times_{H/W}(H/{W\circ})
\,\,(\cong {\rm{Spec}}\,Z(U_\zeta))\qquad
(B^-g,k,t)\mapsto(k,[t^2]))
\]
(see Corollary \ref{cor:center}).
Let $\alpha_0\in\Delta^+$ be the highest root and set
$\tilde{\Pi}=\{\alpha_i\mid i\in I\}\cup\{\alpha_0\}$.
Set
\begin{align*}
H_{ur}&=\{
t\in H\mid
\alpha\in\tilde{\Pi},\,
{\theta_\alpha(t)}^{2\ell}=1\,\Longrightarrow\,
{\theta_\alpha(t)}^2=\zeta^{-(2\rho,\alpha)}\},\\
{\mathcal V}_{ur}&=
\{(B^-g, k,t)\in{\mathcal V}\mid
t\in H_{ur}\}.
\end{align*}
By Brown-Gordon \cite{BG} $U_\zeta\otimes_{Z(U_\zeta)}{\mathcal O}_{{\mathcal V}_{ur}}$ is an Azumaya algebra of rank $\ell^{2|\Delta^+|}$ on ${\mathcal V}_{ur}$.
On the other hand, since $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is a quotient of $\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}$, $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is also locally generated by $\ell^{2|\Delta^+|}$ sections.
Hence we obtain
\[
U_\zeta\otimes_{Z(U_\zeta)}{\mathcal O}_{{\mathcal V}_{ur}}
\cong
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}|_{{\mathcal V}_{ur}}
\]
by Lemma \ref{lem:Azumaya-generality} below.
In particular, $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}|_{{\mathcal V}_{ur}}$ is an Azumaya algebra of rank $\ell^{2|\Delta^+|}$.
\begin{lemma}
\label{lem:Azumaya-generality}
Let $X$ be an algebraic variety over ${\mathbb{C}}$, and let $f:{\mathcal A}\to{\mathcal A}'$ be a homomorphism of ${\mathcal O}_X$-algebras.
Assume that
${\mathcal A}$ is an Azumaya algebra of rank $n^2$ on $X$
and that
${\mathcal A}'$ is coherent and locally generated by $n^2$ sections as an ${\mathcal O}_X$-module.
Assume also that the fiber ${\mathcal A}'(x)$ is not zero for any $x\in X$.
Then $f$ is an isomorphism.
\end{lemma}
\begin{proof}
For each $x\in X$
consider the ${\mathbb{C}}$-algebra homomorphism
$f_x:{\mathcal A}(x)\to{\mathcal A}'(x)$ for the fibers.
Then $f_x$ is a non-zero homomorphism since it sends $1_{{\mathcal A}(x)}$ to $1_{{\mathcal A}'(x)}$ which is non-zero by ${{\mathcal A}'(x)}\ne\{0\}$.
Hence by the simplicity of ${\mathcal A}(x)$ we conclude that $f_x$ is injective.
On the other hand we have $\dim{\mathcal A}(x)=n^2$ and $\dim{\mathcal A}'(x)\leqq n^2$ by our assumption.
It follows that $f_x$ is an isomorphism for any $x\in X$.
Define ${\mathcal A}''$ by the exact sequence
\[
{\mathcal A} \rightarrow {\mathcal A}'\rightarrow {\mathcal A}''\rightarrow 0.
\]
Then we have ${\mathcal A}''(x)=\{0\}$ for any $x\in X$ by the surjectivity of $f_x$.
Hence ${\mathcal A}''=\{0\}$ by Nakayama's lemma.
It follows that $f$ is an epimorphism.
Let us show that $f$ is a monomorphism.
We may assume that $X$ is affine and ${\mathcal A}$ is free.
Set $R=\Gamma(X,{\mathcal O}_X)$, $A=\Gamma(X,{\mathcal A})$, $A'=\Gamma(X,{\mathcal A}')$.
We need to show that the homomorphism $F:A\to A'$ of $R$-modules corresponding to $f$ is injective.
Note that $A$ is isomorphic to $R^{n^2}$.
By the injectivity of $f_x$ for $x\in X$ the homomorphism
$F_{\mathfrak{m}}:R/{\mathfrak{m}}\otimes_RA\to R/{\mathfrak{m}}\otimes_RA'$ is injective for any maximal ideal ${\mathfrak{m}}$ of $R$.
Hence by the commutative diagram
\[
\begin{CD}
A&@>{F}>>&A'\\
@VVV&&@VVV\\
R/{\mathfrak{m}}\otimes_RA
&
@>>{F_{\mathfrak{m}}}>
&
R/{\mathfrak{m}}\otimes_RA'
\end{CD}
\]
and $\mathop{\rm Ker\hskip.5pt}\nolimits(A\to R/{\mathfrak{m}}\otimes_RA)={\mathfrak{m}} A$
we have
\[
\mathop{\rm Ker\hskip.5pt}\nolimits(F)\subset\bigcap_{\mathfrak{m}}\Gm A
\cong\bigcap_{\mathfrak{m}}\Gm R^{n^2}
=(\bigcap_{\mathfrak{m}}\Gm R
)^{n^2}=\{0\}.
\]
\end{proof}
For $\mu\in\Lambda$ we define $t_\mu\in H$ by
\[
\theta_\lambda(t_\mu)=\zeta^{(\lambda,\mu)}\qquad(\lambda\in\Lambda).
\]
Note that we have $t_\mu^{\ell}=1$.
We consider the automorphism
\[
\xi_\mu:{\mathcal V}\to{\mathcal V}
\qquad
(B^-g, k,t)\mapsto (B^-g, k, t_\mu t))
\]
of the algebraic variety ${\mathcal V}$.
\begin{lemma}
For $\mu\in\Lambda$
the ${\mathcal O}_{\mathcal V}$-algebras
$\xi_\mu^*\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ and $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$
are isomorphic locally on ${\mathcal B}$.
\end{lemma}
\begin{proof}
We will show
\[
\xi_\mu^*\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}|_{p_{\mathcal V}^{-1}{\mathcal B}_w}
\cong
\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}|_{p_{\mathcal V}^{-1}{\mathcal B}_w}
\]
for any $w\in W$.
It is sufficient to verify that there exists a ${\mathbb{C}}$-algebra automorphism of
$\Theta_{w}^{-1}{D_\zeta}^{(\ell)}(0)$
which induce
\[
\xi_\mu^*:\Theta_{w}^{-1}Z{D_\zeta}^{(\ell)}(0)
\to
\Theta_{w}^{-1}Z{D_\zeta}^{(\ell)}(0)
\]
under the identification $p_{\mathcal V}^{-1}{\mathcal B}_w
={\rm{Spec}}\,\Theta_{w}^{-1}Z{D_\zeta}^{(\ell)}(0)$.
Note that
\[
\Theta_{w}^{-1}{D_\zeta}^{(\ell)}(0)
\cong
\tilde\Theta_{w}^{-1}{D_\zeta}(0).
\]
Hence the problem is to construct an automorphism of the ${\mathbb{C}}$-algebra $\tilde\Theta_{w}^{-1}{D_\zeta}(0)$ satisfying
\[
\varphi\mapsto\varphi,\quad
\partial_z\mapsto\partial_z,\quad
\sigma_\lambda\mapsto\zeta^{(\lambda,\mu)}\sigma_\lambda
\quad(\varphi\in \Theta_{w}^{-1}A_1(0), z\in Z_{Fr}(U_\zeta), \lambda\in\Lambda).
\]
Take $c\in A_\zeta(\mu)_{w^{-1}\mu}\setminus\{0\}
\subset\tilde{\Theta}_e$
Then the automorphism
\[
\tilde\Theta_{w}^{-1}{D_\zeta}(0).
\to
\tilde\Theta_{w}^{-1}{D_\zeta}(0)\qquad
(P\mapsto c^{-1} P c)
\]
satisfies the desired property.
\end{proof}
By
\[
\bigcup_{\mu\in\Lambda}\xi_\mu({\mathcal V}_{ur})={\mathcal V}
\]
we conclude that $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is an Azumaya algebra of rank $\ell^{2|\Delta^+|}$ on ${\mathcal V}$.
Recall that there exists a surjection
$\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}\to\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$.
By the arguments above there exists locally a generator system of $\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}$ consisting of $\ell^{2|\Delta^+|}$ sections which gives a (local) free basis of $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$.
Hence $\tilde{{\mathcal D}}'_{{\mathcal B}_\zeta}\to\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ is an isomorphism.
The proof of Theorem \ref{thm:Azumaya} is now complete.
By Theorem \ref{thm:Azumaya} and Lemma \ref{lem:Fr1} we have the following.
\begin{corollary}
\label{cor:Azumaya}
$\omega^*D'_\zeta\cong\omega^* D_\zeta$.
\end{corollary}
\subsection{}
Define
\[
\delta:{\mathcal V}\to K\times_{H/W} H
\]
by $\delta(B^-g,k,t)=(k,t)$, where $H\to H/W$ is given by $t\mapsto[t^{2\ell}]$.
In the rest of this section we will prove the following result.
\begin{theorem}
\label{thm:split-Azumaya}
For any $(k, t)\in K\times_{H/W}H$,
the restriction of
$\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$ to ${\delta^{-1}(k,t)}$ is a split Azumaya algebra.
\end{theorem}
We need some preliminaries.
\begin{lemma}
\label{lem:Azumaya-generality2}
Let ${\mathcal A}$ be an Azumaya algebra on an algebraic variety $X$.
Assume that ${\mathcal M}$ is a locally free right ${\mathcal A}$-module of rank one.
Then $\mathop{{\mathcal{E}}nd}\nolimits_{\mathcal A}({\mathcal M})$ is an Azumaya algebra whose rank is the same as that of ${\mathcal A}$.
Moreover, if ${\mathcal A}$ is a split Azumaya algebra, then $\mathop{{\mathcal{E}}nd}\nolimits_{\mathcal A}({\mathcal M})$ is also a split Azumaya algebra.
\end{lemma}
\begin{proof}
Let $V$ be a finite-dimensional vector space over a field $k$ and regard $M=\mathop{\rm{End}}\nolimits_k(V)$ as a right $\mathop{\rm{End}}\nolimits_k(V)$-module by the right multiplication.
Then the left multiplication of $\mathop{\rm{End}}\nolimits_k(V)$ induces
a canonical isomorphism
\begin{equation}
\label{eq:End}
\mathop{\rm{End}}\nolimits_{\mathop{\rm{End}}\nolimits_k(V)}(M)\cong\mathop{\rm{End}}\nolimits_k(V)
\end{equation}
of $k$-algebras.
Hence the first half of our theorem holds.
Let us show the second half.
Assume ${\mathcal A}=\mathop{{\mathcal{E}}nd}\nolimits_{{\mathcal O}_X}({\mathcal W})$ for a locally free ${\mathcal O}_X$-module ${\mathcal W}$.
Then it is sufficient to show $\mathop{{\mathcal{E}}nd}\nolimits_{{\mathcal A}}({\mathcal M})\cong\mathop{{\mathcal{E}}nd}\nolimits_{{\mathcal O}_X}({\mathcal M}\otimes_{\mathcal A}{\mathcal W})$.
This also follows from \eqref{eq:End}.
\end{proof}
\begin{proposition}
\label{prop:Azumaya-translation}
For any $\mu\in\Lambda$
there exists a locally free right $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$-module
${\mathcal L}_\mu$
of rank one such
\[
\xi_\mu^*\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}
\cong
\mathop{{\mathcal{E}}nd}\nolimits_{\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}}
({\mathcal L}_\mu).
\]
\end{proposition}
\begin{proof}
Note that $D_\zeta$ is a $\Lambda$-graded $D_\zeta$-bimodule by the left and the right multiplications.
Define $D_\zeta[\mu]$ to be the $\Lambda$-graded $D_\zeta$-bimodule which coincides with $D_\zeta$ as a $D_\zeta$-bimodule and the grading is given by $(D_\zeta[\mu])(\lambda)=D_\zeta(\lambda+\mu)$.
Then we obtain a $\Lambda$-graded $D_\zeta^{(\ell)}$-bimodule
\[
D_\zeta[\mu]^{(\ell)}
=\bigoplus_{\lambda\in\Lambda}
(D_\zeta[\mu])(\ell\lambda)
\qquad
((D_\zeta[\mu]^{(\ell)})(\lambda)
=D_\zeta(\mu+\ell\lambda)).
\]
Note that the left and the right actions of $ZD_\zeta^{(\ell)}$ on $D_\zeta[\mu]^{(\ell)}$ are different.
In fact we have
\[
z\cdot P=P\cdot\tilde{\xi}_\mu(z)
\qquad(z\in ZD_\zeta^{(\ell)}),
\]
where $\tilde{\xi}_\mu:ZD_\zeta^{(\ell)}\to ZD_\zeta^{(\ell)}$ is the algebra automorphism corresponding to $\xi_\mu$.
Regard $D_\zeta[\mu]^{(\ell)}$ as a $ZD_\zeta^{(\ell)}$-module by the right action and consider its localization ${\mathcal L}_\mu$ on ${\mathcal V}$.
Then the right action of $D_\zeta^{(\ell)}$ on $D_\zeta[\mu]^{(\ell)}$ induces
a right $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$-module structure of ${\mathcal L}_\mu$.
Moreover, ${\mathcal L}_\mu$ is a locally free right $\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$-module
of rank one.
On the other hand
regard $D_\zeta^{(\ell)}$ as a $\Lambda$-graded $ZD_\zeta^{(\ell)}$-algebra by the modified $ZD_\zeta^{(\ell)}$-module structure given by
\[
z\circ P=\tilde{\xi}_\mu(z)P\quad(z\in ZD_\zeta^{(\ell)},\,\,P\in D_\zeta^{(\ell)}).
\]
Then the localization of $D_\zeta^{(\ell)}$ on ${\mathcal V}$ with respect to the modified $\Lambda$-graded $ZD_\zeta^{(\ell)}$-algebra structure coincides with $\xi_\mu^*\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}$.
Hence we obtain an ${\mathcal O}_{\mathcal V}$-algebra homomorphism
\[
\xi_\mu^*\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}
\to
\mathop{{\mathcal{E}}nd}\nolimits_{\tilde{{\mathcal D}}_{{\mathcal B}_\zeta}}
({\mathcal L}_\mu)
\]
induced by the left multiplication of $D_\zeta^{(\ell)}$ on $(D_\zeta[\mu])^{(\ell)}$.
By Theorem \ref{thm:Azumaya}, Lemma \ref{lem:Azumaya-generality2} and Lemma \ref{lem:Azumaya-generality} this is an isomorphism.
\end{proof}
Let us finish the proof of Theorem \ref{thm:split-Azumaya}.
By Brown-Gordon \cite{BG} the assertion holds when $t\in H_{ur}$.
The general case is reduced to this case by Proposition \ref{prop:Azumaya-translation} and Lemma \ref{lem:Azumaya-generality2}.
|
\section{Introduction, Results and Summary}
\label{s:0}
%
Supersymmetry has been studied for almost four decades in the physics literature, and about as long using the superspace formalism\cite{rSSSS1,r1001,rPW,rWB,rBK}. Nevertheless, a complete classification of off-shell representations is still in progress, even just on the worldline; see Refs.\cite{r6-1,r6-3,r6-3.2} and\cite{rFKS,rFS1,rFS2,rKRT,rBKMO,rBG,rBKLS}. These results indicate that there is an overabundance of even the restricted class of so-called {\em\/adinkraic\/} supermultiplets, wherein each supercharge transforms each component field into precisely one other component field (or its derivative), not a linear combination of those. The plentitude of these findings\cite{r6-3,r6-3.2} dwarfs the efforts from earlier literature\cite{r1001,rPW,rWB,rHKLR,rBK,rHSS,rGSS}.
In 1+1-dimensional spacetime|worldsheet|models, it is possible to define superfields that are only {\em\/partly on-shell\/} (on {\em\/half-shell\/}\cite{rHP1}), such as:
\begin{equation}
\textbf{lefton}:\quad D_-\boldsymbol{\L}=0=\Db_-\boldsymbol{\L},\quad\To\quad \vd_\mm\boldsymbol{\L}=0.
\label{e:L}
\end{equation}
Such superfields are off-shell on the left-handed light-cone worldline (0-brane) embedded within the worldsheet, but are not off-shell in the usual sense of $1{+}1$-dimensional worldsheet field theory, and were therefore said to be {\em\/on half-shell\/}\cite{rHP1}. This unique feature of worldsheet field theory of course further complicates any classification of supermultiplets and their superfield representation, but also provides opportunities not possible in other dimensions.
It is our present purpose to demonstrate that unidexterous Lagrange multiplier superfields may be used in two complementary roles, to define:
({\small\bf1})~certain proper off-shell superfields which conventional super-differential constraining cannot (see Section~\ref{s:Strict}), and
({\small\bf2})~partially on-shell superfields, with an unequal number of off-shell bosonic and fermionic component fields (see Section~\ref{s:Limp}). Section~\ref{s:Coda} summarizes these results and conclusions.
Adopting the notation of Refs.\cite{r1001,rHSS}, we list here a few basic definitions and results for completeness.
$(2,2)$-extended supersymmetry on the $(1,1)$-dimensional worldsheet may be studied in $(1,1|2,2)$-superspace notation, wherein the supersymmetry charges $Q_\pm$ and $\Qb_\pm$, superderivatives $D_\pm,\Db_\pm$ and light-cone worldsheet derivatives $\vd_{=\mkern-10mu|\mkern3mu},\vd_\mm$ satisfy:
%
\begin{subequations}
\label{e:SuSy}
\begin{alignat}{3}
\big\{\, Q_- \,,\, \Qb_- \,\big\}&=2i\vd_\mm, &\qquad
\big\{\, Q_+ \,,\, \Qb_+ \,\big\}&=2i\vd_{=\mkern-10mu|\mkern3mu}, \label{eSusyQ}\\
\big\{\, D_- \,,\, \Db_- \,\big\}&=2i\vd_\mm, &\qquad
\big\{\, D_+ \,,\, \Db_+ \,\big\}&=2i\vd_{=\mkern-10mu|\mkern3mu}, \label{eSusyD}\\
\vd_\mm&\Defl(\vd_\t{-}\vd_\s), &\qquad
\vd_{=\mkern-10mu|\mkern3mu}&\Defl(\vd_\t{+}\vd_\s),
\end{alignat}
%
where $H{=}i\hbar\vd_\t$ and $p{=}-i\hbar\vd_\s$ are the worldsheet Hamiltonian and linear momentum, respectively. All other (anti)commutators among these operators vanish. In this notation, all operators\eq{e:SuSy} are eigen-operators of the Lorentz transformation and the number of ``$\pm$'' subscripts counts the additive eigenvalue in units of $\inv2\hbar$. So, the Lorentz-eigenvalue (``spin'') of $Q_\pm$ and $\Qb_\pm$ is $\pm\inv2\hbar$, of $\vd_{=\mkern-10mu|\mkern3mu}$ is $+\hbar$, and of $\vd_\mm$ is $-\hbar$.
\end{subequations}
As customary in superspace\cite{rSSSS1,r1001,rPW,rWB,rBK}, we use the super-derivatives $D_+,\Db_+,D_-,\Db_-$. With a nodding familiarity of {\em\/Adinkras\/}\cite{rA,r6-1,r6-3,r6-3.2} and similar diagrams\cite{rFre,rSIg0,rSIg}, we depict their algebra, up to $\vd_{=\mkern-10mu|\mkern3mu},\vd_\mm$ factors, following Ref\cite{rHSS}:
\begin{equation}
\vC{\begin{picture}(140,52)
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
%
\put(70,47.5){\cB{$\inv4[\Db_+,D_+][\Db_-,D_-]$}}
%
\put(21,37.5){\cB{$\inv2[\Db_+,D_+]D_-$}}
\put(54,37.5){\cB{$\inv2[\Db_+,D_+]\Db_-$}}
\put(86.5,37.5){\cB{$\inv2[\Db_-,D_-]D_+$}}
\put(119,37.5){\cB{$\inv2[\Db_-,D_-]\Db_+$}}
%
\put(4.75,25){\cB{$\inv2[\Db_+,D_+]$}}
\put(31,24.5){\cB{$D_+D_-$}}
\put(57.5,24.5){\cB{$\Db_+D_-$}}
\put(83.5,24.5){\cB{$\Db_-D_+$}}
\put(110,24.5){\cB{$\Db_-\Db_+$}}
\put(135.75,25){\cB{$\inv2[\Db_-,D_-]$}}
%
\put(21,12){\cB{$D_+$}}
\put(53.5,12){\cB{$\Db_+$}}
\put(86.5,12){\cB{$D_-$}}
\put(119,12){\cB{$\Db_-$}}
%
\put(70,2){\cB{$\Ione$}}
\end{picture}}
\label{e:Ds}
\end{equation}
The Adinkras of Refs.\cite{rA,r6-1,r6-3,r6-3.2} depict representations of $N$-extended worldline supersymmetry, which is a subalgebra of\eq{e:SuSy} when $N\leq4$. Similarly here, following edges in\eqs{e:Ds}{e:Xs} {\em\/upward\/} corresponds to the supersymmetry action, where red/green/blue/orange edges correspond, respectively, to $\C1{D_+}$,- $\C2{\Db_+}$,- $\C3{D_-}$- and $\C4{\Db_-}$-action. Since $\Db_\pm=(D_\pm)^\dagger$, following an edges {\em\/downward\/} corresponds to the conjugate super-derivative action of the anti-color (green is anti-red, \etc), and incurs an additional $i\vd_{=\mkern-10mu|\mkern3mu}$ (for red/green edges) or $i\vd_\mm$ (for blue/orange edges) worldsheet derivative:
\begin{equation}
\vC{\begin{picture}(30,25)(0,2)
\put(0,0){\includegraphics[width=30mm]{Red.pdf}}
\put(7,10){\C1{$D_+$}}
\put(18,17){\C2{$\Db_+$}}
\put(27,7){$i\vd_{=\mkern-10mu|\mkern3mu}$}
\end{picture}}
\qquad\textit{vs.}\quad
\vC{\begin{picture}(30,25)(0,2)
\put(0,0){\includegraphics[width=30mm]{Green.pdf}}
\put(7,10){\C2{$\Db_+$}}
\put(18,17){\C1{$D_+$}}
\put(27,7){$i\vd_{=\mkern-10mu|\mkern3mu}$}
\end{picture}}
\qquad\textit{etc.}
\label{e:UpDown}
\end{equation}
In\eq{e:Ds}, the node-labels clarify the action along any particular edge, \ie, between two nodes.
The components of any superfield|including superfield expressions and equations|are obtained by ``projecting'': applying the basis elements\eq{e:Ds} of the exterior algebra generated by $D_+.\Db_+,D_-,\Db_-$ and then setting the nilpotent superspace coordinates to zero. For an {\em\/intact\/}\ft{Herein, ``intact superfields'' are Salam-Strathdee complex superfields\cite{rSSSS1}, subject to neither constraint nor projection, all of which have a 4-cubical chromotopology\cite{r6-3} for worldsheet $(2,2)$-supersymmetry.} superfield $\boldsymbol{X}$, we denote the so-obtained components:
\begin{equation}
\vC{\begin{picture}(140,52)
\put(-15,45){superfield $\boldsymbol{X}$:}
\put(-15,40){\small(intact)}
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
%
\put(70,47.5){\cB{$\mathscr{X}$}}
%
\put(21,37.5){\cB{$\mathcal{X}_+$}}
\put(54,37.5){\cB{$\X_+$}}
\put(86.5,37.5){\cB{$\mathcal{X}_-$}}
\put(119,37.5){\cB{$\X_-$}}
%
\put(4.75,25){\cB{$X_{=\mkern-10mu|\mkern3mu}$}}
\put(31,24.5){\cB{$X$}}
\put(57.5,24.5){\cB{$\acute{X}$}}
\put(83.5,24.5){\cB{$\grave{X}$}}
\put(110,24.5){\cB{$\bar{X}$}}
\put(135.75,25){\cB{$X_\mm$}}
%
\put(21,12){\cB{$\c_+$}}
\put(53.5,12){\cB{$\x_+$}}
\put(86.5,12){\cB{$\c_-$}}
\put(119,12){\cB{$\x_-$}}
%
\put(70,2){\cB{$x$}}
\end{picture}}
\label{e:Xs}
\end{equation}
Comparing\eq{e:Xs} with\eq{e:Ds} shows that, for example,
\begin{equation}
x\Defl\boldsymbol{X}|,\quad
\x_+\Defl\Db_+\boldsymbol{X}|,\quad
\bar{X}\Defl\Db_-\Db_+\boldsymbol{X}|,\quad
\mathcal{X}_+\Defl\inv2[\Db_+,D_+]D_-\boldsymbol{X}|,\quad\etc
\label{e:fewXs}
\end{equation}
These diagrams reveal the underlying 4-cubical (tesseract) {\em\/chromotopology\/}\cite{r6-3}\ft{For our purposes, the 4-cube (tesseract) is evident in the diagrams\eq{e:Ds} and\eq{e:Xs}, the edge-colors being correlated with the four dimensions of the 4-cube, and also the four operators, $D_+,\Db_+,D_-,\Db_-$.} of the supermultiplet represented by the superfield $\boldsymbol{X}$ and encode all supersymmetry relations. For example, using\eq{e:UpDown},
\begin{equation}
\C2{\Qb_+}\,x=\x_+,\quad
\C1{Q_+}\x_+=i(\vd_{=\mkern-10mu|\mkern3mu} x)-X_{=\mkern-10mu|\mkern3mu},\quad
\C2{\Qb_+}\,\x_+=0,\quad\etc,
\label{e:QXs}
\end{equation}
where the minus sign in $\C1{Q_+}\x_+=\C1{Q_+}\C2{\Qb_+}x=-\inv2[\C2{\Qb_+},\C1{Q_+}]x+\inv2\{\C2{\Qb_+},\C1{Q_+}\}x$ is encoded by dashing the $\x_+$--$X_{=\mkern-10mu|\mkern3mu}$ edge, as in Ref.\cite{r6-1}.
Depicting superfields in this manner reveals all the relations between its component fields|and indeed the whole structure of the superfield|in a correct but intuitive way that is not as easily conveyed by formulae.
\section{Strictly Constrained Off-Shell Superfields}
\label{s:Strict}
%
There are two ways of constraining superfields: ({\small\bf1})~by defining them as satisfying a system of super-differential constraints, and ({\small\bf2})~by the inclusion of Lagrange multiplier terms in the Lagrangian.
Exploring systematically the former of these two methods, Ref.\cite{rHSS} defines, among many others, an ``almost unidexterous non-minimal superfield'' by imposing the second order super-differential condition:
\begin{equation}
[\Db_-,D_-]\boldsymbol{A}=0.
\label{e:AL}
\end{equation}
Somewhat akin to\eq{e:L}, some of the components of this superfield vanish, while others satisfy worldsheet differential constraints:
\begin{subequations}
\begin{align}
\{A_\mm;\mathcal{A}_-,@_-;\mathscr{A}\}&=0,\label{e:AL0}\\
\vd_\mm\,\{\a_-,{\rm a}_-;A,\acute{A},\grave{A},\bar{A};\mathcal{A}_+,@_+\}&=0, \label{e:L8}\\
\vd_\mm^{~2}\,\{a;\a_+,{\rm a}_+;A_{=\mkern-10mu|\mkern3mu}\}&=0.\label{e:AL4}
\end{align}
\end{subequations}
with components defined akin to\eqs{e:Ds}{e:fewXs}, see\eq{e:[2]Ys}, and the indicated operators applied on each member of the indicated collection.
Owing to the worldsheet differential constraints\eqs{e:L8}{e:AL4}, the $(1|4|5|2|0)$-component superfield defined to satisfy the super-differential constraint\eq{e:AL} then is not off-shell. In turn, the mapping that restricts an intact superfield to the one satisfying \Eq{e:AL} then is not a {\em\/strict homomorphism of off-shell supermultiplets\/}, as defined in Ref.\cite{r6-3.2}. The component fields of $\boldsymbol{A}$ satisfying\eq{e:AL} are:
\begin{equation}
\vC{\begin{picture}(135,52)
\put(-5,0){\includegraphics[width=150mm]{Spindle3.pdf}}
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
\put(-18,48){$[\Db_-,D_-]\boldsymbol{A}=0$:}
\put(-18,43){\small(is not off-shell)}
\put(0,5){\small\Eq{e:AL4}}
\put(120,5){\small\Eq{e:L8}}
\put(120,45){\small\Eq{e:AL0}}
%
\put(70,47.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,37.5){\cB{$\2{\mathcal{A}_+}$}}
\put(54,37.5){\cB{$\2{@_+}$}}
\put(86.5,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
\put(119,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
%
\put(4.75,25){\cB{$\2{\2{A_{=\mkern-10mu|\mkern3mu}}}$}}
\put(31,24.5){\cB{$\2{A}$}}
\put(57.5,24.5){\cB{$\2{\acute{A}}$}}
\put(83.5,24.5){\cB{$\2{\grave{A}}$}}
\put(110,24.5){\cB{$\2{\bar{A}}$}}
\put(135.75,25){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,12){\cB{$\2{\2{\a_+}}$}}
\put(53.5,12){\cB{$\2{\2{{\rm a}_+}}$}}
\put(86.5,12){\cB{$\2{\a_-}$}}
\put(119,12){\cB{$\2{{\rm a}_-}$}}
%
\put(70,2){\cB{$\2{\2a}$}}
\end{picture}}
\label{e:[2]Ys}
\end{equation}
where the unnamed open/closed nodes represent bosonic/fermionic expressions in terms of the other, named component fields. Simple underlining denotes a first order constraint\eq{e:L8}, whereas double underlining denotes the second-order constraint\eq{e:AL4}. There are no off-shell component fields in the $(1|4|5|2|0)$-component superfield\eq{e:[2]Ys}. The worldsheet differential constraints\eqs{e:L8}{e:AL4} obstruct using such a superfield in the path-integral formalism for quantum field theory.
By contrast, the non-minimal chiral (a.k.a.\ complex linear) superfield, defined below\eqs{e:CLM}{e:Ts}, is a proper, $(1|4|5|2|0)$-component off-shell superfield, as are also the non-minimal twisted chiral superfield, their Hermitian conjugates, and the more familiar chiral and twisted-chiral superfields.
\ping
The lefton superfields, defined by the pair of super-differential constraints in\eq{e:L}, are also not off-shell. Nevertheless, it turns out that precisely this superfield {\em\/on half-shell\/}\cite{rHP1} may be employed to eliminate the highlighted unnamed nodes indicated in\eq{e:[2]Ys}|without incurring any worldsheet differential constraints on the remaining component fields. That is, the (on half-shell) lefton superfield $\boldsymbol{\L}$ may be used to restrict an off-shell $(1|4|6|4|1)$-component intact superfield\eq{e:Xs} to a proper off-shell $(1|4|5|2|0)$-component superfield\eq{e:LYs}.
This then also defines a strict morphism between these off-shell supermultiplets by the previously unexplored means of unidexterous Lagrange multipliers.
To see this, consider the Lagrangian term
\begin{equation}
\mathscr{L}_{\boldsymbol{\L}\boldsymbol{Y}}~\Defl~
\inv4[\Db_+,D_+][\Db_-,D_-]\,\boldsymbol{\L}\,\boldsymbol{Y}|
=L_{=\mkern-10mu|\mkern3mu} Y_\mm + \l_+\Y_- + \bl_+\mathcal{Y}_- + \ell\mathscr{Y},
\label{e:LAL}
\end{equation}
where
\begin{equation}
\ell\Defl\boldsymbol{\L}|,\qquad
\l_+\Defl D_+\boldsymbol{\L}|,\qquad
\bl_+\Defl \Db_+\boldsymbol{\L}|,\qquad
L_{=\mkern-10mu|\mkern3mu}\Defl \inv2[\Db_+,D_+]\boldsymbol{\L}|.
\end{equation}
are the components of the lefton superfield, $\boldsymbol{\L}$. With the standard assignment of mass-dimensions $[\boldsymbol{\L}]=0=[\boldsymbol{Y}]$, it follows that no kinetic term for $\boldsymbol{\L}$ is possible, so that all components of $\boldsymbol{\L}$ are {\em\/auxiliary\/}: their Euler-Lagrange equations are not differential\cite{rIPd,rHP1}. Since they occur linearly in\eq{e:LAL} and nowhere else, integrating out the components of $\boldsymbol{\L}$ imposes the simple constraints:
\begin{equation}
Y_\mm~=~\Y_-~=~\mathcal{Y}_-~=~\mathscr{Y}~=~0,\qquad\textit{and nothing else!}
\label{e:4}
\end{equation}
This defines a $\boldsymbol{\L}$-constrained version of $\boldsymbol{Y}$, which|unlike\eq{e:[2]Ys}|is a fully off-shell $(1|4|5|2|0)$-dimen\-si\-o\-nal representation of $(1,1|2,2)$-supersymmetry, spanned by the remaining, otherwise unconstrained component fields:
\begin{equation}
\vC{\begin{picture}(140,52)
\put(-15,45){Constr.\ by $\mathscr{L}_{\boldsymbol{\L}\boldsymbol{Y}}:$}
\put(-15,40){\small(is off-shell)}
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
%
\put(70,47.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,37.5){\cB{$\mathcal{Y}_+$}}
\put(54,37.5){\cB{$\Y_+$}}
\put(86.5,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
\put(119,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
%
\put(4.75,25){\cB{$Y_{=\mkern-10mu|\mkern3mu}$}}
\put(31,24.5){\cB{$Y$}}
\put(57.5,24.5){\cB{$\acute{Y}$}}
\put(83.5,24.5){\cB{$\grave{Y}$}}
\put(110,24.5){\cB{$\bar{Y}$}}
\put(135.75,25){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,12){\cB{$\h_+$}}
\put(53.5,12){\cB{$\y_+$}}
\put(86.5,12){\cB{$\h_-$}}
\put(119,12){\cB{$\y_-$}}
%
\put(70,2){\cB{$y$}}
\end{picture}}
\label{e:LYs}
\end{equation}
All named component fields are unaffected by the Lagrange multiplier constraining and remain off-shell. Thus, unidexterous constraining by means of the term\eq{e:LAL} turns out to define an off-shell superfield\eq{e:LYs}, which cannot be defined by means of a super-differential constraint such as\eq{e:AL}. The eliminated components, depicted by the highlighted and unnamed nodes, are connected by $\Qb_+,Q_+$ action into a quadrilateral. These features are exactly analogous to the non-minimal chiral superfield, defined below\eqs{e:CLM}{e:Ts}, its twisted variant and their Hermitian conjugates\cite{rHSS}.
The mapping $\boldsymbol{X}\to\boldsymbol{Y}$ from an intact superfield\eq{e:Xs} to the one\eq{e:LYs} constrained by means of the Lagrangian term\eq{e:LAL} is then a strict homomorphism of off-shell supermultiplets, as defined in Ref.\cite{r6-3.2}. The mapping $\boldsymbol{X}\to\boldsymbol{A}$ specified by the second order super-constraint\eq{e:AL} is not.
\section{Unequal Off-Shell Bosonic and Fermionic Components}
\label{s:Limp}
%
Consider a {\em\/non-minimal chiral\/} (a.k.a.\ complex-linear) superfield $\boldsymbol{\Q}$, defined by
\begin{subequations}
\label{e:CLM}
\begin{gather}
\Db_-\Db_+\boldsymbol{\Q}=0,\\[2mm]
\vC{\begin{picture}(140,52)(-5,0)
\put(-15,45){$\Db_-\Db_+\boldsymbol{\Q}=0:$}
\put(-15,40){\small(is off-shell)}
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
%
\put(70,47.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,37.5){\cB{$\t_+$}}
\put(54,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
\put(86.5,37.5){\cB{$\t_-$}}
\put(119,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
%
\put(4.75,25){\cB{$T_{=\mkern-10mu|\mkern3mu}$}}
\put(31,24.5){\cB{$T$}}
\put(57.5,24.5){\cB{$\acute{T}$}}
\put(83.5,24.5){\cB{$\grave{T}$}}
\put(110,24.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
\put(135.75,25){\cB{$T_\mm$}}
%
\put(21,12){\cB{$\q_+$}}
\put(53.5,12){\cB{$\vq_+$}}
\put(86.5,12){\cB{$\q_-$}}
\put(119,12){\cB{$\vq_-$}}
%
\put(70,2){\cB{$t$}}
\end{picture}}
\label{e:Ts}
\end{gather}
\end{subequations}
The component fields eliminated by the super-differential constraint (depicted as highlighted, unnamed nodes) are connected into a quadrilateral by $Q_+,Q_-$, \ie, $D_+,D_-$ action\cite{rFGH}.
Then, the inclusion of the Lagrangian term
\begin{equation}
\mathscr{L}_{\boldsymbol{\L}\boldsymbol{\Q}}\Defl\inv4[\Db_+,D_+][\Db_-,D_-]\,\boldsymbol{\L}\,\boldsymbol{\Q}|
=L_{=\mkern-10mu|\mkern3mu} T_\mm + i\l_+(\vd_\mm\vq_+) + \bl_+\t_-
\label{e:L-CLM}
\end{equation}
induces, through the equations of motion for $\boldsymbol{\L}$, only the constraints:
\begin{equation}
T_\mm~=~(\vd_\mm\vq_+)~=~\t_-~=~0.
\label{e:3}
\end{equation}
This turns the $(1|4|5|2)$-dimensional supermultiplet $\boldsymbol{\Q}$ into a $(1|4|4|1)$-dimensional one, where however one of the lowest-level fermions, $\vq_+$, has been rendered a left-mover, which is also the effect of its standard Dirac equation. This $\vq_+$ component field is therefore on-shell, whereas all other $(1|3|4|1)$-components remain off-shell. Therefore, this doubly constrained supermultiplet consists of 5 off-shell bosons and 4 off-shell fermions, coupled with one on-shell fermion:
\begin{equation}
\vC{\begin{picture}(140,52)
\put(-15,45){$\Db_-\Db_+\boldsymbol{\Q}=0$~\&~$\mathscr{L}_{\boldsymbol{\L}\boldsymbol{\Q}}:$}
\put(-15,40){\small(is not off-shell)}
\put(-5,0){\includegraphics[width=150mm]{Spindle.pdf}}
%
\put(70,47.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,37.5){\cB{$\t_+$}}
\put(54,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
\put(86.5,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
\put(119,37.5){\CB{\includegraphics[width=4mm]{Fermion.pdf}}}
%
\put(4.75,25){\cB{$T_{=\mkern-10mu|\mkern3mu}$}}
\put(31,24.5){\cB{$T$}}
\put(57.5,24.5){\cB{$\acute{T}$}}
\put(83.5,24.5){\cB{$\grave{T}$}}
\put(110,24.5){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
\put(135.75,25){\CB{\includegraphics[width=4mm]{Boson.pdf}}}
%
\put(21,12){\cB{$\q_+$}}
\put(53.5,12){\CB{$\2{\vq_+}$}}
\put(26,5){\small on shell}
\put(21,1){\small (the only one)}
\put(40,5){\vector(2,1){10}}
\put(86.5,12){\cB{$\q_-$}}
\put(119,12){\cB{$\vq_-$}}
%
\put(70,2){\cB{$t$}}
\end{picture}}
\label{e:LTs}
\end{equation}
This time, the eliminated component fields (depicted by highlighted, unnamed nodes) are connected into two overlapping quadrilaterals. In addition, a single component field ($\vq_+$) is constrained on-shell by the combination of super-differential and Lagrange multiplier constraining; it occurs away from the eliminated quadrilaterals. The non-highlighted named component fields remain unconstrained and off-shell, but the single on-shell component $\vq_+$ prevents the entire superfield from being off-shell. A somewhat similar situation was noted in Ref.\cite{r6-4}, for which case however a coupling with another superfield was found to relax both of them off-shell. Whether such a completely off-shell extension of\eq{e:LTs} exists remains an open question.
\section{Conclusions}
\label{s:Coda}
%
The preceding examples show that unidexterous ({\em\/on half-shell\/}\cite{rHP1}) Lagrange multiplier superfields can be used to define:
\begin{enumerate}\itemsep=-3pt\vspace{-4mm}
\item proper off-shell superfields such as the new one%
\eq{e:LYs}, which cannot be defined by conventional super-differential constraining, as in\eq{e:[2]Ys};
\item a strict homomorphism between the intact\eq{e:Xs} and the constrained off-shell superfield\eq{e:LYs};
\item partially on-shell superfields, with unequal numbers of bosonic/fermionic off-shell component fields, such as\eq{e:Ts}.
\end{enumerate}\vspace{-3mm}
Whereas this last feature does not seem to have an evident application, it is fairly novel and merits further study, perhaps \`a la Ref.\cite{r6-4}.
On the other hand, the ability to define novel off-shell superfields such as\eq{e:LYs} is much more likely to have applications in constructions of novel worldsheet Lagrangians and models.
In turn, the existence of the proper supersymmetry morphism~$(\ref{e:Xs}\to(\ref{e:LYs})$ may help a more rigorous understanding of off-shell supersymmetry representation theory.
\bigskip
\paragraph{Acknowledgment:}
I am indebted to the generous support by the Department of Energy through the grant DE-FG02-94ER-40854, and wish to thank for the recurring hospitality and resources provided by
the Physics Department of the University of Central Florida, Orlando, and
the Physics Department of the Faculty of Natural Sciences of the University of Novi Sad, Serbia, where part of this work was completed.
\input SubRep.bbl
\end{document}
|
\section{Introduction}
Computer simulation of quantum mechanical systems is an indispensable tool
in all physical sciences dealing with nanoscale phenomena.
Except for specific and rare cases~\cite{verstraete08}, classical
computers have not been able to efficiently simulate quantum systems, as in all known techniques at
least one of the
computational resources required to perform the simulation scales
exponentially with the size of the system being simulated.
Numerous classical approximative methods, such as density functional
theory~(DFT~\cite{ParrYang}) and quantum Monte Carlo
(QMC~\cite{WMH09}) have been developed to address various
aspects of the efficiency problem, but no known polynomially scaling methods are universally
applicable. Each suffers from particular deficiencies
such as the fermionic sign problem of QMC or the approximate
exchange-correlation functionals of DFT.
Quantum computers on the other hand, as conjectured by
Feynman~\cite{Feynman1982}, may be used to simulate other quantum
mechanical systems efficiently.
Feynman's conjecture was subsequently proven and
expanded leading to the rapidly growing area
of study known as
\textit{quantum simulation}~\cite{Lloyd1996,Lidar1999,Somma2002b,
Laflamme2004,ADLH2005,Brown2006,Lanyon2008Sub,whitfield-2010}.
Quantum simulation is expected to be able to produce classically
unattainable results in feasible run times, using only a modest number of
quantum bits.
For example, calculating the ground state energy of the water molecule to
the level of precision necessary for experimental predictions
($\approx 1$ kcal/mol) --- a problem barely solvable on current
supercomputers~\cite{chan_exact_2003} --- would require roughly~128
coherent quantum bits
on a gate-model quantum computer~\cite{ADLH2005}.
However, this calculation would also require
on the order of billions of quantum gates and thus quantum error correction
would
be essential~\cite{CBMG09}. The large number of gates and qubits needed
for
practical applications of
quantum simulation has lead current experimental work to
fine tune small systems~\cite{Lanyon2008Sub,FSG+08,GKZ09}.
Theoretical quantum simulation falls into two main categories:
\textit{dynamic
evolution}
and \textit{static properties}. Both categories
rely
heavily on the Trotter decomposition to handle
non-commuting terms in the Hamiltonian when mimicking
the unitary time propagator of the system to be simulated.
To approximate evolution under a Hamiltonian $H = \sum_{i=1}^k H_i$
consisting of $k$~non-commuting but local terms~$\{H_i\}_{i=1}^k$,
one applies the sequence of unitary gates
$\{e^{-iH_it/n}\}_{i=1}^k$ a total of $n$ times.
As number of repetitions~$n$
tends to infinity the error caused by the non-commutativity
vanishes and the approximation converges to the exact
result~\cite{Lloyd1996}. If each time slice requires a constant number of
gates independent of the
parameter
$t/n$, then reducing the error by repeating the sequence $n$ times can become
prohibitively expensive for high accuracy applications.
Constructing a practical method of quantum simulation is a significant
challenge (see in particular the recent work~\cite{CBMG09} which
analyzes fault tolerance in a gate-model quantum simulation algorithm).
Moving away from the gate model may offer a less
resource-intensive, and consequently a more feasible solution.
This letter addresses the problem by presenting a
hybrid model of quantum simulation,
consisting of an adiabatically controlled simulation
register coupled to a single gate-model readout qubit.
Our scheme can simulate the ground state properties
of arbitrary spin graph Hamiltonians.
Furthermore, it allows the measurement of the expectation value of any
constant $k$-local observable using a $k+1$-local measurement Hamiltonian.
Even in the presence of considerable noise at least some
information can be extracted.
In most quantum computing architectures, the natural interactions are
2-body. However, under certain conditions $k$-body
interactions can be well approximated using techniques such as
perturbative Hamiltonian gadgets~\cite{kempe2006,oliveira2008,biamonte-2008-78}
or average Hamiltonian methods~\cite{waugh1968approach}. In particular, reference~\cite{PhysRevLett.101.070503} considered the mapping between a given $n$-qubit target Hamiltonian with $k$-body interactions onto a simulator Hamiltonian with two-body interactions.
\textbf{Structure of this paper:}
We will continue by giving an overview of the simulator design,
including algorithm initialization, adiabatic evolution and
measurement. One of the key motivations for the proposed model is its
resistance to certain types of noise which is outlined in Section~\ref{sec:noise},
and we conclude in Section~\ref{sec:conclusion}.
The supporting material in the Appendix further details the specifics of our method
(including the full numeric simulation of the technique in the
presence of noise).
\section{Simulator overview}
We consider the simulation of systems represented by finite collections
of spins acted on by a time-independent Hamiltonian described by a graph
$G = (V, E)$ --- e.g.~~the Heisenberg and Ising models.
Each graph vertex~$v\in V$ corresponds to a spin acted on by a local
Hamiltonian $L_{v}$, and each edge~$e\in E$ to a
pairwise interaction~$K_{e}$ between the involved vertices.
The Hamiltonian~$H_{T}$ we wish to simulate is given by
\begin{equation}
H_{T} = \sum_{v \in V} L_v +\sum_{e \in E} K_e.
\end{equation}
\begin{comment}
\begin{figure}[h!]
\includegraphics[width=0.45\textwidth]{spin_graph}
\caption{Examples of sparse spin graphs.
a) 1D and b) 2D Heisenberg models:
$L_{v_i} = \frac{1}{2} h_i \sigma^z_i$,
$K_{e_{ij}} = -\frac{1}{2} \left(
J_{x} \sigma^x_i \sigma^x_j
+J_{y} \sigma^y_i \sigma^y_j
+J_{z} \sigma^z_i \sigma^z_j\right)$.
c) (pseudo) Random sparse graph representing a spin glass.}
\label{fig:spingraph}
\end{figure}
\end{comment}
The simulator consists of an adiabatically controlled
simulation register~$S$ with Hamiltonian~$H_S$, and a probe register~$P$
which will be acted on by gate operations and measured projectively.
We will engineer the probe~$P$ such that it behaves as a controllable
two-level system with orthonormal basis $\{\ket{p_0}, \ket{p_1}\}$.
Without loss of generality, the probe Hamiltonian can be expressed as
$H_P = \delta \projector{p_1}$, where~$\delta$ is the spectral gap
between the probe's ground and first excited states.
\textit{Initialization:}
We will first set the simulation register Hamiltonian~$H_S$
to~$H_{S,I}$ and prepare $S$ and $P$ in their respective
ground states. The Hamiltonian~$H_{S,I}$ has a simple
ground state which can be (i) computed classically and (ii) prepared
experimentally in polynomial time --- such as a classical approximation to
the
ground
state of the simulated system.
The simulator Hamiltonian is thus initially given by
\begin{equation}
H_0 = H_{S,I} \otimes \openone_P + \openone_S \otimes H_P.
\end{equation}
\begin{comment}
chemical systems
This section explains some methods to prepare an input state to the
simulation algorithm using adiabatic quantum
computation~\cite{FGGS00,WBL02,ADLH2005}. For our purposes,
The critical insight here is to pick an initial state that has some
overlap with the actual ground state of the physical system we wish to
simulate.
For example, such an approximation could be the Hartree-Fock
approximate wave function. This is the output of a classical
algorithm returning, in accessible time, a computational basis state
where the qubits which correspond to occupied molecular orbitals are
in the state $\ket{1}$ with the remaining qubits in state
$\ket{0}$. Such states are readily prepared, and often have sufficient
overlap to the actual ground state of the Hamiltonian we wish to
simulate.
When $|\bra{\phi_0^{HF}}\phi_0^{FCI}\rangle|=1-\epsilon$, where
$\epsilon$ is small, the phase estimation algorithm can be applied to
return the ground state energy estimate as the state of the system
will often collapse to $\ket{\phi_0^{FCI}}$ when measured in the $
{H}_{FCI}$ basis. To increase the algorithms efficiency, after the
system is prepared in state $\ket{\phi_0^{HF}}$, the Hamiltonian $
{H}_{FCI}$ is slowly applied and an adiabatic evolution into the
actual ground state can occur~\cite{FGGS00,WBL02,ADLH2005}. Consider a smooth one-parameter family
of adiabatic path Hamiltonians
\begin{equation}
{H(s)}=(1-s) {H}_{HF}+s {H}_{FCI},
\label{eq:adiabatic_path}
\end{equation}
for monotonic $s\in[0,1]$. This was the adiabatic path proposed
in~\cite{ADLH2005}. Other paths may be used as in
reference~\cite{PVAA+08} but in this study we restrict our attention
to evolution of the form~\ref{eq:adiabatic_path}.
\end{comment}
\textit{Adiabatic evolution:}
According to the adiabatic theorem~\cite{FGGS00},
a quantum system prepared in an energy eigenstate will remain near
the corresponding instantaneous eigenstate of the time-dependent
Hamiltonian governing the evolution if there are no level crossings
and the Hamiltonian varies slowly enough.
By adjusting the simulator parameters, we adiabatically change
$H_S$ from $H_{S,I}$ to $H_{S,T}$, the fully interacting Hamiltonian of the
system to be
simulated.
\begin{comment}
\begin{equation}
H_S(s) = (1-s) H_{S,I} +s H_{S,T}, \quad s \in [0,1]
\end{equation}
\end{comment}
Let us denote the ground state of~$H_{S,T}$ as~$\ket{s_{0}}$.
At the end of a successful adiabatic evolution $P$~is still in its ground
state~$\ket{p_0}$,
and $S$ is in (a good approximation to) the ground state of the simulated
system, $\ket{s_0}$. Hence the simulator is now in the ground state
$\ket{\text{g}} = \ket{s_0}\otimes\ket{p_0}$
of its instantaneous Hamiltonian
\begin{equation}
H_1 = H_{S,T} \otimes \openone_P + \openone_S \otimes H_P.
\end{equation}
The computational complexity of preparing ground states of quantum systems
has
been studied~\cite{kempe2006,oliveira2008,biamonte-2008-78}. It is
possible to
prepare a desired ground state efficiently provided that the gap between
the
ground and excited states is
sufficiently large~\cite{FGGS00}. This depends on the adiabatic path
taken. In general, finding the ground state energy of a Hamiltonian, even when restricted to certain simple models, is known to be complete for QMA, the quantum analogue of
the class NP~\cite{kempe2006,oliveira2008,biamonte-2008-78}.
Hence, it is considered unlikely that even a quantum computer would
be able to do this in polynomial time. In fact there are physical systems such as spin glasses
in nature which may never settle in their ground state.
However,
it is hoped (but not proven) that a host of realistic systems will satisfy the
property of having a large energy gap on even simple linear adiabatic paths
and thus be amenable to quantum
simulation algorithms which rely on adiabatic state preparation.
\textit{Measurement:}
The measurement procedure begins by bringing $S$~and $P$ adiabatically
into interaction.
The simulator Hamiltonian becomes
\begin{equation}
H_2 = H_1 + \underbrace{A \otimes \ketbra{p_1}{p_1}}_{H_{SP}},
\end{equation}
where the operator~$A$ corresponds to an observable of the simulated system
that
commutes with~$H_{S,T}$.
Assuming that $A$ can be decomposed into a sum of two-local operators,
the interaction term $H_{SP}$ involves three-local interactions. These terms can be implemented using either
Hamiltonian gadget techniques or average Hamiltonians (see the supporting material for details).
Let us use
$a_0 := \bra{s_0} A \ket{s_0}$
to denote the expectation value we wish to measure.
If $a_0 +\delta \ge 0$,
then $\ket{\text{g}}$ is also the ground state
of~$H_2$ (see the supporting material for proof).
At time $t=0$ we apply a Hadamard gate to the measurement probe
which puts it into a superposition of its two lowest states.
This is no longer an eigenstate of $H_2$, and the system will
evolve as
\begin{equation}
\ket{\psi(t)} = \frac{1}{\sqrt{2}} \ket{s_0} \otimes (\ket{p_0} +
e^{-i \omega t}\ket{p_1}),
\end{equation}
where $\omega := (a_0+\delta)/\hbar$. We have thus encoded the
quantity $a_0$, a property of the ground state of~$H_{S,T}$, into the
time dependence of the probe~$P$.
After a time $t$, we again apply a Hadamard pulse to the probe,
resulting in the state
\begin{equation}\label{eqn:phase}
\ket{\psi(t)} = \ket{s_0} \otimes \left(\cos \left(\omega t /2\right)
\ket{p_0}
+
i \sin \left(\omega t /2\right) \ket{p_1}\right),
\end{equation}
and then measure the probe. The probability of finding it in the state
$\ket{p_0}$ is
\begin{equation}
\label{eqn:prob}
P_0(t) = \frac{1}{2}\left(1 + \cos(\omega t)\right) =
\cos^2\left(\omega t /2\right).
\end{equation}
If we have non-demolition measurements (see e.g.~~\cite{siddiqi2005}) at
our disposal, then measuring the probe does not disturb the state of
the simulator which can
be reused for another measurement.
One repeats the measurement with different values of~$t$ until
sufficient statistics have been accumulated to reconstruct~$\omega$
and hence~$a_0$ --- this is reminiscent of Ramsey
spectroscopy~\cite{Ram63} and hence should seem natural to
experimentalists. In essence, we have performed Kitaev's phase estimation
algorithm~\cite{Kitaev1995}, using the interaction
Hamiltonian~$H_{SP}$ instead of the controlled unitary.
If the ground state subspace of~$H_{S,T}$ is degenerate and overlaps
more than one eigenspace of~$A$, or
the simulation register~$S$ has excitations to higher
energy states at the beginning of the measurement phase,
the probability of finding the probe in the state $\ket{p_0}$ is given
by a superposition of harmonic modes. For example, for a thermalized state
with inverse temperature~$\beta$, we obtain
\begin{equation}
\label{eqn:prob2}
P_0(t) = \frac{1}{2}\left(1 +\frac{1}{\sum_{xy} e^{-\beta E_x}} \sum_{kl} e^{-\beta E_k} \cos(\omega_{k,l} t)\right),
\end{equation}
where the first summation index runs over the energy eigenstates and
the second over the eigenstates of
$A$ in which energy has value~$E_k$, and $\omega_{k,l} = (a_{k,l}+\delta)/\hbar$.
\section{Effects of noise}
\label{sec:noise}
To assess the effects of noise on the simulation method, we performed
a numerical simulation of the simplest nontrivial implementation of the
hybrid simulator, consisting of two simulator qubits and one probe
qubit, with a randomly chosen~$H_T$.
Each qubit was coupled to its own bosonic heat bath with an Ohmic
spectral density using the Born-Markov approximation~\cite{BP}.
The qubit-bath couplings were chosen such that the resulting single-qubit
decoherence times $T_1$ and $T_2$ are compatible with
recent superconducting flux qubit experiments with fully tunable
couplings (see for example~\cite{niskanen2007}).
The noise model is described further in the supporting material.
The observable~$A$ was chosen to be $H_{S,T}$, the simulated
Hamiltonian itself, which means that the ground state energy was being
measured.
We simulated measurements on 40 evenly distributed values of the time
delay~$t$. At each $t_i$ we performed 50 measurements, and averaged
the results to get an estimate of~$P_0(t_i)$.
An exponentially decaying scaled cosine function was then fitted to
this set of data points to obtain an estimate for~$\omega$ and thus for~$s_0$.
The fit was done using the MATLAB {\sc lsqnonlin} algorithm, guided
only by fixed order-of-magnitude initial guesses for the parameter values.
\begin{figure}[h]
\includegraphics[width=0.70\textwidth]{new_ramsey}
\caption{Measurement procedure under Markovian noise.
The continuous curve represents the probability of finding
the probe in the state~$\ket{p_0}$, the circles
averaged measurement results and the dotted line a least squares fit to them.
$\hbar \omega_0 = h \cdot 25$~MHz is the energy scale of the Hamiltonians involved.
}
\label{fig:noise}
\end{figure}
The results of the simulation are presented in~\Figref{fig:noise}.
The noise, together with the slight nonadiabaticity of
the evolution, cause excitations out of the ground state
which result in a signal consisting of multiple harmonic
modes. However, the ground state mode still dominates and with a
realistic level of noise and relatively modest
statistics we are able to
reconstruct $\omega$ to a relative precision of
better than~$0.01$.
This includes both the uncertainty due to finite statistics, and the
errors introduced by the environment through decoherence and the Lamb-Stark
shift.
In an experimental implementation there may also be other noise
sources not considered here related to the measurement process itself,
as well as systematic (hardware) errors in the quantum processor
(qubits, couplers etc.).
\section{Conclusion}\label{sec:conclusion}
The presented simulation scheme significantly differs from gate-model
methods in
that it does not require millions of coherent gate
operations. Instead, an adiabatic control sequence is used, which
requires much less complicated control hardware and is likely to be more
robust against errors. A simple experimental implementation could be
feasible with present-day technology.
A direct comparison to a gate-model simulator using phase estimation to
extract the result is difficult because such a design would have to
use quantum error correction to be useful at the noise levels
we have considered.
At the measurement stage we do require single-qubit gate operations
and measurements, but only
on the probe qubit. Furthermore, these operations are relatively
simple to implement compared to a full Trotter decomposition of the
simulated Hamiltonian.
It is well known that fault tolerance in the adiabatic model depends
on the energy difference between the Hamiltonian's ground and first
excited states.
Several methods for the adiabatic model have been proposed which could in principle be used to increase
the fault tolerance of our protocol~\cite{Jordan2006,Lidar2008,AAN09}.
In order to simulate a system of $n$~qubits with a 2-local Hamiltonian
described by the graph~$G$,
ideally our scheme requires
one probe qubit for the readout and
$n$~simulation qubits.
Additionally, if Hamiltonian gadgets are used to implement 3-local
interactions,
one ancilla qubit is required for each two-local term in the
simulated observable (represented by an edge in the
corresponding graph).
In practice the number of ancillas may be slightly higher if more involved
types of gadgets are used to implement the 3-local interactions.
The total number of qubits required thus scales as $O(n)$ for sparse
graphs and $O(n^2)$ for maximally connected ones.
\begin{acknowledgments}
We thank Patrick Rebentrost, Man-Hong Yung, Viv Kendon and Terry Rudolph for helpful discussions.
This work received funding from the Faculty of Arts and Sciences at
Harvard University (JDB and VB), EPSRC grant EP/G003017/1 (JDB), Merton College (JF)
and Academy of Finland (VB).
VB visited Oxford using financial support from the EPSRC grant.
AA-G and VB thank DARPA under the Young Faculty Award N66001-09-1-2101-DOD35CAP, the Camille and
Henry Dreyfus Foundation, and the Sloan Foundation for support.
AA-G and JDW thank ARO under contract~{W911-NF-07-0304}.
\end{acknowledgments}
\bibliographystyle{apsrev}
|
\section{Introduction}
The monthly global temperature data are known to be strongly fluctuating.
Actually, the fluctuations are of the same order as the trend itself (see figures 1 and 2).
While the nature of the trend is widely discussed (in relation to the global warming) the
nature of these strong fluctuations is still quite obscure. The problem has also a technical
aspect: detrending is a difficult task for such strong fluctuations.
In order to solve this
problem a wavelet regression detrending method was used in present investigation. Then a
spectral analysis of the detrended data reveals rather surprising nature of the strong
global temperature fluctuations. Namely, the detrended fluctuations of the global temperature
for the last century (the instrumental monthly data are available at Ref. \cite{12data}, see also
Ref. \cite{s}) are completely dominated by so-called one-third subharmonic resonance to annual
forcing.
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig1.eps} \vspace{-4.5cm}
\caption{The monthly global temperature data (dashed line) for the period 1880-2009. The solid curve
(trend) corresponds to a wavelet (symmlet) regression of the data. }
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig2.eps} \vspace{-5.5cm}
\caption{The wavelet regression detrended fluctuations from the data shown in Fig. 1.}
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig3.eps} \vspace{-4.5cm}
\caption{Spectrum of the wavelet regression detrended fluctuations of
the monthly global temperature anomaly (land and ocean combined).}
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig4.eps} \vspace{-4.5cm}
\caption{The monthly global Sea Surface Temperature Anomalies (dashed line) for the period 1850-2008yy
(the data taken from Ref.\cite{sea-data}). The solid curve (trend) corresponds to a wavelet
(symmlet) regression of the data. }
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig5.eps} \vspace{-4.5cm}
\caption{The same as in Fig. 3 but for the wavelet regression detrended fluctuations of the global Sea Surface Temperature Anomalies.}
\end{figure}
\section{Subharmonic resonance to annual forcing}
There are many well known reasons for asymmetry in response of the North and South
Hemispheres to solar forcing: dominance of water in the Southern Hemisphere against
dominance of land in the Northern one, topographical imbalance of land (continents) and
oceans in the Northern Hemisphere due to continental configuration, seasonality and vegetation
changes are much more pronounce on land than on ocean surface, and anthropogenically induced asymmetry of the last century. This asymmetry results
in {\it annual} asymmetry of global heat budget and, in particular,
in annual fluctuations of the global temperature. Nonlinear responses are expected as a
result of this asymmetry.
Figure 1 shows (as dashed line) the instrumental monthly global temperature data
(land and ocean combined) for the period 1880-2009, as presented at the NOAA site \cite{12data}.
The solid curve (trend) in the figure corresponds to a wavelet (symmlet) regression of the data
(cf Refs. \cite{sw},\cite{o}). Figure 2 shows corresponding detrended fluctuations, which produce
a statistically stationary set of data. Most of the regression methods are linear in responses.
At the nonlinear nonparametric wavelet regression one chooses a relatively small number of wavelet
coefficients to represent the underlying regression function. A threshold method is used to keep or
kill the wavelet coefficients. In this case, in particular, the Universal (VisuShrink) thresholding
rule with a soft thresholding function was used. At the wavelet
regression the demands to smoothness of the function being estimated are relaxed considerably in comparison
to the traditional methods.
Figure 3 shows a spectrum of the wavelet regression detrended data
calculated using the maximum entropy method (because it provides an optimal spectral
resolution even for small data sets). One can see in this figure a small peak corresponding
to a one-year period and a huge well defined peak corresponding to a three-years period.
In order to understand underlying physics of the very characteristic picture shown in the Fig. 3
let us imagine a forced excitable system with a large amount of loosely coupled degrees of freedom
schematically represented by Duffing oscillators (which has become a classic model for analysis of
nonlinear phenomena and can exhibit both deterministic and chaotic behavior \cite{ot}-\cite{b}
depending on the parameters range) with a wide range of the
natural frequencies $\omega_0$ (it is well known \cite{tw} that oscillations with a wide range of
frequencies are supported by ocean and atmosphere, cf also Ref. \cite{bkb}):
$$
\ddot{x} + \omega_0^2 x +\gamma \dot{x} +\beta x^3 = F \sin\omega t \eqno{(1)}
$$
where $\dot{x}$ denotes the temporal derivative of $x$, $\beta$ is the strength of nonlinearity, and
$F$ and $\omega$ are characteristic of a driving force. It is known (see for instance Ref. \cite{nm})
that when $\omega \approx 3\omega_0$ and $\beta \ll 1$ the equation (1) has a resonant solution
$$
x(t) \approx a \cos\left(\frac{\omega}{3}t + \varphi \right) + \frac{F}{(\omega^2-\omega_0^2)}
\cos \omega t \eqno{(2)}
$$
where the amplitude $a$ and the phase $\varphi$ are certain constants.
This is so-called one-third subharmonic resonance with the driving frequency
$\omega$ corresponding to the {\it annual} NS-asymmetry of the solar forcing
(the huge peak in Fig. 3 corresponds to the first term in the right-hand side of the Eq. (2)).
For the considered system of the oscillators an effect of synchronization can take place
and, as a consequence of this synchronization, the characteristic peaks in the spectra
of partial oscillations coincide \cite{nl}.
It can be useful to note, for the global climate modeling, that the odd-term subharmonic resonance
is a consequence of the reflection symmetry of the natural nonlinear oscillators
(invariance to the transformation $x \rightarrow -x$, cf. also Ref. \cite{mj}).
\begin{figure} \vspace{-1cm}\centering
\epsfig{width=.45\textwidth,file=fig6.eps} \vspace{-4cm}
\caption{Defect of autocorrelation function versus $\tau$ (in ln-ln scales) for the wavelet regression detrended fluctuations of the global Sea Surface Temperature Anomalies. The straight line is drawn in order to indicate
scaling Eq. (3). }
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig7.eps} \vspace{-2cm}
\caption{Spectrum of sea surface height fluctuations \cite{zw} (TOPEX/
Poseidon and ERS-1/2 altimeter measurements) in the logarithmic scales.
The profound peak corresponds to the annual cycle.
The dashed straight line is drawn in order to indicate correspondence to the scaling Eq. (3).}
\end{figure}
\section{Role of Rossby waves}
The fluctuations of oceanic temperature cause certain variations of the sea surface height.
These variations are intermixed with the sea surface height variations caused by the oceanic planetary Rossby waves.
The oceanic planetary Rossby waves play an important role in the response of the global
ocean to the forcing (see, for instance, Refs. \cite{pl},\cite{zw}) and they are of fundamental importance
to ocean circulation on a wide range of time scales (it was also suggested that the Rossby
waves play a crucial role in the initiation and termination of the $El~Ni\tilde{n}o$ phenomenon, see also below).
Therefore, they present a favorable physical background for the global subharmonic resonance.
For that reason it is interesting to look also separately at global sea surface temperature anomalies.
These data for time range 1850-2008yy are shown in Fig. 4 (the monthly data were taken from
Ref. \cite{sea-data}, see also Ref. \cite{s2}). The solid curve (trend) in the figure
corresponds to the wavelet (symmlet) regression of the data.
Figure 5 shows a spectrum of the wavelet regression detrended data
calculated using the maximum entropy method. The spectrum seems to be
very similar to the spectrum presented in Fig. 3.
Since the high frequency part of the spectrum is corrupted by strong fluctuations (the Nyquist frequency
equals 0.5 [$mon^{-1}$]), it is interesting to look at corresponding autocorrelation function $C(\tau)$ in order to understand what happens on the monthly scales. It should be noted that scaling of defect of the autocorrelation
function can be related to scaling of corresponding spectrum:
$$
1-C(\tau) \sim \tau^{\alpha}~~~~~~\Leftrightarrow~~~~~~~~ E(f) \sim f^{-(1+\alpha)} \eqno{(3)}
$$
Figure 6 shows the defect of the autocorrelation function in ln-ln scales in order to estimate the scaling
exponent $\alpha \simeq 0.6 \pm 0.04$ (the straight line in this figure indicates the scaling Eq. (3)).
The existence of oceanic Rossby waves was confirmed rather recently by NASA/CNES TOPEX/Poseidon satellite
altimetry measurements. Corresponding to these measurements spectrum of the sea surface height
fluctuations, calculated in Ref. \cite{zw} with {\it daily} resolution (see also \cite{zfw}),
is shown in Figure 7. The dashed straight line in this figure is drawn in order to indicate
correspondence to the scaling Eq. (3): $1+\alpha \simeq 1.6$.
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig8.eps} \vspace{-4.5cm}
\caption{The monthly Global-SST ENSO index (dashed line) for the period 1850-2008yy (the data taken from Ref. \cite{sst-enso}, the index is in hundredths of a degree Celsius) . The solid curve (trend) corresponds to a wavelet (symmlet) regression of the data. }
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig9.eps} \vspace{-4.5cm}
\caption{The same as in Fig. 3 but for the wavelet regression detrended fluctuations
of the Global-SST ENSO index.}
\end{figure}
\begin{figure} \vspace{-0.5cm}\centering
\epsfig{width=.45\textwidth,file=fig10.eps} \vspace{-4.5cm}
\caption{The same as in Fig. 6 but for the wavelet regression detrended fluctuations
of the Global-SST ENSO index. }
\end{figure}
\section{El~Nino phenomenon}
The Rossby waves (together with Kelvin waves) and a strong atmosphere-ocean feedback provide physical
background for the $El~Ni\tilde{n}o$ phenomenon (see, for instance, Ref. \cite{tzi} and references therein).
Figure 9 shows spectrum for the wavelet detrended fluctuations
of the so-called Global-SST ENSO index (Fig. 8), which captures the low-frequency part
of the $El~Ni\tilde{n}o$ phenomenon (the monthly data are available at Ref. \cite{sst-enso}). The annual
forcing can come from the oceanic Rossby waves (cf Fig. 7).
To support this relationship we show in figure 10 defect of autocorrelation function calculated using the
wavelet detrended fluctuations from Fig. 8.
The ln-ln scales have been used in Fig. 10 in order to estimate the scaling
exponent $\alpha \simeq 0.6 \pm 0.03$ (the straight line in this figure indicates the scaling Eq. (3),
cf Figs. 6 and 7). Using these observations one can suggest that the $El~Ni\tilde{n}o$ phenomenon
has the one-third subharmonic resonance as a background. \\
The data were provided by National Climatic Data Center at NOAA and by Joint Institute for the Study
of the Atmosphere and Ocean. I also acknowledge that a software provided
by K. Yoshioka was used at the computations.
|
\section{Introduction}
In this paper we study the generalized Ornstein- Uhlenbeck (OU) equation:
\begin{equation}\label{e12}
\left\{ \begin{array}{ll}{{\partial \over {\partial t}}}X_t(x)=
-m X_t(x)+\eta(t)&,\, m>0
\\ X_0(x)= \phi(x),\,\,\,(t,\phi(x))\,\in ]0,\,\infty[\times {\mathbb R}^d
\end{array}
\right.
\end{equation}
Where $\eta=\eta(t)$ is a L\'evy noise (more details will be given in section 2).\\
Generalized Ornstein-Uhlenbeck processes has been introduced by Barndorff-Nielsen (1998). However by replacing the continuous time $t$ by a discrete one, we introduce the generalized random process, then by the use of the well known Bochner-Minlos theorem, see eg. \cite{GV}, we construct the L\'evy noise $\eta$.\\
A graphical calculus will be applied to the generalized OU-process in this paper. In fact we will introduce a special types of graphs called rooted trees with two types of leaves, see.eg \cite{BS1, BS2} an analytic value will be given to each rooted trees and therefore a graphical representation of the OU-process will be recalled in this paper.\\
The procedure of given each graph a numerical value is called "perturbation theory". However our model of representing the solution of equation (\ref{e12}) in terms of rooted trees with two types of leaves is not restricted to such equation in fact one can generalize equation (\ref{e12}) and make it more complicated, e.g by introduction of non linear terms, e.g for force $F$ of gradient type, $F=\nabla V$ we obtain a non-linear SDE and one can ask the same questions as before, again expansion into graphs is possible. Generalizing further we pass from SDE's driven by L\'evy noise to SPDE's.\\
Graphical representation of the OU process seems to be of a great importance in fact they can be used for modeling biological processes such as neuronal response, see e.g \cite{Fu}, also in Mathematical Finance, see. eg \cite{BN}, the modeling of the dynamics of interest rates and volatilities of asset prices where each notions will have a specific graph representation.\\
The remainder of the article is organized as follows:\\
The next section will be reserved to the construction of L\'evy noise, we will develop a interesting model which help the reader to understand the other sections of our paper.\\
In section 3 we recall a theorem which determine the solution of equation (\ref{e12}), the distribution of the generalized OU process will be given as well as a particular case when $\eta$ will be Gaussian noise.\\
In section 4 we introduce the rooted trees with two types of leaves. We will develop an algorithm which gives a numerical value to each rooted tree, in the last part of this section we will recall a graphical representation of the generalized OU process.
\section{ Construction of the L\'evy noise}
This section is devoted to the construction of the L\'evy noise $\eta$, so far we introduce the L\'evy characteristic $\psi$. By taking the time to be discrete the Fourier transform of the joint distribution of $\eta$ will converge to a function which depend of $\psi$, the generalized random process will be given as well. Hence by the use of the well known Bochner Minlos theorem we recall a theorem which gives the L\'evy noise.\\
It is well known that an infinite divisible probability distribution $P$ is a probability distribution which satisfy the existence of a probability distribution $P_n$ such that $P=P_n\ast\cdot\cdot\cdot \ast P_n$ (n times).\\
Let $P$ be an infinitely divisible probability distribution. By L\'evy-Khinchine theorem
(see e.g. \cite{DA}), the Fourier transform. or
characteristic function, of $P$ satisfies:\\
1. $C_{P}(t)=\int_{{\mathbb R}}e^{i\langle
s\,, t\rangle}\,dP(s)=e^{\psi(t)},\,\,t\in {\mathbb R}$, where $\psi$ is the
L\'evy characteristic of $P$ which is uniquely represented by ,see.
eg Lukacs \cite{Lu},
\begin{equation} \label{1.aa eqa} \psi(t)=iat-{{\sigma^2
t^2}\over 2}+z\int_{{\mathbb R}\setminus \{0\}}\left(e^{ist}-1\right) \,
dM(s),\,\,\,\,\, \forall \,\,t \in\,\, {\mathbb R}.
\end{equation}
with $M$ satisfies
$\int_{{\mathbb R}\setminus\{0\}}\min(1,s^2)dM(s)<\infty$.\\
We can explain the meaning of (\ref{1.aa eqa}) as follows:\\
If $\sigma^2=M=0,$ then $\psi(t)=iat$ and the process $\{X(t),\,\,t\geq 0\}$ is simply deterministic motion, here $a$ is called the drift and it's physical interpretation is the velocity of this motion.\\Now if only $M=0$ so that $\psi(t)=iat-{1 \over 2}{\sigma^2}t^2$ and therefore $C_{P}(t)=e^{[iat-{1 \over 2}{\sigma^2}t^2]}$ this is a characteristic function of a Gaussian random variable $X(t)$ having mean $ta$ and covariance $t\sigma$. The process $\{X(t),\,\,t\geq 0\}$ in this case is known as the Brownian motion.\\The last case is when $a\neq 0,\,\sigma \neq 0$ and $M \neq 0$, in this case the process $\{X(t),\,\,t\geq 0\}$ can be represented by $X(t)= bt+\sqrt{a}\,B(t)+\,N(t)$ where $b$ is the drift, $B(t)$
the Brownian motion and $N(t)$ will be recognized as a poisson process with intensity $\lambda$ taking values in $\{nh\,\,n\in {\mathbb N}\}.$\\
In the following we denote by $F$ the probability distribution on ${\mathbb R}_+$ of a given stochastic process.
\begin{Def}\label{d4}
A Generalized stochastic process $\eta(t)$ is a stochastic process such that the following proprieties are satisfied:
\begin{itemize}
\item $F(\eta(t))=F(\eta(s))\,,\,\,\forall\,t,\,s\,\geq\,0$
\item $\eta(t)$ is independent of $\eta(s),$ i.e $F(\eta(t)\eta(s))=F(\eta(t))\otimes F(\eta(s))$
\end{itemize}
\end{Def}
From now our stochastic process $\eta$ will be a generalized stochastic process and
we would like to construct a generalized L\'evy noise:\\
Let us start with a model, where the continuous time $t\,\in\,{\mathbb R}_+$
is replaced by discrete time $t\,\in\,{1\,\over n}{\mathbb N}.$ In this case one can thus
model the noise $\{\eta^{(n)}(t)\}_{t\,\in\,{1\,\over
n}{\mathbb N}}$ by a collection of i.i.d. random variables such that
the "global random fluctuations" that the noise gives should not
depend on the lattice scale $1 \over n$ fixed in our model, i.e to
model the noise independently of the lattice scale, we need that:
\begin{eqnarray}
F(\eta^{(1)}(1))&=&F\left(\eta^{(n)}({1 \over n})+
\eta^{(n)}({2 \over n}) +\cdot\cdot\cdot +\eta^{(n)}({n \over n})
\right)\nonumber\\&=& \left(F(\eta^{(n)}(1)) \right)^{\ast\,n}
\end{eqnarray}
Here $F(\eta^{(1)}(1))$ is the random fluctuation of the unit interval.\\
\begin{Def}The Fourier transform of $\eta^{(n)}(t)$ is given by
\begin{equation}
{\cal{F}}({\eta^{(n)}})(f)(t)=e^{{\psi(f(t))}\over n},\,
f(t)\,\in\,{\mathbb R}^d
\end{equation}
\end{Def}
Where $\psi$ is the L\'evy characteristic given in equation (\ref{1.aa eqa}).
\begin{prop}The Fourier transform of the joint distribution of
${\eta^{(n)}}({1 \over n}),\,{\eta^{(n)}}({2 \over n}),\,... $ converges to $e^{\int_{{\mathbb R}_+}\,\psi(f(t))\,dt},$ as $n \longrightarrow\,\infty.$
\end{prop}
\noindent{\bf Proof.} \,we have:
\begin{eqnarray}
{\cal{F}}\left(\eta^{(n)}({1 \over n}).\,\eta^{(n)}({2 \over
n}).\,...\right)(f({1 \over n}).\,f({2 \over
n}).\,...)&=&\dprod_{t\,\in\,{{1\,\over n}{\mathbb N}}}\,e^{\psi(f(t))\over
n}\nonumber\\&=&e^{\sum_{t\,\in\,{{1\,\over n}{\mathbb N}}}\,{\psi(f(t))\over
n}}\nonumber
\end{eqnarray}
It is now easy to see that the last expression converges to
$e^{\int_{{\mathbb R}_+}\,\psi(f(t))\,dt},$ as $n\longrightarrow\,\infty.$\\ {\hspace*{15cm
$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ }}\\
Now it turns out that, though the above limit exists, it does not
give us the definition of $\eta(t)$ as a random function of $t$, as
the limit:\\
\begin{equation}\label{e9}
F\left((\eta^{(n)}(1))\right)=\delta_0
\end{equation}
shows that the fluctuations of the random variable $\eta^{(n)}(t)$
vanish for $n\,\longrightarrow\,\infty.$\\
The solution of this apparent paradox lies in the fact, that
$\eta(t)$ is not a stochastic process in the classical sense, but rather a distribution in $t$ then a function of $t.$
\begin{Def}\label{d5}
A generalized random process or a random field is a mapping:$$
\phi:(\Omega,\,{\mathcal{B}},\,P)\,\longrightarrow\,{\mathcal{S}}'({\mathbb R},\,{\mathbb R}^d)$$
such that $\omega\,\longrightarrow\,\langle\,\phi(w),\,f \rangle$ is
measurable and for $f_n\,\longrightarrow\,f$ in
${\mathcal{S}}({\mathbb R},\,{\mathbb R}^d),\,
F(\phi(f_n))\,\longrightarrow\,F(\phi(f))$ in low, i.e.
\begin{eqnarray}
\lim_{n\,\longrightarrow\,\infty}\int_{\Omega}F(\phi(f_n))\,dP
&=&\lim_{n\,\longrightarrow\,\infty}\int_{{\mathbb R}^d}F(x)\,F(\phi(f_n))(x)\,dx
\nonumber\\&=&\int_{{\mathbb R}^d}F(x)\,F(\phi(f))(x)\,dx\nonumber\\&=&\int_{\Omega}F(\phi(f))\,dP\nonumber
\end{eqnarray}
\end{Def}
\begin{theo}\label{t1}(P. L\'evy) Let $X$ and $X_n,\,n\,\in\,{\mathbb N}$ be a
stochastic processes then:\\$X_n\,\longrightarrow\,X$ in low if and
only if ${\cal F}(X_n)(k)\,\longrightarrow
{\cal F}(X)(k),\,\forall\,k\,\in\,{\mathbb R}^d.$
\end{theo}
\begin{Def}\label{d6}
Two random fields $\phi_1$ and $\phi_2$ are equivalent in low if all
finite dimensional distributions of $\phi_1$ and $\phi_2$
coincide,\\i.e. If
$\forall\,f_1,...,f_n\,\in\,{\cal S}({\mathbb R},\,{\mathbb R}^d)$ and
$\forall\,{\cal{A}}_1,...,{\cal{A}}_n\,\in\,{\cal{B}}({\mathbb R})$
$$P\left(\phi_1(f_1)\in {\cal{A}}_1,...,\phi_1(f_n)\in\,{\cal{A}}_n\right)=
P\left(\phi_2(f_1)\in
{\cal{A}}_1,...,\phi_2(f_n)\in\,{\cal{A}}_n\right)$$ which means
that:
\begin{equation}F\left(\phi_1(f_1)\cdot\cdot\cdot\phi_1(f_n)
\right)=F\left(\phi_2(f_1)\cdot\cdot\cdot\phi_2(f_n)
\right)\end{equation}
\end{Def}
But how can one construct such random field?\\
We note by ${\cal B}$ the $\sigma$-algebra generated by the cylinder
sets of ${\cal S}'({\mathbb R}^d)$. Then $({\cal S}'({\mathbb R}^d),\,{\cal B})$ is a measurable space.\\
We define a characteristic functional on ${\cal S}({\mathbb R}^d)$, as a
functional $C:\,{\cal S}({\mathbb R}^d)\,\longrightarrow\,{\mathbb C}$ such that :\\
\begin{enumerate}\item $C(0)=1.$\\\item $C$ is continuous on $S({\mathbb R}^d)$;\\
\item $C$ is positive-definite, i.e. $\forall\,z_1,\,...,\,z_n\,\in \,{\mathbb C}, \forall
n\,\in\,{\mathbb N},\,f_1,\,...,\,f_n\,\in\,{\cal{S}}({\mathbb R}^d)$
\begin{equation}
\ds_{l,\,j\,=0}^{n}\,z_l\,\bar{z}_{j}\,{\cal C}(f_l-f_j)\,\geq0\,.
\end{equation}
\end{enumerate}
In the following we choose $(\Omega,\,{\cal B})=({\cal S'},\,{\cal
B}({\cal S}')).$\\
By the well known Bochner Minlos theorem, we know that for
a given a Characteristic functional ${\cal C}$, there exist a
unique (up to equivalence in low) random field $\phi$ such that:
\begin{equation}\label{e10}C(f)=\mathbb{E}[e^{i\,\phi(f)}]=\int_{\Omega}
e^{i\,\phi(f)(\omega)} \,dP(\omega),\,\forall \,\,f \,\,\in
\,{\cal S}({\mathbb R},\,{\mathbb R}^d)\end{equation}
\begin{theo}\label{t2}
Let $\psi$ be a L\'evy characteristic given by the representation (\ref{1.aa eqa}) then
there exist a unique measure $\eta$ on ${\cal S}'$ such that:
\begin{equation}\label{e11} C(f)= \int_{{\cal S}'}
e^{i\,\langle\,\omega,\,f\rangle} \,d\eta(\omega)=\exp\left( \int_{{\mathbb R}^d} \psi(f(x))\,dx
\right),\,\,\,\,x\,\,\in {\mathbb R}^d.
\end{equation}
\end{theo}
\noindent{\bf Proof.}
It is easy to see that the right hand side of (\ref{e11}) is a characteristic equation, now Bochner Minlos theorem concludes. \\ {\hspace*{15cm
$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ }}
\begin{Def}
The measure $\eta$ given by theorem (\ref{t2}) is called white measure with L\'evy characteristic $\psi$, and $({\cal S'}({\mathbb R}^d),\,\beta,\,\eta)$ will be the generalized white noise space associated with $\psi$.
The associated coordinate process
\begin{equation}\label{EQ1} \eta :S({\mathbb R}^d)\times (S'(\Gamma),\,{\cal B},\,\eta)\longrightarrow\,{\mathbb C},~~\eta(f,\,\xi)=\langle f,\,\xi\rangle~\forall f\in {\cal S}({\mathbb R}^d),~ \xi\in {\cal S}'(\Gamma)
\end{equation}
is called a L\'evy noise.
\end{Def}
\section{Generalized Ornstein-Uhlenbeck processes}
In this section the generalized stochastic process $\eta$ will be taken as a L\'evy noise in the sense of the previous section. We recall a result which gives the distribution of the generalized OU process, this later is known as "Mehler's Formula."\\
When there is no confusion we note $X(t)$ by $X_t$.
\begin{Def}\label{d1} The stochastic process $(X_t)_{t\geq 0}$ verifying equation (\ref{e12}), where $\eta$ is a L\'evy noise is called the generalized Ornstein-Uhlenbek (OU) process.
\end{Def}
\begin{theo}\label{pr1}
The solution of the generalized OU- equation (\ref{e12}) is given by:
\begin{equation}\label{e13}
X_t(x)=e^{-m\,t}\,\phi(x)+\int_0^t e^{-m\,(t-t')}\,\eta(t')\,dt'
\end{equation}
\end{theo}
\noindent{\bf Proof.}
The solution of the Homogeneous equation associated to equation (\ref{e12}) is given by\\
$X_{t,C}= C\,e^{-m\,t},$ now by the method of variation of parameters one thus get
\begin{equation}\label{e14}
X_t(x)=\int_0^t e^{-m\,(t-t')}\,\eta(t')\,dt'+\,C\,e^{-m\,t}
\end{equation}
The result now follows by taking $X_0(x)=C=\phi(x).$
\\{\hspace*{15cm $~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ }}\\
The problem now is how to define the integral in theorem (\ref{pr1})?\\
Let $e_j=(0,..., 1,0...)\,\in\,{\mathbb R}^d$, we put
\begin{equation}\label{e13}
\chi_{m,\,t}^j(t')=e_j.\,{\bf 1}_{t'\leq\,t}\,\,e^{-m\,(t-t')}
\end{equation}
and
\begin{equation}\label{e13}
X^j(t)=e^{-m\,t}\,\phi^j\,+\,\eta(\chi_{m,\,t}^j)
\end{equation}
This looks better, but $\chi_{m,\,t}^j$ is not a test function in
${S}({\mathbb R},\,{\mathbb R}^d).$\\
Let
\begin{equation}\label{e14}
\chi_{m,\,t,\,n}^j(t')=e_j\,\chi^n(t-t')\,e^{-m\,(t-t')}
\end{equation}
Here $\chi^n$ is a sequence of $\mathcal{C}^\infty-$test functions that
approximate the function from below.
\begin{prop}\label{p1}
The sequences $\{X^j_n\}_{n\,\geq 1}$ converges in $L^2(\Omega,\,P)$, as $n\,\longrightarrow\,\infty.$
\end{prop}
\noindent {\bf Proof.}\\
\begin{eqnarray}
\mathbb{E}\left[\left(X^j_{m,n}(t)-X^j_{m,n'}(t)\right)^2\right]\nonumber&=&\mathbb{E}\left[\eta \left(\chi^j_{m,t,n}-\chi^j_{m,t,n'}\right)^2\right]
\\ \nonumber&=& \alpha\,\int_0^\infty\,\left(\chi^j_{m,t,n}-\chi^j_{m,t,n'}\right)^2\,dt-\beta\,\left(\int_0^\infty\,(\chi^j_{m,t,n}-\chi^j_{m,t,n'})\,dt\right)^2
\end{eqnarray}
where $\alpha,\,\beta$ are constants, now by the use of the dominated convergence theorem it is clear that the last expression goes to $0$ as $n,n$ goes to infinity.\\ \hspace*{14cm} $~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ \\
Let ${\bf F}$ be the distribution of the generalized OU process, then the following result holds:
\begin{theo}\label{p2}
The distribution $\bf{F}$ of the generalized OU process is given by:
\begin{equation} {\bf F}(X_t)={\bf F}(e^{-mt}X_0)\ast {\bf F}(\eta(\chi_{m,t})) \end{equation}
\end{theo}
\noindent {\bf Proof.} The Fourier transform of $X_t$ is given by:
\begin{eqnarray}
{\cal F}(X_t)(p)\nonumber&=&\exp\left(i\,\langle X_0,\,p\rangle{e^{-{mt}}}+\,\int_0^t\,\psi( {\bf{1}}_{t'\leq t}e^{-m(t-t')}p)dt'\right)
\\ \nonumber&=& \exp\left(i\,\langle X_0,\,p\rangle{e^{-{mt}}}\, +\,\int_0^t\,\psi(e^{-m(t-t')}p)dt'\right)
\end{eqnarray}\\
\hspace*{14cm}$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ \\
\begin{cor}\label{c1}
1)- Let $\eta$ be a Gaussian noise, then the distribution of the OU process is given by:
\begin{equation}\label{RAH1}
{\bf F}(X_t)(x)={(detA)^{-{1\over 2}} {({{2\,\pi}\over m} (1-e^{-2mt}))^{-{d\over
2}}}}\,\exp{\left\{-{1 \over 2}{m\,\over {1-e^{-2mt}}}\langle (x-e^{-mt}X_0),\,D^{-1}(x-e^{-mt}X_0)\rangle
\right\}}dx
\end{equation}
where $A=\left({D\over m}(1-e^{-2mt})\right)^{-1}.$\\
2)- The distribution of the Brownian motion is given by:
\begin{equation}\label{RAH2}
{\bf F}(B_t)(x)={(detA)^{-{1\over 2}} {({{4\,\pi}})^{-{d\over
2}}}}\,\exp{\left\{-{1 \over {4t}}\langle (x-X_0),\,D^{-1}(x-X_0)\rangle
\right\}}dx
\end{equation}
\end{cor}
\noindent {\bf Proof.}
1)- If the noise $\eta$ is taken to be Gaussian noise, then $\psi(k)=-\langle k,\,Dk\rangle$, thus:
\begin{eqnarray}
\int_0^t\,\psi(e^{-m(t-t')}k)dt'\nonumber&=&-\langle k,\,Dk\rangle \int_0^t\,e^{-2m(t-t')}k\,dt'\\ \nonumber&=&-\langle k,\,Dk\rangle {1 \over{2m}}(1-e^{-2mt})
\end{eqnarray}
Hence by applying the Fourier integral, one thus obtains:
\begin{eqnarray}
{\bf F}(X_t)(x)\nonumber&=&{\sqrt{detA}\over {(2\,\pi)^{d\over
2}}}\,\exp{\left\{-{1 \over 2}\langle (x-e^{-mt}X_0),\,A(x-e^{-mt}X_0)\rangle
\right\}}dx\\ \nonumber&=&{(detA)^{-{1\over 2}} {({{2\,\pi}\over m} (1-e^{-2mt}))^{-{d\over
2}}}}\,\exp{\left\{-{1 \over 2}{m\,\over {1-e^{-2mt}}}\langle (x-e^{-mt}X_0),\,D^{-1}(x-e^{-mt}X_0)\rangle
\right\}}dx
\end{eqnarray}
Here $A=\left({D\over m}(1-e^{-2mt})\right)^{-1}.$\\
2)- If $m\,\longrightarrow 0$ the right hand side of equation (\ref{RAH1}) converges to
\begin{equation}{(detA)^{-{1\over 2}} {({{4\,\pi}})^{-{d\over
2}}}}\,\exp{\left\{-{1 \over {4t}}\langle (x-X_0),\,D^{-1}(x-X_0)\rangle
\right\}}dx
\end{equation}
which the distribution of the Brownian motion.
\\ \hspace*{14cm}$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ \\
\begin{rem}
1. The expression of ${\bf F}(X_t)$ given by equation (\ref{RAH1}) is known as "Mehler's Formula".\\
2. For the Gaussian case, i.e, $\psi(k)=-\langle k, D k\rangle$ one thus obtain the Heat equation:
\begin{equation}\label{e13}
\left\{ \begin{array}{ll}{{\partial P_t}(x) \over {\partial t}}=
\Delta\,P_t(x)
\\ P_0(x)= f(x)
\end{array}
\right.
\end{equation}
3. The formula given by equation (\ref{RAH2}) is the semigroup of the Brownian motion. Such a formula for the non Gaussian case is missing\,!\\
\end{rem}
If now $\psi(k)$ is taken to be of the more general type, i.e, $\psi(k)$ is a L\'evy-Khinchine function, see e.g \cite{DA}, one thus get:
\begin{eqnarray}
\label{1.6eqa}
{{\partial } \over {\partial t}}P_t(x)&=&-\sum_{j=1}^d a_j{\partial\over\partial {x_j}}P_t(x)+ \sum_{j,l=1}^d D_{jl}{\partial^2 \over\partial {x_j}\partial{x_l}}P_t(x)\nonumber\\
&&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+z\int_{{\mathbb R}^d\setminus\{0\} } [P_t(x+y)-P_t(x)]\, dr(y).
\end{eqnarray}
Here $a\in{\mathbb R}^d$, $D$ is a real and positive semidefinite $d\times d$ matrix, $z\geq 0$ and $r$ is a probability measure on ${\mathbb R}^d\setminus \{0\}$ s.t. its Fourier transform is entire analytic. The well-known interpretation of $a$
is the drift vector, $D$ determines the diffusion part,
whereas $z$ and $r$ give the frequency and distribution of jumps, respectively.\\
$\psi$ is called the generator of the (distribution of the) L\'evy process and the function \\$k\,\longrightarrow\,\psi(ik)$ is called the symbol of $\psi.$
\section{graph expansion of the generalized OU process}
This section will be reserved to the graph applications to equations of type (\ref{e12}). We recall a graphical representation of the generalized OU process by a special type of graphs called rooted trees. Let us start with some useful tools which help the reader to understand our formalism.\\
Let $G(t,\,{ x})$ be the Green function which satisfies :
\begin{equation}\label{green}
\left\{ \begin{array}{ll}{{\partial G}(t,\,{ x}) \over {\partial
t}}= -m G(t,\,{ x})+ \delta( x)&,
\\ G(t,\,{\bf x})=0,\,\,\, \,\,\,t<0\,\,\,\,\,\,\,\,\,
\end{array}
\right.
\end{equation}
Here $\delta(x)$ is the Dirac distribution.
\begin{prop}\label{P1}
The analytic solution of the generalized OU equation (\ref{e12}) is given by:
\begin{equation}\label{E1}
X_t=G\ast\,\eta(x)+G\,\star\,f(x)
\end{equation}
\end{prop}
\noindent {\bf Proof.}
From the generalized OU equation we obtain
\begin{equation}
\begin{array}{ll}{{({\partial \over {\partial
t}}}+\,m)}\,X_t(x)=\,\eta(x)
\end{array}
\end{equation}
Now the result hold by applying the Green function given by equation (\ref{green}).
\\ \hspace*{14cm}$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ \\
\begin{Def}\label{TR1}
A tree is a graph in which any two vertices are connected by exactly one simple path. In other words, any connected graph without cycles is a tree.\footnote{The various kinds of trees used as data structures in computer science are not really trees in this sense, but rather, types of ordered directed trees.}
A tree is called a rooted tree if one vertex has been designated the root, in which case the edges have a natural orientation, towards or away from the root.\\
Each vertex of a tree $T$ connected with only one edge is called leaf.
\end{Def}
In our model we consider a rooted tree $T$ with two types of leaves. The set of all trees with two types of leaves and $i$ inner vertices will be denoted by ${\cal T}(i).$\\
Now it's of a great importance to ask about a method which gives an analytic value to each rooted tree $T \,\in\,{\cal T}(i)$. The following algorithm named as "Feynman rule" gives a numerical value to each rooted tree with two types of leaves:
\begin{Def}\label{D2}
Let ${\cal R}(T,\,\eta)$ be the analytic value of a rooted tree $T \,\in\,{\cal T}(i)$, then ${\cal R}$ is obtained as follows:\\
- Assign $r=(t,x)$ to the root of the tree $T$.\\
- For each edge of the tree $T$ multiply by $G$, where $G$ is the green function given by equation (\ref{green}).\\
- For each leaf of type one multiply by the noise $\eta$.\\
- For each leaf of type two multiply by the initial condition $f$.\\
- Integrate with respect to the Lebesgue measure $dx$.
\end{Def}
Definitions (\ref{TR1}) and (\ref{D2}) can be summarized by the following table and graph:
\begin{table}[t]
\centerline{
\begin{tabular}{l|c|c|c|c|c|}
\hline
root & inner vertices& leaf 1& leaf 2&edge \\ \hline
$\times$&$\bullet$ &$\diamond$& $\otimes$&$\longrightarrow$ \\
\hline
\end{tabular}}
\caption{\label{1tab} Different types of vertices.}
\end{table}
\begin{figure}
\psset{xunit=1cm,yunit=1cm,runit=1cm,shortput=tab}
\begin{pspicture}(-4,-1)(1,1)
\cnode*(2,0){2.5pt}{C_1}
\dotnode[dotstyle=x,dotscale=2](1.2,0){D}\ncline{C_1}{D}
\dotnode[dotstyle=otimes,dotscale=2](2.9,0.6){F_1}
\ncline{A_1}{F1}\ncline{C_1}{F1}
\dotnode[dotstyle=diamond,dotscale=2](2.9,-0.6){F_2}
\ncline{B}{B_2}\ncline{B_2}{B_1}
\ncline{C_1}{F_1} \ncline{C_1}{F_2}
\end{pspicture}
\caption{\label{bg.fig}
Construction of a rooted tree with two types of leaves and one inner vertex.}
\end{figure}
\begin{theo}\label{T1}
Let $T_j \in {\cal T}(i),\,j\,\in {\mathbb N}$. The solution of the generalized OU equation (\ref{e12}) is given by a sum over all rooted trees that are evaluated according to the rule given in definition (\ref{D2}), i.e.
\begin{equation}
X_t= \ds_{T_j\in {\cal T}(i)}\,{\cal R}(T_j,\,\eta)
\end{equation}
\end{theo}
\noindent {\bf Proof.}
From proposition (\ref{P1}) one thus get:
\begin{eqnarray}
\label{1.Eeqa} X_t&=&G\ast\,\eta(x)+G\,\star\,f(x)\nonumber\\&=&\GA+\GB\nonumber\\&=&{\cal R}(T_1,\,\eta)+{\cal R}(T_2,\,\eta)
\end{eqnarray}
\\ \hspace*{14cm}$~~\mbox{\hfil\vrule height6pt width5pt depth-1pt}$ \\
The graphical representations introduced previously can be applied for more complicated OU equations, e.g by replacing the right hand side of equation (\ref{e12}) by $-m\,X^p+\,\eta$, in this case our graph formalism still true and the rooted trees will contain more inner vertices, hence in the rule given by definition (\ref{D2}) we will integrate over all vertices.
|
\section{Introduction}
Let $G$ be a group acting on a finite set $\Omega$ with its orbits $\Omega_1,\dots,\Omega_n$ and its permutation character $\pi=\sum_{i=1}^{n}\pi_i$ where $\pi_i(g):=|\set{\alpha\in \Omega_i \mid \alpha^g=\alpha}|$ for $g\in G$. One may think what happens if $\pi_i=\pi_j$ for all $1 \leq i,j\leq n$ and can say that the number of orbits of $G$ on $\Omega_i\times \Omega_j$ by its entry-wise action is constant for all $1\leq i,j\leq n$, which motivate us to define the following concepts whose terminology is due to \cite{Inp2009}.
\begin{definition} \rm{Let $V$ be a finite set and $\mathcal{R}$ a set of nonempty binary relations on $V$. The pair $\mathcal{C}=(V,\mathcal{R})$ is called a
\textit{coherent configuration} (for short \textit{scheme}) on $V$ if the following conditions hold:}
\begin{enumerate}[(C1)]
\item $\mathcal{R}$ forms a partition of the set $V\times V$.
\item $\Delta_V:=\set{(v,v)~|~ v\in V}$ is a union of certain relations from $\mathcal{R}$.
\item For every $R\in\mathcal{R}$,\quad$R^t:=\set{(v,u)\mid(u,v)\in R}\in \mathcal{R}$.
\item For every $R,S,T\in\mathcal{R}$, the size of $\set{w\in V\mid (u,w)\in R,\ (w,v)\in S}$ does not depend on the
choice of $(u,v)\in T$ and is denoted by $c_{RS}^{T}$.
\end{enumerate}
We say that the elements of $V$ are {\it points} and those of $\mathcal{R}$ are \it{basis relations}.
\end{definition}
\noindent Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme and $\emptyset\neq X\subseteq V$. We say that $X$ is a {\it fiber} of $\mathcal{C}$ if $\Delta_X=\set{(x,x)\mid x\in X}\in\mathcal{R}$. We denote by $\Fib(\mathcal{C})$ the set of all fibers of $\mathcal{C}$.
\begin{definition} Let $m,n$ and $r$ be positive integers. We say that a scheme $\mathcal{C}$ is an {\it$(m,n,r)$-scheme} if the following conditions hold:
\begin{enumerate}[(i)]
\item $|\set{R\in\mathcal{R}\mid R\subseteq X\times Y}|=r$ for all $X,Y\in\Fib(\mathcal{C})$.
\item $|X|=m$ for all $X\in\Fib(\mathcal{C})$.
\item $|\Fib(\mathcal{C})|=n$.
\end{enumerate}
A scheme $\mathcal{C}$ is called {\it $r$-balanced} ~if {\rm (i)} holds, and {\it balanced} if it is $r$-balanced for some $r$. In Section
\ref{Section:characterization balanced} we will show that (i) implies (ii).
\end{definition}
Let us return to the topic in the first paragraph. Note that the orbits of $G$ on $\Omega\times \Omega$ form the basis relations of a scheme called
the \textit{2-orbit scheme} of $G$ on $\Omega$ and its fibers are $\Omega_1,\dots,\Omega_n$. Furthermore, if $\pi_i=\pi_j$ for all $1\leq i,j\leq n$,
then the 2-orbit scheme of $G$ on $\Omega$ is balanced.
We denote by $\mathcal{P}(\mathcal{C})$ the set of all central primitive idempotents of the adjacency algebra of $\mathcal{C}$ (see Section \ref{Section:Preliminaries} for details). The following theorem shows a characterization of balanced schemes in terms of their central primitive idempotents.
\begin{theorem}\label{THM:main1}
Let $\mathcal{C}$ be a scheme. Then $\mathcal{C}$ is balanced if and only if for each $X\in\Fib(\mathcal{C})$ the mapping $\mathcal{P}(\mathcal{C})\longrightarrow\mathcal{P}(\mathcal{C}_X)$ {\rm($P\mapsto P_X$)} is bijective with $n_P=|\Fib(\mathcal{C})|n_{P_X}$.
\end{theorem}
One may conclude that $|\mathcal{P}(\mathcal{C})|=r$ if $\mathcal{C}$ is $r$-balanced and $r\leq 5$ (see Corollary \ref{cor:r-balanced,p(C)=r}).
The following theorem deals with the converse argument for $r=1,2$.
\begin{theorem}\label{THM:main2} Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme. Then the following hold:
\begin{enumerate}[(i)]
\item $|\mathcal{P}(\mathcal{C})|=1$ if and only if $\mathcal{C}$ is $1$-balanced.
\item $|\mathcal{P}(\mathcal{C})|=2$ if and only if $\mathcal{C}=\mathcal{C}_1\boxplus \mathcal{C}_2$ where $\mathcal{C}_i$ is $i$-balanced.
\end{enumerate}
\end{theorem}
We have the following constructions of balanced schemes (see Sections \ref{Section:characterization balanced}, \ref{Section:reduced} for the details):
\begin{enumerate}[(i)]
\item Let $U$ be a union of fibers of $\mathcal{C}$. Then the restriction of $\mathcal{C}$ to $U$ is $r$-balanced if $\mathcal{C}$ is $r$-balanced.
\item If~$\mathcal{C}_i$ $(i=1,2)$ is an $(m_i,n_i,r_i)$-scheme, then $\mathcal{C}_1\bigotimes \mathcal{C}_2$ is an $\big(m_1m_2,n_1n_2,r_1r_2\big)$-scheme.
\end{enumerate}
We say that a balanced scheme $\mathcal{C}$ is {\it reduced} if there exist no $X,Y\in\Fib(\mathcal{C})$ such that $\mathcal{C}_{X\cup Y}\simeq \mathcal{C}_X\bigotimes \mathcal{T}_2$ where $\mathcal{T}_2$
is a $(1,2,1)$-scheme (in Section \ref{Section:reduced} you will see another equivalent condition for a scheme to be reduced).
Any $r$-balanced scheme is obtained by the restriction of the tensor product of a reduced $r$-balanced scheme and a $1$-balanced scheme
(see Theorem \ref{THM:equivalence}). Now we focus our attention on reduced balanced schemes. It seems a quite difficult problem to find possible
$n$ such that there exists a reduced $(m,n,r)$-scheme for given $m$ and $r$. Actually, D.~G. Higman asked if there exists a reduced $(m,3,3)$-scheme for some $m$ (see \cite[Section 8, p.229]{Higman1987}).
Furthermore, H. Weilandt conjectured that a permutation group of prime degree has at most two inequivalent permutation representations (see \cite{Cam1972}). The following theorem shows that $n=1$ is a unique case under certain assumptions.
\begin{theorem}\label{THM:main3} Let $\mathcal{C}$ be a reduced $(m,n,r)$-scheme and $p$ a prime. Then we have the following:
\begin{enumerate}[(i)]
\item \label{THM:m<2r} If $m<2r$, then $n=1$.
\item \label{THM:p-val} If $p\nmid m$ and $\mathcal{C}_X$ is $p$-valanced for some $X\in\Fib(\mathcal{C})$, then $n=1$.
\end{enumerate}
\end{theorem}
The preceding theorem is applied to characterize $(m,n,r)$-schemes up to $m\leq 11$ as follows.
\begin{theorem}\label{THM:main4} If $\mathcal{C}$ is a reduced $(m,n,r)$-scheme and $m\leq11$, then $n\leq2$.
\end{theorem}
Let us show the organization of this article. In Section \ref{Section:Preliminaries} we prepare some terminologies related to schemes. Section \ref{Section:characterization balanced} is devoted to balanced schemes. First we investigate the features of balanced
schemes. Indeed, we shall characterize a balanced scheme in terms of its central primitive idempotents and we prove Theorem \ref{THM:main1}.
Secondly we shall characterize schemes with at most two central primitive idempotents and we prove Theorem \ref{THM:main2}.
In Section \ref{Section:reduced} we shall extend the notion of inequivalent permutation representations to schemes. Namely, we shall define reduced $(m,n,r)$-schemes and then introduce some examples and known constructions of them to support our theory.
Finally in Section \ref{Section:Enumeration}, first we prove Theorem \ref{THM:main3}, secondly we shall enumerate reduced $(m,n,r)$-schemes for $m\leq11$ in order to prove Theorem \ref{THM:main4}.
\section{Preliminaries}\label{Section:Preliminaries}
According to \cite{Inp2009} we prepare some terminologies related to schemes. For the remainder of this section we assume that $\mathcal{C}=(V,\mathcal{R})$ is a scheme.
One can see that $V=\bigcup_{X\in\Fib(\mathcal{C})}X$~(disjoint union) and
\begin{equation}\label{qtn:partition of R}
\mathcal{R}=\bigcup_{X,Y\in \text{Fib}(\mathcal{C})} \mathcal{R}_{X,Y}\quad\quad(\text{disjoint union}),
\end{equation}
where $\mathcal{R}_{X,Y}:= \{R\in\mathcal{R}\mid R\subseteq X\times Y\}$. We shall denote $\mathcal{R}_{X,X}$ by $\mathcal{R}_X$.
Let $X,Y\in\Fib(\mathcal{C})$ and $R$ be a non-empty union of basis relations in $\mathcal{R}_{X,Y}$. For $(x,y)\in R$ we set $R_{out}(x)=\set{u\mid(x,u)\in R}$ and $R_{in}(y)=\set{v\mid(v,y)\in R}$. We shall denote the size of $R_{out}(x)$ and that of $R_{in}(y)$ by $d_R$ and $e_R$, respectively. It is easy to see that
\begin{equation}\label{qtn:degree in&out}
|X|d_R=|R|=|Y|e_R.
\end{equation}
For $\mathcal{D}\subseteq\mathcal{R}$ we define $d_\mathcal{D}:=\sum_{R\in\mathcal{D}}d_R$ as well as $e_\mathcal{D}:=\sum_{R\in\mathcal{D}}e_R$. For instance $d_{\mathcal{R}_{X,Y}}=|Y|$ and $e_{\mathcal{R}_{X,Y}}=|X|$.
For all $X,Y\in\Fib(\mathcal{C})$, we define the multi-set $d_{X,Y}:=\big\{d_R\mid R\in\mathcal{R}_{X,Y}\big\}$.
Note that $d_R=e_R$ for each $R\in\mathcal{R}$ if and only if $|X|=|Y|$ for all $X,Y\in\Fib(\mathcal{C})$.
A scheme $\mathcal{C}$ is called \textit{half-homogeneous} if the latter condition holds.
If $\mathcal{C}$ is a half-homogeneous scheme, then $d_R$ ($=e_R$) is called the \textit{degree} or the \textit{valency} of $R$. Given a prime $p$ a
half-homogeneous scheme $\mathcal{C}$ is called {\it $p$-valenced} if the degree of each basis relation of $\mathcal{C}$ is a power of $p$.
A basis relation $R\in\mathcal{R}$ is called \textit{thin} if $d_R=e_R=1$ and a scheme $\mathcal{C}$ is called a \textit{homogeneous scheme} or (\textit{association scheme}) if $|\text{Fib}(\mathcal{C})|=1$
or equivalently, if $\Delta_V\in \mathcal{R}$ (for more details regarding association schemes we refer to \cite{Zieschang2005}).
Given $X\in\Fib(\mathcal{C})$ the pair $\mathcal{C}_X=(X,\mathcal{R}_X)$ is a homogeneous scheme called the \textit{homogeneous component} of $\mathcal{C}$ corresponding to $X$.
\begin{definition} For each $R\in\mathcal{R}$ we define a $\set{0,1}$-matrix $A_R$ whose rows and columns are simultaneously indexed by the elements of $V$ such that
the $(u,v)$-entry of $A_R$ is one if and only if $(u,v)\in R$. Then $A_R$ is called the {\it adjacency matrix} of $R$. Note that the subspace of
$\Mat_V(\mathbb{C})$ spanned by $\set{A_R\mid R\in\mathcal{R}}$ is a subalgebra, called the {\it adjacency algebra} of $\mathcal{C}$ and denoted by $\mathcal{A}(\mathcal{C})$.
\end{definition}
Let $\mathcal{A}$ be the adjacency algebra of $\mathcal{C}=(V,\mathcal{R})$. Then the set $\set{A_R\mid R\in\mathcal{R}}$ which is a basis of $\mathcal{A}$ satisfies the following conditions:
\begin{enumerate}
\item[(C$^{'}$1)] $J_V=\displaystyle\sum_{R\in\mathcal{R}}A_R \in \mathcal{A}$ where $J_V$ is the matrix whose entries are all one.
\item[(C$^{'}$2)] The identity matrix $I_V \in \mathcal{A}$.
\item[(C$^{'}$3)] $A_{R^t}=A_R^t$ for every $R\in\mathcal{R}$ where $A_R^t$ is the transpose of $A_{R}$.
\item[(C$^{'}$4)] For every $R, S \in \mathcal{R}$,\quad$A_RA_S=\displaystyle\sum_{T\in\mathcal{R}}c_{RS}^{T}A_T$.
\end{enumerate}
A scheme is called \textit{trivial} if all its fibers are singletons. We denote a trivial scheme on $n$ points by $\mathcal{T}_n$. Note that $\mathcal{A}(\mathcal{T}_n)\cong\Mat_n(\mathcal{C})$ and it is easy to see that a scheme is trivial if and only if it is 1-balanced.
By $\Fib^*(\mathcal{C})$ we mean the set of all non-empty unions of fibres of $\mathcal{C}$. Given $U\in \Fib^*(\mathcal{C})$ we set $\mathcal{R}_U:=\set{R_U\mid R\in\mathcal{R}}$ where $R_U=R\cap (U\times U)$. Then the pair $\mathcal{C}_U=(U,\mathcal{R}_U)$ is a scheme on $U$ called the {\it restriction} of $\mathcal{C}$ to $U$. Note that $\mathcal{C}_U$ is homogeneous whenever $U\in\Fib(\mathcal{C})$.\\
Given $U,U'\in\Fib^\ast(\mathcal{C})$ we define $\mathcal{A}_{U,U'}$ to be the subspace of $\mathcal{A}$ spanned by the set $\set{A_R\mid R\in\mathcal{R}, R\subseteq U\times U'}$. Then the following hold:
\begin{enumerate}[(i)]
\item Given $U\in\Fib^\ast(\mathcal{C})$ we have $\mathcal{A}_U=I_U\mathcal{A} I_U\cong \mathcal{A}(\mathcal{C}_U)$ where $I_U:=\sum_{\substack{X\in\Fib(\mathcal{C}),\\mathfrak{X}\subseteq U}}A_{\Delta_X}$.
\item For all $X,Y,Z,W\in\Fib(\mathcal{C})$, $\mathcal{A}_{X,Y}\mathcal{A}_{Z,W}\subseteq\delta_{YZ}\mathcal{A}_{X,W}$.
\item $\mathcal{A}=\bigoplus_{X,Y\in\Fib(\mathcal{C})}\mathcal{A}_{X,Y}$.
\end{enumerate}
A basis relation $S$ of $\mathcal{C}$ is called \textit{symmetric} if $S^t=S$ and $\mathcal{C}$ is called \textit{symmetric}
if each basis relation of $\mathcal{C}$ is symmetric; and $\mathcal{C}$ is called \textit{commutative} if $c_{RS}^T =c_{SR}^T$ for all $R, S,T\in \mathcal{R}$.
This is equivalent to $A_RA_S=A_SA_R$ for all $R, S \in \mathcal{R}$. It is known that symmetric schemes are commutative and that the converse does not hold.
Furthermore, one can see that a commutative scheme is a homogeneous one.
\begin{lemma}[{\rm \cite[(4.2)]{Higman1975}, \cite[Theorem 4.5.1]{Zieschang2005}}]\label{Lemma:r<5} If $\mathcal{C}=(V,\mathcal{R})$ is a homogeneous scheme and $|\mathcal{R}|\leq 5$, then $\mathcal{C}$ is commutative.
\end{lemma}
Given $R,S\in\mathcal{R}$ the \textit{complex product} of them is defined to be $RS=\big\{T\in\mathcal{R}\mid c_{RS}^T>0\big\}$ and the \textit{relational product}
$R\circ S$ is defined as follows.
$$R\circ S:=\Big\{(u,v)\mid \exists~~ w\in V; (u,w)\in R, (w,v)\in S\Big\}.$$ Note that $R\circ S=\bigcup_{T\in RS}T$ and $d_{R\circ S}=d_{RS}$.
\begin{lemma}\label{Lemma:constants} Let $\mathcal{C}$ be a scheme and $X,Y,Z\in\Fib(\mathcal{C})$. Then for all $R\in\mathcal{R}_{X,Y}$, $S\in\mathcal{R}_{Y,Z}$ and $T\in\mathcal{R}_{X,Z}$ the following
hold:
\begin{enumerate}[(i)]
\item\label{qtn:dRdS} $d_Rd_S=\displaystyle\sum_{T\in\mathcal{R}_{X,Z}}c_{RS}^{T}d_T$.
\item\label{qtn:lcm} $c_{RS}^Td_T=c_{TS^t}^Rd_R=c_{R^tT}^Sd_S$~and~${\rm lcm}(d_R,d_S)\mid c_{RS}^Td_T$.
\item\label{qtn:dgree-constant} $d_R=\displaystyle\sum_{S\in \mathcal{R}_{Y,Z}}c_{RS}^T$, $e_R=\displaystyle\sum_{S\in \mathcal{R}_{Y,Z}}c_{R^tT}^S$, $c_{RS}^T\leq {\rm min}\set{d_R,e_S}$ and $R\mathcal{R}_{Y,Z}=\mathcal{R}_{X,Z}$.
\item\label{qtn:Kronecker} $d_R\delta_{SR^t}=c_{RS}^{\Delta_X}$ and $e_R\delta_{SR^t}=c_{SR}^{\Delta_Y}$ where $\delta$ denotes the Kronecker's delta.
\item\label{qtn:d_RS} $d_S\leq d_{RS}\leq d_Rd_S$ and $e_R\leq e_{RS}\leq e_Re_S$.
\item\label{qtn:dT<dR} If $RS=T$ and $d_T\leq d_R$, then $R=TS^t$.
\item\label{qtn:symmetric-dR=2} If $d_R=2$, then $RR^t=\{\Delta_X,S\}$ where $S$ is a symmetric basis relation of $\mathcal{R}_{X}$ with $d_S\leq2$.
\item\label{qtn:gcd} $|RS|\leq {\rm gcd}(d_R,d_S)$.
\end{enumerate}
\end{lemma}
\begin{proof} The proof is done by the same procedure as \citep[Lemma 1.4.2, 1.4.3, 1.5.2, 1.5.3, 1.5.6]{Zieschang2005}.
\end{proof}
\begin{lemma}\label{Lemma:d_L_S} Let $S\in\mathcal{R}_{X,Y}$ and $L_S:=\set{R\in\mathcal{R}_X\mid RS=\set{S}}$. Then
$$d_{L_S}\mid {\rm gcd}\big(|X|,e_S\big).$$
\end{lemma}
\begin{proof} Let $y\in Y$ and $x\in S_{in}(y)$. The condition $RS=\set{S}$ shows that $\bigcup_{R\in L_S}R_{in}(x)\subseteq S_{in}(y)$ and $\bigcup_{R\in L_S}R$ is an
equivalence relation on $X$. Since $y\in Y$ and $x\in S_{in}(y)$ are arbitrarily taken, all equivalence classes have the same size $d_{L_S}$. It follows that $d_{L_S}$ divides both $d_S$ and $|X|$.
\end{proof}
\begin{lemma}\label{Lemma:constant2} Let $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$ and $R,S\in\mathcal{R}_{X,Y}$ with $R\neq S$. Then $T\in R^tR\cap S^tS$ for some $T\in\mathcal{R}_Y$ with
$T\neq\Delta_Y$ if and only if $c_{RS^t}^{T'}\geq2$ for some $T'\in\mathcal{R}_X$.
\end{lemma}
\begin{proof} Let us prove the necessity. By the assumption $c_{R^tR}^{T}\neq0$ and $c_{S^tS}^{T}\neq0$. Taking $(y,y')\in T$ (of course $y\neq y'$) there exist
$x,x'\in X$ such that $(x,y),(x,y')\in R$ and $(x',y),(x',y')\in S$. On the other hand, there exists $T'\in\mathcal{R}_X$ such that $(x,x')\in T'$. It follows that $c_{RS^t}^{T'}\geq2$ (see the Figure \ref{Fig:constant2}). Sufficiency follows from Figure \ref{Fig:constant2} since $c_{RS^t}^{T'}\geq2$ implies that $y\neq y'$.
\begin{figure}[H]
\begin{displaymath}
\xymatrix{
&\VRT{y} \ar@{<-}[dl]_{\scriptscriptstyle\mathit{R}}\ar[dd] \ar@{<-}[dr]^{\scriptscriptstyle\mathit{S}} & \\
\VRT{x} \ar[dr]_{\scriptscriptstyle\mathit{R}} \ar[rr]& & \VRT{x'} \\
&\VRT{y'}\ar@{<-}[ur]_{\scriptscriptstyle\mathit{S}}&}
\end{displaymath}\label{fig:6}
\vspace{-5mm}
\renewcommand{\captionlabeldelim}{.}
\caption{}\label{Fig:constant2}
\end{figure}
\end{proof}
Let $U,U'\in\Fib^*(\mathcal{C})$ such that $U\cap U'=\emptyset$ and $V=U\cup U'$. Then we say that $\mathcal{C}$ is the {\it internal direct sum} of $\mathcal{C}_U$ and $\mathcal{C}_{U'}$ if $|\mathcal{R}_{X,Y}|=1$ for all $X,Y\in\Fib(\mathcal{C})$ with $X\subseteq U$ and $Y\subseteq U'$. In this case we shall write $\mathcal{C}=\mathcal{C}_U\boxplus\mathcal{C}_{U'}$.
Let $\mathcal{C}_i=(V_i,\mathcal{R}_i)$ $(i=1,2)$ be schemes.
We set
$$
\mathcal{R}_1\otimes\mathcal{R}_2=\small\{R_1\otimes R_2\mid R_1\in\mathcal{R}_1,\ R_2\in\mathcal{R}_2\big\},
$$
where $R_1\otimes R_2=\Big\{\big((u_1,u_2),(v_1,v_2)\big)\mid (u_1,v_1)\in
R_1,\ (u_2,v_2)\in R_2\Big\}$. Then $\mathcal{C}=\big(V_1\times V_2,\mathcal{R}_1\otimes\mathcal{R}_2\big)$ is a scheme called the {\it tensor product} of $\mathcal{C}_1$ and $\mathcal{C}_2$ and
denoted by $\mathcal{C}_1\bigotimes\mathcal{C}_2$. One can see that $\Fib(\mathcal{C})=\Fib(\mathcal{C}_1)\times\Fib(\mathcal{C}_2)$.
An \textit{isomorphism} from $\mathcal{C}_1$ to $\mathcal{C}_2$ is defined to be
a bijection $\psi:V_1\cup\mathcal{R}_1\longrightarrow V_2\cup\mathcal{R}_2$ such that for all $u,v\in V_1$ and $R\in\mathcal{R}_1$, $(u,v)\in R$ if and only if $\big(\psi(u),\psi(v)\big)\in\psi(R)$. We say that $\mathcal{C}_1$ is \textit{isomorphic} to $\mathcal{C}_2$ and denote it by $\mathcal{C}_1\simeq \mathcal{C}_2$ if there exists an isomorphism from $\mathcal{C}_1$ to $\mathcal{C}_2$.
Let $\mathcal{A}$ be the adjacency algebra of $\mathcal{C}$. Since $\mathcal{A}$ is closed under the complex conjugate transpose map, $\mathcal{A}$ is semisimple. By the Wedderburn theorem $\mathcal{A}$ is isomorphic to a direct sum of full matrix algebras over~$\mathbb{C}$:
\begin{equation}\label{qtn:Wedderburn}
\mathcal{A}=\bigoplus_{P\in\mathcal{P}(\mathcal{C})}\mathcal{A} P\cong \bigoplus_{P\in\mathcal{P}(\mathcal{C})}\Mat_{n_P}(\mathbb{C}),
\end{equation}
where $\mathcal{P}(\mathcal{C})$ is the set of central primitive idempotents of $\mathcal{A}$, $n_P$ is a positive integer and $\Mat_{n_P}(\mathbb{C})$ is the full matrix algebra of complex $n_P\times n_P$ matrices.\\
A comparison of dimensions of the left- and right-hand sides of (\ref{qtn:Wedderburn}) shows that
\begin{equation}
|\mathcal{R}|=\sum_{P\in\mathcal{P}(\mathcal{C})}n_P^2.
\end{equation}
It is known that $\mathcal{C}$ is commutative if and only if $n_P=1$ for each $P\in\mathcal{P}(\mathcal{C})$.\\
Denote by $\mathbb{C}^V$ the natural $\mathcal{A}$-module spanned by the elements of $V$. As $I_V=\sum_{P\in\mathcal{P}(\mathcal{C})}P$ we have\\
\begin{equation}\label{decom}
\mathbb{C}^V=\bigoplus_{P\in\mathcal{P}(\mathcal{C})}P\mathbb{C}^V.
\end{equation}
For each $P\in\mathcal{P}(\mathcal{C})$ we set $m_P:=\dim_\mathbb{C}(P\mathbb{C}^V)/n_P$. Then the decomposition~(\ref{decom}) shows that
\begin{equation}\label{qtn:size of V}
|V|=\sum_{P\in\mathcal{P}(\mathcal{C})}m_Pn_P.
\end{equation}
The numbers $m_P$ and $n_P$ are called the {\it multiplicity} and the {\it degree} of $P$. Set $P_0=\sum_{X} J_X/|X|$ where $X$ runs over the fibers of the
scheme~$\mathcal{C}$ and $J_X=\sum_{R\in\mathcal{R}_X}A_R$. Note that $P_0$ is a central primitive idempotent of the algebra~$\mathcal{A}$, which is called {\it principal}.
It is known that
\begin{equation}\label{qtn:m_P0n_P_0}
(m_{P_0},n_{P_0})=(1,|\Fib(\mathcal{C})|).
\end{equation}
If $\mathcal{C}$ is homogeneous, then $P_0=J_V/|V|$ and $m_{P_0}=n_{P_0}=1$.
Below for $X\in\Fib^*(\mathcal{C})$ and $P\in\mathcal{P}(\mathcal{C})$ put $P_X=PI_X$ and set
$$\mathcal{P}_X(\mathcal{C})=\Big\{P\in \mathcal{P}(\mathcal{C})\mid\ P_X\ne 0\Big\}~~\text{and}~~ \Supp(P)=\Big\{X\in\Fib(\mathcal{C})\mid\ P_X\ne 0\Big\}.$$
\begin{theorem}[{\rm \cite[Proposition 2.1]{Inp1999}}] \label{THM:embeding} Let
$\mathcal{C}=(V,\mathcal{R})$ be a scheme. Then the following hold:
\begin{enumerate}
\item[{\rm~(i)}] For each $X\in\Fib^\ast(\mathcal{C})$ the mapping $P\mapsto P_X$ induces a bijection between $\mathcal{P}_X(\mathcal{C})$ and $\mathcal{P}(\mathcal{C}_X)$.
\item[{\rm ~(ii)}] For all $P\in\mathcal{P}(\mathcal{C})$ and $X\in \Supp(P)$, $n_P=\sum_{X\in \Supp(P)}n_{P_X}$ and $m_P=m_{P_X}$ .
\end{enumerate}
\end{theorem}
\begin{lemma}\label{Lemma:union-support} Let
$\mathcal{C}=(V,\mathcal{R})$ be a scheme. Then the following hold:
\begin{enumerate}[(i)]
\item $\mathcal{P}(\mathcal{C})=\mathcal{P}_X(\mathcal{C})$ for each $X\in\Fib(\mathcal{C})$ if and only if $\Supp(P)=\Fib(\mathcal{C})$ for each $P\in\mathcal{P}(\mathcal{C})$.
\item $\Supp(P)\neq\emptyset$ for each $P\in\mathcal{P}(\mathcal{C})$, and
\begin{equation}\label{qtn:union-support}
\mathcal{P}(\mathcal{C})=\displaystyle\bigcup_{X\in\Fib(\mathcal{C})}\mathcal{P}_X(\mathcal{C}).
\end{equation}
Besides, $\mathcal{P}(\mathcal{C})=\mathcal{P}_U(\mathcal{C})\cup\mathcal{P}_{U'}(\mathcal{C})$ where $U,U'\in\Fib^*(\mathcal{C})$ with $U\cap U'=\emptyset$ and $V=U\cup U'$.
\end{enumerate}
\end{lemma}
\begin{proof} (i) Let $X\in\Fib(\mathcal{C})$ and $P\in\mathcal{P}(\mathcal{C})$. Then $P\in\mathcal{P}_X(\mathcal{C})$ if and only if $X\in\Supp(P)$.
This completes the proof.\\
(ii) Let $P\in\mathcal{P}(\mathcal{C})$ such that $\Supp(P)=\emptyset$. Then for all $X\in\Fib(\mathcal{C})$, $PI_X=0$ and then
$P=PI_V=\sum_{X\in\Fib(\mathcal{C})} PI_X=0$, a contradiction. Therefore, $\Supp(P)\neq\emptyset$. Let $P\in\mathcal{P}(\mathcal{C})$, as $\Supp(P)\neq\emptyset$, there exists
$X\in\Fib(\mathcal{C})$ such that $PI_X\neq0$. This means that $P\in\mathcal{P}_X(\mathcal{C})$ and the proof of (\ref{qtn:union-support}) is completed.
Let $P\in\mathcal{P}(\mathcal{C})$. Then $P\in\mathcal{P}_X(\mathcal{C})$ for some $X\in\Fib(\mathcal{C})$. Since $V=U\cup U'$, $X\subseteq U$ or $X\subseteq U'$. It follows that $P\in\mathcal{P}_U(\mathcal{C})$
or $P\in\mathcal{P}_{U'}(\mathcal{C})$. This completes the proof.
\end{proof}
\begin{proposition} [{\rm \cite[p.223]{Higman1987}, \cite[p.22 (8.1)]{Higman1975}}]\label{Prop:Higman-d_XY} Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme.
Then the following hold:
\begin{enumerate}[(i)]
\item Let $X,Y\in\Fib^*(\mathcal{C})$ such that $X\cap Y=\emptyset$ and $V=X\cup Y$. Then
$$\dim_{\mathbb{C}}(\mathcal{A}_{X,Y})=\sum_{P\in \mathcal{P}_X\bigcap \mathcal{P}_Y}n_{P_X}n_{P_Y}.$$
\item For all $X,Y\in\Fib(\mathcal{C})$, $|\mathcal{R}_{X,Y}|=\displaystyle\sum_{P\in \mathcal{P}_X\bigcap \mathcal{P}_Y}n_{P_X}n_{P_Y}$.
\end{enumerate}
\end{proposition}
\begin{lemma}\label{Lemma:direct sum1}
Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme with the adjacency algebra $\mathcal{A}(\mathcal{C})$. If $U,U'\in\Fib^\ast(\mathcal{C})$ such that $U\cap U'=\emptyset$, then
\begin{equation}\label{qtn:direct sum1}
|\Fib(\mathcal{C}_U)||\Fib(\mathcal{C}_{U'})|\leq \dim_{\mathbb{C}}(\mathcal{A}_{U,U'}).
\end{equation}
Furthermore, the equality holds if and only if $\mathcal{C}_{U\cup U'}=\mathcal{C}_U\boxplus\mathcal{C}_{U'}$.
\end{lemma}
\begin{proof} It is clear that $\Fib(\mathcal{C}_U)=\set{X\in\Fib(\mathcal{C})\mid X\subseteq U}$ as well as $\Fib(\mathcal{C}_{U'})=\set{Y\in\Fib(\mathcal{C})\mid Y\subseteq U'}$. By the assumption,
$$
\mathcal{A}_{U,U'}=\bigoplus_{\substack{X\in\Fib(\mathcal{C}_U),\\Y\in\Fib(\mathcal{C}_{U'})}}\mathcal{A}_{X,Y}.
$$
On the other hand, $\dim_{\mathbb{C}}(\mathcal{A}_{X,Y})=|\mathcal{R}_{X,Y}|\geq1$ for each $X\in\Fib(\mathcal{C}_U)$ and $Y\in\Fib(\mathcal{C}_{U'})$. It follows that
$$|\Fib(\mathcal{C}_U)||\Fib(\mathcal{C}_{U'})|\leq\sum_{\substack{X\in\Fib(\mathcal{C}_U),\\Y\in\Fib(\mathcal{C}_{U'})}}|\mathcal{R}_{X,Y}|=\dim_{\mathbb{C}}(\mathcal{A}_{U,U'}).$$
The equality holds if and only if $|\mathcal{R}_{X,Y}|=1$ for all $X\in\Fib(\mathcal{C}_U)$ and $Y\in\Fib(\mathcal{C}_{U'})$ which is exactly the definition of internal direct sums.
\end{proof}
\begin{lemma}\label{Lemma:direct sum2} Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme with the principal idempotent $P_0$ and let $U,U'\in\Fib^\ast(\mathcal{C})$ such that $U\cap U'=\emptyset$ and $V=U\cup U'$.
Then $\mathcal{C}=\mathcal{C}_U\boxplus\mathcal{C}_{U'}$ if and only if
$\mathcal{P}_U(\mathcal{C})\cap\mathcal{P}_{U'}(\mathcal{C})=\set{P_0}$.
\end{lemma}
\begin{proof} Let us prove the sufficiency first. It is clear that $P_0\in \mathcal{P}_U(\mathcal{C})\cap\mathcal{P}_{U'}(\mathcal{C})$. By Lemma \ref{Lemma:direct sum1} and Proposition \ref{Prop:Higman-d_XY}~(i)
we have $$|\Fib(\mathcal{C}_U)||\Fib(\mathcal{C}_{U'})|=\sum_{P\in \mathcal{P}_U\cap\mathcal{P}_{U'}}n_{P_U}n_{P_{U'}}.$$
Since $n_{P_{0U}}=|\Fib(\mathcal{C}_U)|$ and $n_{P_{0U'}}=|\Fib(\mathcal{C}_{U'})|$, it follows that
\begin{equation}
\mathcal{P}_U(\mathcal{C})\cap\mathcal{P}_{U'}(\mathcal{C})=\set{P_0}.
\end{equation}
Conversely, if $\mathcal{P}_U(\mathcal{C})\cap\mathcal{P}_{U'}(\mathcal{C})=\set{P_0}$, then by Proposition \ref{Prop:Higman-d_XY}~(i), $$\dim_{\mathbb{C}}(\mathcal{A}_{U,U'})=|\Fib(\mathcal{C}_U)||\Fib(\mathcal{C}_{U'})|.$$ It follows from Lemma \ref{Lemma:direct sum1} that $\mathcal{C}=\mathcal{C}_U\boxplus\mathcal{C}_{U'}$.
\end{proof}
\section{Characterization of balanced schemes}\label{Section:characterization balanced}
\noindent\textbf{Proof of Theorem \ref{THM:main1}:}\\
First we prove the necessity. Let $X,Y\in\Fib(\mathcal{C})$. By Proposition \ref{Prop:Higman-d_XY},
$|\mathcal{R}_{X,Y}|=\sum_{P\in{\mathcal{P}_X\cap\mathcal{P}_Y}} n_{P_X}n_{P_Y}$. By the Cauchy-Schwarz inequality we have
\begin{eqnarray*}
|\mathcal{R}_{X,Y}|^2=\big(\sum_{P\in \mathcal{P}_X\cap\mathcal{P}_Y} n_{P_X}n_{P_Y} \big)^2\label{Cauchy-Schwarz}
&\leq& \sum_{P\in \mathcal{P}_X\cap\mathcal{P}_Y} n_{P_X}^2 \sum_{P\in\mathcal{P}_X\cap\mathcal{P}_Y}n_{P_Y}^2\\
&\leq& \sum_{P\in\mathcal{P}_X} n_{P_X}^2 \sum_{P\in\mathcal{P}_Y}n_{P_Y}^2\\
&=&|\mathcal{R}_{X}||\mathcal{R}_{Y}|=|\mathcal{R}_{X,Y}|^2.
\end{eqnarray*}
This implies that
$$\big(\sum_{P\in{\mathcal{P}_X\cap\mathcal{P}_Y}} n_{P_X}n_{P_Y}\big)^2=\sum_{P\in\mathcal{P}_X} n_{P_X}^2\sum_{P\in\mathcal{P}_Y}n_{P_Y}^2.$$
It follows that $\mathcal{P}_X(\mathcal{C})=\mathcal{P}_Y(\mathcal{C})$ and thus applying Lemma \ref{Lemma:union-support}~(i) we have $\mathcal{P}(\mathcal{C})= \mathcal{P}_X(\mathcal{C})$. Consequently, the mapping $\mathcal{P}(\mathcal{C})\rightarrow \mathcal{P}(\mathcal{C}_X)~(P\mapsto P_X)$ is well-defined and bijective by Theorem \ref{THM:embeding}. Since the equality holds in the Cauchy-Schwarz inequality,
we have $\Big\langle n_{P_X}|P\in\mathcal{P}(\mathcal{C})\Big\rangle=\alpha\Big\langle n_{P_Y}|P\in\mathcal{P}(\mathcal{C})\Big\rangle$.
However, $\alpha=1$ since $|\mathcal{R}_X|=|\mathcal{R}_Y|$. Hence, $n_{P_X}=n_{P_Y}$ for all $P\in\mathcal{P}(\mathcal{C})$. Therefore, by Theorem \ref{THM:embeding} and
Lemma \ref{Lemma:union-support}~(ii),
$$n_P=\sum_{X\in\Supp(P)}n_{P_X}=\sum_{X\in\Fib(\mathcal{C})}n_{P_X}=|\Fib(\mathcal{C})|n_{P_X}.$$
Now let us prove the sufficiency. Given $X,Y\in\Fib(\mathcal{C})$ the assumption along with Theorem \ref{THM:embeding} assert that $\mathcal{P}(\mathcal{C})=\mathcal{P}_X(\mathcal{C})=\mathcal{P}_Y(\mathcal{C})$ and $n_{P_X}=n_{P_Y}$ for each $P\in\mathcal{P}(\mathcal{C})$. On the other hand, by Preposition \ref{Prop:Higman-d_XY}~(ii), we have
$$|\mathcal{R}_{X,Y}|=\displaystyle\sum_{P\in{\mathcal{P}_X\cap\mathcal{P}_Y}} n_{P_X}n_{P_Y}=
\sum_{P\in\mathcal{P}(\mathcal{C})} n_{P_X}^2=|\mathcal{R}_X|.$$
Hence, $\mathcal{C}$ is balanced.\hspace{9.8cm}$\square$
\begin{corollary}\label{cor:r-balanced,p(C)=r} Let $\mathcal{C}$ be an $r$-balanced scheme. If $\mathcal{C}_X$ is commutative for some $X\in\Fib(\mathcal{C})$, then $|\mathcal{P}(\mathcal{C})|=r$. For instance, $|\mathcal{P}(\mathcal{C})|=r$ if $r\leq 5$.
\end{corollary}
\begin{proof} Let $X\in\Fib(\mathcal{C})$. Since $\mathcal{C}_X$ is commutative, $|\mathcal{P}(\mathcal{C}_X)|=|\mathcal{R}_X|=r$. By Theorem \ref{THM:main1},
$|\mathcal{P}(\mathcal{C})|=|\mathcal{P}(\mathcal{C}_X)|=r$. In particular, if $r\leq 5$, then by Lemma \ref{Lemma:r<5}, $\mathcal{C}_X$ is commutative and thus $|\mathcal{P}(\mathcal{C})|=r$.
\end{proof}
\noindent\textbf{Proof of Theorem \ref{THM:main2}~(i):}\\
Let $X\in\Fib(\mathcal{C})$. By Theorem \ref{THM:embeding}, $|\mathcal{P}(\mathcal{C}_X)|=1$. On the other hand, $\mathcal{C}_X=(X,\mathcal{R}_X)$ is a homogeneous scheme, so $|X|=m_{P_{0X}}n_{P_{0X}}=1$, by (\ref{qtn:m_P0n_P_0}).
Hence, every fiber of $\mathcal{C}$ is a singleton and thus $\mathcal{C}$ is trivial. Conversely, the adjacency algebra of a trivial scheme is the full matrix algebra and
thus it has only one central primitive idempotent.\hspace{+8cm}$\square$\\
In order to prove Theorem \ref{THM:main2}~(ii), we need the following preparations.
\begin{lemma}[{\rm \cite[(4.2)]{Higman1975}}]\label{Lemma:p=2homoge} Let $\mathcal{C}=(V,\mathcal{R})$ be a homogenous scheme. Then $|\mathcal{P}(\mathcal{C})|=2$ if and only if $|\mathcal{R}|=2$.
\end{lemma}
\begin{lemma}\label{Lemma:p=2} Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme. If $\mathcal{P}(\mathcal{C})=\set{P_0,P_1}$ with $P_0\neq P_1$, then the following hold:
\begin{enumerate}
\item [\rm~(i)] $X\notin \Supp(P_1)$ if and only if $|X|=1$.
\item [\rm~(ii)] $|\mathcal{R}_X|=
\begin{cases}
2 & \quad \text{if}~~X\in \Supp(P_1) \\
1 & \quad \text{if}~~X \notin \Supp(P_1)
\end{cases}.$
\end{enumerate}
\end{lemma}
\begin{proof} ~(i) Since $I_V=P_0+P_1$, $P_1=\sum_{X\in\Fib(\mathcal{C})}(I_X-J_X/|X|)$. Let $X\in\Fib(\mathcal{C})$. Then $X\notin \Supp(P_1)$ if and only if $0=P_1I_X=I_X-J_X/|X|$ if and only if $|X|=1$.\\
~(ii) If $X\in \Supp(P_1)$, then $P_{1}I_X\neq0$ and by
Theorem \ref{THM:embeding}, $|\mathcal{P}(\mathcal{C}_X)|=2$. Since $\mathcal{C}_X=(X,\mathcal{R}_X)$ is homogeneous, it follows from Lemma \ref{Lemma:p=2homoge} that $|\mathcal{R}_X|=2$.
If $X\notin \Supp(P_1)$, then by (i), we have $|X|=1$. It follows that
$|\mathcal{R}_X|=1$.
\end{proof}
\begin{lemma}\label{Lemma:half2} Let $\mathcal{C}=(V,\mathcal{R})$ be a scheme. If $\mathcal{P}(\mathcal{C})=\set{P_0,P_1}$ with $P_0\neq P_1$, then the following hold:
\begin{enumerate}
\item [\rm~(i)] $n_{P_1}=|\Supp(P_1)|$ and $|X|=1+m_{P_1}$ for each $X\in \Supp(P_1)$.
\item [\rm~(ii)] $|\mathcal{R}_{X,Y}|=2$ for each $X,Y\in\Supp(P_1)$.
\end{enumerate}
\end{lemma}
\begin{proof} (i) Let $X\in\Supp(P_1)$. By Lemma \ref{Lemma:p=2}, $|\mathcal{R}_X|=2$ and thus by Lemma \ref{Lemma:r<5}, $\mathcal{C}_X$ is commutative. By Theorem \ref{THM:embeding} we have
$$n_{P_1}=\sum_{X\in\Supp(P_{1})}n_{P_{1X}}=|\Supp(P_{1})|.$$
Thus (\ref{qtn:size of V}) implies that $|X|=1+m_{P_1}$.\\
\rm~(ii) Let $X,Y\in\Supp(P_1)$. Then by Lemma \ref{Lemma:p=2}~(ii), $|\mathcal{R}_Y|=|\mathcal{R}_X|=2$. It follows from Lemma \ref{Lemma:p=2homoge} that
$\mathcal{P}_X(\mathcal{C})\cap\mathcal{P}_Y(\mathcal{C})=\mathcal{P}(\mathcal{C})$. Therefore, Proposition \ref{Prop:Higman-d_XY}~(ii)
implies that $|\mathcal{R}_{X,Y}|=2.$
\end{proof}
\noindent\textbf{Proof of Theorem \ref{THM:main2}~(ii):} Let $\mathcal{P}(\mathcal{C})=\set{P_0,P_1}$ and set $U:=\bigcup_{X\in\Supp(P_1)}X$ and $U':=V\setminus U.$
If $X\in \Supp(P_1)$ and $Y\notin \Supp(P_1)$,
then $|\mathcal{R}_{X,Y}|=1$, since $|Y|=1$ by Lemma \ref{Lemma:p=2}. This implies that
$$\mathcal{C}=\mathcal{C}_U\boxplus \mathcal{C}_{U'}.$$
Note that by Lemma \ref{Lemma:p=2}~(ii) and Lemma \ref{Lemma:half2}~(ii), $\mathcal{C}_U$ is 2-balanced and $\mathcal{C}_{U'}$ is 1-balanced.
Conversely, by Lemma \ref{Lemma:direct sum2} and Corollary \ref{cor:r-balanced,p(C)=r},
$|\mathcal{P}(\mathcal{C})|=|\mathcal{P}(\mathcal{C}_1\boxplus\mathcal{C}_2)|=|\mathcal{P}(\mathcal{C}_1)|+|\mathcal{P}(\mathcal{C}_2)|-1=|\mathcal{P}(\mathcal{C}_2)|=2$.\hspace{+4cm}$\square$\\
\begin{corollary}\label{Corollary:half-homo} If $\mathcal{C}$ is a balanced scheme, then the following hold:
\begin{enumerate}[(i)]
\item $\mathcal{C}$ is half-homogeneous.
\item For every $X,Y\in\Fib(\mathcal{C})$, $\mathcal{A}_X$ and $\mathcal{A}_Y$ are isomorphic as $\mathbb{C}$-algebras.
\end{enumerate}
\end{corollary}
\begin{proof}
(i) Let $X\in\Fib(\mathcal{C})$ and consider the scheme $\mathcal{C}_X=(X,\mathcal{R}_X)$. It follows from Theorem \ref{THM:main1} that the mapping $\mathcal{P}(\mathcal{C})\longrightarrow \mathcal{P}(\mathcal{C}_X)$
{\rm($P\mapsto P_X$)} is bijective with $n_P=|\Fib(\mathcal{C})|n_{P_X}$. By (\ref{qtn:size of V}) and Theorem \ref{THM:embeding}~(ii), the size of $X$ is computed as follows.
\begin{eqnarray*}
|X|=\sum_{P\in\mathcal{P}(\mathcal{C})}n_{P_X}m_{P_X}=\frac{1}{|\Fib(\mathcal{C})|}\sum_{P\in\mathcal{P}(\mathcal{C})}n_{P}m_{P}=\frac{|V|}{|\Fib(\mathcal{C})|}.
\end{eqnarray*}
Hence, the size of each fiber is constant and thus $\mathcal{C}$ is half-homogeneous.\\
(ii) Let $X,Y\in\Fib(\mathcal{C})$ and $P\in\mathcal{P}(\mathcal{C})$. Then by Theorem \ref{THM:main1}, $n_{P_X}=n_{P_Y}$. It follows from (\ref{qtn:Wedderburn}) that
$$\mathcal{A}_X=\bigoplus_{P\in\mathcal{P}(\mathcal{C})}\Mat_{n_{P_X}}(\mathbb{C})\cong\bigoplus_{P\in\mathcal{P}(\mathcal{C})}\Mat_{n_{P_Y}}(\mathbb{C})=\mathcal{A}_Y.$$
\end{proof}
Given a scheme $\mathcal{C}$ we define a relation $E_{\mathcal{C}}$ on $\Fib(\mathcal{C})$ as follows.
\begin{equation}\label{qtn:EC}
E_{\mathcal{C}}:=\Set{(X,Y)\in\Fib(\mathcal{C})\mid \exists R\in \mathcal{R}_{X,Y}~;~d_R=e_R=1}.
\end{equation}
\begin{lemma} $E_{\mathcal{C}}$ is an equivalence relation on $\Fib(\mathcal{C})$.
\end{lemma}
\begin{proof} For each $X\in\Fib(\mathcal{C})$, $\Delta_X$ is a thin basis relation in $\mathcal{R}_X$ and thus $E_{\mathcal{C}}$ is reflexive. If $R\in\mathcal{R}_{X,Y}$ is thin,
then $R^t\in\mathcal{R}_{Y,X}$ is also thin and then $E_{\mathcal{C}}$ is symmetric. Let $X,Y,Z\in\Fib(\mathcal{C})$ and $R\in \mathcal{R}_{X,Y}, S\in \mathcal{R}_{Y,Z}$ such that $d_R=d_S=1$
and $e_R=e_S=1$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:d_RS}) that $RS$ is a thin basis relation in $\mathcal{R}_{X,Z}$ and thus $E_{\mathcal{C}}$ is
transitive.
\end{proof}
\begin{theorem}\label{THM:equivalence} Let $\mathcal{C}$ be an $(m,n,r)$-scheme with the equivalence relation $E_{\mathcal{C}}$. If $\mathfrak{X}$ is a transversal of $E_{\mathcal{C}}$ in $\Fib(\mathcal{C})$, then $\mathcal{C}$ is a restriction of $\mathcal{C}_{U_{\mathfrak{X}}}\bigotimes\mathcal{T}_n$ where $U_{\mathfrak{X}}=\bigcup_{X\in\mathfrak{X}}X$ and $\mathcal{T}_n$ is a $(1,n,1)$-scheme.
\end{theorem}
\begin{proof}
Let $I_n:=\set{1,\dots,n}$ and $E_n:=\set{e_{ij}~\mid~ 1\leq i,j \leq n}$ where for $1\leq i,j\leq n$, $e_{ij}=\set{(i,j)}$. Then $\mathcal{T}_n=(I_n,E_n)$.
Let $\mathfrak{X}=\set{X_1,\dots, X_s}$ and suppose that for each $i\in\set{1,2,\dots,s}$, $E_{\mathcal{C}}(X_i)=\set{X_{i1},X_{i2},\dots,X_{im_i}}$ where
$X_{i1}:=X_i$ and $X_{ij}$'s are distinct fibers. In this case, $V=\bigcup_{i=1}^{s}\bigcup_{j=1}^{m_i}X_{ij}$.
For all $i\in\set{1,\dots,s}$ and $j\in\set{1,\dots,m_i}$, there exists $R_{ij}\in\mathcal{R}_{X_{i1},X_{ij}}$ with $d_{R_{ij}}=1$.
Therefore, there exists a bijection $R_{ij}:X_{i}\longrightarrow X_{ij}, (x_i\mapsto x)$ where $x$ is the unique element of $X_{ij}$ such that $(x_i,x)\in R_{ij}$.
Indeed, $R_{ij}(X_{i})=X_{ij}$. Thus, for each $x\in V$, there exist unique $i\in\set{1,\dots,s}$ and $j\in\set{1,\dots,m_i}$ such that
$R_{ij}(X_{i})=X_{ij}$ and $x\in R_{ij}(X_{i})$. We claim that the map $\psi$ defined as follows is a monomorphism.
\begin{eqnarray*}
\psi:V\cup\mathcal{R}&\longrightarrow&\big(U_{\mathfrak{X}}\times I_n\big)\cup(\mathcal{R}_{U_{\mathfrak{X}}}\bigotimes E_n).\\
x&\mapsto& (x_i,j); \quad \quad R_{ij}(x_i)=x, \\
R&\mapsto& R_{ij}RR_{kl}^t\otimes e_{jl}; \quad R\in\mathcal{R}_{X_{ij},X_{kl}}
\end{eqnarray*}
Note that $\psi$ is injective, since $R_{ij}$ is a bijection for all $i\in\set{1,\dots,s}$ and $j\in\set{1,\dots,m_i}$.
Let $(x,y)\in R$ and $R\in\mathcal{R}_{X_{ij},X_{kl}}$. Then there exists $(x_i,y_k)\in X_i\times X_k$ such that $R_{ij}(x_i)=x$ and $R_{kl}(y_k)=y$.
This means that $(x_i,y_k)\in R_{ij}RR_{kl}^t$. It follows that, $(\psi(x),\psi(y))=((x_i,j),(y_k,l))\in\psi(R)$.
\end{proof}
The following is an immediate consequence of the preceding theorem.
\begin{corollary}\label{Cor:main T.D}
Let $\mathcal{C}=(V,\mathcal{R})$ be an $(m,n,r)$-scheme. Then $\mathcal{C}\simeq \mathcal{C}_X\otimes \mathcal{T}_n$ for $X\in\Fib(\mathcal{C})$ if and only if $E_{\mathcal{C}}$ is trivial.
\end{corollary}
\section{Reduced $(m,n,r)$-schemes}\label{Section:reduced}
\begin{definition} An $(m,n,r)$-scheme $\mathcal{C}$ is called \textit{reduced }if its equivalence relation $E_{\mathcal{C}}$ is discrete.
\end{definition}
\begin{remark} Note that by Corollary \ref{Cor:main T.D}, a balanced scheme $\mathcal{C}$ is {\it reduced} if and only if there exist no $X,Y\in\Fib(\mathcal{C})$ such that $\mathcal{C}_{X\cup Y}\simeq \mathcal{C}_X\bigotimes \mathcal{T}_2$ where $\mathcal{T}_2$ is a $(1,2,1)$-scheme.
\end{remark}
\begin{exmpl}\label{Example:Symmetric design} Let $(X,\mathcal{B},\mathcal{I})$ be a symmetric design with the set $X$ of points, the set $\mathcal{B}$ of blocks and the incidence relation $\mathcal{I}\subseteq X\times\mathcal{B}$. Set $V=X\cup \mathcal{B}$ (disjoint union) and define the relations
$R_i~(i=1,\ldots,8)$ on $V$ as follows.
\begin{align*}
R_1&=\Delta_X,\quad & R_2&=\Delta_{\mathcal{B}}, &
R_3&=(X\times X)\setminus\Delta_X,\quad & R_4&=(\mathcal{B}\times \mathcal{B})\setminus\Delta_\mathcal{B},\\
R_5&=\mathcal{I},\quad & R_6&=R_5^t,\quad &
R_7&=(X\times \mathcal{B})\setminus\mathcal{I},\quad & R_8&=R_7^t.
\end{align*}
It is known that $(V,\set{R_i}_{i=1}^{8})$ is an $(m,2,2)$-scheme where $m=|X|$.
\end{exmpl}
In \cite{Cam1972} and \cite{Higman-SRD95} they mentioned $(m,n,2)$-schemes and $(m,2,3)$-schemes as linked symmetric designs and
strongly regular designs of the second kind, respectively.
\begin{exmpl}[{\rm \cite{Polla}, \cite[p.6, Example~(i)]{Cam1972}}]
The split extension of the translation group of the vector space $\GF(2^t)^{2k}$ by the
symplectic group ${\rm Sp}(2k,2^t)$ has $2^t$ pairwise inequivalent doubly transitive representations of degree $2^{2kt}$ with the same character.
This gives a reduced $(2^{2kt},2^t,2)$-scheme.
\end{exmpl}
\begin{lemma}[{\rm \cite[Chapter 4]{Cam2001}}]\label{Lemma:Cameron}
Let $G={\rm PGL}(t,q)$ and $\Omega_k$ the set of $k$-dimensional subspaces of the vector space ${\GF}(q)^t$.
Let $\pi_k$ denote the permutation character of $G$ on $\Omega_k$. Then we have the following:\\
For each $k\leq \frac{t}{2}$ there exist irreducible characters $\chi_0,\chi_1,\dots,\chi_k$ of $G$ with $\chi_0 =1_G$ such that
\begin{equation}\label{k=r-1}
\pi_{t-k}=\pi_k=\sum_{i=0}^{k}\chi_i.
\end{equation}
\end{lemma}
\begin{exmpl}\label{Cor:existance}
Let $r$ and $t$ be positive integers such that $r-1\leq \frac{t}{2}$. Then by Lemma \ref{Lemma:Cameron}, the 2-orbit scheme of~${\rm PGL}(t,q)$ on $\Omega_{r-1}\cup \Omega_{t-r+1}$ is a reduced $\big(\genfrac{[}{]}{0pt}{}{t}{r-1}_{q},2,r\big)$-scheme, say $\mathcal{C}$. Moreover, as the actions of ${\rm PGL}(t,q)$ on both $\Omega_{r-1}$ and $\Omega_{t-r+1}$ are multiplicity free, both $\mathcal{C}_{\Omega_{r-1}}$ and $\mathcal{C}_{\Omega_{t-r+1}}$ are commutative and hence by Corollary \ref{cor:r-balanced,p(C)=r}, $|\mathcal{P}(\mathcal{C})|=r$.
\end{exmpl}
\begin{lemma}\label{Lemma:reduced construct} Let $\mathcal{C}_i$ be an $(m_i,n_i,r_i)$-scheme for $i=1,2$.
Then $\mathcal{C}_1\bigotimes\mathcal{C}_2$ is an $\big(m_1m_2,n_1n_2,r_1r_2\big)$-scheme. Furthermore, $\mathcal{C}_1\bigotimes\mathcal{C}_2$ is reduced if
and only if both $\mathcal{C}_1$ and $\mathcal{C}_2$ are reduced.
\end{lemma}
\begin{proof} The first statement is obtained by the definition of $\mathcal{C}_1\bigotimes\mathcal{C}_2$. Let $R_i$ be a basis relation of $\mathcal{C}_i$ for $i=1,2$. Then $R_1\otimes R_2$ is thin if and only if both $R_1$ and
$R_2$ are thin. This implies that $\mathcal{C}_1\bigotimes\mathcal{C}_2$ is reduced if and only if both $\mathcal{C}_1$ and $\mathcal{C}_2$ are reduced.
\end{proof}
\noindent\textbf{Problem 1:} Given an odd prime $p$ does there exist any reduced $(m,3,p)$-scheme for some $m$?\\
The following problem is inspired from a conjecture by H. Weilandt on permutation representations (see \cite{Cam1972}, Remark \ref{Remark:(m,n,2)}
and Lemma \ref{Lemma:(m,n,2)a,b}).\\
\noindent\textbf{Problem 2:}
If $\mathcal{C}$ is a reduced $(p,n,r)$-scheme for some $r$ and prime $p$, then $n\leq2$.
\section{Enumeration of $(m,n,r)$-schemes for $m\leq 11$}\label{Section:Enumeration}
\noindent\textbf{Proof of Theorem \ref{THM:main3}~(i):} Let $\mathcal{C}$ be a reduced $(m,n,r)$-scheme and $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$. Then $2\leq d_R$ for each
$R\in\mathcal{R}_{X,Y}$ and
$$2|\mathcal{R}_{X,Y}|\leq\sum_{R\in\mathcal{R}_{X,Y}}d_R=m,$$
a contradiction.
In order to prove Theorem \ref{THM:main3}~(ii) we need the following lemma.
\begin{lemma}\label{Lemma:coprime} Let $\mathcal{C}$ be an $(m,n,r)$-scheme and $X,Y,Z\in\Fib(\mathcal{C})$. If $T\in\mathcal{R}_{X,Y}$ such that $d_T$ is prime to
$\prod_{R\in\mathcal{R}_{Y,Z}} d_R$, then $d_{Y,Z}$ coincides with $d_{X,Z}$ as multi-sets and $d_T\leq\min\big\{d_R\mid R\in \mathcal{R}_{Y,Z}\big\}$.
\end{lemma}
\begin{proof} For each $R\in\mathcal{R}_{Y,Z}$, ${\rm gcd}(d_T,d_R)=1$. By Lemma \ref{Lemma:constants}~(\ref{qtn:gcd}), $|TR|=1$ and we may define the following map.
\begin{eqnarray*}
\psi:\mathcal{R}_{Y,Z}&\longrightarrow&\mathcal{R}_{X,Z}\\
R&\mapsto& S;\quad TR=\set{S}.
\end{eqnarray*}
By Lemma \ref{Lemma:constants}, $\psi$ is surjective. Since $|\mathcal{R}_{Y,Z}|=|\mathcal{R}_{X,Z}|$, $\psi$ must be a bijection.
Consequently, $\sum_{R\in\mathcal{R}_{Y,Z}}d_R=\sum_{S\in\mathcal{R}_{X,Z}}d_{S}=\sum_{R\in\mathcal{R}_{Y,Z}}d_{TR}$. On the other hand, by Lemma \ref{Lemma:constants}~(\ref{qtn:d_RS}), $
d_R\leq d_{TR}$ for each $R\in\mathcal{R}_X$ and thus $d_R=d_{TR}$ for each $R\in\mathcal{R}_{Y,Z}$. Furthermore, by Lemma \ref{Lemma:constants}~(\ref{qtn:d_RS}),
$d_T\leq d_{TR}=d_R$ for each $R\in\mathcal{R}_{Y,Z}$.
\end{proof}
\noindent\textbf{Proof of Theorem \ref{THM:main3}~(ii):} Let $\mathcal{C}$ be a reduced $(m,n,r)$-scheme and let $X\in\Fib(\mathcal{C})$ such that $\mathcal{C}_X$ is $p$-valanced. Clearly $m=\sum_{T\in\mathcal{R}_{X,Y}}d_T$ where $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$. Since $p\nmid m$, so there exists $T\in\mathcal{R}_{X,Y}$ such that $p\nmid d_T$. Since $\mathcal{C}_X$ is $p$-valenced, $d_T$ is prime to
$\prod_{R\in\mathcal{R}_{X}} d_R$. As $d_{\Delta_X}=1$, it follows from Lemma \ref{Lemma:coprime} that
$d_T\leq\min\big\{d_R\mid R\in \mathcal{R}_X\big\}=1,$
a contradiction.~~~~~\hspace{12cm}$\square$
\begin{lemma}\label{Lemma:(m,n,2)symmetric} Let $\mathcal{C}$ be an $(m,n,2)$-scheme and $R\in\mathcal{R}_{X,Y}$ where $X,Y\in\Fib(\mathcal{C})$. Then $d_R(d_R-1)=\lambda(m-1)$ for some $\lambda\in \mathbb{N}$.
\end{lemma}
\begin{proof}
Let $\mathcal{C}$ be an $(m,n,2)$-scheme and $X,Y\in\Fib(\mathcal{C})$. For each $R\in\mathcal{R}_{X,Y}$ we have by Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),
$$A_RA_{R^t}=\sum_{S\in\mathcal{R}_X}c_{RR^t}^S A_S=d_R I_X+ c_{RR^t}^{\Delta_X^c}(J_X-I_X),$$
where $\Delta_X^c=(X\times X)\setminus\Delta_X$. It follows that $R\in\mathcal{R}_{X,Y}$ is regarded as the incident relation of a symmetric
$(m, d_R,\lambda)$-design where $\lambda=c_{RR^t}^{\Delta_X^c}$. A basic property of symmetric deigns implies that $d_R(d_R-1)=\lambda(m-1)$.
\end{proof}
\begin{remark}\label{Remark:(m,n,2)} Let $m$ and $t$ be positive integers and $q$ an odd prime power such that
\begin{equation}\label{qtn:(m,n,2)}
m-1=2^tq.
\end{equation}
Then there are exactly four $d\in\set{1,\dots,m-1}$ such that $d(d-1)\equiv0$ $({\rm mod} ~2^tq)$ by an elementary number theoretical argument.
It follows that if $\mathcal{C}$ is a reduced $(m,n,2)$-scheme, then $d_{X,Y}$ is uniquely determined for all $X,Y\in\Fib(\mathcal{C})$. In particular, if m is prime
satisfying (\ref{qtn:(m,n,2)}), then there is $\gamma \in \set{2,\dots,m-2}$ such that $d_{X,Y}=\set{\gamma, m-\gamma}$ and ${\rm gcd}(\gamma,m-\gamma)=1$ for all $X,Y\in\Fib(\mathcal{C})$.
\end{remark}
\begin{lemma}\label{Lemma:(m,n,2)a,b} Let $\mathcal{C}$ be a reduced $(m,n,2)$-scheme. Suppose that $d_{X,Y}=\set{a,b}$ with ${\rm gcd}(a,b)=1$ for all $X,Y\in\Fib(\mathcal{C})$.
Then $n\leq2$.
\end{lemma}
\begin{proof} Suppose that $X,Y$ and $Z$ are distinct fibers of $\mathcal{C}$ and let $\mathcal{R}_{X,Y}=\set{R,R'}$, $\mathcal{R}_{Y,Z}=\set{S,S'}$,
$\mathcal{R}_{X,Z}=\set{T,T'}$ so that $d_R=d_S=d_T=a<b=d_{R'}=d_{S'}=d_{T'}$.
By Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:dgree-constant}), $a^2=d_Rd_S=\alpha a+\beta b$ such that $a\mid b\beta$ and
$\beta <a$. Since ${\rm gcd}(a,b)=1$, it follows that $\beta=0$ and $\alpha=a$. This implies that $c_{RS}^{T}=a=d_R$. It follows that
\begin{equation}\label{qtn:(a,b)=1}
R_{out}(x)\subseteq S_{in}(z),
\end{equation}
where $(x,z)\in T$. Now we take $y_1,y_2\in R_{out}(x)$ so that $y_1\neq y_2$. It follows from (\ref{qtn:(a,b)=1}) that $T_{out}(x)\subseteq S_{out}(y_1)\cap S_{out}(y_2)$.
This fact along with Lemma \ref{Lemma:constants}~(\ref{qtn:dgree-constant}) assert that $a=c_{SS^t}^{\Delta_Y^c}$ where
$\Delta_Y^c=(Y\times Y)\setminus\Delta_Y$. Therefore, by Lemma \ref{Lemma:(m,n,2)symmetric}, $a(a-1)=a(a+b-1)$. It follows that $ab=0$, a contradiction. This completes the proof.
\end{proof}
\begin{lemma} \label{Lemma:r=2}Let $\mathcal{C}$ be a reduced $(m,n,2)$-scheme. If $m-1$ is a prime power, then $n=1$.
\end{lemma}
\begin{proof} Let $p$ be prime such that $m-1=p^t$ for some $t$. In this case, $p$ does not divide $m$ and we are done by Theorem \ref{THM:main3}~(ii).
\end{proof}
\begin{lemma}\label{Lemma:m=2r} Let $\mathcal{C}$ be a reduced $(m,n,r)$-scheme and $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$. If $m=2r$, then the following hold:
\begin{enumerate}
\item [\rm~(i)] For each $T\in\mathcal{R}_{X,Y}$, $d_T=2$.
\item [\rm~(ii)] For each $R\in\mathcal{R}_{X}$, $d_R\in\{1,2,4\}$ and
$$|\{R\in\mathcal{R}_{X}\mid d_R=1\}|=2|\{R\in\mathcal{R}_{X}\mid d_R=4\}|.$$
\end{enumerate}
\end{lemma}
\begin{proof} (i) Let $\mathcal{C}$ be a reduced $(m,n,r)$-scheme and $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$. Then as $d_T\geq2$ for each $T\in\mathcal{R}_{X,Y}$, it follows from
$m=\sum_{T\in\mathcal{R}_{X,Y}}d_T$ that $2r\leq m$ and the equality holds if and only if $d_T=2$ for each $T\in\mathcal{R}_{X,Y}$.\\
~(ii) Let $R\in\mathcal{R}_X$ and $T\in\mathcal{R}_{X,Y}$. Then by Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:dgree-constant}), there exist none-negative integers $\alpha$ and $\beta$ such that $2d_R=d_Rd_T=\alpha d_S+\beta d_{S'}=2\alpha+2\beta$ and $\alpha, \beta\leq2$.
This implies that $d_R\leq4$. By Lemma \ref{Lemma:coprime}, $d_R\in\{1,2,4\}$.
We set $k_i:=|\{R\in\mathcal{R}_{X}\mid d_R=i\}|$ for $i\in\{1,2,4\}$. Since $k_1+k_2+k_4=|\mathcal{R}_X|=|\mathcal{R}_{X,Y}|$, it follows that $m=k_1+2k_2+4k_4=2(k_1+k_2+k_4)$. Therefore, $k_1=2k_4$.
\end{proof}
\begin{lemma} \label{Lemma:m prim Hanaki} For each $(m,n,r)$-scheme, if $m$ is prime, then $r-1$ divides $m-1$.
\end{lemma}
\begin{proof} Let $X\in\Fib(\mathcal{C})$ and consider the homogeneous component $(X,\mathcal{R}_X)$. Since $|X|=m$ is prime, by \cite[Theorem 3.3]{Han_Uno2006} $d_R=d$ for all
$R\in\mathcal{R}_X$ with $R\neq\Delta_X$. Then $m-1=\sum_{\substack{R\in\mathcal{R}_X,\\\Delta_X\neq R}}d_R=(r-1)d.$
\end{proof}
\begin{lemma}\label{Lemma:symmetric-odd} Let $\mathcal{C}=(V,\mathcal{R})$ be an $(m,n,r)$-scheme. If $m$ is odd, then
each non-reflexive symmetric basis relation of $\mathcal{C}$ has even degree.
\end{lemma}
\begin{proof} Let $S\in\mathcal{R}_X$ be symmetric for some $X\in\Fib(\mathcal{C})$. Since $S\neq\Delta_X$, $|S|$ is even. By (\ref{qtn:degree in&out}), $|S|=d_Sm$
and thus $d_S$ is even.
\end{proof}
\begin{lemma}[{\rm \cite[(3.2)]{Higman-SRD95}}]\label{Lemma:Higman symmetric} Let $\mathcal{C}$ be a reduced $(m,n,3)$-scheme.
Then $\mathcal{C}_X$ is symmetric for each $X\in\Fib(\mathcal{C})$.
\end{lemma}
\noindent\textbf{Proof of Theorem \ref{THM:main4}:}
So far in this section we have been preparing some lemmas which will be applied to enumerate reduced $(m,n,r)$-schemes for $m$ up to $11$.
The entries $(r,m)$ such that $m<2r$ are eliminated by Theorem \ref{THM:main3}~(i) whereas $(2,m)$'s are eliminated by
Lemma \ref{Lemma:r=2} except $(2,7)$ and $(2,11)$. If $\mathcal{C}$ is a reduced $(m,n,2)$-scheme with $m\in\set{7,11}$, then by
Remark \ref{Remark:(m,n,2)} and Lemma \ref{Lemma:(m,n,2)a,b}, $n\leq 2$. Thus we have eliminated the first row of
Table \ref{Tab:enume1}.
Applying Lemma \ref{Lemma:m=2r} for $(r,m)=(5,10)$ we obtain that $d_{X}=\set{1,1,2,2,4}$. According to
\cite{Nomia, Han_Miya} there are no homogeneous schemes with $d_{X}=\set{1,1,2,2,4}$ where we can prove this fact by a
theoretical way.
The entries $(4,11)$ and $(5,11)$ are eliminated by Lemma \ref{Lemma:m prim Hanaki} whereas $(3,7)$ and $(3,11)$ are eliminated by
Lemmas \ref{Lemma:Higman symmetric} and \ref{Lemma:symmetric-odd}.
An $(r,m)$-entry of the following table is denoted by $*$ if there exists no $(m,1,r)$-scheme.
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|cc|}
\hline
\backslashbox{$r$}{$m$} & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 \\\hline
2 & 1 & 1 & 1 & $\leq 2$ & 1 & 1 & 1 & $\leq$ 2 \\ \hline
3 & 1 & 1 & 1 & 1 & $\leq 2$ & 1 & 1 & 1 \\ \hline
4 & 1 & * & 1 & 1 & $\leq 2$ & 1 & 1 & * \\ \hline
5 & * & 1 & * & * & 1 & 1 & * & * \\
\hline
\end{tabular}
\renewcommand{\captionlabeldelim}{.}
\caption{}\label{Tab:enume1}
\end{table}
\newpage
The following is the list of $(r,m)$, $\sum_{i=1}^{r} a_i$ and $\sum_{i=1}^{r} b_i$ where $m=\sum_{i=1}^{r} a_i=\sum_{i=1}^{r} b_i$, $1=a_1\leq a_2\leq\dots\leq a_r$ and $2\leq b_1\leq b_2\dots\leq b_r$ such that $d_X=\set{a_1,\dots,a_r}$ and $d_{X,Y}=\set{b_1,\dots,b_r}$ for some $(m,1,r)$-scheme $(X,\mathcal{R}_X)$ not satisfying the assumption of Theorem \ref{THM:main3} (see \cite{Han_Miya, Nomia}).
\begin{table}[H]
\centering
\begin{tabular}{|l|l|c|c|}
\hline
$(r,m)$ & $\sum_{R\in\mathcal{R}_X}d_R$ & $\sum_{R\in\mathcal{R}_{X,Y}}d_R$& \\
\hline
(3,6) &1+1+4 & 2+2+2& Not occur by Lemma \ref{Lemma:(6,n,3)-scheme} \\
\hline
(3,8) &1+1+6& 2+2+4& Not occur by Lemma \ref{Lemma:reduced(8,n,3)-scheme1}\\\cline{3-4}
& & 2+3+3& $n\leq2$ by Lemma \ref{Lemma:reduced(8,n,3)-scheme2} \\\cline{2-2}\cline{3-4}
&1+3+4& 2+2+4& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) \\\cline{3-4}\cline{3-4}
& & 2+3+3& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) \\
\hline
(3,9) &1+2+6&2+2+5& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &2+3+4& Not occur by Lemma \ref{Lemma:$(9,n,3)$-scheme} \\\cline{3-4}
& &3+3+3& Not occur (see Lemma \ref{Lemma:$(9,n,3)$-scheme}) \\\cline{3-4}\cline{2-2}
\hline
(3,10) &1+1+8& 2+3+5& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& & 3+3+4& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& & 2+4+4& Not occur by Lemma \ref{Lemma:$(10,n,3)$-scheme}~(i) \\\cline{3-4}
& & 2+2+6& Not occur by Lemma \ref{Lemma:$(10,n,3)$-scheme}~(i) \\\cline{3-4} \cline{2-2}
&1+3+6&2+3+5& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &3+3+4& Not occur by Lemma \ref{Lemma:$(10,n,3)$-scheme}~(ii) \\\cline{3-4}
& &2+4+4& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &2+2+6& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) \\\cline{2-2}\cline{3-4}
&1+4+5&2+3+5& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &3+3+4& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &2+4+4& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) \\\cline{3-4}
& &2+2+6& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) \\\cline{3-4}
\hline
(4,8) &1+1+2+4&2+2+2+2& $n\leq2$ (see Lemma \ref{Lemma:reduced (8,n,4)-scheme2}) \\
\hline
(4,9) &1+1+1+6&2+2+2+3& Not occur by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) and Lemma \ref{Lemma:symmetric-odd}\\\cline{2-2}\cline{3-4}
&1+2+3+3&2+2+2+3& Not occur by Lemma \ref{Lemma:$(9,n,4)$-scheme} \\
\hline
(4,10) &1+2+2+5&2+2+2+4& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &2+2+3+3& Not occur by Lemma \ref{Lemma:coprime} \\\cline{2-2}\cline{3-4}
&1+1+4+4&2+2+3+3& Not occur by Lemma \ref{Lemma:coprime} \\\cline{3-4}
& &2+2+2+4& Not occur by Lemma \ref{Lemma:$(10,n,4)$-scheme} \\
\hline
\end{tabular}
\renewcommand{\captionlabeldelim}{.}
\caption{}\label{Tab:enume3}
\end{table}
\newpage
\begin{lemma} \label{Lemma:(6,n,3)-scheme} If $\mathcal{C}$ is a reduced $(6,n,3)$-scheme such that $d_{X}=\set{1,1,4}$ for some $X\in\Fib(\mathcal{C})$, then
$d_{X,Y}\neq\set{2,2,2}$ for each $Y\in\Fib(\mathcal{C})$ with $Y\neq X$.
\end{lemma}
\begin{proof}
Suppose by the contrary that $d_{X,Y}=\set{2,2,2}$ for some $Y\neq X$. By Lemma \ref{Lemma:m=2r}, $d_{Y}=\set{1,1,4}$.
Taking $R,S\in\mathcal{R}_{X,Y}$ with $R\neq S$ we obtain from Lemma \ref{Lemma:constants} (\ref{qtn:symmetric-dR=2}) that $R^tR=S^tS=\set{\Delta_Y,T}$ where
$T\in \mathcal{R}_Y$ with $T\neq\Delta_Y$ and $d_T=1$. By Lemma
\ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:Kronecker}), $4=d_Rd_{S^t}=\alpha+4\beta$ for some non-negative integers $\alpha,\beta\leq2$. This implies $\alpha=0$ and $\beta=1$, which contradicts Lemma \ref{Lemma:constant2}.
\end{proof}
\begin{lemma} \label{Lemma:reduced(8,n,3)-scheme1}
If $\mathcal{C}$ is a reduced $(8,n,3)$-scheme such that $d_X=\set{1,1,6}$ for some $X\in\Fib(\mathcal{C})$, then we have the following:
\begin{enumerate}[(i)]
\item For each $Y\in\Fib(\mathcal{C})$ with $Y\neq X$, $d_{X,Y}\neq\set{2,2,4}$. Indeed, $d_{X,Y}=\set{2,3,3}$ for each $Y\in\Fib(\mathcal{C})$.
\item Let $\mathcal{R}_{X,Y}=\set{R,S,S'}$ such that $d_{R}=2$ and $d_{S}=d_{S'}=3$. Let $T\in\mathcal{R}_X$ with $T\neq\Delta_X$ and $d_{T}=1$.
Then $TR=\set{R}$, $TS=\set{S'}$ and $TS'=\set{S}$.
\end{enumerate}
\end{lemma}
\begin{proof} (i) Suppose by the contrary that $d_{X,Y}=\set{2,2,4}$ for some $Y\in\Fib(\mathcal{C})$, and take $R\in\mathcal{R}_{X}$ and $S\in\mathcal{R}_{X,Y}$ so that
$d_R=6$ and $d_S=2$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:dgree-constant}) that for some non-negative
integers $\alpha,\beta, \gamma$ we have
$$12=d_Rd_S=2\alpha +2\beta + 4\gamma ,\quad ~6\mid2\alpha,~6\mid2\beta,~6\mid4\gamma \quad\alpha,\beta,\gamma\leq 2.$$
This implies that $\gamma=0$ and $12=2\alpha +2\beta\leq8$, a contradiction.\\
~(ii) As $d_T=1$, Lemma \ref{Lemma:constants}~(\ref{qtn:d_RS}), (\ref{qtn:gcd}) asserts that $d_{TR}=2$ and $TR=\set{R}$,
since $R$ is the unique basis relation in $\mathcal{R}_{X,Y}$ of degree 2.
By the same observation $d_{TS}=3$ and $TS\in\mathcal{R}_{X,Y}$. If $TS=\set{S}$, then by Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}), $c_{TS}^{S}=1$ and Lemma \ref{Lemma:constants}~(\ref{qtn:lcm}) implies that
$c_{SS^t}^{T}=3$. Therefore, applying Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}) we have $9=d_Sd_{S^t}=3+3+c_{SS^t}^{T'}6$ where $T'\in\mathcal{R}_X$ with $d_{T'}=6$, a contradiction.
\end{proof}
\begin{lemma} \label{Lemma:reduced(8,n,3)-scheme2}
If $\mathcal{C}$ is a reduced $(8,n,3)$-scheme such that $d_X=\set{1,1,6}$ for some $X\in\Fib(\mathcal{C})$, then $n\leq2$.
\end{lemma}
\begin{proof} Suppose by the contrary that $X,Y$ and $Z$ are distinct fibers of $\mathcal{C}$. Then by Lemma \ref{Lemma:reduced(8,n,3)-scheme1},
$d_{X,Y}=d_{Y,Z}=d_{X,Z}=\set{2,3,3}$. Let $R\in\mathcal{R}_{X,Y}$ and $S\in\mathcal{R}_{Y,Z}$ with $d_R=2$ and $d_S=3$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:dgree-constant}), there exist non-negative integers $\alpha,\beta, \gamma$ such that
$$6=d_Rd_S=2\alpha+3\beta+3\gamma,\quad 6\mid2\alpha,\quad\alpha\leq 2.$$
This implies $\alpha=0$ and $RS=\set{S'}$ where $S'\in\mathcal{R}_{Y,Z}$ with $d_{S'}=3$. Since $c_{RS}^{S'}=2$, by Lemma \ref{Lemma:constant2}, $R^tR\cap SS^t=\set{\Delta_Y,T}$ for some $T\in\mathcal{R}_Y$ with $d_T=1$. Thus $9=d_Sd_{S'}=3+3+6\alpha$. It follows that $3=6\alpha$ , a contradiction.
\end{proof}
\begin{lemma}\label{Lemma:$(9,n,3)$-scheme} Let $\mathcal{C}$ be a reduced $(9,n,3)$-scheme and $d_{X}=\set{1,2,6}$ for some fiber $X$.
Then $d_{X,Y}\notin\big\{\set{2,3,4},\set{3,3,3}\big\}$ for each $Y\in\Fib(\mathcal{C})$.
\end{lemma}
\begin{proof}
Suppose that $d_{X,Y}=\set{2,3,4}$ and take the basis relations $R\in\mathcal{R}_X$ and $S\in\mathcal{R}_{X,Y}$ so that $d_R=6$ and $d_S=2$.
It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:dgree-constant}),
there exist non-negative integers $\alpha,\beta, \gamma$ such that
$$12=d_Rd_S=2\alpha +3\beta + 4\gamma,\quad~6\mid2\alpha,~6\mid3\beta,~6\mid4\gamma, \quad\alpha,\beta,\gamma\leq 2.$$
This implies that $\gamma=0$ and $12=2\alpha +3\beta\leq 10$, a contradiction.
Suppose that $d_{X,Y}=\set{3,3,3}$ for some $Y\in\Fib(\mathcal{C})$ and take distinct $R,S\in \mathcal{R}_{X,Y}$. By
Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}),
for some non-negative integers $\alpha,\beta$ we have $9=d_Rd_{S^t}=2\alpha+6\beta=2(\alpha+3\beta)$, a contradiction.
\end{proof}
\label{Lemma:constant3} Let $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$ and $R,S,S'\in\mathcal{R}_{X,Y}$. Then $R^tR\cap S^tS'\neq\emptyset$ if and only if $RS^t\cap RS'^{t}\neq\emptyset$. We use this fact in the proof of the following lemma.
\begin{lemma}\label{Lemma:$(10,n,3)$-scheme} Let $\mathcal{C}$ be a reduced $(10,n,3)$-scheme. Then the following hold:
\begin{enumerate}[(i)]
\item If $d_{X}=\set{1,1,8}$ for some $X\in\Fib(\mathcal{C})$, then $d_{X,Y}\notin\set{\set{2,2,6},\set{2,4,4}}$ for each $Y\in\Fib(\mathcal{C})$.
\item If $d_{X}=\set{1,3,6}$ for some $X\in\Fib(\mathcal{C})$, then $d_{X,Y}\neq\set{3,3,4}$ for each $Y\in\Fib(\mathcal{C})$.
\end{enumerate}
\end{lemma}
\begin{proof} (i) Suppose that $d_{X,Y}=\set{2,2,6}$ for some $Y\in\Fib(\mathcal{C})$. Take $R,S\in\mathcal{R}_{X,Y}$ with $R\neq S$ and $d_R=d_S=2$. By Lemma
\ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}),~(\ref{qtn:dgree-constant}), $4=d_Rd_{S^t}=\alpha+8\beta$ for some non-negative integers $\alpha,\beta$ with $\alpha,\beta\leq 2$. It follows that $\alpha=0$ and $4=8\beta$, a contradiction.
Suppose that $\mathcal{R}_{X,Y}=\set{R,S,S'}$
such that $d_R=2$ and $d_S=d_{S'}=4$. By Lemma~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}), $8=d_Rd_{S^t}=\alpha+8\beta$ for some non-negative integers
$\alpha,\beta$ with $4\mid \alpha \leq2$. This implies that $\alpha=0$ and then $\beta=1$. Therefore, $RS^t=\set{T'}$ where $T'\in\mathcal{R}_X$ with $d_{T'}=8$.
By the same observation, $RS'^t=\set{T'}$. Therefore, $T\in S^tS'\cap R^tR$ where $T\neq\Delta_Y$. On the other hand,
by Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}), $d_T=1$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:Kronecker}),~(\ref{qtn:dgree-constant}) that
for some non-negative integers $\alpha,\beta$ we have $16=d_{S^t}d_{S'}=\alpha+8\beta$ with $4\mid\alpha$ and $0<\alpha\leq4$.
This implies that $\alpha=4$ and thus $12=8\beta$, a contradiction.
(ii) Take $R\in\mathcal{R}_X$ and $S\in\mathcal{R}_{X,Y}$ with $d_R=3$ and $d_S=4$.
By Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:lcm}),~(\ref{qtn:dgree-constant}),
for some non-negative integers $\alpha,\beta,\gamma$ we have $12=d_{R}d_{S}=3\alpha+3\beta+4\gamma$ with $12\mid\alpha$ and $12\mid\beta$ and $\alpha,\beta,\gamma\leq3$. This implies that $\alpha=\beta=0$ and $\gamma=3$.
Hence $RS=\set{S}$. By Lemma \ref{Lemma:d_L_S}, $d_{L_S}\mid\rm gcd(10,4)=2$ which is a contradiction since $d_{L_S}>d_R=3$.
\end{proof}
\begin{lemma} \label{Lemma:reduced (8,n,4)-scheme1}
Let $\mathcal{C}$ be a reduced $(8,n,4)$-scheme such that $d_X=\set{1,1,2,4}$ for some $X\in\Fib(\mathcal{C})$. Then for all $X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$, there exists $R\in\mathcal{R}_{X,Y}$ such that $RR^t=\set{\Delta_X,S}$ \rm{(}resp. $R^tR=\set{\Delta_Y,S'}$\rm{)} where $S$ is the unique basis relation in $\mathcal{R}_X$ with $d_S=2$
\rm{(}resp. $S'$ is the unique basis relation in $\mathcal{R}_Y$ with $d_{S'}=2$\rm{)}.
\end{lemma}
\begin{proof} Let $\mathcal{C}$ be a reduced $(8,n,4)$-scheme such that $d_X=\set{1,1,2,4}$ for some $X\in\Fib(\mathcal{C})$. Then $d_{X,Y}=\set{2,2,2,2}$ for all
$X,Y\in\Fib(\mathcal{C})$ with $X\neq Y$.
Let $T\in\mathcal{R}_X$ with $T\neq\Delta_X$ and $d_T=1$. Then $d_{TR}=2$ and $|TR|=1$. Suppose that $TR=\set{R}$ for each $R\in\mathcal{R}_{X,Y}$. Then $T\notin RS^t$ for all $R,S\in\mathcal{R}_{X,Y}$ with $R\neq S$. Thus by Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:dgree-constant}),
$4=d_Rd_{S^t}=2\alpha+4\beta$ for some none-negative integers $\alpha,\beta$ with $\alpha\leq2$. By Lemma \ref{Lemma:constant2}, $\beta=0$ and $\alpha=2$.
This implies that $R\mathcal{R}_{Y,X}\subsetneq\mathcal{R}_X$, which contradicts Lemma \ref{Lemma:constants}~(\ref{qtn:dgree-constant}). Thus there exists $R\in\mathcal{R}_{X,Y}$ such that $TR\neq\set{R}$. Equivalently, $T\notin RR^t$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:symmetric-dR=2}) that $RR^t=\set{\Delta_X, S}$ where $S$ is the unique basis relation in $\mathcal{R}_X$ with $d_S=2$
\end{proof}
\begin{lemma}\label{Lemma:reduced (8,n,4)-scheme2}If $\mathcal{C}$ is a reduced $(8,n,4)$-scheme such that $d_X=\set{1,1,2,4}$ for some $X\in\Fib(\mathcal{C})$, then $n\leq 2$.
\end{lemma}
\begin{proof}
Suppose by the contrary that $X,Y$ and $Z$ are distinct fibers of $\mathcal{C}$. Then by Lemma \ref{Lemma:reduced (8,n,4)-scheme1}, there exist $R\in\mathcal{R}_{X,Y}$ and $T\in\mathcal{R}_{Y,Z}$ such that $R^tR=TT^t=\set{\Delta_X, S}$ where $S$ is the unique basis relation in $\mathcal{R}_Y$ with $d_S=2$. It follows from
Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}) that $c_{TT^t}^S=c_{R^tR}^S=1$. Let $(y,y')\in S$. Then there exists $(x,z)\in X\times Z$ such that
$R_{in}(y)\cap R_{in}(y')=\set{x}$ and $T_{out}(y)\cap T_{out}(y')=\set{z}$. As $d_T=2$, we may assume that $T_{out}(y)=\set{z,z_1}$ and
$T_{out}(y')=\set{z,z_2}$. Note that $z_1\neq z_2$, otherwise $c_{TT^t}^S \geq2$, a contradiction. This means that $(R\circ T)_{out}(x)=\set{z,z_1,z_2}$
and thus $d_{RT}=d_{R\circ T}=3$, which is a contradiction since $d_{RT}$ must be a sum of degrees in $d_{X,Z}=\set{2,2,2,2}$.
\vspace{-3mm}
\begin{figure}[H]
\begin{displaymath}
\xymatrix{
&\VRT{x}\ar[dl]_{\scriptscriptstyle\mathit{R}} \ar[dr]^{\scriptscriptstyle\mathit{R}}& \\
\VRT{y}\ar[d]_{{\scriptscriptstyle\mathit{T}}}\ar[dr]^{{\scriptscriptstyle\mathit{T}}} \ar[rr]_{{\scriptstyle\mathit{s}}}& & \VRT{y'}\ar[dl]_{{\scriptscriptstyle\mathit{T}}}\ar[d]^{{\scriptscriptstyle\mathit{T}}} \\
\VRT{z_1}& \VRT{z} & \VRT{z_2}}
\end{displaymath}\label{fig:6}
\vspace{-4mm}
\renewcommand{\captionlabeldelim}{.}
\caption{}\label{Fig:(8,n,4)-scheme}
\end{figure}
\end{proof}
\begin{exmpl} The association scheme \verb"as16 No.122" as in {\rm \cite{Han_Miya}} induces the thin residue fission {\rm(see \cite[Proposition 3.1]{28points})}, which is a reduced $(8,2,4)$-scheme whose relational matrix is
$$
\left(
\begin{array}{cccccccc|cccccccc}
0 & 1 & 2 & 2 & 3 & 3 & 3 & 3 & 4 & 4 & 5 & 5 & 6 & 7 & 6 & 7 \\
1 & 0 & 2 & 2 & 3 & 3 & 3 & 3 & 4 & 4 & 5 & 5 & 7 & 6 & 7 & 6 \\
2 & 2 & 0 & 1 & 3 & 3 & 3 & 3 & 5 & 5 & 4 & 4 & 6 & 7 & 7 & 6 \\
2 & 2 & 1 & 0 & 3 & 3 & 3 & 3 & 5 & 5 & 4 & 4 & 7 & 6 & 6 & 7 \\
3 & 3 & 3 & 3 & 0 & 1 & 2 & 2 & 6 & 7 & 6 & 7 & 4 & 4 & 5 & 5 \\
3 & 3 & 3 & 3 & 1 & 0 & 2 & 2 & 7 & 6 & 7 & 6 & 4 & 4 & 5 & 5 \\
3 & 3 & 3 & 3 & 2 & 2 & 0 & 1 & 6 & 7 & 7 & 6 & 5 & 5 & 4 & 4 \\
3 & 3 & 3 & 3 & 2 & 2 & 1 & 0 & 7 & 6 & 6 & 7 & 5 & 5 & 4 & 4 \\\hline
4' & 4' & 5' & 5' & 6' & 7' & 6' & 7' & 0' & 1' & 2' & 2' & 3' & 3' & 3' & 3' \\
4' & 4' & 5' & 5' & 7' & 6' & 7' & 6' & 1' & 0' & 2' & 2' & 3' & 3' & 3' & 3' \\
5' & 5' & 4' & 4' & 6' & 7' & 7' & 6' & 2' & 2' & 0' & 1' & 3' & 3' & 3' & 3' \\
5' & 5' & 4' & 4' & 7' & 6' & 6' & 7' & 2' & 2' & 1' & 0' & 3' & 3' & 3' & 3' \\
6' & 7' & 6' & 7' & 4' & 4' & 5' & 5' & 3' & 3' & 3' & 3' & 0' & 1' & 2' & 2' \\
7' & 6' & 7' & 6' & 4' & 4' & 5' & 5' & 3' & 3' & 3' & 3' & 1' & 0' & 2' & 2' \\
6' & 7' & 7' & 6' & 5' & 5' & 4' & 4' & 3' & 3' & 3' & 3' & 2' & 2' & 0' & 1' \\
7'' & 6' & 6' & 7'' & 5' & 5' & 4' & 4' & 3' & 3' & 3' & 3' & 2' & 2' & 1' & 0' \\
\end{array}
\right).
$$
Also the thin residue fission of the association scheme \verb"as16 No.51" as in {\rm\cite{Han_Miya}}, is a reduced $(8,2,3)$-scheme.
\end{exmpl}
\begin{lemma}\label{Lemma:$(9,n,4)$-scheme} Let $\mathcal{C}$ be a reduced $(9,n,4)$-scheme such that $d_{X}=\set{1,2,3,3}$ for some $X\in\Fib(\mathcal{C})$. Then $d_{X,Y}\neq\set{2,2,2,3}$ for each $Y\in\Fib(\mathcal{C})$.
\end{lemma}
\begin{proof} Suppose by the contrary that $R_1,R_2,R_3\in \mathcal{R}_{X,Y}$ with $d_{R_i}=2$ for $i\in\set{1,2,3}$. For all $i,j\in\set{1,2,3}$ with $i\neq j$, by
Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}) we have $4=d_{R_i}d_{R_j^t}=2\alpha+3\beta+3\gamma$. This implies that $\beta=\gamma=0$ and $\alpha=2$. Hence, for all $i,j\in\set{1,2,3}$ with $i\neq j$, $R_iR_j^t=\set{T}$ where $T\in\mathcal{R}_X$ with $d_T=2$. It follows from Lemma \ref{Lemma:constants}~(\ref{qtn:dT<dR}) that $\set{R_1}=TR_2=\set{R_3}$, a contradiction.
\end{proof}
\begin{lemma}\label{Lemma:$(10,n,4)$-scheme} Let $\mathcal{C}$ be a reduced $(10,n,4)$-scheme such that $d_{X}=\set{1,1,4,4}$ for some $X\in\Fib(\mathcal{C})$.
Then $d_{X,Y}\neq\set{2,2,2,4}$ for each $Y\in\Fib(\mathcal{C})$.
\end{lemma}
\begin{proof} Suppose by the contrary that $d_{X,Y}=\set{2,2,2,4}$ for some $Y\in\Fib(\mathcal{C})$ with $Y\neq X$. According to \cite{Nomia, Han_Miya},
$d_Y\in\set{\set{1,1,4,4},\set{1,2,2,5}}$. It follows from Lemma \ref{Lemma:coprime} that $d_Y=\set{1,1,4,4}$. Take $R,S\in \mathcal{R}_{X,Y}$ with $R\neq S$ and $d_{R}=d_S=2$. By Lemma \ref{Lemma:constants}~(\ref{qtn:dRdS}),~(\ref{qtn:Kronecker}),~(\ref{qtn:dgree-constant}), there exist non-negative
integers $\alpha,\beta,\gamma$ such that
$$4=d_Rd_{S^t}=\alpha +4\beta+4\gamma,\quad \alpha,\beta, \gamma\leq 2.$$
This implies that $4\mid\alpha$. Hence, $\alpha=0$ and $\beta+\gamma=1$ which contradicts Lemma \ref{Lemma:constant2}.
\end{proof}
\noindent\textbf{Acknowledgements}\\
The authors would like to thank I. Ponomarenko, E. Bannai and A. Ivanov for their valuable comments to this article, especially internal direct sums, $(m,2,r)$-schemes and tensor products, respectively.
|
\section{Introduction}
\label{sec:Introduction}
Discrete rearranging patterns include cellular patterns, for instance liquid foams, biological tissues and grains in polycrystals; assemblies of particles such as beads, granular materials, colloids, molecules and atoms; and interconnected networks \cite{Outils}. Many of these disordered materials display elastic and plastic properties, so that the stress tensor can rotate and is not necessarily aligned with the strain rate tensor; in models this effect is included in objective derivatives \cite{Kolymbas2003}.
Use of simplified geometries, {\it e.g.} in a rheometer, allows a first characterization of the material through measurements of shear stress.
An overshoot in the shear stress is seen during the first loading in materials such as polymers \cite{Larson1999}, granular materials \cite{Xu2006}, and emulsions \cite{Lahtinen1988}. For liquid foams this effect has been observed in a plate-plate rheometer \cite{Khan1988} and in simulations \cite{Okuzono1995,KablaPreprintSimul}. It is unclear whether this is due to a change in the material's structure, or a tensorial effect of shear; but nevertheless the overshoot is an essential ingredient in a recent model \cite{weaire_preprint} of the strain-rate discontinuity in the cylindrical Couette foam flow experiments of ref. \cite{lauridsen2004}. Such an overshoot results in mechanical bistability: two different values of strain correspond to the same value of stress between the plateau and the maximum, and can thus coexist.
Here, we investigate the elastic regime and elasto-plastic transition in a fully tensorial model. To describe the
mechanical behaviour we use a formalism adapted for discrete rearranging disordered patterns
which enables us to quantify rotational
effects and to test the relevant parameters \cite{Outils}.
We use as an example a sheared liquid foam \cite{Abdelkader1999,Debregeas2001,Lauridsen2002,Pratt2003,Janiaud2005,Wang2006,Janiaud2006,Cheddadi2008}.
Although a liquid foam consists only of gas bubbles surrounded by liquid walls, it exhibits a complex mechanical behaviour. It is elastic for small strains, plastic for large strains and flows at large strain rates \cite{Weaire1999,Saint-Jalmes1999,Hohler2005}. This behavior is useful for numerous applications such as ore separation, oil extraction, foods and cosmetics. The individual objects, namely the bubbles, are easily identified, which makes a liquid foam (or alternatively an emulsion, made of droplets) a model for the study of other complex fluids.
This paper is organized as follows. In Section \ref{sec:Simulations}, we
simulate the quasistatic 2D flow of a foam in a Couette shear geometry; we
explain how we perform and represent the measurements. In Sec.
\ref{sec:Model}, we present our equations, and discuss the specific
effects due to the use of tensors, such as the overshoot. Sec.
\ref{sec:ComparisonSimulationModel} compares the model and the simulation,
and extracts the relevant information. Sec. \ref{sec:Practical} presents
applications to practical situations, {\it i.e.} how and when to use the
model. Sec. \ref{sec:Conclusions} summarizes our findings. An Appendix
explains the notation and provides the detailed equations.
\section{Simulations}
\label{sec:Simulations}
We simulate numerically a 2D foam flowing in a linear Couette shear geometry. Simulations of dry foams offer several advantages: (i) the parameters are homogeneous (liquid fraction, bubble area) and controlled (no diffusion-driven coarsening or film rupture); (ii) the yield strain is of order of 0.3, which is large enough to observe a full tensorial elastic regime while small enough that plastic effects can be easily observed; (iii) all physical quantities can be easily measured.
\begin{table}
\centering
\begin{tabular}{|c||c|c|c|c|}
\hline
symbol & $\Phi_{\rm eff}$ & $\delta A/A$ & geometry& $\gamma_{max}$\\
\hline
\hline
{\large \color{black}$\blacktriangledown$} or {\large \color{black}$\triangledown$} & $9.7.10^{-5}$ & 0 & fully periodic & $\pm$ 2\\
\hline
{\color{blue}${\mathbf{\times}}$} or {\color{blue}${\mathbf{+}}$} & $3.9.10^{-4}$ & 0 & fully periodic & $\pm$ 2\\
\hline
{\color{green}$\blacklozenge$} or {\color{green}$\lozenge$} & $3.9.10^{-4}$ & 0.025 & fully periodic & $\pm$ 2\\
\hline
{\color{cyan}$\blacktriangle$} or {\color{cyan}$\triangle$} & $3.9.10^{-4}$ & 0.66 & fully periodic & $\pm$ 2\\
\hline
{\Large \color{red}$\bullet$} or {\Large \color{red}$\circ$} & $3.5.10^{-4}$ & 0 & confined & $\pm$ 2.5\\
\hline
{\color{magenta}$\blacksquare$} or {\color{magenta}$\square$} & $3.5.10^{-4}$ & 0.66 & confined & $\pm$ 2.5\\
\hline
\end{tabular}
\caption{
Characteristics of simulated foams. The different columns correspond to the symbols used in Figs.
\ref{Fig:PlaPlasticLimit} and \ref{Fig:PlaOvershootUs}, effective liquid fraction \cite{Raufaste2007}, area dispersity, boundary conditions and maximal amplitude of the cycles.
}\label{Tab:DefSimul}
\end{table}
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(8,12)(0.0,0.0)
\put(0,8.1){\includegraphics[width=4cm]{refFP_1_desordre.eps}}
\put(4.5,8.1){\includegraphics[width=4cm]{refFP_800_desordre.eps}}
\put(0,4){\includegraphics[width=4cm]{refFP_5600_desordre.eps}}
\put(4.5,4){\includegraphics[width=4cm]{refFP_6100_desordre.eps}}
\put(2,-0.1){\includegraphics[width=4cm]{refFP_7200_desordre.eps}}
\put(-0.5,8.1){\large{\textbf{(1)}}}
\put(4,8.1){\large{\textbf{(2)}}}
\put(-0.5,4){\large{\textbf{(3)}}}
\put(4,4){\large{\textbf{(4)}}}
\put(1.5,-0.1){\large{\textbf{(5)}}}
\end{picture}
\caption{
Example of 2D foam simulation. Pictures are successive snapshots of a quasi-statically sheared, fully periodic foam. Numbers correspond to those of Figs. \ref{Fig:integral_strain} and \ref{Fig:SimulationRepresentation}. Bubbles with 6 neighbours are displayed in white, otherwise in gray.}
\label{Fig:PlaStructureRefFP}
\end{figure}
\subsection{Methods}
\label{sec:NumericalApproach}
Several ideal, two-dimensional, dry foams \cite{Weaire1999} are simulated
(Fig. \ref{Fig:PlaStructureRefFP} and Table \ref{Tab:DefSimul}).
We use the Surface Evolver \cite{brakke92} in a mode in which each film is
represented as a circular arc. The value of surface tension is taken equal to 1 throughout, without loss of generality. A realistic foam structure is found by minimizing the total film length
subject to the constraint of fixed bubble areas, prescribed at the beginning of the simulation.
The simulations are quasistatic, which means that the system has time to relax between successive time steps (increments in applied strain). Relaxation effects are thus neglected and viscosity does not need to be included. The behaviour is expected to be elasto-plastic.
The simulation procedure is as follows. A Voronoi construction of randomly distributed points
\cite{brakke86} (not shown) is first used to generate a fully periodic tessellation of
the plane. To create a confined foam, bubbles at the top and bottom
are sequentially deleted until the required number of bubbles remains.
In each case, the structure is imported into the Surface Evolver and target bubble areas prescribed, either all the same (monodisperse, $\delta A/A = 0$), a small random variation of up to 20\% about monodisperse ($\delta A/A = 0.025$), or equal to the areas given by the Voronoi construction ($\delta A/A = 0.66$).
The initial foam configuration for each simulation ({\it e.g.} label (1) in Fig. \ref{Fig:PlaStructureRefFP}) is found by reducing the total
film length to a local minimum. During this minimization, neighbour swappings (so-called ``T1s" \cite{Weaire1999}) are
triggered by deleting each film that shrinks below a certain critical length $l_c$ and allowing a new film to form to complete the process. The critical length $l_c$ defines and measures an effective liquid fraction, $\Phi_{\rm eff}$ \cite{Raufaste2007}, here chosen to be very dry (Table \ref{Tab:DefSimul}).
One geometry consists of a unit cell of 400 bubbles with fully periodic boundary conditions to eliminate any artefacts due to small sample sizes. The second geometry mimics more closely a real experiment, and consists of 296 bubbles with two parallel bars (about 15 bubble diameters apart) confining the foam and with periodicity in one direction only.
To shear the foams, two different procedures are required. For the periodic foams, one off-diagonal component of the matrix describing the periodicity of the unit cell is adjusted by a small amount \cite{brakke92}. For the confined foams, a small step in strain is applied by moving one of the
confining walls a small distance and then moving all vertices affinely. In each case this is followed by reduction of the film length to a minimum correct to 16 d.p. using conjugate gradient (without biasing the search by introducing any large-scale perturbations of the structure).
\begin{figure}[!h]
\includegraphics[width=8cm]{Uxy_gamma_time_gammadot_6_25e-4_b.eps}
\caption{
Time evolution of the reference
simulation ({\color{blue}$\times$} in Tab. \ref{Tab:DefSimul}). Horizontal axis: time is in arbitrary units, equivalent to the ``cumulated strain" $ \int | \dot{\gamma}|dt$, where $t$ is the time and $ \dot{\gamma}$ is defined up to an arbitrary prefactor; here 2.25 cycles are represented. Vertical axis: all curves represent $U_{xy}$ (left scale) except for the saw-tooth which is the applied $\gamma$ (right scale). Numbers correspond to the pictures in Fig. \ref{Fig:PlaStructureRefFP}. The first step, plotted with a thick solid line, starts at the $\Box$ (indicated also by a number 1) and its end is labelled by number 2: $\gamma=0\to 2$. The second step, plotted with a thin solid line, is from number 2 to 3. The third step, plotted with a middle solid line, starts at the $\lozenge $ (indicated also by a number 3), and extends to number 5: $\gamma=-2\to 2$. Four predictions of the model are plotted as dashed lines (see Fig. \ref{Fig:Pla-h-n} for explanation of the legend); for clarity they are plotted only from 1 to 2 and from 3 to 5: note that from 3 to 4 all predictions are indiscernable from the simulation.
}
\label{Fig:integral_strain}
\end{figure}
Each foam is subjected to at least two ``saw-tooth" shear cycles of amplitude $\gamma_{max}$. Positive and negative steps correspond respectively to shear toward increasing or decreasing imposed strain $\gamma$ (Fig. \ref{Fig:integral_strain})
\begin{figure}[!h]
\includegraphics[width=8cm]{U1U2_ref_it.eps}\\
{\large{\textbf{a)}}}\\
\includegraphics[width=8cm]{Emodulus_montage_it_2.eps}\\
{\large{\textbf{b)}}}
\caption{Validation of hypotheses. Both approximations of section
\ref{sec:Measurements} are tested during the shear of the reference
simulation ({\color{blue}$\times$} in Tab. \ref{Tab:DefSimul}).
a)
Strain eigenvalues $U_1$ {\it vs} $U_2$; black solid line: initial
shear ($1-2$)which anneals the disorder; blue dots: shear cycles ($2-5$); dashed blue line: straight line of slope -1 passing through the
origin.
b)
Deviatoric stress-strain relation: red, $\sigma_{xy}$ {\it vs}
$U_{xy}$; blue, $(\sigma_{xx}- \sigma_{yy})/2$ {\it vs} $(U_{xx}-
U_{yy})/2$; red and blue data almost perfectly overlap (see zoom in
inset); the slope determines $2\mu$, where $\mu$ is the shear modulus.
}
\label{Fig:VerifUTraceLess}
\end{figure}
\subsection{Measurements}
\label{sec:Measurements}
At each step, the positions of the bubble centres and films are recorded.
Tensorial quantities are measured by averaging over all bubbles, as follows \cite{Outils}.
The texture tensor $\tensor{M}=\langle \vec{\ell} \otimes \vec{\ell} \rangle$ is computed statistically as
an average over vector links $\vec{\ell}$ between centres of neighbouring bubbles. We assume here that the reference texture at rest, $\tensor{M}_0$, is isotropic. We thus define it by measuring the average of the determinant of $\tensor{M}$ over the duration of the whole simulation,
$\det(\tensor{M}_0) = \left\langle \det(\tensor{M}) \right\rangle$.
The elastic strain of bubbles expresses the deviation from the reference state, $ \tensor{U}
= \left( \log \tensor{M} - \log \tensor{M}_0 \right)/2$ (Eq. \ref{Eq:Fundament_4_ann}).
This tensor is symmetric by construction ($U_{yx} = U_{xy}$); it can be diagonalized and has two eigenvalues ($U_{1}$, $U_{2}$) in two orthogonal eigendirections.
The simulations (Fig. \ref{Fig:VerifUTraceLess}a) verify that we can reasonably assume its trace to be always close to zero, $U_1+U_2 = U_{xx}+U_{yy}\approx 0$, as is roughly expected for an incompressible material \cite{Outils}. Thus, due to its symmetry and vanishing trace, $\tensor{U}$ has only two independent components:
\begin{equation}\label{Eq:StressMatrix}
\tensor{U} =
\left(
\begin{array}{cc}
U_{xx} & U_{xy}\\
U_{yx} & U_{yy}
\end{array}
\right)
\approx
\left(
\begin{array}{cc}
\frac{U_{xx}-U_{yy}}{2} & U_{xy}\\
U_{xy} & -\frac{U_{xx}-U_{yy}}{2}
\end{array}
\right)
.
\end{equation}
Fig. \ref{Fig:VerifUTraceLess}a shows that our measurement of the elastic strain makes evident the effect of shear-induced shuffling \cite{{QuillietPreprint}}: the annealed foam (blue dots) differs significantly from the initial one (black solid line).
The contribution to the stress of the network of bubble walls is obtained by integrating over all films
\cite{bachelor,coxwhittick}; it yields the
deviatoric ({\it i.e}. traceless) part of the elastic stress tensor $\tensor{\sigma}$.
The trace of the stress, namely the pressure, is unimportant here. The simulations are quasistatic and the viscous part of the stress is not relevant.
We check (Fig. \ref{Fig:VerifUTraceLess}b) that the stress and strain are strongly correlated; that their correlation is linear; and that it is isotropic (the same for $xy$ and $xx-yy$ components) \cite{Asipauskas2003,Cartes}. Half the slope thus defines and measures the elastic shear modulus $\mu$.
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(7,4)(0.0,0.0)
\put(-1,0){\resizebox{4.cm}{!}{\includegraphics{Pla_Defangle4_it.eps}}}
\put(3.5,0){\resizebox{4.3cm}{!}{\includegraphics{Mat_ComposantesTemps2_it.eps}}}
\put(-0.5,0){\large{\textbf{a)}}}
\put(3.5,0){\large{\textbf{b)}}}
\end{picture}
\caption{Representation of the evolution of $\stackrel{=}{U}$. a) \emph{Physical space}: evolution of the point ($U\cos{\theta},U\sin{\theta}$); for completeness we also plot the opposite (and strictly equivalent) point
($-U\cos{\theta},-U\sin{\theta}$). b) \emph{Component space}: trajectory of ($(U_{xx}-U_{yy})/2,U_{xy}$), that is, ($U\cos{2 \theta},U\sin{2 \theta}$).}
\label{Fig:RepresentationPhysicalStress}
\end{figure}
\subsection{Representations}
\label{sec:Representations}
To summarize, $\tensor{M}$, $\tensor{U}$ and $\tensor{\sigma}$ characterize
the current state of the foam.
These three tensors carry here the same information, since $\mu$ appears constant.
In what follows, texture, elastic strain and stress tensors are always aligned.
We choose to display $\tensor{U}$ only, because it is dimensionless, and thus more general: it makes the comparison of different materials easy.
One possibility \cite{Outils} is to represent the traceless tensor $\tensor{U}$ as a circle of radius $U$, with a straight line to indicate the direction $\theta$ of its positive eigenvalue: see thick lines (circle and straight line) in Fig. \ref{Fig:RepresentationPhysicalStress}a. We do not use it here, except in the inset of Fig. \ref{Fig:PlaPlasticLimit}. In fact, it is easier to represent $\tensor{U}$ at a given time by a point, enabling us to plot trajectories.
Its two independent components can be represented in two different but equally useful ways, as follows.
Both representations are equally appropriate in the problem considered here because of the circular symmetry of the yield criterion (see Eq. \ref{Eq:UCriterion}).
First, in the case of a traceless tensor, the absolute value of the two eigenvalues is the same and equal to the amplitude $U$ of the tensor $\tensor{U}$ defined as
\begin{equation}\label{Eq:UPrincipalComponent_3}
U = \sqrt{ \left( U_{xx}-U_{yy} / 2 \right)^2 + U_{xy}^2},
\end{equation}
or equivalently
$U = \left| (U_{1}-U_{2})/2\right| = \left| \left| \tensor{U} \right| \right| / \sqrt{2}$, where $ \left| \left| \tensor{U} \right| \right|
= \left(\Sigma_{ij} \left(U_{ij}\right)^2\right)^{1/2}$ is the euclidian norm of $\tensor{U}$.
We call $\theta$ the direction of the greatest eigenvalue.
We call \emph{physical space} the representation of the parameters ($U$,$\theta$).
It is useful because it shows the evolution of the structure (elongation, orientation),
In particular, we plot the trajectory of the point ($U \cos \theta(\gamma), U \sin \theta(\gamma)$) (Fig. \ref{Fig:RepresentationPhysicalStress}a).
The other representation, which has already been used for foams \cite{KablaPreprintSimul}, is called \emph{component space}. It plots the trajectory of the point ($(U_{xx}(\gamma)-U_{yy}(\gamma))/2, U_{xy}(\gamma)$) (Fig. \ref{Fig:RepresentationPhysicalStress}b).
It is more suitable for comparison with experimental data, since rheometers measure the tangential
stress ($xy$), and sometimes the normal stress difference ($xx-yy$).
These two possible choices are related by
\begin{subeqnarray}
\label{Eq:UPrincipalComponent}
\slabel{Eq:UPrincipalComponent_1}
& U_{xy} & = U\sin{2\theta} \\
\slabel{Eq:UPrincipalComponent_2}
& \displaystyle \frac{U_{xx}-U_{yy}}{2} & = U\cos{2\theta}.
\end{subeqnarray}
Complete data for one simulation are plotted in Fig. \ref{Fig:SimulationRepresentation} and are discussed in the next section.
\newpag
\onecolumngrid
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(14,13)(0.0,0.0)
\put(-1.7,0){\resizebox{10cm}{!}{\includegraphics{Ugamma_it_num.eps}}}
\put(-1.2,0){\large{\textbf{c)}}}
\put(-1.7,7.1){\resizebox{10cm}{!}{\includegraphics{Uxygamma_it_num.eps}}}
\put(-1.2,7.1){\large{\textbf{a)}}}
\put(8.4,0){\resizebox{7.7cm}{!}{\includegraphics{componentspace_it_num.eps}}}
\put(8.5,0){\large{\textbf{d)}}}
\put(8.1,7){\resizebox{7.7cm}{!}{\includegraphics{physicalspace_it_num.eps}}}
\put(8.5,7.1){\large{\textbf{b)}}}
\end{picture}
\caption{Different representations of Fig. \ref{Fig:integral_strain} (same symbols) according to Fig. \ref{Fig:RepresentationPhysicalStress}.
a) $U_{xy}$ versus $\gamma$.
b) \emph{Physical space}. c) $U$ versus $\gamma$. d) \emph{component space}.
The dashed circle in b) and d) has radius $U_Y = 0.34$.
}
\label{Fig:SimulationRepresentation}
\end{figure}
\twocolumngrid
\section{Model}
\label{sec:Model}
\subsection{Implementation}
\subsubsection{Elasticity equations}
\label{sec:Formulation}
As already mentioned, we consider a quasistatic limit in which flow is slow enough that we may neglect viscous stresses. The stress is then related to the elastic strain (Fig \ref{Fig:VerifUTraceLess}b). We don't consider here the effect of external forces, such as gravity or friction on the boundary if the system is confined between glass plates \cite{Wang2006}.
In the present 2D case, classical plasticity \cite{Hill1950} suggests that the material begins to yield when the difference between the stress eigenvalues becomes too large: $(\sigma_{1}-\sigma_{2})^2 = 4 \sigma_{Y}^2$.
The yield stress $\sigma_Y$ separates a domain of pure elasticity from a domain in which the material flows plastically. A complete set of continuous equations (Reuss equations \cite{Hill1950}) can then be derived; $\sigma_{Y}$ is assumed to be constant (no strain hardening). The effect of pressure (trace of the stress) is neglected, which is usually a good first approximation for metals for instance \cite{Hill1950}.
It must be even more appropriate for soft materials, like a foam, for which the shear modulus is several orders of magnitude smaller than the bulk modulus \cite{Weaire1999}. In what follows we prefer to use the component yield criterion \cite{Hill1950}:
\begin{equation}\label{Eq:ComponentYieldCriterion}
\left(\frac{\sigma_{xx}-\sigma_{yy}}{2}\right)^2 + \sigma_{xy}^2 = \sigma_{Y}^2 .
\end{equation}
Equivalently, Eq. \ref{Eq:ComponentYieldCriterion} can be written for $\tensor{U}$, since the deviatoric parts of $\tensor{\sigma}$ and $\tensor{U}$ are proportional (Fig. \ref{Fig:VerifUTraceLess}b). From Eqs. \ref{Eq:UPrincipalComponent_3} and \ref{Eq:ComponentYieldCriterion}, we can write the yield criterion as
\begin{equation}\label{Eq:UCriterion}
U = U_{Y}.
\end{equation}
This is represented by a circle in both physical and component spaces (Fig. \ref{Fig:SimulationRepresentation}bd)
For the example of foams and highly-concentrated emulsions, Marmottant and Graner \cite{MarmottantPinceau} suggested that the transition between elastic and plastic regimes is not sharp, but can be described by a yield function $h$. This function is 0, respectively 1, in the pure elastic, respectively plastic, domain. Between these two limits, both effects are present and the proportion seems to depend mostly on the elastic strain amplitude $U$ (Fig. \ref{Fig:Pla-h-n}a). This assumption was successfully tested on different flow geometries of a 2D foam \cite{Cartes}.
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(8,5)(0.0,0.0)
\put(-0.4,-0.1){\includegraphics[height=5cm]{h_n_it.eps}}
\put(5.8,0){\includegraphics[height=5cm]{limit_cycle_it.eps}}
\put(0,0){\large{\textbf{a)}}}
\put(5.8,0){\large{\textbf{b)}}}
\end{picture}
\caption{(color online) Model. (a) The different yield functions $h$ used as examples in the present paper are power laws $h(U)= (U/U_Y)^n$ with $n=1$ (green, dash-dots), $n=2$ (red, dashes), $n=4$ (blue, dots) and $n=+\infty$ (black, thick dashes, equal to 0 everywhere except at $U=U_Y$ where it is equal to 1).
b) Corresponding limit cycles predicted by the model, plotted in
\emph{component space} (same legend).
}
\label{Fig:Pla-h-n}
\end{figure}
By assuming that the deformation is affine \cite{Outils} and according to the prediction of plasticity (Eq. 22 in \cite{Cartes}), we can then write a tensorial equation of evolution of the texture (Eq. 9 in \cite{Outils}). This dictates the evolution of the texture due to imposed strain (deformation, rotation) and due to relaxation (rearrangements), for details see Eqs. \ref{Eq:Fundament_3_ann},\ref{Eq:PlasticMarmottant}:
\begin{eqnarray}\label{Eq:EvolutionEquation}
\frac{d}{d t} \tensor{M} & = & \tensor{M}.\tensor{\nabla v} + \tensor{\nabla v}^{t}.\tensor{M} \nonumber \\
& & - \left(\frac{\tensor{U}}{U}:\tensor{\nabla v}_{sym}\right)\; {\cal H} \; h\left(\frac{U}{U_{Y}}\right) \frac{\tensor{U}}{U} .\tensor{M} \nonumber\\
& &
\end{eqnarray}
Here $d/dt$ is the Lagrangian derivative in time (including advection); $ \tensor{M}.\tensor{\nabla v} + \tensor{\nabla v}^{t}.\tensor{M} $ is the variation of $\tensor{M}$ due to convection by the velocity gradient $\tensor{\nabla v}$; as explained in appendix \ref{sec:NotationsForTensors} and \ref{sec:TensorComponents}, the notation
${U}:\tensor{\nabla v}_{sym}$ is the scalar product of the elastic strain tensor
with the symmetrized velocity gradient tensor
$\tensor{\nabla v}_{sym}= \left(\tensor{\nabla v} + \tensor{\nabla v}^{t}\right)/2$
\cite{Outils}; conversely, $\tensor{U} .\tensor{M} $ is the usual product of tensors; here
$ {\cal H} = {\cal H}\left(\frac{\tensor{U}}{U}:\tensor{\nabla v}_{sym}\right)$ is the Heaviside function, which is equal to 1 if ${U}:\tensor{\nabla v}_{sym}$ is positive and 0 otherwise.
Eq. \ref{Eq:EvolutionEquation} links the evolution of the foam texture with the elastic strain. It is quasistatic in the sense that the strain is relevant, not the strain-rate. It can be generalized to evolutions quicker than the relaxation times of the structure \cite{Saramito2007}.
Plasticity occurs only when the elastic strain is oriented in the direction of shear, as expressed by the Heaviside function ${\cal H}$ (Appendix \ref{sec:ElastoPlasticComponentEquations}).
The model is continuous and analytic, without fluctuations. The information regarding disorder is recorded in $h$.
Trapped stresses \cite{Larson1999} are recorded in the initial value $\tensor{M}_i$ (or equivalently $\tensor{U}_i$).
The material's yielding criterion is encoded in $U_Y$. The history of the material only plays a role in determining $h$, $\tensor{M}_i$ (or $\tensor{U}_i$), and $U_Y$,
which together fully describe the material. According to the expression of $h$, Eq. \ref{Eq:EvolutionEquation} can be integrated analytically or numerically.
\subsubsection{Simple shear}
\label{sec:Implementation}
To study the structure-evolution equation, {\it i.e.} the competition between elasticity and plasticity, and predict the rheological behaviour, we impose a strain rate $\dot{\gamma}$ on the material.
We take $x$ as the direction of the shear, which gives the following velocity field:
\begin{equation}
\tensor{\nabla v}=
\left(
\begin{array}{cc}
\partial_{x} v_{x} & \partial_{x} v_{y} \\
\partial_{y} v_{x} & \partial_{y} v_{y}
\end{array}
\right)
= \dot{\gamma}
\left(
\begin{array}{cc}
0 & 0 \\
1 & 0
\end{array}
\right)\label{Eq:PlaSimplificationGradV}
\end{equation}
and hence
\begin{equation}
\tensor{\nabla v}_{sym}
= \dot{\gamma}
\left(
\begin{array}{cc}
0 & 1/2 \\
1/2 & 0
\end{array}
\right).\label{Eq:PlaSimplificationGradVsym}
\end{equation}
This factor $1/2$ appears when comparing the scalar and tensorial descriptions
(appendix \ref{sec:AnnScalarLimit}).
In this geometry, the advection term is taken equal to zero and the resulting system of equations is given in Appendix \ref{sec:ElastoPlasticComponentEquations}.
The reference state $\tensor{M}_0$ is considered isotropic and constant throughout the evolution.
We recall that this evolution is quasistatic: $\dot{\gamma}$ appears as a prefactor in the time evolution, Eqs. \ref{Eq:Ann2EquationQuasistatique}. We thus follow the evolution with the strain $\gamma=\int \dot{\gamma}\; dt$, instead of the time.
Tensor operations and the time evolution of $\tensor{M}$ are implemented by a finite difference procedure.
Between two time steps, $\tensor{U}$ is recalculated with Eq. \ref{Eq:Fundament_4_ann}.
\subsection{Predictions}
\label{sec:Predictions}
We now address the resolution of the full elasto-plastic set of equations \ref{Eq:Ann2EquationQuasistatique}.
Our representation underlines the specifically tensorial effects.
\subsubsection{Purely elastic regime}
\label{sec:elasticRegime}
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(8,4)(0.0,0.0)
\put(-0.5,0){\includegraphics[width=4.6cm]{Uvec_elastic_0.3_dt_3200_1of4_it.eps}}
\put(3.8,0){\includegraphics[width=4.9cm]{Ucomponents_elastic_0.3_dt_3200_line_it.eps}}
\put(0,0){\large{\textbf{a)}}}
\put(4.3,0){\large{\textbf{b)}}}
\end{picture}
\caption{Elastic model, represented
in \emph{physical} (a) and \emph{component} (b) spaces. Different initial states are taken: $U_i = 0$ (center, red) and $U_i = 0.3$ (initial points scattered around the circle, black).
Here $n \rightarrow +\infty$: all the trajectories evolve elastically.
For $\dot{\gamma} > 0$,
$U_{xy}$ increases, that is, time evolve upwards on (b).
If $\dot{\gamma} < 0$, these graphs are unchanged, due to their symmetry with respect to the horizontal axis, showing that the purely elastic trajectories are reversible.
}
\label{Fig:ModelCurvesUyElast}
\end{figure}
As a first example of our representation,
we consider here the pure elastic regime. This means that we allow the structure to deform elastically (stretching, contraction), but not to relax plastically (no rearrangements).
Our formalism allows us to describe pure elasticity by computing the elastic strain and its evolution when the material is deformed.
Our formalism extends to large strains, even those of order one (for strains much larger than one, without plasticity, the formalism of large amplitude strain \cite{Labiausse2007} might be preferable).
Fig. \ref{Fig:ModelCurvesUyElast}ab shows trajectories for different initial elastic strains.
For an initially isotropic material, $U_i = 0$, we recover the classical results in the small strain limit: $U \simeq \gamma/2$ and $U_{xy} \simeq \gamma/2$. The Poynting relation \cite{Labiausse2007} thus takes the form of a parabola; we even extend it to an initially anisotropic material, $U_i \neq 0$ (see Fig. \ref{Fig:test} and \ref{Eq:ApproximationComplete}).
For higher strains, $U_{xy}$ is less linear with respect to $\gamma$, due to the rotation of the elastic strain.
In all cases, $U_{xy}$ increases monotonically. Note that this is not the case for $(U_{xx}-U_{yy})/2$ nor for $U$.
When the structure is aligned perpendicularly to the shearing direction ($\tensor{U}:\tensor{\nabla v}_{sym} < 0$), it contracts ($U$ decreases) under shear, until it aligns with the shear.
When the structure is aligned parallel to the shearing direction ($\tensor{U}:\tensor{\nabla v}_{sym} > 0$), it stretches ($U$ increases) under shear.
Since $\tensor{U}$ is a tensor, it can continuously decrease, change direction and increase again without ever vanishing (as opposed to a scalar, which can change sign only when it is equal to zero).
For instance, a trajectory which starts with a direction opposed to that of shear has first a decreasing $U$ (contraction, with $U_{xy}$ negative and increasing), then an increasing $U$
(stretching, with $U_{xy}$ positive and increasing), then a constant $U$ (yielding, with a rotation of $\tensor{U}$ towards the plastic limit).
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(7,5)(0.0,0.0)
\put(0,-0.2){\resizebox{7cm}{!}{\includegraphics{Usat_2_it.eps}}}
\put(1.4,0.8){\resizebox{5.5cm}{!}{\includegraphics{Pla_Uellipse.eps}}}
\end{picture}
\caption{Plastic limit for $\dot{\gamma}>0$. Model of $\theta_Y$ versus $U_Y$ (solid line)
and corresponding representations of $\tensor{U}$ as circles with straight lines, indicating the direction of positive eigenvalue (inset), for several values of $U_Y$.
Simulation points (same symbols as in Table \ref{Tab:DefSimul}) are plotted for comparison.}
\label{Fig:PlaPlasticLimit}
\end{figure}
\subsubsection{Plastic Limit}
\label{par:modelPlasticLimit}
The yield strain is the amplitude of the strain when the material yields, that is, a scalar number.
The {\it plastic limit} is defined as the elastic strain tensor $\tensor{U}$ obtained after an infinitely long shearing ($\gamma \rightarrow +\infty$). The amplitude of this tensor is that of the yield strain. Its direction is obtained by solving Eq. \ref{Eq:EvolutionEquation}
when its left-hand side equals 0, $h$ equals 1, and ${\cal H}$ equals 1:
\begin{subeqnarray}
\label{Eq:PlaSolutionplastique}
\label{Eq:PlaSolutionplastique_1}
U & = & U_Y,\\
\label{Eq:PlaSolutionplastique_2}
\cos\theta & = & \frac{1}{\sqrt{1+e^{-4 U_Y}}},\\
\label{Eq:PlaSolutionplastique_3}
\sin\theta & = & \mathtt{sign}(\dot{\gamma})\frac{1}{\sqrt{e^{4 U_Y}+1}}
.
\end{subeqnarray}
This plastic limit is represented on Fig. \ref{Fig:PlaPlasticLimit}.
It shows that the larger $U_Y$, the less aligned $\tensor{U}$ is with respect to $\tensor{\nabla v}_{sym}$. This tensorial effect is strong because $\theta_Y$ decreases quickly with $U_Y$. The scalar limit corresponds to $\theta \simeq 45^\circ$, as discussed in Sec. \ref{sec:ComparisonScalarTensorial}.
\subsubsection{Transient regime}
\label{sec:TransientRegim}
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(8,12)(0.0,0.0)
\put(0,8){\includegraphics[width=4cm]{Uvec_plasticINF_0.3_dt_3200_1of4_2_it_arrow.eps}}
\put(4,8){\includegraphics[width=4cm]{Ucomponents_plasticINF_0.3_dt_3200_line1of4_2_it_arrow.eps}}
\put(0,4){\includegraphics[width=4cm]{Uvec_plastic2_0.3_dt_3200_1of4_2_it_arrow.eps}}
\put(4,4){\includegraphics[width=4cm]{Ucomponents_plastic2_0.3_dt_3200_line1of4_2_it_arrow.eps}}
\put(0,0){\includegraphics[width=4cm]{Quadrant1_it.eps}}
\put(4,0){\includegraphics[width=4cm]{Quadrant2_it.eps}}
\put(0,8){\large{\textbf{a)}}}
\put(4,8){\large{\textbf{b)}}}
\put(0,4){\large{\textbf{c)}}}
\put(4,4){\large{\textbf{d)}}}
\put(0,0){\large{\textbf{e)}}}
\put(4,0){\large{\textbf{f)}}}
\end{picture}
\caption{Elasto-plastic model.
Representation of the model for $\dot{\gamma} > 0$ in \emph{physical} (a,c,e) and \emph{component} (b,d,f) spaces.
If $\dot{\gamma} < 0$, the vertical axis of all these graphs should be inverted.
(a-b) For $n \rightarrow +\infty$, all the trajectories evolve elastically ($U_{xy}$ increases, that is, time evolves in the direction of the arrows) up to the yield strain, here taken as $U_Y=0.3$, then evolve plastically ($U$ constant) towards
the plastic limit (blue point) which corresponds to a specific angle, see Eq.
\ref{Eq:PlaSolutionplastique}. If the trajectory reaches the yield strain on the left of the plastic limit, $U_{xy}$ passes through a maximum (overshoot).
(c-d) For $n = 2$, plasticity appears more progressively, thus smoothening the transition between elastic and plastic regimes, and decreasing (or even suppressing) the overshoot.
(e-f) Summary of the mechanical behavior: ``stretching" and ``contraction" according to the direction of $\tensor{U}$ with respect to shear. Here ``plasticity on" or ``plasticity off" refers to the Heaviside function in the last term of Eq. \ref{Eq:EvolutionEquation}, when the plasticity is progressive ($n$ finite); when $n$ increases the ``plasticity on" zone narrows, and for $n$ infinite it is reduced to the limit circle.
}
\label{Fig:ModelCurvesUyPlast}
\end{figure}
We now consider the shearing of an initially anisotropic elastic strain, $\tensor{U}_i\neq \tensor{0}$. In \emph{physical} or \emph{component space}, the state of the material is initially situated on an elastic trajectory and must arrive at the plastic limit point (Fig. \ref{Fig:ModelCurvesUyPlast}a-d). At this limit point, an increase of elastic strain is immediately transformed into plastic strain. Plasticity may occur only if $\tensor{U}:\tensor{\nabla v}_{sym} > 0$.
Graphically, both in \emph{physical space} and in \emph{component
space} (Fig. \ref{Fig:ModelCurvesUyPlast}), the plastic limit is
represented by the point where a trajectory reaches perpendicularly the
circle at $U = U_Y$ (the elastic strain increases along the
tangent of the trajectory, the plastic strain relaxes towards the centre of
the circle: to balance each other, they must be parallel).
The shape of the yield function $h$ then determines how the material reaches the plastic limit. For simplicity, we take $h$ as a power law function: $h = (U/U_Y)^n$ (Fig. \ref{Fig:Pla-h-n}a). Two examples of the resulting behaviour are plotted on Fig. \ref{Fig:ModelCurvesUyPlast}.
The limit $n \rightarrow \infty$ is intuitive: the material follows the elastic trajectory up to $U = U_Y$; then $U$ is fixed and the plastic limit is reached by describing an arc of a circle
in \emph{physical} and \emph{component spaces} (Fig. \ref{Fig:ModelCurvesUyPlast}ab).
For other cases (finite $n$) plasticity occurs earlier (Fig. \ref{Fig:ModelCurvesUyPlast}ef) and trajectories converge to the plastic limit (Fig. \ref{Fig:ModelCurvesUyPlast}cd).
The behaviour changes qualitatively if the sign of $\dot{\gamma}$ is abruptly reversed. Unlike the elastic term, the plasticity term is irreversible due to the Heaviside function $\cal H$ in
Eq. \ref{Eq:EvolutionEquation}.
This leads to an inversion of the plastic domain in \emph{physical} and \emph{component spaces} (Fig. \ref{Fig:ModelCurvesUyPlast}ef) and to a new plastic limit position ($\theta_Y \rightarrow -\theta_Y$, Eq. \ref{Eq:PlaSolutionplastique}). This hysteretic effect is shown on
Fig. \ref{Fig:SimulationRepresentation}
by reversing $\dot{\gamma}$ once the plastic limit is reached. For high $n$, the new plastic limit is quickly reached, since the two plastic limits are on the same elastic trajectory.
If we perform alternate sign changes of $\dot{\gamma}$, we observe that the material is stuck in a limit trajectory
(Fig. \ref{Fig:Pla-h-n}b).
This trajectory is almost insensitive to $h$ and therefore close to the elastic trajectory joining the two plastic limits. This has an important consequence: once in the plastic regime, the elastic strain (and thus the stress) can not be totally relaxed if we only reverse the shearing direction. This is examined in more detail below (Fig. \ref{Fig:RelaxationCycles}).
\begin{figure}[!h]
a)
\includegraphics[width=7cm]{zoom_overshoot_it.eps}\\
b)
\includegraphics[width=7cm]{overshoot4_it.eps}\\
c)
\includegraphics[width=7cm]{overshoot4_zoom_it.eps}\\
\caption{
Overshoot, defined as the difference between the maximum of the $U_{xy}$ versus $\gamma$ curve, and the value of the plateau which follows this maximum.
Calculations are performed in the case of an initially isotropic structure ($U_i = 0$).
(a) Zoom over the maxima of the curves of
Fig. \ref{Fig:SimulationRepresentation}a.
(b) Maxima for different $h$ functions (same legend as Fig. \ref{Fig:Pla-h-n}), and plateau value (starred line, which is the same for all $h$ functions), plotted {\it vs} $U_Y$.
Simulation points
are plotted for comparison, with the same symbols as in Table \ref{Tab:DefSimul}; closed symbols and {\color{blue}$+$} correspond to the maximal value (averaged over a few successive points) during the first shear step; open symbols and {\color{blue}$\times$} correspond to the averaged value along the plateau of the last shear step; $U_Y$ is measured as the plateau value of $U$.
(c) Zoom of (b).
}
\label{Fig:PlaOvershootUs}
\end{figure}
\subsubsection{Overshoot}
\label{sec:Overshoot}
As observed for $n \rightarrow +\infty$ (Fig. \ref{Fig:ModelCurvesUyPlast}b),
the overshoot is due to the transition from an elastic trajectory to the plastic limit.
The structure itself has no overshoot: $U$ increases monotonically.
The overshoot appears in the tangential strain $U_{xy}$: it is a purely tensorial effect due to a rotation of the structure.
In fact, $U_{xy}$ increases in all elastic trajectories.
Upon reaching $U=U_Y$ there is a sudden transition to the plastic regime. For trajectories to the left of the plastic limit (Fig. \ref{Fig:ModelCurvesUyPlast}b),
we see that $U_{xy}$ decreases towards the plastic limit. The overshoot corresponds to the difference between the maximum value of $U_{xy}$
(where the trajectory meets the circle) and the plateau value (plastic limit).
>From Fig. \ref{Fig:ModelCurvesUyPlast}b,
we observe a tiny overshoot for the normal stress difference if
the trajectories reach the $U=U_Y$ circle to the right of the plastic limit. In that case,
the trajectories move towards the right in the elastic regime, then towards the left in the plastic regime.
Such trajectories correspond to the right of Fig. \ref{Fig:ModelCurvesUyPlast}b, that is, a structure with a large trapped normal stress difference
$U_{xx}-U_{yy}$.
For smaller $n$, the elastic-plastic transition is smoother and the overshoot is reduced.
The overshoot amplitude for different $h$ and $U_Y$ is plotted on Fig. \ref{Fig:PlaOvershootUs},
for the case of an initially isotropic structure ($U_i = 0$).
The overshoot increases with $U_Y$ because the plastic limit moves away from the initial elastic trajectory.
\section{Comparison between simulations and the model}
\label{sec:ComparisonSimulationModel}
\subsection{Plastic limit}
\label{sec:ResultPlasticLimit}
The model (section \ref{sec:TransientRegim}) predicts that after a few cycles a limit trajectory is reached. This trajectory is also displayed in simulations by Kabla and Debregeas \cite{KablaPreprintSimul}.
The plastic limit can thus be evaluated in simulations: $U_Y$ is estimated by averaging $U$ on the last plateau
(using any of Fig. \ref{Fig:SimulationRepresentation}a-d);
similarly, $\theta_Y$ is estimated by averaging $\theta$ on the last plateau of Fig. \ref{Fig:SimulationRepresentation}b or d.
On Fig. \ref{Fig:PlaPlasticLimit}, results from simulations are compared with the model. Agreement is good, and the model captures tensorial effects, especially because the measured $\theta_Y$ deviate much from the $45^\circ$ scalar limit.
Taking larger $l_c$ in the simulations favours neighbour swappings (``T1s"), thus it corresponds to an increased effective liquid fraction. As expected \cite{{Saint-Jalmes1999},Raufaste2007}, we see a decrease in $U_Y$. However, since the effective liquid fraction we are simulating remains in a very dry range ($<4 \; 10^{-4}$), it does not influence much $U_Y$, which thus varies over a narrow range ($0.26-0.37$).
Reaching higher $U_Y$ is possible with other materials, but not with disordered 2D foams. Reaching lower $U_Y$ is possible (and usual) in experiments on disordered wet foams, but not in the present simulations where the algorithm would require adaptation at high $l_c$.
\subsection{Yield strain and yield function }
\label{sec:EvolutionOfTheStructure}
Simulation results fluctuate, due to the limited number of bubbles (discrete description), while model curves are smooth, corresponding to the limit of a large number of bubbles (continuous material description).
There is a qualitative agreement, which is good enough to deduce $U_Y$ and $h$ approximately.
For instance, Fig. \ref{Fig:SimulationRepresentation} compares a simulation with models using various $h$ functions.
We observe that $n \approx 2$
(red dashes on Fig. \ref{Fig:SimulationRepresentation})
describes well the simulation during the first positive shearing step, and $n \approx 4$
(blue dotted lines on Fig. \ref{Fig:SimulationRepresentation})
during the second one.
Similarly, $U_Y$ is deduced from the plateau value of $U$ (Figs. \ref{Fig:PlaPlasticLimit} and \ref{Fig:PlaOvershootUs}).
In practice, in a first approximation, it is enough to consider $U_Y$ and $h$ as constant.
Their variations are small and thus have a small effect on the foam rheology.
However, these variations do exist.
For instance, in this example of Fig. \ref{Fig:SimulationRepresentation}, $n$ (and thus $h$) evolves throughout the simulation, revealing that the structure
evolves too; $h$ seems to be sensitive to the (topological) disorder of the foam \cite{{QuillietPreprint}}. Here $U_Y$ is constant, but there are other cases (data not shown, see \cite{RaufasteThese}) where, due to the decrease of the topological disorder during the shearing, $U_Y$ decreases.
More generally, a real foam is constantly evolving under the effect of drainage, coarsening
\cite{Hohler2005}, or
shuffling \cite{{QuillietPreprint}}.
These effects should probably have to be considered in future models, which would try to predict $U_Y$ and $h$, based on the average and fluctuations of the structure, respectively.
\subsection{Overshoot }
\label{sec:DiscOvershoot}
We can now identify two distinct physical mechanisms which can cause a stress overshoot in shear experiments of elasto-plastic materials.
The first one is an {\it orientation} effect, suggested in Sec. \ref{sec:Overshoot}.
$U$ increases monotonically, but if $U_Y$ is large enough then the rotation of $\tensor{U}$ under shear implies that the tangential shear strain $U_{xy}$ passes through a maximum.
This purely tensorial effect is absent from scalar models.
Under certain additional conditions on the initial elastic strain $U_i$, which are also described correctly only when taking into account the tensorial aspects, the normal strain difference $U_{xx}-U_{yy}$ too passes through a maximum.
Fig. \ref{Fig:PlaOvershootUs} shows a comparison between the model and the simulations. Given that in the range of simulated $U_Y$ the overshoot is tiny and difficult to extract from the fluctuations,
the agreement is surprisingly good. In most foam experiments, where $U_Y$ is even lower,
this effect should be too small to be measurable.
The second one is outside of the scope of the present paper.
It is due to an evolution of the {\it structure} itself during the first shear step (see section \ref{sec:EvolutionOfTheStructure}).
This might be invoked to explain the larger overshoot of the data corresponding to the confined simulations (red and pink), as well as most experimental observations (such as that of ref. \cite{Khan1988}).
\section{Practical applications}
\label{sec:Practical}
\subsection{Comparison between scalar and tensorial representations}
\label{sec:ComparisonScalarTensorial}
As long as the applied shear keeps a constant direction, and the elastic strain remains much smaller than 1,
its eigenvectors correspond to that of $\tensor{\nabla v}_{sym}$. That is, they are at $45^\circ$ to the direction of shear.
This is called the {\it scalar approximation},
and it considerably simplifies the study of the mechanical behaviour.
In that case,
a single (scalar) number is enough to fully describe the elastic strain.
This scalar number can equally well be chosen as the amplitude $U$, or the eigenvalue $U_1$, or the tangential shear strain $U_{xy}$, among others. To switch from one choice to the other requires
care regarding the prefactors \cite{Cartes}: this is often a source of confusion in the literature, especially regarding the definition and value of the yield strain.
The link between the simplified (scalar) and complete (tensorial) equations is detailed in Appendix \ref{sec:AnnScalarLimit}, using $2U_{xy}$ as a scalar.
If $U_Y \ll 1$, which is the case for wet foams and emulsions, then $U$ remains always much smaller than 1, and $\theta_Y \simeq 45^\circ$, so that the scalar approximation holds, see Fig. 19 in ref. \cite{Cartes}
(except if the direction of the shear changes, in 2D or in 3D).
In that limit tensorial effects such as normal differences or stress overshoot are negligible.
Quantitatively, $U_{xy}$ is linked to $\sin(2 \theta)$ (Eq. \ref{Eq:UPrincipalComponent_1}).
This implies that a difference of 10\% between the scalar and tensorial equations is reached when $\sin(2 \theta_Y) = 0.9$, which corresponds to $U_Y = 0.23$ (Eq. \ref{Eq:PlaSolutionplastique}).
Very dry foams, such as those simulated here, are slightly above this limit: a tensorial model is therefore useful.
\begin{figure}[!h]
a)
\includegraphics[width=7cm]{components_cycle_it.eps}\\
b)
\includegraphics[width=7cm]{Unormal_cycle_it.eps}\\
\caption{Shearing cycles to remove trapped stresses. Here $U_i = 0$ for the first step and $U_Y = 0.3$, $n \rightarrow +\infty$. Between two steps: the direction of shearing is turned $90^\circ$ clockwise (blue dots); the amplitude is decreased from $\gamma =2$ to 0 in 15 steps (black dashes); simultaneously, the direction of shear is turned and its amplitude is decreased (red solid line).
a) Component space. b) $(U_{xx} - U_{yy})/2$ versus time ($\left|\dot{\gamma}\right| = 2.5 \times 10^{-3}$s$^{-1}$).}
\label{Fig:RelaxationCycles}
\end{figure}
\subsection{Trapped strains and stresses}
\label{sec:TrappedStrainsStresses}
A dry foam is a material with sufficiently high $U_Y$ that normal stresses may exist even when the material is at rest \cite{Larson1999,Kraynik2003,Janiaud2005}. To relax such residual (or ``trapped") stresses,
we should first shear the foam enough to reach the plastic stage, so that plastic rearrangements anneal the disorder.
We then must perform cycles of shear.
If the direction of shear is kept constant, and the shear simply reversed, the foam asymptotically reaches a limit trajectory, and the stress is not relaxed.
Decreasing the amplitude of the shear cycle does not enable to leave this limit trajectory (black dashes in Fig. \ref{Fig:RelaxationCycles}ab).
Kraynik \emph{et al.} simulated dry 3D foam and applied shearing cycles (actually uniaxial contractions) of amplitude $\approx 0.2$ in different directions, rotated by $90^\circ$; this procedure decreases the trapped stress by a factor of around 2, which does not improve with more cycles (Fig. 7 of \cite{Kraynik2003}).
Here we propose a reproducible procedure based on section \ref{sec:TransientRegim}, which couples shearing cycles in different directions and decreasing amplitudes, as follows:
\begin{itemize}
\item The amplitude $\gamma_i$ of the first step is large enough to completely reach the plastic stage: $\gamma_i \gg 2 U_Y$.
\item At each step, the shearing direction is rotated by $90^\circ$
and the shearing amplitude is decreased.
\item The decrease in amplitude between successive steps is smaller than $2U_Y/5$, ensuring there are at least 5 steps between $2 U_Y$ and 0
({\it i.e.} the total number of steps is at least $5 \gamma_i / 2 U_Y$).
\end{itemize}
The red solid line in Fig. \ref{Fig:RelaxationCycles}ab shows that the normal stress difference decreases more at each cycle; for instance, 6 cycles yield a decrease by a factor of 10, apparently without saturating.
In 3D the procedure is the same, rotating the shearing direction successively along the $x$, $y$ and $z$ axes \cite{Kraynik2003}.
This procedure is easy to apply to simulations, especially of fully periodic foams.
In experiments, a special set-up should be built: in 2D, it can be a rubber frame in the spirit of refs. \cite{Abdelkader1999,QuillietPreprint}, if the four corners can be independently displaced.
\begin{figure}[!h]
\setlength{\unitlength}{1cm}
\centering
\begin{picture}(7,4.5)(0.0,0.0)
\put(-0.7,-0.1){\resizebox{4.5cm}{!}{\includegraphics{sursautUy1_Uxy_it.eps}}}
\put(3.5,0){\resizebox{4.5cm}{!}{\includegraphics{sursautUy1_Ucomponent_it.eps}}}
\put(-0.5,0){\large{\textbf{a)}}}
\put(4,0){\large{\textbf{b)}}}
\end{picture}
\caption{Representations of the model for $U_Y = 1$.
a)
$U_{xy}$ versus $\gamma$ for the first step and different yield functions, as in Fig. \ref{Fig:Pla-h-n}a.
b) Representation of the model for $\dot{\gamma} > 0$ in \emph{component spaces}
for $n \rightarrow +\infty$,
as in Fig. \ref{Fig:ModelCurvesUyPlast}b.
}
\label{Fig:BigovershootU1}
\end{figure}
\begin{figure}[!h]
\centering
\setlength{\unitlength}{1cm}
\begin{picture}(8,4.5)(0.0,0.0)
\put(0,-0.5){\resizebox{8cm}{!}{\includegraphics{elasticity_parabola_2_it.eps}}}
\end{picture}
\caption{Representations of the model for small strains. $U_{xy}$ {\it vs} $U_{n}=(U_{xx}-U_{yy})/2$. Black: exact model. Blue: first parabolic approximation $U_{n}=U^i_{n}+ U_{xy}^2$. Red: complete parabolic approximation (Eq. \ref{Eq:ApproximationComplete}). }
\label{Fig:test}
\end{figure}
\subsection{Materials with low and high $U_Y$}
For practical purposes we plot the reference curves for two limiting types of materials: those with $U_Y$ much higher or much lower than 0.23.
Fig. \ref{Fig:BigovershootU1} shows the example of $U_Y=1$. The plastic limit corresponds to a small angle $\theta_Y$, resulting in a strong overshoot.
Fig. \ref{Fig:test} shows that for small strains in the elastic regime, all curves can be expressed using a single parameter; for instance, as here, $U^i_{n}$, which is the normal elastic strain $U_{n}=(U_{xx}-U_{yy})/2$ at zero tangential shear ($U_{xy}^i=0$). A rough parabolic approximation, and a refined one (Eq. \ref{Eq:ApproximationComplete}) are plotted here for $U_Y=0.3$. For smaller $U_Y$, this approximation is good over its whole range of validity (namely the elastic regime), but this range is smaller.
\section{Summary}
\label{sec:Conclusions}
We propose a continuous model of the elasticity and plasticity of disordered, discrete materials such as cellular patterns (for instance liquid foams or emulsions) and assemblies of particles (for instance colloids). It is based on statistical quantities including (i) the elastic strain $\tensor{U}$, a dimensionless quantity measurable on images, which facilitates the comparison between different experiments or models, and makes apparent the effect of shear on the material's structure; (ii) the yield strain $U$, a classical criterion for the transition between reversible, elastic and irreversible, plastic regimes; (iii) and the yield function $h(U/U_Y)$, which describes how progressive this transition is, by measuring the relative proportion of elastic and plastic deformation. They suffice to relate the discrete scale with the collective, global scale. At this global scale, the material behaves as a continuous medium; it is described with tensors such as strain, stress and velocity gradient. We give the differential equations which predict the elastic and plastic behaviour. The model is fully tensorial and thus general, in 2D or in 3D.
We study in detail the case of simple shear. An original representation, suitable for 2D incompressible materials, is introduced to follow the evolution of the material during shear.
Since $\tensor{U}$ is a tensor, it has an orientation and an amplitude, which both evolve under shear.
It can continuously decrease its amplitude, change direction and increase again its amplitude without ever vanishing (as opposed to a scalar, which can change sign only when it is equal to zero).
Predictions of the model regarding orientation and stretching are plotted. They include a rotation of the structure, which can induce an overshoot of the shear strain or shear stress (and a smaller, rarer overshoot in normal stress differences) even without overshoot in the elastic strain amplitude.
This purely tensorial effect exists if $U_Y$ is at least of order of 0.3. Independently, the shear can also induce a change in the material's structure, sometimes resulting in a (purely scalar) overshoot in the modulus of the elastic strain.
The model extends a classical plasticity criterion to disordered media. It can be solved numerically and yields testable predictions. We successfully compare them with carefully converged quasistatic simulations of shear cycles in 2D foams: the elastic strain increases, saturates and reverses.
From this comparison between model and simulation we determine $U_Y$ and estimate $h$. This method is similar to that which we have used in experiments to extract $U_Y$ \cite{Cartes,Cheddadi2008}, and a rough estimate of $h$. We still lack a model to predict $U_Y$ and $h$. Both quantities evolve throughout the simulation, probably due to the evolution of the foam's internal structure, as well as the disorder and fluctuations. In short, the material obeys a continuous description determined by its average properties, while $U_Y$ and $h$ account for the effect at large scale of its fluctuations.
All quantities involved in the model are directly measurable, as tensors, in the current state of the material; this includes trapped stresses which we discuss (we also explain how to relax them):
the history of the sample which led to this current state plays no other direct, explicit role.
We explain how and when to use the model in practice, and provide a set of curves and analytical approximations, including a discussion and an extension of the Poynting relation. At low strain, typically below 0.2, tensorial effects vanish and an approximate scalar simplification holds.
\section*{Acknowledgements}
We thank C. Quilliet, B. Dollet, S. Attai Talebi and other participants in the Grenoble Foam Mechanics Workshop 2008 for stimulation and useful discussions. We thank K. Brakke for his development and maintenance of the Surface Evolver code. FG thanks Alexandre Kabla for critical reading of the manuscript and recalling the link between overshoot and bistability. SJC thanks the British Council Alliance programme, CNRS and EPSRC (EP/D048397/1, EP/D071127/1) for financial support and UJF for hospitality during the period in which this work was conceived. CR thanks the Alliance programme for having supported one visit to Aberystwyth University, project 15154XB \emph{Foam rheology in two dimensions}.
|
\section{\bf Introduction}
Since in the Standard Model (SM) the lepton flavor violating (LFV)
interactions are extremely suppressed, any observation of the LFV
processes would serve as a robust evidence for new physics beyond
the SM. These LFV processes, which have been searched in various
experiments \cite{exp2,exp3,exp4-l-lr}, can be greatly enhanced in
new physics models like supersymmetry \cite{LFC-1,rrmutau-susy} and
the topcolor-assisted models (TC2)
\cite{tc2-rev,lfv-tc2,eemutau-tc2}. Such enhancement can be several
orders to make them potentially accessible at future collider
experiments.
Due to its rather clean environment, the proposed International Linear
Collider (ILC) will be an ideal machine to probe new physics. In
such a collider, in addition to $e^+ e^-$ collision, we can also
realize $\gamma \gamma$ collision with the photon beams generated by the backward
Compton scattering of incident electron- and laser-beams. The LFV interactions
in TC2 model will induce various processes at the ILC, such as
the productions of $\tau\bar \mu$, $\tau \bar e$ and $\mu \bar e$
via $e^+ e^-$.
It is noticeable that the productions of $\tau\bar \mu$, $\tau \bar e$ and
$\mu \bar e$ in $\gamma \gamma$ collision have not been studied in
the framework of the TC2. Such LFV productions in $\gamma \gamma$
collision may be more important than in $e^+ e^-$ collision
\cite{eemutau-tc2} collision since these productions are a good
probe for new physics because it is essentially free of any SM
irreducible background.
It is also noticeable that all these LFV
processes at the ILC involve the same part of the parameter space of
the TC2.
Therefore, it is necessary to comparatively study all these processes
to find out which process is best to probe the TC2 model.
We in this work will study the LFV processes $\gamma\gamma \to \ell_i\bar \ell_j$
($i\neq j$ and $\ell_i = e,~\mu~\tau $) induced by the extra $U(1)$ gauge boson
$Z'$ in TC2 models. We calculate the production rates
to figure out if they can reach the sensitivity of the photon-photon collision
of the ILC.
The work is organized as follows. We will briefly discuss the TC2
model in Section II, giving the new couplings which will be involved
in our calculation. In Section III and IV we give the calculation
results and compare with the results in the SUSY. Finally, the
conclusion is given in Section V.
\section{About TC2 model}
There are many kinds of new physics scenarios predicting new particles, which
can lead to significant LFV signals. For example, in the minimal
supersymmetric SM, a large $\nu_\mu - \nu_\tau$ mixing leads to clear LFV
signals in slepton and sneutrino production and in the decays of neutralinos
and charginos into sleptons and sneutrinos at hadron colliders and lepton
colliders \cite{vv-mixing}. The non-universal U(1) gauge bosons $Z'$, which are
prediced by
various specific models beyond the SM, can lead to the large tree-level flavor
changing(FC) couplings. Thus, these new particles may have significant
contributions to some LFV processes \cite{z'couple}.
The key feature of TC2 models \cite{tc2-rev} and flavor-universal
TC2 models \cite{K_1} is that the large top quark mass is
mainly generated by topcolor interactions at a scale of order
1 TeV . The topcolor interactions may be flavor non-universal (as in
TC2 models) or flavor-universal (as in flavor-universal TC2 models).
However, to tilt the chiral condensation in the $t\bar{t}$ direction
and not form a $b\bar{b}$ condensation, all of these models need a
non-universal extended hypercharge group $U(1)$. Thus, the existence
of the extra $U(1)$ gauge bosons $Z^{\prime}$ is predicted. These
new particles treat the third generation fermions (quarks and
leptons) differently from those in the first and second generations,
namely, couple preferentially to the third generation fermions.
After the mass diagonalization from the flavor eigenbasis into the
mass eigenbasis, these new particles lead to tree-level FC
couplings. The flavor-diagonal couplings of the extra $U(1)$ gauge
bosons $Z^{\prime}$ to ordinary fermions, which are related to our
calculation, can be written as \cite{tc2-rev,exp-tc2}:
\begin{equation}
{\cal L}=-\frac{1}{2}g_{1}\{\tan\theta^{\prime}\{(\bar{e}_{L}\gamma^{\mu}
e_{L}+2\bar{e}_{R}\gamma^{\mu}e_{R} + \bar{\mu}_{L}\gamma^{\mu}
\mu_{L}+2\bar{\mu}_{R}\gamma^{\mu}\mu_{R}) +
\cot\theta' (\bar{\tau}_{L}\gamma^{\mu}
\tau_{L}+2\bar{\tau}_{R}\gamma^{\mu}\tau_{R})\}\cdot Z'_{\mu}
\end{equation}
where $g_{1}$ is the ordinary hypercharge gauge coupling constant.
$\theta^{\prime}$ is the mixing angle and
$\tan\theta^{\prime}=\frac{g_{1}}{2\sqrt{\pi K_{1}}}$ where $K_1$ is
the coupling constant.
The flavor-changing couplings of the extra $U(1)$ gauge bosons
$Z^{\prime}$ to ordinary fermions, which are related to our
calculation, are given in the followings: \cite{tc2-rev,exp-tc2}:
\begin{equation}
{\cal L}=-\frac{1}{2}g_{1}\{K_{\mu e}(\bar{e}_{L}\gamma^{\mu}
\mu_{L}+2\bar{e}_{R}\gamma^{\mu}\mu_{R})+k_{\tau\mu}(\bar{\tau}_{L}
\gamma^{\mu}\mu_{L}+2\bar{\tau}_{R}\gamma^{\mu}\mu_{R})
+k_{\tau e}(\bar{\tau}_{L}
\gamma^{\mu}e_{L}+2\bar{\tau}_{R}\gamma^{\mu}e_{R}) \}\cdot Z'_{\mu},
\end{equation}
where $k_{\mu e}$, $k_{\tau e}$ and $k_{\tau\mu}$ are the flavor mixing
factors. Since the new gauge boson $Z'$ couples preferentially to the third
generation, the factor $K_{\mu e}$ are negligibly small, so in the
following estimation, we will neglect the $\mu- e$ mixing, and consider only
the flavor changing coupling processes $\gamma\gamma \to \tau\bar\mu$ and $\gamma\gamma \to
\tau \bar e$
Note that the difference between the $Z'\tau\bar\mu$ and $Z'\tau\bar
e$ couplings lies only in the flavor mixing factor $K_{\tau\mu}$
and $K_{\tau e}$ and the masses of the final state leptons. Since the
non-universal gauge boson $Z'$ treats the fermions in the third generation
differently from those in the first and second generations and treats the
fermions in the first same as those in the second generation, so in the
following calculation, we will assume $K_{\tau\mu} = K_{\tau e}$.
Then what makes the discrepancy of the cross sections of the two
channels $\gamma\gamma \to \tau\bar \mu$ and $\gamma\gamma \to \tau
\bar e$ is only the masses of the final state particles. Considering
the large mass $M_{Z'}> 1$ TeV, for simplicity, we will take
$M_\tau=M_\mu=M_e=0$ in the following discussion, i.e., assuming the
cross sections of the two channels $\gamma\gamma \to \tau\bar \mu$
and $\gamma\gamma \to \tau\bar e$ are equal to
each other.
\section{Calculation}
The Feynman diagram of the LFV processes $ \gamma\gamma \to
\ell_i\ell_j$ ($i\neq j$ and $\ell_i = e,~\mu,~\tau$) induced by the
extra U(1) $Z'$ is shown in
Fig.~\ref{fig1}. There are only t- and u- channel contributions,
the latter not shown in Fig~\ref{fig1}, but We can calculate them by
exchanging the two photons.
\begin{figure}[tbh]
\begin{center}
\epsfig{file=fig1.ps,width=8cm} \caption{ Feynman diagrams
contributing to the process $\gamma\gamma \to \ell_i\ell_j$ in TC2
models.} \label{fig1}
\end{center}
\end{figure}
Note that there is no s-channel contribution to the LFV processes.
As we know, in the SM production of on-shell $Z$ boson at a
photon-photon collider (or $Z$ decays into $\gamma\gamma$) is
strictly forbidden by angular momentum conservation and Bose
statistics, which is the predict of the famous Laudau-Yang Theorem.
This theorem is still effective to our case, since that the two real
photons cannot be in a state with angular momentum $J=1$ regardless
of on-shell or off-shell bosons, so the s-channel contribution with
two real photons to an extra $Z'$ vanishes automatically.
The electroweak gauge bosons $\gamma$ and $Z$ can not couple to
$\tau\bar e$, $\mu\bar e$ and $\tau\bar \mu$, so we need not consider
the interference effects between the $\gamma$, $Z$ and $Z^{\prime}$
on the cross section of the process $\gamma\gamma \to \ell_i
\bar\ell_j$($i\neq j$ and $\ell_i = e,~\mu,~\tau$). In TC2 models
the gauge invariant amplitude of $\gamma\gamma \to \tau\bar\mu(\bar
e)$ induced by the extra boson $Z'$ is given by
\begin{eqnarray}
{\cal M}=\frac{1}{2} ~\bar{u}_\tau\Gamma^{\mu\nu}P_Lv_\mu
~\epsilon_\mu(\lambda_1)\epsilon_\nu(\lambda_2)
\end{eqnarray}
where the $\Gamma^{\mu\nu}$ is defined same as that in
\cite{rrtc}. These amplitudes contain the Passarino-Veltman
one-loop functions, which are calculated by using LoopTools
\cite{Hahn}.
Since the photon beams in $\gamma\gamma$ collision are generated
by the backward Compton scattering of the incident electron- and
the laser-beam, the events number is obtained by convoluting the
cross section of $\gamma\gamma$ collision with the photon beam
luminosity distribution:
\begin{eqnarray}
N_{\gamma \gamma \to \ell_i \bar\ell_j}&=&\int d\sqrt{s_{\gamma\gamma}}
\frac{d\cal L_{\gamma\gamma}}{d\sqrt{s_{\gamma\gamma}}}
\hat{\sigma}_{\gamma \gamma \to \ell_i \bar\ell_j}(s_{\gamma\gamma})
\equiv{\cal L}_{e^{+}e^{-}}\sigma_{\gamma \gamma \to \ell_{i} \bar\ell_{j}}(s)
\end{eqnarray}
where $d{\cal L}_{\gamma\gamma}$/$d\sqrt{s}_{\gamma\gamma}$ is the photon-beam luminosity
distribution and $\sigma_{\gamma \gamma \to \ell_i \bar\ell_j}(s)$ (
$s$ is the squared center-of-mass energy of $e^{+}e^{-}$ collision) is defined as
the effective cross section of $\gamma \gamma \to \ell_{i} \bar\ell_{j}$.
In optimum case, it can be written as \cite{photon collider}
\begin{eqnarray}
\sigma_{\gamma \gamma \to \ell_i \bar\ell_j}(s)&=&
\int_{\sqrt{a}}^{x_{max}}2zdz\hat{\sigma}_{\gamma \gamma \to \ell_{i} \bar\ell_{j}}
(s_{\gamma\gamma}=z^2s) \int_{z^{2/x_{max}}}^{x_{max}}\frac{dx}{x}
F_{\gamma/e}(x)F_{\gamma/e}(\frac{z^{2}}{x})
\end{eqnarray}
where $F_{\gamma/e}$ denotes the energy spectrum of the back-scattered photon for the
unpolarized initial electron and laser photon beams given by
\begin{eqnarray}
F_{\gamma/e}(x)&=&\frac{1}{D(\xi)}\left[1-x+\frac{1}{1-x}-\frac{4x}{\xi(1-x)}
+\frac{4x^{2}}{\xi^{2}(1-x)^{2}}\right]
\end{eqnarray}
with
\begin{eqnarray}
D(\xi)&=&(1-\frac{4}{\xi}-\frac{8}{\xi^{2}})\ln(1+\xi)
+\frac{1}{2}+\frac{8}{\xi}-\frac{1}{2(1+\xi)^{2}}.
\end{eqnarray}
Here $\xi=4E_{e}E_{0}/m_{e}^{2}$ ($E_{e}$ is the incident electron
energy and $E_{0}$ is the initial laser photon energy) and
$x=E/E_{E}$ with $E$ being the energy of the scattered photon moving
along the initial electron direction. The definitions of parameters
$\xi$, $D(\xi)$ and $x_{max}$ can be found in Ref.\cite{photon
collider}. In our numerical calculation, we choose $\xi=4.8$,
$D(\xi)=1.83$ and $x_{max}=0.83$.
\section{Numerical Results and Discussions}
As for the involved SM parameter, we take \cite{pdg}
\begin{eqnarray}
m_{\mu}=0.106{\rm ~GeV}, m_{\tau}=1.777{\rm ~GeV}, m_{b}=4.2{\rm ~GeV},
\alpha=1/137,\sin^2\theta_W=0.223
\end{eqnarray}
The TC2 parameters concerned in this process are $K_{\tau e}$,
$K_{\tau\mu}$, $K_{e\mu}$, $K_1$ and the mass of the extra gauge
boson $M_Z'$. $K_{e\mu}$ is very small, about $10^{-3}$, we will not
consider the $e-\mu$ conversion processes. In our calculation, we
have assumed $K_{\tau\mu}=K_{\tau e} \simeq \lambda \simeq 0.22$
\cite{exp-tc2,tc2-cla}, which $\lambda$ is the Wolfenstein
parameter \cite{Wolfenstein}. It has been shown that the vacuum
tilting (the topcolor interactions only condense the top quark but
not the bottom quark), the coupling constant $K_{1}$ should satisfy
certain constraint, i.e. $K_{1}\leq 1$ \cite{K_1}. The limits on the
$Z'$ mass $M_Z'$ can be obtained via studying its effects on various
experimental observables \cite{exp-tc2}. Ref.\cite{m_z'}, for
example, has been shown that to fit the electroweak mearsurement
data, the $Z'$ mass $M_Z'$ must be larger than $1$
TeV. As numerical estimation, we choose the center-of-mass
energy $\sqrt{s}=500$ and $1000$ GeV, to observe the different
behavior in the two energy area, and take the $M_Z'$ and $K_1$ as
free parameters. Finally, Note that the charge conjugate $\bar \tau
\mu(e)$ production channel are also included in our numerical study.
\def\figsubcap#1{\par\noindent\centering\footnotesize(#1)}
\begin{figure}[bht
\label{modes}
\begin{center}
\hspace{-0.25cm}
\parbox{6.05cm}{\label{modes}\epsfig{figure=fig2-a.ps,width=6.25cm}
\figsubcap{a}\label{modes}}
\hspace*{0.2cm}
\parbox{6.05cm}{\epsfig{figure=fig2-b.ps,width=6.25cm}
\figsubcap{b}\label{masses}}
\caption{ The cross section $\sigma$ of the LFV process $\gamma\gamma
\to\tau\bar\mu(\bar e)$ as a function of the gauge boson $Z'$
mass $M_{Z'}$ for $K_{1}=0.2$, $0.6$ and $1.0$ with
(a) $\sqrt{s}=500GeV$ (b)$\sqrt{s}=1000GeV$. \label{fig2} }
\end{center}
\end{figure}
\def\figsubcap#1{\par\noindent\centering\footnotesize(#1)}
\begin{figure}[bht
\label{modes}
\begin{center}
\hspace{-0.25cm}
\parbox{6.05cm}{\label{modes}\epsfig{figure=fig3-a.ps,width=6.25cm}
\figsubcap{a}\label{modes}}
\hspace*{0.2cm}
\parbox{6.05cm}{\epsfig{figure=fig3-b.ps,width=6.25cm}
\figsubcap{b}\label{masses}}
\caption{The cross section $\sigma$ of the LFV process $\gamma\gamma
\to\tau\bar\mu(\bar e)$ as a function of the parameter $K_1$
with the gauge boson $Z'$ mass $M_{Z'}=1$ , $1.5$ and $2.5$ TeV for
(a) $\sqrt{s}=500GeV$ (b)$\sqrt{s}=1000GeV$. \label{fig3} }
\label{modmass}
\end{center}
\end{figure}
In Fig.~\ref{fig2} we plot the production cross section $\sigma$ of
the LFV process $\gamma\gamma \to \ell_i \bar\ell_j$ as a function
of $M_{Z}$ for three values of the parameter $K_{1}$: $K_{1}=0.2$,
$0.6$, and $1.0$. We can see from Fig.2 that the production cross
section $\sigma$ increases as $K_{1}$ increasing and strongly
suppressed by large $M_{Z'}$. This situation
is slightly different from
the result of $e^+e^-\to \ell_i \bar\ell_j$ in \cite{eemutau-tc2},
in which from Fig.1 we can see the cross section of $e^+e^-\to
\bar\mu\tau$ increases with $K_1$ decreasing. The reason is that
the $Z'\tau\bar\tau$ coupling involved in the process $\gamma\gamma
\to \tau\ell_i$($\ell_i=e$ or $\mu $) is proportional to
$1/\tan\theta^\prime\sim \sqrt{K_1}$, while the $Z'e^+e^-$ contains
$\tan\theta^\prime\sim \frac{1}{\sqrt{K_1}}$ and $\tan\theta'<<1$.
We can feel from this point the spirit of the technicolor models: to
give the natural top quark mass, the third generation is singled out
from the former two ones, so that it always shows distinct features.
The background for $\gamma \gamma \to e \bar{\tau}$ comes from
$\gamma\gamma \to \tau^{+}\tau^{-} \to
\tau^{-}\nu_{e}\bar{\nu}_{\tau}e^{+}$, $\gamma\gamma \to W^{+}W^{-}
\to \tau^{-}\nu_{e}\bar{\nu}_{\tau}e^{+}$ and $\gamma\gamma \to
e^{+}e^{-}\tau^{+}\tau^{-}$ \cite{rrmutau-susy}, and we make
kinematical cuts \cite{l-prod-ILC}: $|\cos\theta_\ell|<0.9$ and
$p^{\ell}_{T}>20{\rm ~GeV}$ ($\ell=e,\mu$), to enhance the ratio of
signal to background. With these cuts, the background cross sections
from $\gamma\gamma \to \tau^{+}\tau^{-} \to
\tau^{-}\nu_{e}\bar{\nu}_{\tau}e^{+}$, $\gamma\gamma \to W^{+}W^{-}
\to \tau^{-}\nu_{e}\bar{\nu}_{\tau}e^{+}$ and $\gamma\gamma \to
e^{+}e^{-}\tau^{+}\tau^{-}$ at $\sqrt{s}=500$ GeV are suppressed
respectively to $9.7\times 10^{-4}$ fb, $1.0\times 10^{-1}$ fb and
$2.4\times 10^{-2}$ fb (see Table I of \cite{l-prod-ILC}). To get
the $3 \sigma$ observing sensitivity with $3.45 \times 10^2$
fb$^{-1}$ integrated luminosity \cite{tesla}, the production rates
of $\gamma \gamma \to \tau\bar{e}, \tau \bar{\mu}$ after the cuts
must be larger than $2.5\times 10^{-2}$ fb \cite{l-prod-ILC}. We see
from Fig.\ref{fig2} that under the current bounds from $\ell_i \to
\ell_j \gamma$\cite{exp4-l-lr} and $\mu \to 3 e$\cite{z'couple}, the
LFV couplings in TC2 models can still large enough to enhance the
productions $\gamma\gamma \to e\bar{\tau}, \mu \bar{\tau}$ to the $3
\sigma$ sensitivity and may be detected in the future ILC colliders.
Finally note that we in Fig.2 only show the results of the channels
with the $\tau$ lepton in the final states, i.e., $\gamma\gamma \to
\tau\bar{\mu}$, $\tau\bar{e}$.
Fig.3 shows that the cross section of the LFV processes as a
function of $K_1$ for $ M_Z'= 1$, $1.5$,
$2.5$ TeV. We can see more clearly that the cross
section is increasing as the $K_1$ increasing.
\def\figsubcap#1{\par\noindent\centering\footnotesize(#1)}
\begin{figure}[bht
\label{modes}
\begin{center}
\hspace{-0.25cm}
\parbox{6.05cm}{\epsfig{figure=fig4-a.ps,width=6.25cm}
\figsubcap{a}}
\hspace*{0.2cm}
\parbox{6.05cm}{\epsfig{figure=fig4-b.ps,width=6.25cm}
\figsubcap{b}}
\caption{The dependence of the cross section $\sigma$ of the LFV process
$\gamma\gamma \to\tau\bar\mu(\bar e)$ on the center-of-mass energy
$\sqrt{s}$ for $M_{Z'}=1$, $1.5$, and $2.5$ TeV with (a)
$k_{1}=0.2$, (b) $k_{1}=0.6$. \label{fig4} }
\end{center}
\end{figure}
We also show the cross sections of $\gamma \gamma \to \ell_i
\bar\ell_j$ as a function of center-of-mass energy $\sqrt{s}$ of the
ILC in Fig.4. We see that with the increasing of the center-of-mass
energy, the cross sections of these processes are not compressed,
instead of becoming larger. This is different with the results in
\cite{eemutau-tc2}, since, as mentioned above, the contribution of
the $\gamma\gamma \to \ell_i\bar\ell_j$ are the the results of t-
and u-channels, while in the processes $e^+e^- \to
\ell_i\bar\ell_j$, the s-channel contribution
decreases with the increasing $\sqrt{s}$ when the center-of-mass
energy of the processes arrives at the critical value\cite{eemutau-tc2}.
Actually, we can also feel the larger cross section with larger
center of mass from Fig.2 and Fig.3.
\null \noindent {\small Table I: Theoretical predictions for the
$\ell_i\bar\ell_j$ ($i\neq j$) productions at $\gamma\gamma$
collision at the ILC. SUSY and TC2 predictions are the optimum
values. The collider energy is $500$ GeV.} \vspace*{0.1cm}
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
&~~SUSY ~~&~~~~TC2~~ \\
\hline ~~$\sigma(\gamma\gamma \to \tau\bar\mu)$ &~~${\cal O}
(10^{-2})$ fb & ~~~~$1$ fb \\ \hline ~~$\sigma(\gamma\gamma \to
\tau\bar e)$ &~~${\cal O} (<10^{-1})$ fb & ~~~~$1$ fb\\ \hline
~~$\sigma(\gamma\gamma \to \mu\bar e)$ ~~~~ &~~${\cal O} (<10^{-3})$
fb ~~~~& ~~~~$10^{-3}$ fb~~~~\\ \hline
\end{tabular}
\end{center}
As discussed in the former sections, motivated by the fact that
any process that is forbidden or strongly suppressed within the SM
constitutes a natural laboratory to search for any new physics
effects, LFV processes have been the subject of considerable
interest in the literature. It turns out that they may have large
cross sections, much larger than the SM ones, within some extended
theories such as the R-parity violating MSSM \cite{rrmutau-susy} and
the TC2 models. However, in the R-parity violating MSSM, as
discussed in \cite{rrmutau-susy},
the LFV coupling by the exchange of the squark
is $\lambda_{ijk}\sim 10^{-2}$, much smaller than that of the TC2
models ($ K_{\tau\mu(e)}\sim~ 0.2$). Therefore we can evaluate that
in the SUSY models the sigma of the LFV process $\gamma\gamma \to
\ell_i\ell_j$ is about $2-3$ order smaller than that in the TC2
models, as shown in table. I.
\section{Conclusion}
We have performed an analysis for the TC2-induced LFV productions of
$\tau\bar\mu$ and $\tau \bar e$ via $\gamma \gamma$ collision at the
ILC. We found that in the optimum part of the parameter space, the
production rate of $\gamma \gamma \to \tau\bar{\mu}(\bar e)$ can
reach $1$ fb. This means that we may have $100$ events each year for
the designed luminosity of $100$ fb$^{-1}$/year at the ILC. Since
the SM value of the production rate is completely negligible, the
observation of such $ \tau\bar{\mu}(\bar e)$ events would be a
robust evidence of TC2. Therefore, these LFV processes may serve as
a sensitive probe of TC2.
\vspace{2.5mm}
{\bf \large Acknowledgement}
\vspace{2.5mm}
We would like to thank J. J. Cao, J. M. Yang
and C. P. Yuan for helpful discussion.
|
\section{Introduction}
Classification of objects is typically the first step towards scientific understanding, since it brings order to a previously unorganised set of observational data and provides standardised terms to describe objects. These standardised terms are usually qualitative, but they can also be quantitative which makes them accessible for mathematical analysis. A famous example of a successful classification from the field of astrophysics is the Hertzsprung-Russell diagram, where stars exhibit distinct groups in the colour-magnitude diagram that represent their different evolutionary stages. For the same reason, galaxy classification is an important conceptual step towards understanding physical properties, formation and evolution scenarios of galaxies.
With the advent of modern sky surveys containing millions (e.g. SDSS, COSMOS, PanSTARRS, GAMA) or even billions (e.g. LSST) of galaxies, the classification of these galaxies is becoming more and more problematic. The vast amount of data excludes the hitherto common practice of visual classification and clearly calls for an automated classification scheme that is more efficient and more objective. In this work, we present an algorithm for automated and probabilistic classification, where the classes are discovered automatically, too. However, the intention of this work is not to come up with ''yet another morphological classification scheme'', but rather to demonstrate of how it could be done alternatively to the standard practice of classification in astrophysics. Besides, we are unable to present a full solution to the problem of morphological galaxy classification, since there is still no accepted method for parametrising arbitrary galaxy morphologies \citep[cf.][]{Andrae2010b}. In addition to the lack of convincing classification schemes, this is why many experts are very sceptical about the subject of classifying galaxy morphologies as a whole. As parametrisation of galaxy spectra is more reliable, spectral classifications have become more accepted.\\
\\*
In the remaining part of this introduction, we first give an overview about modern automated classification methods and work out a categorisation of these methods. We describe our parametrisation of galaxy morphologies using shapelets \citep{Refregier2003} in Sect. \ref{sect:shapelets}. In Sect. \ref{sect:soft_clustering_algorithm} we present the algorithm we are using, which has been introduced before by \citet{Yu2005} in the field of pattern recognition. We extensively investigate the reliability of this classification algorithm in Sect. \ref{sect:systematic_tests}. Such a study has not been undertaken by \citet{Yu2005}. In Sect. \ref{sect:worked_example_SDSS} we present a worked example with a small sample of 1,520 bright galaxies from the SDSS. The objects in this sample are selected such that no practical problems with parametrisation arise, as we want to disentangle the problems of classification and parametrisation as much as possible. The aim of this worked example is \textit{not} to do science with the resulting classes or data-to-class assignments, but to demonstrate that such an algorithm indeed produces reasonable results. We conclude in Sect. \ref{sect:conclusions}.
\subsection{Overview about classification methods}
In Table \ref{tab:summary_classification_algorithms} we give an overview of different classification methods and some example algorithms. The two criteria for this categorisation are:
\begin{enumerate}
\item Is the data-to-class assignment probabilistic (soft) or not (hard)?
\item Are the classes specified a priori (classification) or discovered automatically (clustering)?
\end{enumerate}
Not all algorithms fit into this categorisation, namely those that do not directly assign classes to objects (e.g. self-organising maps).
\begin{table}
\begin{tabular}{ccc}
\hline\hline
Type & Classification & Clustering \\
\hline
Hard & nearest neighbour, & $K$-means, \\
& Fisher's linear & spectral clustering, \\
& discriminant analysis & kernel PCA \\
\hline
Soft & na\"ive Bayes, & Gaussian mixture \\
& linear/quadratic & models \\
& discriminant analysis, & \\
& neural networks & \\
\hline
\end{tabular}
\caption{Overview of different classification and clustering algorithms with examples.\newline Soft (probabilistic) algorithms are always model-based, whereas hard algorithms are not necessarily. Soft algorithms can always be turned into hard algorithms, but not vice-versa. The list of example algorithms is not complete.}
\label{tab:summary_classification_algorithms}
\end{table}
The algorithm we are going to present is a soft algorithm, i.e. the data-to-class assignment is probabilistic (cf. next section). The reason is that in the case of galaxy morphologies, it is obvious that the classes will \textit{not} be clearly separable. We rather expect the galaxies to be more or less homogeneously distributed in some parameter space, with the classes appearing as local overdensities and exhibiting potentially strong overlap. As we demonstrate in Sect. \ref{sect:hart_cuts_bias_means}, hard algorithms break down in this case, producing biased classification results. There are physical reasons to expect overlapping classes: First, the random inclination and orientation angles w.r.t. the line of sight induce a continuous transition of apparent axis ratios, apparent steepness of the radial light profiles and ratio of light coming from bulge and disk components. Second, observations of galaxies show that there are indeed transitional objects between different morphological types. For instance, there are transitional objects between early- and late-type galaxies in the ''green valley'' of the colour bimodality \citep[e.g.][]{Strateva2001,Baldry2004}, which is also reproduced in simulations \citep{Croton2006}. Hence, we have to draw the conclusion that hard algorithms are \textit{generically inappropriate} for analysing galaxy morphologies. This conclusion is backed up by practical experience, since even various specialists usually do not agree in hard visual classifications \citep[e.g.][]{Bamford2009}. In fact, the outcome of multi-person visual classifications becomes a probability distribution automatically.
Furthermore, our algorithm is a clustering algorithm, i.e. we do not specify the morphological classes a priori, but let the algorithm discover them. This approach is called ''unsupervised learning'' and it is the method of choice if we are uncertain about the type of objects we will find in a given data sample. In the context of clustering analysis classes are referred to as \textit{clusters}, and we adopt this terminology in this article.
\subsection{Probabilistic data-to-class assignment}
Let $O$ denote an object and $\vec x$ its parametrisation. Furthermore, let $c_k$ denote a single class out of $k=1,\ldots,K$ possible classes, then $\textrm{prob}(c_k|\vec x)$ denotes the probability of class $c_k$ given the object $O$ represented by $\vec x$. This conditional probability $\textrm{prob}(c_k|\vec x)$ is called the \textit{class posterior} and is computed using Bayes' theorem
\begin{equation}\label{eq:def_cluster_posterior}
\textrm{prob}(c_k|\vec x) = \frac{\textrm{prob}(c_k)\,\textrm{prob}(\vec x|c_k)}{\textrm{prob}(\vec x)} \;\textrm{.}
\end{equation}
The marginal probability $\textrm{prob}(c_k)$ is called \textit{class prior} and $\textrm{prob}(\vec x|c_k)$ is called \textit{class likelihood}. The denominator $\textrm{prob}(\vec x)$ acts as a normalisation factor. The class prior and likelihood are obtained from a generative model (Sect. \ref{sect:bipartite_graph_model}). Prior and posterior satisfy the following obvious normalisation constraints
\begin{equation}\label{eq:normalisation_of_cluster_posteriors_priors}
\sum_{k=1}^K \textrm{prob}(c_k) = 1 \quad\textrm{and}\quad \sum_{k=1}^K \textrm{prob}(c_k|\vec x) = 1 \;\textrm{,}
\end{equation}
which ensure that each object is definitely assigned to some class. In the case of hard assignments, both posterior $\textrm{prob}(c_k|\vec x)$ and likelihood $\textrm{prob}(\vec x|c_k)$ are replaced by Kronecker symbols.
\section{Parametrising galaxy morphologies with shapelets \label{sect:shapelets}}
\subsection{Basis functions and expansion}
We parametrise galaxy morphologies in terms of shapelets \citep{Refregier2003}. Shapelets are a scaled version of two-dimensional Gauss-Hermite polynomials and form a set of complete basis functions that are orthonormal on the interval $[-\infty,\infty]$. A given galaxy image $I(\vec x)$ can be decomposed into a linear superposition of basis functions $B_{m,n}(\vec x/\beta)$, i.e.
\begin{equation}\label{eq:linear_superposition}
I(\vec x) = \sum_{m,n=0}^\infty c_{m,n}B_{m,n}(\vec x/\beta) \,\textrm{,}
\end{equation}
where the $c_{m,n}$ denote the expansion coefficients that contain the morphological information and $\beta>0$ denotes a scaling radius. In practice, the number of basis functions we can use is limited by pixel noise, such that the summation in Eq. (\ref{eq:linear_superposition}) stops at a certain maximum order $N_\textrm{max}<\infty$ which depends on the object's signal-to-noise ratio and resolution. This means Eq. (\ref{eq:linear_superposition}) is an approximation only,
\begin{equation}\label{eq:approximated_linear_superposition}
I(\vec x) \approx \sum_{m,n=0}^{N_\textrm{max}} c_{m,n}B_{m,n}(\vec x/\beta) \,\textrm{.}
\end{equation}
We use the C++ algorithm by \citet{Melchior2007} to estimate $N_\textrm{max}$, the scale radius and the linear coefficients, which was shown to be faster and more accurate than the IDL algorithm by \citet{Massey2005}.
\subsection{Problems with shapelet modelling}
It was shown by \citet{Melchior2009a} that the limitation of the number of basis functions in Eq. (\ref{eq:approximated_linear_superposition}) can lead to severe modelling failures and misestimations of galaxy shapes in case of objects with low signal-to-noise ratios. They identified two origins of these biases: First, the Gaussian profile of shapelets does not match the true profiles of galaxies, which are typically much steeper. Second, the shapelet basis functions are intrinsically spherical, i.e. they have problems in modelling highly eccentric objects. However, in this demonstration we consider only galaxies with high signal-to-noise ratios, where we can use many basis functions such that the impact of these biases is negligible. We demonstrate this in Fig. \ref{fig:example_shapelet_models_of_3_SDSS_galaxies}, where we show the shapelet reconstructions of a face-on disk, an edge-on disk and an elliptical galaxy drawn from the sample presented in Sect. \ref{sect:analysis_small_set}. The reconstruction of the face-on disk galaxy (top row) is excellent, leaving essentially uncorrelated noise in the residuals. However, the reconstructions of the edge-on disk galaxy (centre row) and the elliptical galaxy (bottom row) exhibit ring-like artefacts that originate from the steep light profiles of the elliptical and the edge-on disk along the minor axis. Such modelling failures appear systematically and do \textit{not} introduce additional scatter into the results, i.e. similar galaxies are affected in a similar way. However, since shapelet models do not capture steep and strongly elliptical galaxies very well, we are aware that our algorithm has less dicriminatory power for galaxies of this kind.
\begin{figure*}
\includegraphics[width=5.5cm]{FON_image}
\includegraphics[width=5.5cm]{FON_model}
\includegraphics[width=5.5cm]{FON_residuals}
\\*
\includegraphics[width=5.5cm]{EON_image}
\includegraphics[width=5.5cm]{EON_model}
\includegraphics[width=5.5cm]{EON_residuals}
\\*
\includegraphics[width=5.5cm]{ELL_image}
\includegraphics[width=5.5cm]{ELL_model}
\includegraphics[width=5.5cm]{ELL_residuals}
\caption{Examples of shapelet models of three galaxies from SDSS ($g$ band).}
Shown are the original images (left column), the shapelet models (centre column) and the residuals (right column) of a face-on disk galaxy (top row), an edge-on disk galaxy (centre row) and an elliptical galaxy (bottom row). Note the different plot ranges of the residual maps. The shapelet decomposition used $N_\textrm{max}=16$, i.e. 153 basis functions.
\label{fig:example_shapelet_models_of_3_SDSS_galaxies}
\end{figure*}
\subsection{Distances in shapelet space}
The coefficients form a vector space and we denote them as vectors $\vec x$. In first-order approximation, these coefficient vectors are independent of the size of the object, which was encoded by the scale radius $\beta$. Moreover, we can also make $\vec x$ invariant against the image flux, since Eq. (\ref{eq:linear_superposition}) implies that for a constant scalar $\alpha\neq 0$ the transformation $\vec x\rightarrow \alpha\vec x$ changes the image flux by this same factor of $\alpha$. Therefore, if we demand $\vec x\cdot\vec x=1$, then differing image fluxes will have no impact on the shapelet coefficients. This implies that morphologies are a \textit{direction} in shapelet coefficient space and the corresponding coefficient vectors lie on the surface of a hypersphere with unit radius. We can thus measure distances between morphologies on this surface via the angle spanned by their (normalised) coefficient vectors,
\begin{equation}\label{eq:def:spherical_distance}
d(\vec x_1,\vec x_2)=\sphericalangle\left(\vec x_1,\vec x_2\right) = \arccos\left(\vec x_1\cdot\vec x_2\right) \,\textrm{.}
\end{equation}
Employing the polar representation of shapelets \citep{Massey2005}, we can apply rotations and parity flips to shapelet models. We can estimate the object's orientation angle from the second moments of its light distribution \citep[e.g.][]{Melchior2007} and then use this estimate to align all models. This ensures invariance of the coefficients against random orientations. Additionally, we can break the degeneracy between left- and right-handed morphologies by applying parity flips such that the distance of two objects is minimised. These transformations in model space do not suffer from pixellation errors and increase the local density of similar objects in shapelet space.
\section{Soft Clustering Algorithm\label{sect:soft_clustering_algorithm}}
We now present the soft clustering algorithm of \citet{Yu2005}. Before we explain the details, we want to give a brief outline of the general method. The basic idea is to assign similarities to pairs of objects, so we first explain how to measure similarities of galaxy morphologies and what a similarity matrix is. These pairwise similarities are then interpreted by a probabilistic model, which provides our generative model. We also present the algorithm that fits the model to the similarity matrix.
\subsection{Estimating similarities \label{sect:optimal_similarity_measure}}
Instead of analysing the data in shapelet space, we compute a \textit{similarity matrix} by assigning similarities to any two data points. This approach is an alternative to working directly in the sparsely populated shapelet space or employing a method for dimensionality reduction. If we have $N$ data points $\vec x_n$, then this similarity matrix will be an $N\times N$ symmetric matrix. It is this similarity matrix to which we are going to apply the soft clustering analysis.
Based on the pairwise distances in shapelet coefficient space (Eq. (\ref{eq:def:spherical_distance})), we estimate pairwise similarities up to a constant factor as
\begin{equation}\label{eq:def:similarity_measure}
W_{mn} \propto 1 - \frac{(d(\vec x_m,\vec x_n)/d_\textrm{max})^\alpha}{s} \;\textrm{.}
\end{equation}
Here $d_\textrm{max}$ denotes the maximum distance between any two objects in the given data sample, while the exponent $\alpha>0$ and $s>1$ are free parameters that tune the similarity measure. We explain how to choose $\alpha$ and $s$ in Sect. \ref{sect:impact_non-optimal_simi}. This definition ensures that $0< W_{mn}\leq 1$ and that the maximum similarities are self-similarities for which $d(\vec x_m,\vec x_m)=0$. Note that this similarity measure is invariant under changes of size, flux, orientation, and parity of the galaxy morphology.
\subsection{Similarity matrices and weighted undirected graphs}
Square symmetric similarity matrices have a very intuitive interpretation: They represent a weighted undirected graph. Figure \ref{fig:similarity_matrix_as_graph} shows a sketch of such a graph. The data points $\vec x_n$ are represented symbolically as nodes $x_n$. The positions of these nodes are usually arbitrary, it is neither necessary nor helpful to arrange them according to the true locations of the data points in parameter space. Any two data nodes $x_m$ and $x_n$ are connected by an edge, which is assigned a weight $W_{mn}$. Obviously, all the weights $W_{mn}$ form an $N\times N$ matrix $ W$, and if this matrix is symmetric, i.e. $W_{mn}=W_{nm}$, the edges will have no preferred direction. In this case, the weighted graph is undirected. In graph theory the matrix of weights $ W$ is called \textit{adjacency matrix}, and we can interpret the similarity matrix as adjacency matrix of a weighted undirected graph.
\begin{figure}\centering
\begin{minipage}[t]{0.45\textwidth}
\begin{minipage}[t]{1.0\textwidth}\centering
\includegraphics[width=0.75\textwidth]{similarity_matrix_as_graph}
\end{minipage}
\caption{Sketch of a weighted undirected graph.}
\label{fig:similarity_matrix_as_graph}
The data nodes $x_n$ are connected by edges. For the sake of visibility, only edges connecting $x_1$ are shown. The edges are undirected and weighted by the similarity of the two connected nodes.
\end{minipage}
\end{figure}
Inspecting Fig. \ref{fig:similarity_matrix_as_graph}, we now introduce some important concepts. First, we note that there is also an edge connecting $x_1$ with itself. This edge is weighted by the ''self-similarity'' $W_{11}$. These self-similarities $W_{nn}$ are usually non-zero and have to be taken into account in order to satisfy normalisation constraints (cf. Eq. (\ref{eq:normalisation_of_W})). Second, we define the \textit{degree} $d_n$ of a data node $x_n$ as the sum of weights of all edges connected with $x_n$, i.e.
\begin{equation}
d_n = \sum_{m=1}^N W_{mn} \;\textrm{.}
\end{equation}
We can interpret the degree $d_n$ to measure the connectivity of data node $x_n$ in the graph. For instance, we can detect outlier objects by their low degree, since they are very dissimilar to all other objects. Third, we note that we can rescale all similarities by a constant scalar factor $C>0$ without changing the pairwise relations. Hence, we acquire the normalisation constraint
\begin{equation}\label{eq:normalisation_of_W}
\sum_{m,n=1}^N W_{mn} = \sum_{n=1}^N d_n = 1 \;\textrm{.}
\end{equation}
This constraint ensures the normalisation of the probabilistic model we are going to set up for our soft clustering analysis of the similarity matrix.
\subsection{Bipartite-graph model \label{sect:bipartite_graph_model}}
We need a probabilistic model of the similarity matrix $ W$ that can be interpreted in terms of a soft clustering analysis. Such a model was proposed by \citet{Yu2005}. As similarity matrices are closely related to graphs, this model is motivated from graph theory, too. The basic idea of this model is that the similarity of two data points $\vec x_m$ and $\vec x_n$ is induced by both objects being members of the same clusters. This is the basic hypothesis of any classification approach: Objects from the same class are more similar than objects from different classes.
\begin{figure}\centering
\begin{minipage}[t]{0.45\textwidth}
\begin{minipage}[t]{1.0\textwidth}\centering
\includegraphics[width=0.5\textwidth]{bipartite_graph_sketch}
\end{minipage}
\caption{Sketch of a bipartite graph.}
The bipartite graph contains two sets of nodes, $\mathcal X=\{x_1,\ldots,x_N\}$ and $\mathcal C=\{c_1,\ldots,c_K\}$. Edges connect nodes from different sets only and are weighted by an adjacency matrix $ B$. Not all edges are shown.
\label{fig:bipartite_graph_sketch}
\end{minipage}
\end{figure}
In detail, we model a weighted undirected graph (Fig. \ref{fig:similarity_matrix_as_graph}) and its similarity matrix by a \textit{bipartite graph} (Fig. \ref{fig:bipartite_graph_sketch}). A bipartite graph is a graph whose nodes can be divided into two disjoint sets $\mathcal X=\{x_1,\ldots,x_N\}$ of data nodes and $\mathcal C=\{c_1,\ldots,c_K\}$ of cluster nodes, such that the edges in the graph only connect nodes from different sets. Again, the edges are weighted and undirected, where the weights $B_{nk}$ form an $N\times K$ rectangular matrix $ B$, the bipartite-graph adjacency matrix. The bipartite-graph model for the similarity matrix then reads
\begin{equation}\label{eq:bipartite_graph_model}
\hat W_{mn} = \sum_{k=1}^K \frac{B_{nk}B_{mk}}{\lambda_k} \;\textrm{,}
\end{equation}
with the cluster priors $\lambda_k = \sum_{n=1}^N B_{nk}$. A detailed derivation is given in the following section. This model induces the pairwise similarities via two-hop transitions $\mathcal X\rightarrow\mathcal C\rightarrow\mathcal X$ \citep[cf.][]{Yu2005}. The numerator accounts for the strength of the connections of both data nodes to a certain cluster. The impact of the denominator is that the common membership to a cluster of small degree is considered more decisive. Obviously, the model defined by Eq. (\ref{eq:bipartite_graph_model}) is symmetric, as the similarity matrix itself. The normalisation constraint on $ W$ as given by Eq. (\ref{eq:normalisation_of_W}) translates via the bipartite-graph model to
\begin{equation}\label{eq:normalisation_B_and_lambda_k}
\sum_{k=1}^K \sum_{n=1}^N B_{nk} = \sum_{k=1}^K \lambda_k = 1 \;\textrm{.}
\end{equation}
These constraints need to be respected by the fit algorithm. Having fitted the bipartite-graph model to the given data similarity matrix, we compute the cluster posterior probabilities
\begin{equation}\label{eq:cluster_posteriors_bipartite-graph_model}
\textrm{prob}(c_k|x_n) = \frac{\textrm{prob}(x_n,c_k)}{\textrm{prob}(x_n)} = \frac{B_{nk}}{\sum_{l=1}^K B_{nl}} \;\textrm{,}
\end{equation}
which are the desired soft data-to-cluster assignments. Obviously, $K$ cluster posteriors are assigned to each data node $x_n$ and the normalisation constraint $\sum_{k=1}^K\textrm{prob}(c_k|x_n)=1$ is satisfied.
\subsection{Mathematical derivation \label{sect:deviation_bgm}}
Here we give a derivation of the bipartite-graph model of Eq. (\ref{eq:bipartite_graph_model}) that is more detailed than in \citet{Yu2005}. The ansatz is to identify the similarity $\hat W_{mn}$ with the joint probability
\begin{equation}\label{eq:W_as_joint_probability}
\hat W_{mn} = \textrm{prob}(x_m,x_n) \;\textrm{.}
\end{equation}
This interprets $\hat W_{mn}$ as the probability to find $x_m$ and $x_n$ in the same cluster. Eq. (\ref{eq:normalisation_of_W}) ensures the normalisation $\sum_{m,n=1}^N \textrm{prob}(x_m,x_n)=1$. As we do not know which particular cluster induces the similarity, we have to marginalise over all cluster nodes in Fig. \ref{fig:bipartite_graph_sketch},
\begin{equation}\label{eq:swtich_2_bipartite_graph_model}
\textrm{prob}(x_m,x_n) = \sum_{k=1}^K \textrm{prob}(x_m,x_n,c_k) \;\textrm{.}
\end{equation}
With this marginalisation we have switched from the weighted undirected graph to our bipartite-graph model. Applying Bayes' theorem yields
\begin{equation}
\textrm{prob}(x_m,x_n) = \sum_{k=1}^K \textrm{prob}(x_n|c_k)\,\textrm{prob}(x_m,c_k) \;\textrm{,}
\end{equation}
where we have used
\begin{equation}\label{eq:assumption_in_derivation}
\textrm{prob}(x_n|x_m,c_k) = \textrm{prob}(x_n|c_k) \;\textrm{,}
\end{equation}
since $x_m$ and $x_n$ are not directly connected in the bipartite graph, i.e. they are statistically independent. This is the only assumption in this derivation and it implies that \textit{all} statistical dependence is induced by the clusters. Using Bayes' theorem once more yields
\begin{equation}
\textrm{prob}(x_m,x_n) = \sum_{k=1}^K \frac{\textrm{prob}(x_n,c_k)\,\textrm{prob}(x_m,c_k)}{\textrm{prob}(c_k)} \;\textrm{.}
\end{equation}
We identify the bipartite-graph adjacency matrix in analogy to Eq. (\ref{eq:W_as_joint_probability}),
\begin{equation}\label{eq:def:bipartite_graph_adjacency_matrix}
B_{nk} = \textrm{prob}(x_n,c_k) \;\textrm{,}
\end{equation}
with its marginalisation
\begin{equation}\label{eq:def_lambda_k}
\lambda_k = \textrm{prob}(c_k) = \sum_{n=1}^N \textrm{prob}(x_n,c_k) = \sum_{n=1}^N B_{nk} \;\textrm{.}
\end{equation}
The marginalised probabilities $\lambda_k$ are the cluster priors of the cluster nodes $c_k$ in the bipartite graph. Moreover, the $\lambda_k$ are the degrees of the nodes.
\subsection{Fitting the similarity matrix \label{sect:fitting_bipartite-graph_model}}
In order to fit the bipartite-graph model defined by Eq. (\ref{eq:bipartite_graph_model}) to a given similarity matrix, we perform some simplifications. First, we note that we can rewrite Eq. (\ref{eq:bipartite_graph_model}) using matrix notation,
\begin{equation}
\hat{ W} = B\cdot\Lambda^{-1}\cdot B^T \;\textrm{,}
\end{equation}
where $ B$ is the $N\times K$ bipartite-graph adjacency matrix and $\Lambda=\textrm{diag}(\lambda_1,\ldots,\lambda_k)$ is the $K\times K$ diagonal matrix of cluster degrees. This notation enables us to employ fast and efficient algorithms from linear algebra. We change variables by
\begin{equation}
B = H\cdot\Lambda \;\textrm{,}
\end{equation}
where $ H$ is an $N\times K$ matrix. The elements of $ H$ can be interpreted as the cluster likelihoods, since $H_{nk}=\frac{B_{nk}}{\lambda_k}=\frac{\textrm{prob}(x_n,c_k)}{\textrm{prob}(c_k)}=\textrm{prob}(x_n|c_k)$. Using these new variables $ H$ and $\Lambda$, the model $\hat{ W}$ of the data similarity matrix $ W$ is given by
\begin{equation}\label{eq:modified_bipartite-graph_model}
\hat{ W} = H\cdot\Lambda\cdot H^T \;\textrm{,}
\end{equation}
where we have eliminated the matrix inversion and reduced the nonlinearity to some extent. The normalisation constraints of Eq. (\ref{eq:normalisation_B_and_lambda_k}) translate to $ H$ as
\begin{equation}\label{eq:normalisation_constraint_H}
\sum_{n=1}^N H_{nk} = \sum_{n=1}^N \textrm{prob}(x_n|c_k) = 1 \qquad \forall\,k=1,\ldots,K \;\textrm{.}
\end{equation}
The normalisation constraints on $ H$ and $\Lambda$ are now decoupled, and we can treat both matrices as independent of each other. As $ H$ is an $N\times K$ matrix and $\Lambda$ a $K\times K$ diagonal matrix, we have $K(N+1)$ model parameters. In comparison to this number, we do have $\frac{1}{2}N(N+1)$ independent elements in the data similarity matrix due to its symmetry. Hence, a reasonable fit situation requires $\frac{1}{2}N\gg K$ in order to constrain all model parameters.
The data similarity matrix $ W$ is fitted by maximising the logarithmic likelihood function $\log\mathcal L$ of the bipartite-graph model. \citet{Yu2005} give a derivation of this function based on the theory of random walks on graphs. Their result is
\begin{equation}\label{eq:loglik_bipartite_graph_model}
\log\mathcal L(\Theta| W) = \sum_{m,n=1}^N W_{mn}\,\log\textrm{prob}(x_m,x_n|\Theta) \;\textrm{,}
\end{equation}
where $\Theta=\{H_{11},\ldots,H_{NK},\lambda_1,\ldots,\lambda_K\}$ denotes the set of $K(N+1)$ model parameters and $\textrm{prob}(x_m,x_n|\Theta)=\sum_{k=1}^K H_{mk}\lambda_k H_{nk}=\hat W_{mn}$ is the model. If we remember that $W_{mn}=\textrm{prob}(x_m,x_n)$, then we see that $\log\mathcal L$ is the cross entropy of the true probability distribution $W_{mn}=\textrm{prob}(x_m,x_n)$ and the modelled distribution $\hat W_{mn}=\textrm{prob}(x_m,x_n|\Theta)$. Consequently, maximising $\log\mathcal L$ maximises the information our model contains about the data similarity matrix.
Directly maximising $\log\mathcal L$ is too hard, since the fit parameters are subject to the constraints given by Eqs. (\ref{eq:normalisation_B_and_lambda_k}) and (\ref{eq:normalisation_constraint_H}). We use an alternative approach that makes use of the expectation-maximisation (EM) algorithm, which is an iterative fit routine. Given an initial guess on the model parameters, the EM algorithm provides a set of algebraic update equations to compute an improved estimate of the optimal parameters that automatically respects the normalisation. These update equations are \citep{Bilmes1997,Yu2005}
\begin{equation}\label{eq:EM_update_lambda}
\lambda_k^\textrm{new} = \lambda_k \sum_{m,n=1}^N \frac{W_{mn}}{\left( H\cdot\Lambda\cdot H^T\right)_{mn}} H_{mk} H_{nk} \,\textrm{,}
\end{equation}
\begin{equation}\label{eq:EM_update_H}
H_{nk}^\textrm{new}\propto H_{nk}\lambda_k \sum_{m=1}^N \frac{W_{mn}}{\left( H\cdot\Lambda\cdot H^T\right)_{mn}} H_{mk} \,\textrm{.}
\end{equation}
The $H_{nk}^\textrm{new}$ have to be normalised by hand, whereas the $\lambda_k^\textrm{new}$ are already normalised. Each iteration step updates all the model parameters, which has time complexity $\mathcal O(K\cdot N^2)$ for $K$ clusters and $N$ data nodes. We initialise all the cluster degrees to $\lambda_k^0=\frac{1}{K}$, whereby we trivially satisfy the normalisation condition and simultaneously ensure that no cluster is initialised as virtually absent. The $H_{nk}^0$ are initialised randomly and normalised ''by hand''.
Now, we want to briefly discuss the convergence properties of the EM algorithm. It has been shown \citep[e.g.][]{Redner1984} that the EM algorithm is guaranteed to converge to a \textit{local} maximum of $\log\mathcal L$ under mild conditions. Indeed, it was shown that the EM algorithm is monotonically converging, i.e. each iteration step is guaranteed to increase $\log\mathcal L$. Therefore, after each iteration step, we check how much $\log\mathcal L$ was increased compared to the previous step. If $\log\mathcal L$ changed by less than a factor of $10^{-9}$, we will consider the EM algorithm to have converged. This convergence criterion was chosen based on systematic tests like those discussed in Sect. \ref{sect:systematic_tests}. Finally, we note that the fit results are not unique, since the ordering of the clusters is purely random.
\subsection{Estimating the optimal number of clusters \label{sect:estimating_optimal_K}}
In this section we demonstrate how to estimate the optimal number of clusters for a given data set, which is a crucial part of any clustering analysis. It is essential to estimate the optimal number of clusters with due caution. This is a problem of assessing nonlinear models and there are no theoretically justified methods, there are only heuristic approaches. Common heuristics are the Bayesian information criterion
\begin{equation}\label{eq:def_BIC}
\textrm{BIC} = -2\,\log\mathcal L + p\,\log N
\end{equation}
and Akaike's information criterion
\begin{equation}
\textrm{AIC} = -2\,\log\mathcal L + 2p \,\textrm{,}
\end{equation}
where $p$ and $N$ denote the number of model parameters and the number of data points, respectively. As we have seen in Sect. \ref{sect:fitting_bipartite-graph_model}, the bipartite-graph model involves $K(N+1)$ model parameters. Consequently, BIC and AIC are not applicable, since $\log\mathcal L$ is not able to compensate for the large impact of the penalty terms. This inability of $\log\mathcal L$ is likely to originate from the sparse data population in the high-dimensional parameter space. Another tool of model assessment is cross-validation, but this is computationally infeasible in this case.
\begin{figure}
\includegraphics[width=8cm]{demo_SSR_and_mean_angular_changes}
\caption{Estimating the optimal number of clusters for the data sample shown in Fig. \ref{fig:demo_toy_example_data_set}.}
(a) $\textrm{SSR}(K)$ as a function of the number $K$ of clusters. (b) Mean angular changes $\langle\Delta(K)\rangle$ averaged over ten fits.
\label{fig:demo_estimate_optimal_K}
\end{figure}
We now explain how to compare bipartite-graph models of different complexities heuristically, i.e. how to estimate the optimal number of clusters. This heuristic employs the sum of squared residuals
\begin{equation}\label{eq:definition_SSR}
\textrm{SSR}(K) = \sum_{m=1}^N\sum_{n=1}^m \left( \frac{W_{mn} - \sum_{k=1}^K H_{mk}\lambda_k H_{nk} }{W_{mn}} \right)^2 \,\textrm{.}
\end{equation}
The definition puts equal emphasis on all elements. If we left out the denominator in Eq. (\ref{eq:definition_SSR}), the SSR would emphasise deviations of elements with large values, whereas elements with small values would be neglected. However, both large and small values of pairwise similarities are decisive. We estimate the optimal $K$ via the position of a \textit{kink} in the function $\textrm{SSR}(K)$ (cf. Fig. \ref{fig:demo_estimate_optimal_K}). Such a kink arises if adding a further cluster does not lead to a significant improvement in the similarity-matrix reconstruction.
We demonstrate this procedure in Fig. \ref{fig:demo_estimate_optimal_K} by using the toy example of Figs. \ref{fig:demo_toy_example_data_set} and \ref{fig:demo_estimate_optimal_simi_measure}, which is composed of six nicely separable clusters. We fit bipartite-graph models to the similarity matrix shown in Fig. \ref{fig:demo_estimate_optimal_simi_measure}, with $K$ ranging from 1 to 15. The resulting SSR values are shown in panel (a) of Fig. \ref{fig:demo_estimate_optimal_K}. In fact, $\textrm{SSR}(K)$ exhibits two prominent kinks at $K=3$ and $K=6$, rather than a single one. Obviously, for $K=3$, the clustering algorithm groups the four nearby clusters together, thus resulting in three clusters. For $K=6$, it is able to resolve this group of ''subclusters''.
We can construct a more quantitative measure by computing the angular change $\Delta(K)$ of $\log\textrm{SSR}(K)$ at each $K$,
\begin{displaymath}
\Delta(K) = \arctan\left[\log\textrm{SSR}(K-1) - \log\textrm{SSR}(K)\right]
\end{displaymath}
\begin{equation}
- \arctan\left[\log\textrm{SSR}(K) - \log\textrm{SSR}(K+1)\right] \;\textrm{.}
\end{equation}
As $K$ is an integer, $\log\textrm{SSR}(K)$ is a polygonal chain and thus an angular change is well defined. A large positive angular change then indicates the presence of a kink in $\textrm{SSR}(K)$.\footnote{It is not possible to compute the angular change for $K=1$, but this case is not a reasonable grouping anyway under the assumption that there are objects of different types in the given data sample.} However, we can even do better by fitting the similarity matrix several times for each $K$ and averaging the angular changes. The results of the fits differ slightly, since the model parameters are randomly initialised each time. These mean angular changes are shown in panel (b) of Fig. \ref{fig:demo_estimate_optimal_K}, averaged over 20 fits for each $K$. First, for large $K$ the mean angular changes are consistent with zero, i.e. in this domain increasing $K$ decreases $\textrm{SSR}(K)$ but does not improve the fit systematically. Second, for $K=3$ and $K=6$ the mean angular changes deviate significantly from zero. For $K=2$ and $K=4$, the mean angular changes are negative, which corresponds to ''opposite'' kinks in the SSR spectrum and is due to $K=3$ being a very good grouping of the data.
For large $K$ these detections may be less definite due to the flattening of $\textrm{SSR}(K)$. Therefore, we may systematically underestimate the optimal number of clusters. Moreover, this toy example also demonstrates that there may be more than a single advantageous grouping of the data and there may be disadvantageous groupings. If there are multiple detections of advantageous groupings, it may be difficult to judge which grouping is the best. In the worst case, we even may not find any signal of an advantageous grouping, which would either imply that our given sample is composed of objects of the same type or that the data does not contain enough information about the grouping. Unfortunately, this scheme of estimating the optimal number of clusters is extremely inefficient from a computational point of view. This is a severe disadvantage for very large data sets. Moreover, though this heuristic is working well, the significance of the mean angular changes is likely to be strongly influenced by the variance caused by the algorithm's initialisation.
\subsection{Comparison with previous work}
As the work of \citet{Kelly2004,Kelly2005} is very close to our own work, we want to discuss it in some detail and work out the differences. The authors applied a soft clustering analysis to the first data release of SDSS. In \citet{Kelly2004} they decomposed $r$-band images of 3,037 galaxies into shapelets, using the IDL shapelet code by \citet{Massey2005}. In \citet{Kelly2005} they extended this scheme to all five photometric bands $u,g,r,i,z$ of SDSS, thereby also taking into account colour information. Afterwards, they used a principal component analysis (PCA) to reduce the dimensionality of their parameter space. In \citet{Kelly2004} the reduction was from 91 to 9 dimensions and in \citet{Kelly2005} from 455 to 2 dimensions. Then they fitted a mixture-of-Gaussians model \citep{Bilmes1997} to the compressed data, where each Gaussian component represents a cluster. They were able to show that the resulting clusters exhibited a reasonable correlation to the traditional Hubble classes.
Reducing the parameter space with PCA and also using a mixture-of-Gaussians model are both problematic from our point of view. First, PCA relies on the assumption that those directions in parameter space that carry the desired information do also carry a large fraction of the total sample variance. This is neither guaranteed nor can it be tested for in practice. Second, galaxy morphologies are not expected to be normally distributed. Therefore, using a mixture-of-Gaussians model is likely to misestimate the data distribution. Nonetheless, the work by \citet{Kelly2004,Kelly2005} was a landmark, both concerning their usage of a probabilistic algorithm and conceptually, by applying a clustering analysis to the first data release of SDSS.
In contrast to \citet{Kelly2004,Kelly2005}, we do not reduce the dimensionality of the parameter space and then apply a clustering algorithm to the reduced data. We also do not try to model the data distribution in the parameter space, which would be virtually impossible due to its high dimensionality \citep[\textit{curse of dimensionality}, cf.][]{Bellman1961}. Rather, we use a similarity matrix, which has two major advantages: First, we do not rely on any compression technique such as PCA. Second, we cannot make any mistakes by choosing a potentially wrong model for the data distribution, since we model the similarity matrix. There are two sources of potential errors in our method:
\begin{enumerate}
\item Estimation of pairwise similarities (Eq. (\ref{eq:def:similarity_measure})). This is hampered by our lack of knowledge about the metric in the morphological space and it is in some sense similar to mismodelling.
\item Modelling the similarity matrix by a bipartite-graph model. As the only assumption in the derivation of the bipartite-graph model is Eq. (\ref{eq:assumption_in_derivation}), this happens if and only if a significant part of the pairwise similarity is \textit{not} induced by the clusters, but rather by e.g. observational effects. However, any other classification method (automated or not) will have problems in this situation, too.
\end{enumerate}
\section{Systematic Tests\label{sect:systematic_tests}}
In this section we conduct systematic tests using artificial data samples that are specifically designed to investigate the impact of certain effects. First, we demonstrate that hard classification schemes cause problems with subsequent parameter estimation. Furthermore, we investigate the impact of non-optimal similarity measures, two-cluster separation, noise and cluster cardinalities on the clustering results.
\subsection{Overview}
We start by describing the artificial data sets that we are going to use. Furthermore, we describe the diagnostics by which we assess the performance of the clustering algorithm.
The data sets are always composed of two clusters, where the number of example objects drawn from each cluster may be different. The clusters are always designed as $p$-variate Gaussian distributions, i.e.
\begin{equation}
\textrm{prob}(\vec x|\vec\mu,\Sigma) = \frac{\exp\left[ -\frac{1}{2}\left(\vec x-\vec\mu\right)^T\cdot\Sigma^{-1}\cdot\left(\vec x-\vec\mu\right) \right]}{\sqrt{\left(2\pi\right)^p\,\det\Sigma}} \;\textrm{,}
\end{equation}
where $\vec\mu$ and $\Sigma$ denote the mean vector and the covariance matrix, respectively.
Knowing the true analytic form of the underlying probability distributions, we are able to assess the probabilistic data-to-cluster assignments proposed by the clustering algorithm. For two clusters $A$ and $B$, the true data-to-cluster assignment of some data point $\vec x$ to cluster $k=A,B$ is given by the cluster posterior
\begin{equation}\label{eq:def_true_cluster_posteriors}
\textrm{prob}(k|\vec x) = \frac{\textrm{prob}(\vec x|\vec\mu_k,\Sigma_k)}{\textrm{prob}(\vec x|\vec\mu_A,\Sigma_A) + \textrm{prob}(\vec x|\vec\mu_B,\Sigma_B)} \;\textrm{.}
\end{equation}
The numerator $\textrm{prob}(\vec x|\vec\mu_k,\Sigma_k)$ is the cluster likelihood. The cluster priors $\textrm{prob}(A)=\textrm{prob}(B)=\frac{1}{2}$ are flat and cancel out. For a given data set of $N$ objects, these true cluster posteriors are compared to the clustering results using the expectation values of the zero-one loss function
\begin{equation}
\langle\mathcal L_{01}\rangle = \frac{1}{N}\sum_{n=1}^N \left\{\begin{array}{lcl} 0 & \Leftrightarrow & \textrm{prob}_\textrm{fit}(C_n|\vec x_n)\\
& & \;\;>\textrm{prob}_\textrm{fit}(\neg C_n|\vec x_n) \\ 1 & \textrm{else} & \\ \end{array}\right.
\end{equation}
and of the squared-error loss function
\begin{equation}
\langle\mathcal L_\textrm{SE}\rangle = \frac{1}{N}\sum_{n=1}^N \left( \textrm{prob}_\textrm{fit}(C_n|\vec x_n) - \textrm{prob}_\textrm{true}(C_n|\vec x_n) \right)^2 \;\textrm{,}
\end{equation}
where $C_n$ denotes the correct cluster label of object $\vec x_n$ and $\neg C_n$ the false label. The zero-one loss function is the misclassification rate, whereas the squared-error loss function is sensitive to misestimations of the cluster posteriors that do not lead to misclassifications. As the two clusters are usually well separated in most of the following tests, the true maximum cluster posteriors are close to $100\%$. Therefore, misestimation means underestimation of the maximum posteriors, which is quantified by $\sqrt{\langle\mathcal L_\textrm{SE}\rangle}$.
\subsection{Impact of hard cuts on parameter estimation\label{sect:hart_cuts_bias_means}}
\begin{figure}
\includegraphics[width=8cm]{hard_cuts_biasing_means}
\caption{Break-down of hard classifications in case of overlapping clusters.}
Deviation $\hat\mu_A-\mu_A$ of estimated and true means vs. two-cluster separation $\Delta x$ for class A for hard estimator (red line), soft estimator (blue line), and predicted bias of hard estimator for $\Delta x\rightarrow 0$ (dashed line). From 1,000 realisations of data samples we estimated errorbars, which are shown but too small to be visible.
\label{fig:hart_cuts_bias_means}
\end{figure}
In this first test, we want to demonstrate that hard cuts that are automatically introduced when using hard classification or clustering algorithms can lead to systematic misestimations of parameters, i.e. biases. This is a general comment in order to support our claim that hard data-to-class assignments are generically inappropriate for overlapping classes. We are not yet concerned with our soft-clustering algorithm. We use two one-dimensional Gaussians with means $\mu_A$ and $\mu_B$, variable two-cluster separation $\Delta x=\mu_A-\mu_B$, and constant variance $\sigma^2=1$. From each Gaussian cluster we then draw $N=$10,000 objects. From the resulting data sample we estimate the means $\hat\mu_k$ of the two Gaussians and compare with the true means $\mu_k$. The results are averaged over 1,000 realisations of data samples.
Figure \ref{fig:hart_cuts_bias_means} shows the deviations of the estimated from the true means when using a hard cut at $x=0$ (red line) and a weighted mean (blue line). A hard cut at $x=0$ that assigns all data points with $x<0$ to class A and those with $x>0$ to class B is the most reasonable hard classification in this simple example. Once the complete sample is divided into two subsamples for classes A and B, we estimate the usual arithmetic mean
\begin{equation}\label{eq:def:hard_estimator}
\hat\mu_k^\textrm{hard} = \frac{1}{N_k}\sum_{n=1}^{N_k} x_{k,n} \,\textrm{.}
\end{equation}
As Fig. \ref{fig:hart_cuts_bias_means} shows, this estimator is strongly biased in case of overlapping clusters ($\Delta x\rightarrow 0$). In the limit of $\Delta x = 0$, we can predict this bias analytically from the expectation value
\begin{equation}
\langle x\rangle_{A/B} = \mp 2\int_0^\infty dx\,x\,e^{-x^2/2} = \frac{\mp 2}{\sqrt{2\pi}} \approx \mp 0.7979 \,\textrm{,}
\end{equation}
where the integration is only over one half of the parameter space and the factor of 2 arises from both Gaussians contributing the same for $\Delta x = 0$. This bias is shown as dashed line in Fig. \ref{fig:hart_cuts_bias_means}, where for $\Delta x = 0$ also $\mu_{A/B}=\mp \Delta x/2=0$. If we employ the true posteriors defined by Eq. (\ref{eq:def_true_cluster_posteriors}) as weights and use
\begin{equation}\label{eq:def:soft_estimator}
\hat\mu_k^\textrm{soft} = \frac{\sum_{n=1}^N \textrm{prob}(k|x_n)\,x_n}{\sum_{n=1}^N \textrm{prob}(k|x_n)} \,\textrm{,}
\end{equation}
then we will get an unbiased estimate despite the overlap, as is evident from Fig. \ref{fig:hart_cuts_bias_means}. This comparison demonstrates the break-down of hard algorithms in case of overlapping clusters.
\subsection{Impact of non-optimal similarity measures \label{sect:impact_non-optimal_simi}}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{demo_SSR_kink_data}
\includegraphics[width=8cm]{demo_distances}
\end{center}
\caption{Artificial data sample with six clusters (top) and the matrix of pairwise Euclidean distances (bottom).}
Each cluster has an underlying bivariate Gaussian distribution with covariance matrix $\Sigma=\textrm{diag}(1,1)$. We sampled 50 data points from each cluster.
\label{fig:demo_toy_example_data_set}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{demo_estimation_optimal_simi}
\includegraphics[width=8cm]{demo_similarity_matrix}
\caption{Estimating the optimal similarity measure for the example data of Fig. \ref{fig:demo_toy_example_data_set}.}
Top panel: Modified Manhattan distance $C$ (Eq. (\ref{eq:def:Manhattan_dist_constant})) for $s=1.01$ (cyan line), $s=1.03$ (blue line) and $s=1.1$ (red line). For $\alpha\rightarrow 0$ the matrix becomes a step matrix, which is why the constant levels depend on the scale parameter. Bottom panel: The resulting similarity matrix.
\label{fig:demo_estimate_optimal_simi_measure}
\end{figure}
We now explain how to optimise the similarity measure defined in Eq. (\ref{eq:def:similarity_measure}) and what ''optimal'' means. Given the $N\times N$ symmetric matrix of pairwise distances $d(\vec x_m,\vec x_n)$, we can tune the similarity measure by adjusting the two parameters $\alpha$ and $s$. Tuning the similarity measure has to be done with care, since there are two undesired cases: First, for $\alpha\rightarrow\infty$, the resulting similarity matrix approaches a constant, i.e. $W_{mn}=\frac{1}{N^2}$ for all elements, since $d(\vec x_m,\vec x_n)\leq d_\textrm{max}$. This case prefers $K=1$ clusters, independent of any grouping in the data. Second, for $\alpha\rightarrow 0$, the similarity matrix approaches the step matrix defined by
\begin{equation}\label{eq:def:step_matrix}
S_{mn} \propto \left\{\begin{array}{lcl} 1 & \Leftrightarrow & m=n \\ 1-\frac{1}{s} & \Leftrightarrow & m\neq n \end{array}\right. \;\textrm{,}
\end{equation}
which is normalised such that $\sum_{m,n=1}^N S_{mn}=1$. This case prefers $K=N$ clusters. The \textit{optimal} similarity measure should be as different as possible from these two worst cases. We choose $\alpha$ and $s$ such that the modified Manhattan distance to the constant matrix
\begin{equation}\label{eq:def:Manhattan_dist_constant}
C = \sum_{m=1}^N\sum_{n=1}^m \left|W_{mn}-\frac{1}{N^2}\right|
\end{equation}
is large. Figure \ref{fig:demo_estimate_optimal_simi_measure} demonstrates how to tune the similarity measure using the artificial data set from the toy example of Fig. \ref{fig:demo_toy_example_data_set}. The basis is the $N\times N$ symmetric distance matrix shown in the bottom panel of Fig. \ref{fig:demo_toy_example_data_set}. For three different values of $s$, the top panel shows $C$ as functions of $\alpha$. Obviously, $C(\alpha)$ exhibits a maximum and can thus be used to choose $\alpha$. For $s=1.1$ (red curve) the maximum is lowest and so is the distance to a constant matrix. $s=1.01$ exhibits the maximum deviation from a constant matrix, but this choice of $s$ downweights off-diagonal terms in $ W$ according to Eq. (\ref{eq:def:step_matrix}). Thus, we also prefer if $s$ is not too close to 1 and $s=1.03$ (blue curve) is the compromise of the three scale parameters shown in Fig. \ref{fig:demo_estimate_optimal_simi_measure}. Note that the choice of $s$ is not an optimisation but a heuristic. Although the artificial data set of Fig. \ref{fig:demo_toy_example_data_set} and its distance matrix are very special, we experienced that $C(\alpha)$ as shown in Fig. \ref{fig:demo_estimate_optimal_simi_measure} is representative for the general case.
The resulting similarity matrix is shown in the right panel of Fig. \ref{fig:demo_estimate_optimal_simi_measure} and exhibits a block-like structure, since we have ordered the data points in the set. This is just for the sake of visualisation and does \textit{not} affect the clustering results. We clearly recognise six blocks along the diagonal, because the within-cluster similarities are always larger than the between-cluster similarities. Furthermore, we recognise a large block of four clusters in the bottom right corner that are quite similar to each other, whereas the remaining two clusters are more or less equally dissimilar to all other clusters. Consequently, the similarity matrix indeed represents all the features of the data set shown in Fig. \ref{fig:demo_toy_example_data_set}.
We now demonstrate first that the optimal similarity measure indeed captures the crucial information on the data and what happens if we do not use the optimal similarity measure. We use an artificial data set composed of two one-dimensional Gaussian clusters, both with unit variance and two-cluster separation of $\Delta x=3$. We sample 100 example objects from each cluster and compute the matrix of pairwise distances using the Euclidean distance measure. This data set and its distance matrix remain unchanged. For a constant parameter $s=2.25$, we vary the exponent $\alpha$ in the similarity measure defined by Eq. (\ref{eq:def:similarity_measure}). For each value of $\alpha$, we fit bipartite-graph models with $K=1$, 2 and 3 to the resulting similarity matrix, averaging the results over 15 fits each time.
Results of this test are shown in Fig. \ref{fig:impact_non-optimal_simi_measure}. Panel (a) shows the modified Manhattan distance $C$ to a constant matrix. This curve is very similar to Fig. \ref{fig:demo_estimate_optimal_simi_measure}. There is a prominent peak at $\alpha\approx 0.6$, indicating the optimal similarity measure. If the similarity measure is very non-optimal, then the similarity matrix will be close to a constant or step matrix, i.e. it poorly constrains the bipartite-graph model. In this case, we expect to observe overfitting effects, i.e. low residuals of the reconstruction and results with high variance. The computation times are longer, too, since the nonlinear model parameters can exhibit degeneracies thereby slowing down the convergence. Counterintuitively, we seek a large value of SSR in this test, since a similarity matrix which captures well the information content of the data is harder to fit. Indeed, the SSR values shown in panel (c) of Fig. \ref{fig:impact_non-optimal_simi_measure} are significantly lower for non-optimal $\alpha$'s and peak near the optimal $\alpha$. As expected, the mean computation times shown in panel (b) are minimal for the optimal similarity measure. Panel (d) shows how the evidence for two clusters evolves.\footnote{This is the reason why we need to fit bipartite-graph models using $K=1$, 2 and 3, in order to compute the angular change of $\textrm{SSR}(K)$ at $K=2$.} Near the optimal $\alpha$, also the evidence for two clusters shows a local maximum. The misclassification rate shown in panel (e) is insensitive to $\alpha$ over a broad range, but approaches a rate of $50\%$ rather abruptly for extremely non-optimal similarity measures. The squared-error loss shown in panel (f) is more sensitive to non-optimalities. It exhibits a minimum for the optimal $\alpha$ and grows monotonically for non-optimal values.
\begin{figure*}
\includegraphics[width=8cm]{optimality_test_1}
\includegraphics[width=8cm]{optimality_test_2}
\caption{Impact of non-optimal similarity measures on clustering results.}
(a) Modified Manhattan distance $C$ (Eq. (\ref{eq:def:Manhattan_dist_constant})). (b) Mean computation time per fit without errorbars. (c) Mean $\textrm{SSR}(K)$ values of resulting fits for $K=2$. (d) Mean angular change of $\textrm{SSR}(K)$ at $K=2$. (e) Mean misclassification rate (solid line) and 50\% misclassification rate (dashed line). (f) Mean squared-error loss of maximum cluster posteriors.
\label{fig:impact_non-optimal_simi_measure}
\end{figure*}
The most important conclusion to draw from this test is that our method of choosing $\alpha$ and $s$ for the similarity measure defined in Sect. \ref{sect:optimal_similarity_measure} is indeed ''optimal'', in the sense that it minimises both the misclassification rate and the squared-error loss. Additionally, we see that using the optimal similarity measure can also reduce computation times by orders of magnitude.
\subsection{Impact of two-cluster overlap \label{sect:impact_two-cluster_overlap}}
As we have to expect overlapping clusters in the context of galaxy morphologies, we now investigate the impact of the two-cluster overlap on the clustering results. The data sets used are always composed of 100 example objects drawn from two one-dimensional Gaussian clusters, both with unit variance. The two-cluster separation $\Delta x$ is varied from 1 to 1000. For each data set, we compute the matrix of pairwise Euclidean distances and then automatically compute the optimal similarity matrix by optimising $\alpha$ using a constant $s=2.25$ as described in Sect. \ref{sect:impact_non-optimal_simi}. To each similarity matrix we fit bipartite-graph models with $K=1$, 2 and 3 clusters. Furthermore, we fit a $K$-means algorithm with $K=1$, 2 and 3 to each data set in order to compare the results of both clustering algorithms. For each configuration, the results are averaged over 50 fits.
Results of this test are summarised in Fig. \ref{fig:impact_overlap}. Panel (a) shows the mean evidence for two clusters, based on the angular changes in $\textrm{SSR}(K)$ for the bipartite-graph model and the within-cluster scatter for the $K$-means algorithm. For decreasing separation $\Delta x$, i.e. increasing overlap, the evidence for two clusters decreases for both algorithms, as is to be expected.\footnote{Note that these two curves cannot be compared directly. Their agreement for $\Delta x < 20$ is coincidence.} As panel (b) reveals, the misclassification rates for $K$-means and the bipartite-graph model are both in agreement with the theoretically expected misclassification rate expected in the ideal case (black curve). For two one-dimensional Gaussians with means $\pm\frac{\Delta x}{2}$, the theoretical misclassification rate is given by
\begin{equation}
\langle\mathcal L_{01}^\textrm{theo}\rangle = \int_{-\infty}^0 dx\,\textrm{prob}\left(x\left|\mu=+\frac{\Delta x}{2}\right.,\sigma\right) \;\textrm{,}
\end{equation}
which measures the overlap of both Gaussians. In the limit $\Delta x=0$, this yields $\langle\mathcal L_{01}^\textrm{theo}\rangle = \frac{1}{2}$. The explanation for the excellent performance of both $K$-means and bipartite-graph model is, that in this case the clusters have equal cardinalities and are spherical. Nevertheless, the results of the $K$-means are biased due to the hard data-to-cluster assignment. Panel (c) of Fig. \ref{fig:impact_overlap} shows the mean squared-error loss of the bipartite-graph models.\footnote{We do not compare with $K$-means, since $K$-means is a hard algorithm and squared-error loss is no reasonable score function in this case.} First, the general trend is that the squared-error loss increases for decreasing two-cluster separation. This is due to the growing amount of overlap confusing the bipartite-graph model. Second, for $\Delta x\lesssim 4$, the squared-error loss decreases significantly. This effect can be explained as follows: For very small separations, the overlap is so strong that even the true cluster posteriors are both close to 50\%. Therefore, the fitted cluster posteriors scatter around 50\%, too, thereby reducing the squared error. Third, the squared error establishes a constant value of $\langle\mathcal L_\textrm{SE}\rangle\approx 0.045$ at large separations. In this case, the true maximum cluster posteriors are essentially $100\%$, so this corresponds to a systematic underestimation of the maximum posteriors of $\sqrt{\langle\mathcal L_\textrm{SE}\rangle}\approx 21\%$. Due to the large two-cluster separation, this \textit{bias} does not lead to misclassifications, as is evident from panel (b) in Fig. \ref{fig:impact_overlap}. This bias originates from the fact that any two objects have a finite distance and thus a non-vanishing similarity.
\begin{figure}
\includegraphics[width=8cm]{impact_two_cluster_separation}
\caption{Impact of two-cluster overlap on clustering results for $K$-means algorithm and bipartite-graph model.}
(a) Mean angular change of $\textrm{SSR}(K)$ (bipartite-graph model) and within-cluster scatter ($K$-means) at $K=2$. (b) Mean misclassification rates of $K$-means and bipartite-graph model (see text) compared to theoretical prediction. All curves coincide. (c) Mean squared-error loss of bipartite-graph model.
\label{fig:impact_overlap}
\end{figure}
This test further demonstrates that the bipartite-graph model yields convincing results. This is most evident in the misclassification rate, which is in excellent agreement with the theoretical prediction of the best possible score.
\subsection{Impact of noise \label{sect:impact_noise}}
As observational data is subject to noise, we now investigate the response of the clustering results to noise on the similarity matrix. We simulate the noise by adding a second dimension $y$ to the data. The two clusters are bivariate Gaussian distributions, both with $\sigma_x^2=1$ and two-cluster separation of $\Delta x=10$ and $\Delta y=0$. We vary the size of the variance in $y$-direction ranging from $\sigma_y^2=0.1$ to 10000, thereby introducing noise that translates via the Euclidean distance to the similarity matrix. From each cluster 100 example objects are drawn and we fit bipartite-graph and $K$-means models using $K=1$, 2 and 3. The results are averaged over 50 fits for each value of $\sigma_y^2$.
\begin{figure}
\includegraphics[width=8cm]{impact_noise_on_clustering}
\caption{Impact of noise variance $\sigma_y^2$ on clustering results for $K$-means algorithm and bipartite-graph model.}
(a) Mean angular change of $\textrm{SSR}(K)$ (bipartite-graph model) and within-cluster scatter ($K$-means) at $K=2$. (b) Mean misclassification rate of $K$-means and bipartite-graph model. (c) Mean squared-error loss of bipartite-graph model.
\label{fig:impact_noise}
\end{figure}
Results of this test are shown in Fig. \ref{fig:impact_noise}. The evidence for two clusters (panel (a)) rapidly degrades for increasing variance for the bipartite-graph model as well as the $K$-means algorithm, as is to be expected. Inspecting the misclassification rate (panel (b)) reveals that both algorithms are insensitive to $\sigma_y^2$ until a critical variance is reached where both misclassification rates increase abruptly. For the $K$-means algorithm, this break down happens at $\sigma_y^2\approx 30$, whereas the bipartite-graph model breaks down at $\sigma_y^2\approx 40$, which amounts to $\frac{\Delta x}{\sigma_y}\approx 1.6$ in this setup. The evidence for two clusters (panel (a)) rises again for larger variances, although both algorithms have already broken down. This is a geometric effect: With increasing $\sigma_y^2$, the two clusters become more extended in $y$-direction, until it becomes favourable to split the data along $x=0$ rather than $y=0$. This also explains why the misclassification rate is 50\% in this regime. Consequently, the abrupt break-down originates from the setup of this test. Sampling more objects from each cluster might have prevented this effect, but would have increased the computational effort drastically. Moreover, this also demonstrates that isotropic distance measures are problematic. Using e.g. a diffusion distance \citep[e.g.][]{Richards2009} may solve this problem. The break down is less abrupt in the mean squared-error loss (panel (c)), since $\langle\mathcal L_{SE}\rangle$ is also sensitive to posterior misestimation that do not lead to misclassifications.
We conclude that the bipartite-graph model is fairly insensitive to noise of this kind over a broad range, until the setup of this test breaks down.
\subsection{Impact of cluster cardinalities\label{sect:focal_problem}}
Typically different types of galaxy morphologies have different abundancies in a given data sample. For instance, \citet{Bamford2009} observe different type fractions of early-type and spiral galaxies in the Galaxy Zoo project. Therefore, we now investigate how many objects of a certain kind are needed in order to detect them as a cluster. The concept of a number of objects being members of a certain cluster is poorly defined in the context of soft clustering. We generalise this concept by defining the \textit{cardinality} of a cluster $c_k$
\begin{equation}\label{eq:def:cluster_cardinality}
\textrm{card}(k) = \sum_{n=1}^N \textrm{prob}(c_k|\vec x_n) \;\textrm{.}
\end{equation}
This definition satisfies $\sum_{k=1}^K \textrm{card}(k) = N$, since the cluster posteriors are normalised. In the case of hard clustering, Eq. (\ref{eq:def:cluster_cardinality}) is reduced to simple number counts, where the cluster posteriors become Kronecker symbols. We use two clusters, both one-dimensional Gaussians with unit variance and a fixed two-cluster separation of $\Delta x=10$. We then vary the number of objects drawn from each cluster such that the resulting data set always contains 200 objects. For each data set, we compute \textit{two different} similarity matrices: First, we compute the similarity matrix using the optimal $\alpha$ for a constant $s=2.0$, according to the recipe given in Sect. \ref{sect:impact_non-optimal_simi}. This similarity measure is adapted to every data set (\textit{adaptive similarity measure}). Second, we compute the similarity matrix using $\alpha=0.6$ and $s=2.0$, which is the optimal similarity measure for the data set composed to equal parts of objects from both clusters. This similarity measure is the same for all data sets (\textit{constant similarity measure}). To each of the two similarity matrices we fit a bipartite-graph model using $K=2$ and average the results over 50 fits.
\begin{figure}
\includegraphics[width=8cm]{impact_cluster_cardinatlities}
\caption{Impact of cardinalities on clustering results.}
(a) Mean misclassification rate. (b) Mean squared-error loss. (c) Correlation of estimated and true cluster cardinality.
\label{fig:impact_cardinalities}
\end{figure}
The results are summarised in Fig. \ref{fig:impact_cardinalities}. Panel (a) shows the dependence of the misclassification rate on the cardinality of cluster A. For the adaptive similarity measure the bipartite-graph model will break down, if one cluster contributes less than $10\%$ to the data set. For the constant similarity measure it will break down, if one cluster contributes less than $3\%$. The same behaviour is evident from the squared-error loss in panel (b). This problem is caused by the larger group in the data set dominating the statistics of the modified Manhattan distance $C$ defined by Eq. (\ref{eq:def:Manhattan_dist_constant}). This is a failure of the similarity measure, \textit{not} of the bipartite-graph model. The constant similarity measure stays ''focussed'' on the difference between the two clusters and its break-down at $3\%$ signals the limit to which clusters are detectable with the bipartite-graph model.
Panel (c) in Fig. \ref{fig:impact_cardinalities} shows the correlation of the measured cluster cardinality to the true cluster cardinality. For the constant similarity measure, both quantities correlate well. In contrast to this, for the adaptive similarity measure the two quantities do not correlate at all. Again, the adaptive similarity measure is dominated by the larger group, i.e. the similarities between the large and the small group are systematically too high. This leads to a systematic underestimation of the maximum cluster posteriors (cf. panel (b)), since for a two-cluster separation of $\Delta x=10$ the true posteriors are essentially $100\%$ as shown by Fig. \ref{fig:impact_overlap}c. This also affects the cluster cardinalities defined by Eq. (\ref{eq:def:cluster_cardinality}). If the cluster overlap is stronger, then this bias is likely to lead to misclassifications, too.
We conclude that the optimal similarity measure defined in Sect. \ref{sect:impact_non-optimal_simi} fails to discover groups that contribute 10\% or less to the complete data sample. A different similarity measure may solve this problem, but the optimal similarity measure has the advantage of minimising the misclassification rate and the squared-error loss for the discovered groups.
\section{Worked Example with SDSS Galaxies\label{sect:worked_example_SDSS}}
In this section we present our worked example with SDSS galaxies. First, we describe the sample of galaxies we analyse. Before applying the bipartite-graph model to the whole sample, we apply it to a small subsample of visually classified galaxies to prove that it is working not only for simple simulated data but also for real galaxy morphologies. Again, we emphasise that this is just meant as a demonstration, so parametrisation and sample selection are idealised.
\subsection{The data sample by Fukugita et al. (2007) \label{sect:analysis_small_set}}
\citet{Fukugita2007} derived a catalogue of 2,253 bright galaxies with Petrosian magnitude in the $r$ band brighter than $r_P=16$ from the Third Data Release of the SDSS \citep{Abazajian2005}. We analyse only $g$-band imaging of this sample, which is sensitive to HII regions and spiral arm structures. We expect that objects that are bright in $r$ are also bright in the neighbouring $g$-band. Therefore, all these objects have a high signal-to-noise ratio, i.e. the shapelet decomposition can employ a maximum order sufficiently large to reduce possible modelling problems.
Apart from the $g$-band imaging data, we also retrieved further morphological information from the SDSS database, namely Petrosian radii $r_{50}$ and $r_{90}$ containing $50\%$ and $90\%$ of the Petrosian flux, ratios of isophotal semi major and semi minor axis, and the logarithmic likelihoods of best-fitting de Vaucouleurs and exponential-disk profiles. Given the Petrosian radii $r_{50}$ and $r_{90}$ containing $50\%$ and $90\%$ of the Petrosian flux, we define the concentration index in analogy to \citet{Conselice2003},
\begin{equation}\label{eq:concentration_index}
C = 5\,\log\left( \frac{r_{90}}{r_{50}} \right) \;\textrm{.}
\end{equation}
For compact objects, such as elliptical galaxies, this concentration index is large, whereas it is smaller for extended objects with slowly decreasing light profiles, such as disk galaxies.
We then reduce the data sample in three steps: First, we sort out peculiar objects, i.e. objects that are definitely not galaxies, blended objects and objects that were cut in the mosaic. All these objects have no viable galaxy morphologies. This was done by visual inspection of all objects. Second, we decompose all images into shapelets using the same maximum order $N_\textrm{max}=12$ (91 expansion coefficients) for all objects. The shapelet code performs several internal data processing steps, namely estimating the background noise and subtracting the potentially non-zero noise mean, image segmentation and masking of multiple objects, estimating the object centroid position \citep[cf.][]{Melchior2007}. Third, we sort out objects for which the shapelet reconstruction does not provide reasonable models. This is done by discarding all objects whose best fits have a reduced $\chi^2$ that is not in the interval $[0.9,2.0]$. The lower limit is chosen very close to unity, since shapelets have the tendency to creep into the background noise and overfit objects. Setting out from the 2,253 bright galaxies of \citet{Fukugita2007}, the data processing leaves us with 1,520 objects with acceptable $\chi^2$. We check that the morphological information contained in the original data set and the reduced data set does not differ systematically, by comparing the sample distributions of Petrosian radii, axis ratios, concentration indeces, and logarithmic likelihoods of best-fitting deVaucouleur and exponential-disk profiles. All objects are large compared to the point-spread function (PSF) of SDSS, such that a PSF deconvolution as described in \citet{Melchior2009} is not necessary. This means we analyse apparent instead of intrinsic morphologies, but both are approximately the same.
\subsection{Demonstration with three clusters \label{sect:testbed_FON_EON_ELL}}
In this section we apply the soft clustering algorithm by \citet{Yu2005} for the first time to real galaxies. We use a small data set of 84 galaxies, which we visually classified as edge-on disk, face-on disk or ellipticals (28 objects per type). As these 84 galaxies were very large and very bright, we decomposed them anew using a maximum order of $N_\textrm{max}=16$, resulting in 153 shapelet coefficients per object. Figure \ref{fig:example_shapelet_models_of_3_SDSS_galaxies} shows one example object and its shapelet reconstruction for each type. This data set exhibits a strong grouping and we demonstrate that the bipartite-graph model indeed discovers the edge-on disks, face-on disks and ellipticals automatically, without any further assumptions.
The estimation of the number of clusters is shown in Fig. \ref{fig:3_class_testbed}. The mean angular changes in $\textrm{SSR}(K)$ averaged over 20 fits indeed reveal only one significant kink at $K=3$. The lowest value of SSR at $K=3$ is $\textrm{SSR}\approx 48$, which corresponds to an RMS residual (cf. Eq. (\ref{eq:definition_SSR})) of
\begin{equation}\label{eq:mean_residual_from_SSR}
\sqrt{\frac{\textrm{SSR}}{ \frac{1}{2}N(N+1) }} \approx 11.6\% \;\textrm{.}
\end{equation}
The denominator $\frac{1}{2}N(N+1)$ is the number of independent elements in the symmetric similarity matrix.
\begin{figure}
\includegraphics[width=8cm]{three_class_mean_angular_changes}
\caption{Mean angular changes $\langle\Delta(K)\rangle$ of bipartite-graph model for data set composed of edge-on disks, face-on disks and ellipticals.}
\label{fig:3_class_testbed}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{three_class_posterior_plane}
\caption{Cluster posterior space of bipartite-graph model for edge-on disks, face-on disks and ellipticals.}
The triangle defines the subspace allowed by the normalisation constraint of the posteriors. The corners of the triangle mark the points of $100\%$ posterior probability. The * indicates the point where all three posteriors are equal. Colours encode a-priori classifications unknown to the algorithm.
\label{fig:3_class_testbed_posterior_planes}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{PCA_2D_pc_space}
\caption{Comparing Fig. \ref{fig:3_class_testbed_posterior_planes} with results of PCA for edge-on disks, face-on disks and ellipticals.}
Parameter space spanned by the first two principal components. The first principal component carries $\approx 45.2\%$ and the second $\approx 21.4\%$ of the total variance. Colours encode a-priori classifications unknown to the PCA algorithm.
\label{fig:3_class_PCA}
\end{figure}
We conclude from Fig. \ref{fig:3_class_testbed} that the bipartite-graph model indeed favours three clusters. However, we still have to prove that the similarity matrix contains sufficient information on the data and that the bipartite-graph model discovers the correct classes. For $K=3$, the cluster posteriors populate a two-dimensional plane because they are subject to a normalisation constraint. This plane is shown in Fig. \ref{fig:3_class_testbed_posterior_planes}. Indeed, the distribution of cluster posteriors exhibits an excellent grouping of ellipticals, edge-on disks and face-on disks. The three clusters are well separated, apart from two objects labelled as edge-on disks but assigned to the cluster of ellipticals. A second visual inspection of these two ''outliers'' revealed that we had initially misclassified them as edge-on disk. The excellent results are particularly impressive if we remember that we analysed 84 data points distributed in a 153-dimensional parameter space. Moreover, it is very encouraging that the soft clustering analysis did indeed recover the ellipticals, face-on and edge-on disks \textit{automatically}.
In order to get an impression of how good these results actually are, we compare the cluster posterior plane to results obtained from PCA. Therefore, we estimate the covariance matrix $\Sigma$ of the data sample in shapelet-coefficient space and diagonalise it. Only the first 83 eigenvalues of $\Sigma$ are non-zero, since the 84 data objects poorly constrain the $153\times 153$ covariance matrix. The first two principal components carry $66.6\%$ of the total sample variance and Fig. \ref{fig:3_class_PCA} displays the parameter space spanned by them. Obviously, PCA performs well in reducing the parameter space from 153 dimensions down to two, since the ellipticals, face-on and edge-on disks exhibit a good grouping.\footnote{PCA only reduces the parameter space, but does \textit{not} assign classes to objects.} However, the bipartite-graph model provides much more compact and well-separated groups. This is due to the degeneracies we have broken when we computed the minimal spherical distances as described in Sect. \ref{sect:shapelets}. In case of PCA, these degeneracies are unbroken and introduce additional scatter.
In both Figs. \ref{fig:3_class_testbed_posterior_planes} and \ref{fig:3_class_PCA} we notice that the group of ellipticals is significantly more compact than the groups of face-on and edge-on disks. This is caused by three effects: First, as discussed in Sect. \ref{sect:shapelets}, our parametrisation of elliptical galaxies is problematic, thereby introducing common artefacts for all objects of this type. These common features are then picked up by the soft-clustering algorithm. Ironically, the problems of the parametrisation help to discriminate the types in this case. Second, we described in Sect. \ref{sect:shapelets} how to make our morphological distance measure invariant against various random quantities, namely image size, image flux, orientation angle and handedness. However, the distance measure and thereby the similarity measure are \textit{not} invariant against the inclination angle w.r.t. the line of sight, which introduces additional scatter into the clustering results. We expect that the impact of this random effect is smaller for ellipticals than for disk galaxies. Third, disk galaxies usually exhibit complex substructures (e.g. spiral arms or star-forming regions), whereas elliptical galaxies do not. Consequently, the intrinsic morphological scatter of disk galaxies is larger than for ellipticals.
\subsection{Analysing the data set of Fukugita et al. (2007) \label{sect:analysing_Fuku}}
We now present the soft-clustering results from analysing all 1,520 bright galaxies from the reduced data set of \citet{Fukugita2007}. We have chosen the similarity measure with $s=1.02$ and corresponding optimal $\alpha\approx 0.12$, according to Sect. \ref{sect:impact_non-optimal_simi}. The shapes of the curves of the modified Manhattan distances $C(\alpha)$ are of the same generic form as before. Fit results of the similarity matrix for $K$ ranging from 1 to 20 are shown in Table \ref{table:fit_results_small_sample} and Fig. \ref{fig:results_Fuku}. There are significant deviations of the mean angular changes from zero for $K=3$ and $K=8$. The signal at $K=2$ is ignored, since the SSR value is very high (cf. Table \ref{table:fit_results_small_sample}).
\begin{figure}
\includegraphics[width=8cm]{Fuku_mean_angular_changes}
\caption{Estimating the number of clusters in the data set of \citet{Fukugita2007}.}
Mean angular changes $\langle\Delta(K)\rangle$ are averaged over 15 Fits.
\label{fig:results_Fuku}
\end{figure}
\begin{table}
\begin{tabular}{ccc}
\hline\hline
$K$ & minimal SSR & mean angular changes (degrees) \\
\hline
1 & $39,220$ & -- \\
2 & $12,313$ & $14.489\pm 0.047$ \\
3 & $6,146$ & $22.67\pm 0.14$ \\
4 & $4,965$ & $-2.01\pm 0.19$ \\
5 & $3,868$ & $2.76\pm 0.27$ \\
6 & $3,155$ & $2.19\pm 0.61$ \\
7 & $2,676$ & $-0.89\pm 1.17$ \\
8 & $2,254$ & $5.91\pm 0.69$ \\
9 & $2,093$ & $-0.18\pm 0.95$ \\
10 & $1,931$ & $-0.35\pm 1.11$ \\
11 & $1,790$ & $0.24\pm 0.86$ \\
12 & $1,661$ & $1.92\pm 0.52$ \\
13 & $1,593$ & $0.36\pm 0.35$ \\
14 & $1,532$ & $0.03\pm 0.73$ \\
15 & $1,476$ & $0.15\pm 0.94$ \\
16 & $1,430$ & $0.86\pm 0.71$ \\
17 & $1,405$ & $0.04\pm 0.43$ \\
18 & $1,383$ & $0.22\pm 0.39$ \\
19 & $1,365$ & $0.14\pm 0.34$ \\
20 & $1,348$ & -- \\
\hline
\end{tabular}
\caption{Fitting the similarity matrix of 1,520 objects.\newline We present the minimal SSR value out of 15 fits and the mean angular change averaged over 15 fits.}
\label{table:fit_results_small_sample}
\end{table}
First, we investigate the clustering results for $K=3$, where we have $\textrm{SSR}\approx 6,146$ (cf. Table \ref{table:fit_results_small_sample}) corresponding to an RMS residual of $\approx 3.7\%$ (cf. Eq. (\ref{eq:mean_residual_from_SSR})) for the similarity-matrix reconstruction. In Fig. \ref{fig:results_Fuku_K3} we show the top five example objects for each of the three clusters together with a histogram of the distribution of cluster posteriors. Inspecting the example images, we clearly see that the first cluster is obviously composed of face-on disk galaxies, whereas the second cluster contains ellipticals. The third cluster is the cluster of edge-on disk galaxies or disks with high inclination angles. However, a blended object has been misclassified into this cluster, too. There are still some blended objects left that we failed to remove, since when sorting out blended objects we visually inspected the images in reduced resolution. The cluster posteriors for $K=3$ are very informative: First, we notice that objects from cluster 1 have typically very low posteriors in cluster 2 and intermediate posteriors in cluster 3, i.e. face-on disks are more similar to edge-on disks than to ellipticals. Second, objects from cluster 2 have low posteriors in all other clusters. Third, objects from cluster 3 tend to be more similar to objects in cluster 2, i.e. edge-on disks are more similar to ellipticals. This is probably due to the higher light concentration and steep light profiles.
These results demonstrate that the clustering analysis indeed yields reasonable results for realistic data sets. Furthermore, the results for three clusters are very similar to the clustering scheme of Sect. \ref{sect:testbed_FON_EON_ELL}. However, three clusters are not enough to describe the data faithfully. This is evident from the much larger SSR value for $K=3$ compared to $K=8$ and from Fig. \ref{fig:Fuku_posterior_plane_K3}, where we show the resulting cluster posterior space for $K=3$. Large parts of the available posterior space remain empty whereas the central region is crowded. This behaviour is due to the lack of complexity in the bipartite-graph model and strongly suggests that more clusters are necessary.
\begin{figure}
\includegraphics[width=1.6cm]{cluster_1_1}
\includegraphics[width=1.6cm]{cluster_1_2}
\includegraphics[width=1.6cm]{cluster_1_3}
\includegraphics[width=1.6cm]{cluster_1_4}
\includegraphics[width=1.6cm]{cluster_1_5}
${}$\\*
\includegraphics[width=1.6cm]{posteriors_K1_ex1}
\includegraphics[width=1.6cm]{posteriors_K1_ex2}
\includegraphics[width=1.6cm]{posteriors_K1_ex3}
\includegraphics[width=1.6cm]{posteriors_K1_ex4}
\includegraphics[width=1.6cm]{posteriors_K1_ex5}
${}$\\*
\includegraphics[width=1.6cm]{cluster_2_1}
\includegraphics[width=1.6cm]{cluster_2_2}
\includegraphics[width=1.6cm]{cluster_2_3}
\includegraphics[width=1.6cm]{cluster_2_4}
\includegraphics[width=1.6cm]{cluster_2_5}
${}$\\*
\includegraphics[width=1.6cm]{posteriors_K2_ex1}
\includegraphics[width=1.6cm]{posteriors_K2_ex2}
\includegraphics[width=1.6cm]{posteriors_K2_ex3}
\includegraphics[width=1.6cm]{posteriors_K2_ex4}
\includegraphics[width=1.6cm]{posteriors_K2_ex5}
${}$\\*
\includegraphics[width=1.6cm]{cluster_3_1}
\includegraphics[width=1.6cm]{cluster_3_2}
\includegraphics[width=1.6cm]{cluster_3_3}
\includegraphics[width=1.6cm]{cluster_3_4}
\includegraphics[width=1.6cm]{cluster_3_5}
${}$\\*
\includegraphics[width=1.6cm]{posteriors_K3_ex1}
\includegraphics[width=1.6cm]{posteriors_K3_ex2}
\includegraphics[width=1.6cm]{posteriors_K3_ex3}
\includegraphics[width=1.6cm]{posteriors_K3_ex4}
\includegraphics[width=1.6cm]{posteriors_K3_ex5}
\caption{Top example objects for $K=3$ clusters.}
Each row corresponds to a cluster. We also show the distribution of its cluster posteriors beneath each object. Cluster 1 seems to contain face-on disks, cluster 2 compact objects, and cluster 3 edge-on disks.
\label{fig:results_Fuku_K3}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{posterior_space_K3}
\caption{Cluster posterior space for $K=3$.}
\label{fig:Fuku_posterior_plane_K3}
Projected cluster posteriors are displayed 10\% translucent such that their density becomes visible. See Fig. \ref{fig:3_class_testbed_posterior_planes} for an explanation of the topology of this plot.
\end{figure}
For $K=8$ we have $\textrm{SSR}\approx 2,254$ (cf. Table \ref{table:fit_results_small_sample}), which corresponds to an RMS residual of $\approx 2.2\%$ for the similarity-matrix reconstruction. We show ten top example objects for each cluster in Fig. \ref{fig:results_Fuku_K8}. First, we notice that the resulting grouping is excellent. However, it is difficult to understand the differences between some clusters. Clusters 1 and 5 are obviously objects with high ellipticities, e.g. edge-on disks, but what is their difference? Is it the bulge dominance which is much weaker in cluster 1 than in 5? Do the clusters differ in their radial light profiles? What is the difference between clusters 2 and 7 which are both face-on disks? Of particular interest are clusters 3 and 8, where both seem to contain roundish and compact objects. However, the posterior histograms reveal a highly asymmetric relation: Objects from cluster 3 also prefer cluster 8 above all other clusters. Nevertheless, most of the top examples of cluster 8 have extremely low posteriors in cluster 3, i.e. association with cluster 3 is highly disfavoured. Although we cannot explain this result without further investigation, it is interesting that the algorithm picked up such a distinctive signal.
\begin{figure*}
\includegraphics[width=17cm]{fig_results_Fuku_K8}
\caption{Top example objects for $K=8$ clusters.}
Each row corresponds to a cluster. For each object, we also show the histogram of the distribution of its cluster posteriors beneath it. The objects were aligned in shapelet space, not in real space.
\label{fig:results_Fuku_K8}
\end{figure*}
As we have access to the isophotal axis ratio and the concentration index (cf. Eq. (\ref{eq:concentration_index})) for all objects, we investigate their distributions for the clusters. Figure \ref{fig:Fuku_cluster_axis_vs_concentration} shows the mean axis ratios and the mean concentration indices for all eight clusters averaged over the 100 top examples. The cluster with the highest mean axis ratio is cluster 1, which is the cluster of edge-on disk galaxies. The cluster with lowest concentration index is cluster 7, which is the cluster of face-on disk galaxies that exhibit extended smooth light profiles. Clusters 3, 4, 5 and 8 have the largest concentration indices. As is evident from Fig. \ref{fig:results_Fuku_K8}, these clusters are indeed composed of rather compact objects that seem to be elliptical galaxies. However, there is no decisive distinction in Fig. \ref{fig:Fuku_cluster_axis_vs_concentration}. This is not necessarily a flaw in the clustering results, but rather more likely caused by concentration and axis ratio being an insufficient parametrisation scheme \citep[cf.][]{Andrae2010b}.
\begin{figure}
\includegraphics[width=8cm]{Fuku_axis_ratio_concentration_clusters}
\caption{Mean axis ratios and concentration indices for the clusters of Fig. \ref{fig:results_Fuku_K8}.}
Weighted means were computed from the top 100 example objects of each cluster. We show the contours of $1\sigma$ and take into account possible correlations.
\label{fig:Fuku_cluster_axis_vs_concentration}
\end{figure}
It seems like the resulting classification scheme is essentially face-on disk, edge-on disk and elliptical. If we increase the number of clusters, we get further diversification that may be caused by bulge dominance or inclination angles. We emphasise again that our primary goal is to demonstrate that our method discovers morphological classes and provides data-to-class assignments that are reasonable.
\section{Conclusions\label{sect:conclusions}}
We briefly summarise our most important arguments and results:
\begin{itemize}
\item Galaxy evolution, the process of observation, and the experience with previous classification attempts strongly suggest a probabilistic (''soft'') classification. Hard classifications appear to be generically inappropriate.
\item There are two distance-based soft-clustering algorithms that have been applied to galaxy morphologies so far: Gaussian mixture models by \citet{Kelly2004,Kelly2005} and the bipartite-graph model by \citet{Yu2005} presented in this work. The weak points of the Gaussian mixture model are the dimensionality reduction and its assumption of Gaussianity. The weakness of the bipartite-graph model is the definition of the similarity measure.
\item The shapelet formalism, our similarity measure, and the bipartite-graph model produce reasonable clusters and data-to-cluster assignments for real galaxies. The automated discovery of classes corresponding to face-on disks, edge-on disks and elliptical galaxies without any prior assumptions is impressive and demonstrates the great potential of clustering analysis. Moreover, the automatically discovered classes have a qualitatively different meaning compared to pre-defined classes, since they represent to grouping that is preferred by the given data sample itself.
\item Random effects such as orientation angle and inclination are a major obstacle, since they introduce additional scatter into a parametrisation of galaxy morphologies.
\item For data sets containing $N$ galaxies, the computation times scale as $\mathcal O(N^2)$. Nevertheless, we experienced that a clustering analysis is feasible for data sets containing up to $N=10,000$ galaxies \textit{without} employing supercomputers. We conclude that a clustering analysis on a data set of one million galaxies is possible using supercomputers.
\item It is possible to enhance this method by setting up a classifier based on the classes found by the soft-clustering analysis, thereby improving the time complexity from $\mathcal O(N^2)$ to $\mathcal O(N)$.
\item The method presented in this paper is not limited to galaxy morphologies only. For instances, it could possibly be applied to automated star-galaxy classification or AGN detection.
\end{itemize}
The bottom line of this paper is that automatic discovery of morphological classes and object-to-class assignments (clustering analysis) does work and is less prejudiced and time-consuming than visual classifications, though the interpretation of the results is still an open issue. Especially when analysing new data samples for the first time, clustering algorithms are more objective than using pre-defined classes and visual classifications. The advantages of such a sophisticated statistical algorithm justify its considerable complexity.
\begin{acknowledgements}
RA thanks Coryn Bailer-Jones for his comments that have been particularly useful to improve this work. Furthermore, RA thanks Knud Jahnke and Eric Bell for helpful discussions especially on the interpretation of the results. RA is funded by a Klaus-Tschira scholarship. PM is supported by the DFG Priority Programme 1177.
\end{acknowledgements}
\bibliographystyle{aa}
\def\physrep{Phys. Rep.}%
\def\apjs{ApJS}%
\def\apj{ApJ}%
\def\aj{AJ}%
\def\aap{A\&A}%
\def\aaps{A\&AS}%
\def\mnras{MNRAS}%
|
\section{Introduction}
\label{intro}
The amount of influence which the solar magnetic cycle has on Earth's climate continues to be a matter of debate~\citep{lean-3069,2005ClDy25205W,moore-L17705,moore-34918,prsa:1452085,echera-41,li-1465111,2009ClDy321S}. The intriguing coincidence of the Maunder Minimum~\citep{eddy06181976,Luterbacher2001}, a period of exceptionally low solar activity as indicated by a dearth of sunspots observed from 1645 and 1715, and the coldest years of the Little Ice Age~\citep{Grove2001,mann:02}, a period of exceptionally low temperatures as observed in Europe, North America, and elsewhere~\citep{KJK:08291997,TCJ:012001,hmst:9749} ending in the middle of the 19th century, suggests that the temperature on Earth might display a dependence on the level of solar magnetic activity. Elucidating that dependence is this article's primary topic of investigation.
Using the renormalized continuous wavelet transform to provide a running estimate of the solar activity from the international sunspot number~\citep{rwj:jgr02,rwj:astro01}, here we extend its comparison to the yearly mean average temperature provided by the United States Historical Climatology Network, or USHCN~\citep{NDP:019,CDIAC:30,CDIAC:87}, for 1218 stations covering much of North America. The mean average temperature series as given does not come with an error estimate (though one may be derived from the underlying monthly measurements and is the subject of further investigation), thus one unit of variance is assumed throughout. From these data we extract the solar dependence and mean temperature for each station.
Our basic strategy is to replace the abscissa of time in the temperature record for a station with the corresponding level of solar magnetic activity, thereby determining from the parameters of a linear regression that station's temperature's dependence on solar activity as well as the mean temperature without regard to the activity level. Those parameters then become the ordinate against the three abscissas of station coordinates. The distribution of these parameters displays a correlation with the geophysical location of the station, particularly for latitude and elevation but not so much for longitude, and another linear regression reveals their dependence. The latitudinal variation of the solar dependence might be explained by an increased polar influx of energetic particles from the solar wind during periods of greater solar magnetic activity.
\section{Solar magnetic activity}
\label{sec:sma}
To evaluate the level of solar activity from the historical sunspot record, we use the renormalized continuous wavelet transform. The algorithm is discussed in detail in~\citet{rwj:jpamt01}, thus only some highlights are given here. Writing the Morlet wavelet at scale $s>0$ as the product of a real constant $C_s$, a Gaussian window $\Phi_s \equiv \exp (- \eta^2/2)$, and a normalized wave $\Theta_s \equiv \exp (i \omega_1 \eta)$ in terms of the parameter $\eta \equiv (t' - t) / s$, \begin{subeqnarray}
\psi_{s,t}(t') &\equiv& C_s \Phi_{s,t}(t') \Theta_{s,t}(t') \;, \\
&=& \sqrt{2} \pi^{-1/4} s^{-3/2} e^{- \eta^2/2} e^{i \omega_1 \eta} \;,
\end{subeqnarray} where $\omega_1 \approx 2 \pi$ is the central frequency of the mother wavelet at unity scale $s=1$ and zero offset $t=0$ which here spans $2 \chi + 1 = 13$ time units $\Delta t \equiv 1$, the normalization $C_s = (4 / \pi s ^6)^{1/4}$ produces a wavelet with norm $\sqrt{2}/s$ and a symmetric forward and inverse transform pair given a mean-subtracted signal $y(t) \rightarrow y(t) - \langle y(t) \rangle$ of duration $N_t$, \begin{eqnarray}
{\rm CWT}(s,t) &\equiv& \sum_{t'} \psi_{s,t}^*(t') y(t') \;,\\
{\rm ICWT}(t) &\equiv& Re \left[\sum_s \sum_{t'} \psi_{s,t}^*(t') {\rm CWT}(s,t') \Delta s \right] \;,
\end{eqnarray} {\it cf.} Equations (6) and (9) by~\citet{Frick:1997426}. One compensates for the wavelet truncation caused by the finite signal duration by taking $\psi_{s,t} \leftarrow \psi_{s,t} \sqrt{2 / s^2 \vert \psi_{s,t} \vert^2}$ which preserves the wavelet norm. The peak normalized power spectral density ${\rm PSD}(s,t) \equiv \vert \sqrt{2} {\rm CWT} \vert^2$ is twice the square of the amplitude of the CWT and displays in the instant wavelet power ${\rm IWP}_t(s) \equiv {\rm PSD}(s,t)$ response peaks whose integrated area equals the sum of the squared amplitudes for an infinite signal with sinusoidal components of stationary amplitude and period. The integrated instant power ${\rm IIP}(t) \equiv \sum_s {\rm PSD}(s,t) \Delta s$ then gives a running estimate of the instant signal power, and the mean wavelet power ${\rm MWP}(s) \equiv N_t^{-1} \sum_t {\rm PSD}(s,t)$ provides the net power spectrum. We have found satisfactory if not superior performance for data analysis when neglecting the admissibility condition~\citep{Frick:1997426,rwj:astro01} related to the DC response (nonzero mean) of the analyzing wavelet $\sum_{t'} \psi_{s,t}(t') \neq 0$, which remains negligibly small until the wavelet truncation becomes significant.
Since the invention of the telescope, solar astronomers have been able to keep an instrumental record of the magnetic flux erupting from the surface of the sun, which appears as darkened spots. First formulated by Wolf, the international sunspot number $R_i$ by the~\citet{sidc:online} continues the Z\"{u}rich number up until the present day~\citep{2002JAVSO3148H} and is known to display a tighter correlation with sunspot area and radio flux~\citep{hathaway-357} than the sunspot group number $R_g$ by~\citet{hoytschatten:98}. The CWT analysis of the mean-subtracted yearly smoothed sunspot number $R_i$ is presented in Figure~\ref{fig:A}, where the PSD has indicated in white the cone-of-influence, where wavelet compensation (which begins at the more restrictive cone-of-admissibility) becomes significant, as well as the scale of the signal duration at 309 years, beyond which lies the extremely low frequency ELF region. Prominent peaks in the MWP are indicated by the dotted lines. Its IIP gives the spot number activity SNA and outlines the magnitude of $R_i$ occuring at the maxima of the Schwabe cycle roughly every 11 years.
Underlying the sunspot number cycle is the solar magnetic activity discovered by~\citet{1908ApJ28315H} with a period of about 22 years and extending over the entire solar surface~\citep{1961ApJ133572B}. The magnetic field of the spots reverses polarity over one Hale cycle, thus the sunspot number represents in some sense a rectified version of the solar magnetic activity signal. In accord with the investigation by~\citet{BuckandMac:1992,BuckandMac:1993} and after a thorough evaluation of the arguments by~\citet{bracewell-53,bracewell-88}, we apply alternating signs to each of the Schwabe cycles in $R_i$ to form the derectified solar magnetic signal $\pm R_i$. Its CWT analysis, shown in Figure~\ref{fig:B}, displays a much simpler spectral content dominated by the Hale cycle and its third harmonic at about 7 years. Its IIP gives the solar magnetic activity SMA and is approximately double the SNA, with either denoted by SA. Both the SNA and the SMA compare well with the power estimate found by~\citet{cjaa:46578} determined from just the Schwabe cycle. While aware of the work by~\citet{1538:L154}, we maintain the traditional numbering of a tenth Schwabe cycle beginning about 1800.
\section{USHCN temperature record}
\label{sec:hcn}
Unfortunately, the North American historical climate record does not begin shortly after the invention of the thermometer but gradually picks up contributing stations throughout the 19th and early 20th centuries. For this analysis, we utilize the USHCN yearly mean average temperature data~\citep{NDP:019,CDIAC:30,CDIAC:87} as our climate signal. The USHCN stations cover much of North America, as depicted in Figure~\ref{fig:C}, spanning approximately 30 degrees of latitude, 60 degrees of longitude, and 3 kilometers of elevation. The density and extent of the station data reflects the duration of Western occupation as well as the accessibility of the location. The values carry a flag indicating where missing data has been replaced with an estimate from surrounding data, and we will consider both the full data set as well as a restricted set consisting only of values flagged ``good''.
In performing our linear fits, we adopt a prior which is uniform on the angle of the slope rather than its magnitude, summarizing our methodology for those unfamiliar with Bayesian data analysis expressed in terms of conditional probabilities~\citep{Durrett-1994,Sivia-1996}. Using notation $p(M \vert_I D) \equiv {\rm prob}(M \vert D, I)$ reading ``probability of $M$ given $D$ and information $I$'', we further abbreviate $p(M \vert_I D)$ to $p(M \vert D) \equiv \p{M}{D}$ and $p(M \vert I) \equiv \p{M}{}$ when the background information $I$ is unchanging. For model $M$ with parameters $\vec{m} \equiv \{m_j\}$ and data $\vec{D} \equiv \{D_k\}$, Bayes' Theorem allows one to write $\p{M}{\vec{D}} = \p{\vec{D}}{M} \p{M}{} / \p{\vec{D}}{}$ reading ``the posterior for $M \vert \vec{D}$ equals the product of the likelihood for $\vec{D} \vert M$ and the prior for $M$ divided by the evidence for $\vec{D}$'', where the denominator $\p{\vec{D}}{}$ affecting neither parameter estimation nor model selection is often omitted. The best estimate for the model parameters ${\vec{m}}_P$ is given by the maximum of the posterior $\p{\vec{m}}{\vec{D}} \propto \p{\vec{D}}{\vec{m}} \p{\vec{m}}{}$, whose logarithm (base $e$) is written as $L_P = L_L + L_{\vec{m}} - L_{\vec{D}}$, where the log evidence $L_{\vec{D}} = \#_{\vec{D}}$ is a constant.
For a uniform prior $L_{\vec{m}} = \#_{\vec{m}}$, the maximum of the posterior is at the maximum of the likelihood ${\vec{m}}_P = {\vec{m}}_L$, and for independent data $\p{D_k}{D_l, \vec{m}} = \p{D_k}{\vec{m}}$ with Gaussian noise $\{\sigma_k\}$, \begin{equation}
\p{\vec{D}}{\vec{m}} = \prod_k \p{D_k}{\vec{m}} = \prod_k (2 \pi \sigma_k^2)^{-1/2} \exp (-R_k^2/2) \;,
\end{equation} where $R_k = (M_k - D_k) / \sigma_k$ is the weighted residual for the $k$th datum, so that ${\vec{m}}_L$ minimizes the least-squares residual $\chi^2 \equiv \sum_k R_k^2$. For a linear model $M_k \equiv y_k = m x_k + c$ giving the ordinate over an abscissa $\vec{x}$, the solution $\{m,c\}_L$ is found by setting $-\vec{\nabla} L_L = \vec{\nabla} \chi^2/2 = 0$, where \begin{equation}
\frac{1}{2} \vec{\nabla} \chi^2 = \left[ \begin{array}{cc} S_{20} & S_{10} \\ S_{10} & S_{00} \end{array} \right] \left[ \begin{array}{c} m \\ c \end{array} \right] - \left[ \begin{array}{c} S_{11} \\ S_{01} \end{array} \right]
\end{equation} in terms of the weighted sums $S_{ab} \equiv \sum_k x_k^a D_k^b / \sigma_k^2$, and if one takes $x_k \rightarrow x'_k \equiv x_k - \langle \vec{x} \rangle$, the parameter $c_L$ returns the best estimate for the mean $\langle \vec{D} \rangle$ and the covariance matrix will be diagonal so that the error correlation $r_\sigma \equiv \sigma_{mc}^2 / (\sigma_{mm}^2 \sigma_{cc}^2)^{1/2}$ is zero, where the mean is computed with respect to weight $\sum_k \sigma_k^{-2}$. Linearity implies that the shape of the posterior is a Gaussian given solely by the second derivative, \begin{equation}
- \vec{\nabla} \vec{\nabla} L_L = \frac{1}{2} \vec{\nabla} \vec{\nabla} \chi^2 = \left[ \begin{array}{cc} S_{20} & S_{10} \\ S_{10} & S_{00} \end{array} \right] = \left[ \begin{array}{cc} \sigma_{mm}^2 & \sigma_{mc}^2 \\ \sigma_{mc}^2 & \sigma_{cc}^2 \end{array} \right]^{-1} \;,
\end{equation} whose width determines the covariance matrix. With unit variance on the data, the deviation of the parameters depends only upon the values of the abscissa, which we assume are known exactly~\citep{Press-1992,Sivia-1996}.
A prior which is uniform on the magnitude of the slope $m$ will not be uniform on the angle of the slope $\theta$. To show no preference for the slope of a line, one should take a prior which is uniform on its angle, and given unit bearing axes $\vec{x}$ and $\vec{y}$, that angle should be determined on the scale invariant axes $\vec{X}$ and $\vec{Y}$ to be defined. First, one takes $\vec{X} \equiv \vec{x}' / \Delta_X$ for $\Delta_X \equiv {\rm max}(\vert x'_k \vert)$ to scale the abscissa to $[-1,1]$, and then the ordinate $\vec{Y} \equiv \vec{y} / \Delta_Y$ for $\Delta_Y \equiv {\rm max}(\vert D_k - \langle \vec{D} \rangle \vert)$ is not centered but is scaled as if it were. On these axes, the line has slope $m' = m \Delta_X / \Delta_Y$ with angle $\theta = \arctan m'$, and it is on $\theta \in (-\pi/2, \pi/2)$ that we apply the uniform prior $\p{\theta}{} = 1/\pi$ to encompass slopes within the range $m \sim m' \in (-\infty,\infty)$. In terms of $m'$, one writes $\p{\theta}{} = \p{m'}{} \vert {\rm d}m' / {\rm d}\theta \vert$ to find $\p{m'}{} = [\pi (1 + m'^2)]^{-1}$ such that $-L_{m'} = \log (1 + m'^2) + \#_\pi$, and then in terms of $m$ the prior is Lorentzian, \begin{equation}
\p{m}{} = \left \lbrace \frac{\Delta_Y}{\Delta_X} \pi \left[1 + \left( \frac{\Delta_X}{\Delta_Y} m \right)^2 \right] \right \rbrace^{-1} \;,
\end{equation} with normalization $\int \p{m'}{} {\rm d}m' = \int \p{m}{} {\rm d}m = 1$. For comparison, a prior uniform on $m'$ will appear in $\theta$ as $\p{\theta}{} \propto 1 + (\tan \theta)^2$, which clearly displays a preference for a slope of extreme magnitude.
The prior nonuniform on $m \sim m'$ introduces a nonlinearity to the posterior, which becomes $L_P = L_L + L_m + L_c - L_{\vec{D}} = L_L + L_m + \#_{c, \vec{D}}$, as $\p{c}{} = 1/\Delta_c$ over a range of $\Delta_c$. To find $m'_P$, one must solve $\vec{\nabla} L_P = 0$, where \begin{subeqnarray}
-\vec{\nabla} L_P &=& \frac{1}{2} \vec{\nabla} \chi^2 + \vec{\nabla} \log (1+m'^2) \\
&=& \left[ \begin{array}{cc} S'_{20} & S'_{10} \\ S'_{10} & S'_{00} \end{array} \right] \left[ \begin{array}{c} m' \\ c' \end{array} \right] - \left[ \begin{array}{c} S'_{11} \\ S'_{01} \end{array} \right] + \left[ \begin{array}{c} 2 m' / (1 + m'^2) \\ 0 \end{array} \right] \;,
\end{subeqnarray} where $S'_{ab}$ are the weighted sums and $\{m',c'\}$ are the model parameters on the scale invariant axes $\vec{X}$ and $\vec{Y}$. The equation for $c'$ remains linear, leaving a cubic equation to solve for $m'$. The appropriate real root is found between $m'_L$ and 0, and the solution is then scaled $\{m',c'\}_P \rightarrow \{m,c\}_P$, as is the covariance matrix $- [\vec{\nabla} \vec{\nabla} L_P]^{-1}$, which differs from $- [\vec{\nabla} \vec{\nabla} L_L]^{-1}$ only by $S'_{20} \rightarrow S'_{20} + 2 (1 - m'^2) / (1 + m'^2)^2$. The nonlinear term introduces higher order corrections to the shape of the posterior, which remains very close to a Gaussian unless dominated by the prior.
The net effect of the Lorentzian prior is to prevent one from overestimating the magnitude of the slope best supported by the evidence. The estimate for the mean remains unaffected. In the limit of poor data $\vert \vec{\nabla} L_m \vert \gg \vert \vec{\nabla} L_L \vert$, the best estimate for the slope $m_P$ is driven from $m_L$ towards zero. We can see the effect of the prior on a fit to three points with increasing variance in Figure~\ref{fig:D}, where the maximum likelihood estimate is a dashed line. One should consider how often one sees maximum likelihood used for cases such as (c) and (d), where the latter exhibits insufficient evidence for a significant slope.
After so much digression, let us look at some data. For each of the 1218 stations indexed by $j$ we take its temperature record and replace the abscissa of time with the value for the solar activity of that year (normalized by $10^5$). In Figure~\ref{fig:E} we display the linear regression for two locations chosen at random using the SMA and the full data set, and in Table~\ref{tab:A} we give the corresponding parameters for all four combinations of SA and data set. The mean station temperature $T_0$, indicated by the heavy dot in the figure, is also the value of the line \begin{equation}
T^j({\rm SA}) = \Upsilon_{\rm SA}^j \times {\rm SA'} + T_0^j
\end{equation} at the mean of ${\rm SA}$ and indicates the temperature expected at that location without regard to solar magnetic activity. The solar dependence $\Upsilon$ gives the correlation of the mean average temperature with the solar magnetic activity and for any particular station may be in either direction, as seen in Figure~\ref{fig:E}.
There remains to answer the question of how much evidence there is for a nonvanishing solar dependence $\Upsilon$. One cannot simply ask how good is the fit, as the only logical answer is, compared to what? Bayesian data analysis frames the question in terms of model selection using the ratio of posteriors $P \equiv \p{B}{\vec{D}} / \p{A}{\vec{D}} = \p{\vec{D}}{B}\p{B}{} / \p{\vec{D}}{A}\p{A}{} = (\p{B}{} / \p{A}{}) (\p{\vec{D}}{B} / \p{\vec{D}}{A})$, where here model A has the single parameter $T_0$ and model B has two parameters $T_0$ and $\Upsilon$. With no prior preference for either model $\p{B}{} / \p{A}{} = 1$, that ratio becomes \begin{equation}
P \rightarrow \frac{\p{\vec{D}}{B}}{\p{\vec{D}}{A}} = \frac{\int \int \p{\Upsilon}{B} \p{\vec{D}}{T_0,\Upsilon,B} {\rm d}T_0{\rm d}\Upsilon}{\int \p{\vec{D}}{T_0,A} {\rm d}T_0} \;,
\end{equation} where a common factor of $\p{T_0}{} = 1 / \Delta_{T_0}$ has been cancelled, whose logarithm (base 10) we write $P_{10}$. For the remaining prior factor we restrict the range of the angle to $\pi/4$ so that $\p{\Upsilon}{B} = 4 \p{m}{}(m \rightarrow \Upsilon)$ by arguing that any child could draw a line through a given set of dots with a slope correct to within a slice of $\pi$, and in Table~\ref{tab:B} we give heuristic descriptions of several prior ranges which could be assigned. The question essentially boils down to determining how much to penalize model B for having an extra parameter, and we feel that half a quadrant is sufficient, as without the prior factor model B would always be preferred. In the numerator we take the quadratic approximation (which is exact for model A), $\p{\vec{D}}{B} \simeq 2 \pi \sigma_{\Upsilon \Upsilon} \sigma_{T_0 T_0} \exp (-\chi^2_{T_0,\Upsilon}/2) \p{\Upsilon}{B}$ evaluated at the posterior maximum $\{T_0, \Upsilon\}_P$, after verifying its accuracy numerically. In Figure~\ref{fig:F} we display $P_{10}$ for all stations using the SMA and the full data set, and in Table~\ref{tab:C} we give its mean and median for both SA and data sets. We admit a preference for the SMA on the basis of the polarity associated with the Hale cycle, and we state subjectively that there is sufficient evidence to ascribe physical significance to the solar dependence $\Upsilon$.
\section{Results}
\label{sec:reslt}
Collecting our parameters into two sets $\{T_0^j\}$ and $\{\Upsilon^j\}$, we proceed by plotting the ordered pairs $(T_0^j, X_i^j)$ and $(\Upsilon^j, X_i^j)$ for $i \in \{1,2,3\}$ and $\vec{X}^j \equiv ({\rm LAT}^j, {\rm LONG}^j, {\rm ELEV}^j)$ and perform another linear regression. With so many points contributing, the prior here makes not much difference, but we maintain its use in our evaluation. Beginning with the mean station temperature $T_0$, we plot its dependence against station location $\vec{X}^j$ in Figure~\ref{fig:G} for the SMA and full data set, and in Table~\ref{tab:D} are the results of the linear regression for all the combinations. The column labelled $\sqrt{\langle \chi^2 \rangle}$ gives the rms weighted residual, {\it ie} the expected distance between the fit and a data point. We see that $T_0$ displays the tightest correlation with latitude, as confirmed by the magnitude of Pearson's $r \equiv r_P$. Its dependence on latitude and elevation is not unexpected, and its dependence on longitude is minor, most likely the result of correlation among the $\vec{X}^j$ representing the surface of the Earth. The units (hm) of elevation are chosen so that its range is approximately equal to that of latitude, thus the relative dependence given by their slopes may be compared directly. We see that the decrease in $T_0$ attributed to 1 degree of increasing latitude equates roughly to that attributed to 2 hm of increasing elevation and that the magnitudes of their $r_P$ are in about the same ratio. The choice of solar activity signal does not affect $T_0$, and the choice of full or good temperature data has little impact on the results.
The solar dependence $\Upsilon$ also displays a dependence on location, which again is strongest for latitude. Figure~\ref{fig:H} displays the SMA and full data set results, and Table~\ref{tab:E} gives the entire set of results. The station averaged $\Upsilon$ is consistently positive, and we note a paucity of negative values at the lowest latitudes surveyed. Dots for stations with little evidence for a slope are shaded darker and confined to a band around $\Upsilon = 0$. The solar magnetic activity SMA results display a tighter correlation with location than those of the spot number activity SNA, and its units were chosen to make the visible range of $\Upsilon$ roughly the same as that of $T_0$ so that their relative slopes may also be compared. In these units, the dependence of $\Upsilon$ on latitude is about a third the magnitude and opposite in direction to that for $T_0$, and its $r_P$ has been reduced also by about a third. Its dependence on elevation is a third of that for latitude, and its variation with longitude is negligible. The tighter correlation and smaller errors and residual indicate that the solar magnetic activity derived from the derectified sunspot number provides a more physical signal than the activity derived from the usual (rectified) sunspot number. The choice of temperature data set does impact the results, with the full data set indicating a greater solar dependence.
\section{Discussion}
\label{discus}
Comparison of our work to others' is made difficult because so few investigators consider the derectified sunspot number in their analysis of solar magnetic activity and most employ a cross coherence analysis looking for cycles common to both the historical sunspot record and various indicators of climate~\citep{barnston-1295,zhaohan-42189,piscaron-1661,weber-917,barlyaeva-2009,2009ClDy321S}. Our methodology here is quite different, eliciting a relationship between the net solar magnetic activity derived from the instant power carried by the derectified sunspot signal and the temperature recorded across much of North America for almost a century and a half. What that data lacks in duration compared to the Central England Temperature record~\citep{rwj:jgr02,rwj:astro01} it more than makes up for in breadth of geophysical location. We note that our results are in accord with other investigators~\citep{fcl-11011991,lean-3069,echera-41} and contrast with those who have found negligible evidence of solar forcing on Earth's climate~\citep{moore-L17705,moore-34918,prsa:1452085,li-1465111}. Most authors, when speaking of the influence of solar forcing on Earth's climate, are actually discussing the effect of changing some model parameters in a {\it simulation} of meteorological systems~\citep{shind-4094S,2005ClDy25205W}, whereas here we are discussing a correlation found between independent instrumental observations from the historical record.
While several mechanisms have been proposed to explain the possible coupling between solar activity and terrestrial climate~\citep{svens-8122,lean-3069,hameedlee-32}, notably variation in irradiance, modulation of cosmic ray influx, and changes in global circulation, we would like to consider an alternative explanation. From the dependence of $\Upsilon$ on latitude, we see that the degree of solar dependence increases with distance from the equator. Thus, any proposed model for coupling should account for the greater dependence on solar magnetic activity towards the polar regions. One effect of increased solar activity is an increased solar wind~\citep{webb94-4201}, whose temperature $10^5$K$\sim 10$eV is sufficient for ionization to the plasma state. Those charged particles follow the lines of Earth's geomagnetic field emanating from the poles to produce the auroral illuminations and heating of the upper atmosphere. By the well known conservation law, that energy must be deposited somewhere, and we propose that during times of increased solar activity, an increased influx of energy bearing solar wind particles produces a preferential heating towards the polar regions. Whether that prediction will be supported by more detailed evaluations remains to be investigated. The first step in such an evaluation would be to verify the relation between the energy flux from the solar wind and the solar magnetic activity, such as one can do for the 10.7cm radio flux~\citep{rwj-jaa01}, to put the historical sunspot record onto an axis bearing units relevant to geophysical analysis. Then, with an estimate of the historical thermal flux variation, one could look for correlations with various climate indicators.
\section{Conclusions}
\label{concl}
A running estimate of the net solar magnetic activity is derived from the integrated instant power of the historical sunspot record, derectified to account for the opposite polarities of adjacent Schwabe cycles, using the renormalized continuous wavelet transform. The net solar magnetic activity provides the abscissa for a linear regression using a Lorentzian prior of temperatures observed at stations covering much of North America for a duration exceeding one century. A relationship is found between the level of solar activity and the observed temperature characterized by the slope of the linear regression which itself depends upon geophysical location. The solar dependence is found to increase with latitude, suggesting a coupling mechanism to the polar regions of thermal flux from the solar wind.
\section*{Acknowledgements}
International sunspot number provided by the SIDC-team, World Data Center for the Sunspot Index, Royal Observatory of Belgium, Monthly Report on the International Sunspot Number, online catalogue of the sunspot index: http://www.sidc.be/sunspot-data/, 1700--2008. Temperature data provided by the United States Historical Climatology Network, National Climatic Data Center, National Oceanic and Atmospheric Administration, http://cdiac.ornl.gov/epubs/ndp/ushcn/ushcn.html. AlphaWavelet Toolbox Version 1.0 is available at http://www.alphawaveresearch.com.
|
\section{Introduction}
Magnetic monopoles, the topological soliton solutions of
Yang-Mills-Higgs gauge theories in three space dimensions with
particle-like properties, have been the subject of considerable
interest over the years. BPS monopoles, arising from a limit in
which the the Higgs potential is removed but a remnant of this
remains in the boundary conditions, satisfy the first order
Bogomolny equation
$$B_i=\frac{1}{2}\sum_{j,k=1}\sp3\epsilon_{ijk}F\sp{jk}=D_i\Phi$$
and have merited particular attention (see \cite{ms04} for a recent
review). This focus is in part due to the ubiquity of the Bogomolny
equation. Here $F_{ij}$ is the field strength associated to a gauge
field $A$, and $\Phi$ is the Higgs field. We shall focus on the case
when the gauge group is $SU(2)$. The Bogomolny equation may be
viewed as a dimensional reduction of the four dimensional self-dual
equations upon setting all functions independent of $x_4$ and
identifying $\Phi=A_4$; they are also encountered in supersymmetric
theories when requiring certain field configurations to preserve
some fraction of supersymmetry. The study of BPS monopoles is
intimately connected with integrable systems. Nahm gave a transform
of the ADHM instanton construction to produce BPS monopoles
\cite{nahm82} and the resulting Nahm's equations have Lax form with
corresponding spectral curve $\hat{\mathcal{C}}$. This curve,
investigated by Corrigan and Goddard \cite{cg81}, was given a
twistorial description by Hitchin \cite{hitchin82} where the same
curve lies in mini-twistor space, $\hat{\mathcal{C}}\subset$
T$\mathbb{P}\sp1$. Just as Ward's twistor transform relates
instanton solutions on $\mathbb{R}\sp4$ to certain holomorphic
vector bundles over the twistor space $\mathbb{CP}\sp3$, Hitchin
showed that the dimensional reduction leading to BPS monopoles could
be made at the twistor level as well and was able to prove that all
monopoles could be obtained by this approach \cite{hitchin83}
provided the curve $\hat{\mathcal{C}}$ was subject to certain
nonsingularity conditions. Bringing methods from integrable systems
to bear upon the construction of solutions to Nahm's equations for
the gauge group $SU(2)$ Ercolani and Sinha \cite{es89} later showed
how one could solve (a gauge transform of) the Nahm equations in
terms of a Baker-Akhiezer function for the curve
$\hat{\mathcal{C}}$.
Although many general results have now been obtained few explicit
solutions are known. This is for two reasons, each coming from a
transcendental constraint on the curve $\hat{\mathcal{C}}$. The
first is that the curve $\hat{\mathcal{C}}$ is subject to a set of
constraints whereby the periods of a meromorphic differential on
the curve are specified. This type of constraint arises in many
other settings as well, for example when specifying the filling
fractions of a curve in the AdS/CFT correspondence. Such constraints
are transcendental in nature and until quite recently these had only
been solved in the case of elliptic curves (which correspond to
charge $2$ monopoles). In \cite{bren06,bren07} they were solved for
a class of charge $3$ monopoles using number theoretic results of
Ramanujan. The second type of constraint is that the linear flow on
the Jacobian of $\hat{\mathcal{C}}$ corresponding to the integrable
motion only intersects the theta divisor in a prescribed manner. In
the monopole setting this means the Nahm data will yield regular
monopole solutions but a similar constraint also appears in other
applications of integrable systems. In Hitchin's approach (reviewed
below) this may be expressed as the vanishing of a real one
parameter family of cohomologies of certain line bundles,
$H^0(\hat{\mathcal{C}},L^{\lambda}(n-2))=0$ for $\lambda\in(0,2)$.
Viewing the line bundles as points on the Jacobian this is
equivalent to a real line segment not intersecting the theta divisor
$\Theta$ of the curve. Indeed there are sections for $\lambda=0,2$
and the flow is periodic (mod $2$) in $\lambda$ and so we are
interested in the number of times a real line intersects $\Theta$.
While techniques exist that count the number of intersections of a
complex line with the theta divisor we are unaware of anything
comparable in the real setting and again solutions have only been
found for particular curves \cite{bren09}. Thus the application of
integrable systems techniques to the construction of monopoles and
(indeed more generally) encounters two types of problem that each
merit further study.
The present paper will use symmetry to reduce these problems to
ones more manageable. Long ago monopoles of charge $n$ with cyclic
symmetry $\texttt{C}_n$ were shown to exist \cite{or82} and more
recently such monopoles were reconsidered \cite{hmm95} from a
variety of perspectives. The latter work indeed considered the
case of monopoles with more general Platonic symmetries and for
the case of tetrahedral, octahedral and icosahedral symmetry
(where such monopoles exist) the curves were reduced to elliptic
curves. (See \cite{hs96a, hs96b, hs97} for development of this
work.) Our first result is to strengthen work of Sutcliffe
\cite{sut96}. Motivated by Seiberg-Witten theory Sutcliffe gave an
ansatz for $\texttt{C}_n$ symmetric monopoles in terms of $su(n)$
affine Toda theory. The spectral curve $\hat{\mathcal{C}}$ of a
$\texttt{C}_n$ symmetric monopole yields an $n$-fold unbranched
cover of the hyperelliptic spectral curve $\mathcal{C}$ of the
affine Toda theory, a spectral curve that arises in Seiberg-Witten
theory describing the pure gauge $\mathcal{N}=2$ supersymmetric
$su(n)$ gauge theory. (We shall recall some properties of the Nahm
construction and this relation between curves in section 2.)
Sutcliffe's ansatz (section 3) shows how solutions to the affine
Toda equations yield cyclically symmetric monopoles. Our first
result proves that any cyclically symmetric monopole is gauge
equivalent to Nahm data given by Sutcliffe's ansatz, and so
obtained from the affine Toda equations. We mention that Hitchin
in an unpublished note had, prior to Sutcliffe, observed that
cyclic charge 3 monopoles were equivalent to solutions of the
affine Toda equations. The remainder of this paper shows that the
relation between the Nahm data and the affine Toda system is much
closer than simply that they yield the same equations of motion.
The solution of an integrable system is typically expressed in
terms of the straight line motion on the Jacobian of the system's
spectral curve. Such a line is determined both by its direction
and a point on the Jacobian. We shall show that both the direction
(given by the Ercolani-Sinha vector, section 4) and point relevant
for monopole solutions (section 5) are obtained as pull-backs of
Toda data. This connection is remarkable and ties the geometry
together in a very tight manner. Section 6 recalls a theorem of
Accola and Fay that holds in precisely this setting, showing how
the theta-functional solutions of the monopole reduce to precisely
the theta-functional solutions of Toda. At this stage we have
reduced the problem of constructing cyclically symmetric monopoles
to one of determining hyperelliptic curves that satisfy the
transcendental constraints described above. Though more manageable
the problems are still formidable and a construction in the charge
$3$ setting will be described elsewhere \cite{bde10}. We conclude
with a discussion.
\section{Monopoles}
We shall briefly recall the salient features for constructing
$su(2)$ monopoles of charge $n$. We begin with Nahm's construction
\cite{nahm82}. In generalizing the ADHM construction of instantons
Nahm established an equivalence between nonsingular monopoles and
what is now referred to as Nahm data: three $n\times n$ matrices
$T_i(s)$ with $s\in[0,2]$ satisfying
\begin{itemize}
\item[\bf{N1}] Nahm's equation
\begin{equation}
\frac{dT_i}{ds}=\frac{1}{2}\sum_{j,k=1}^3\epsilon_{ijk}[T_j,T_k],
\end{equation}
\item[\bf{N2}] $T_i(s)$ is regular for $s\in(0,2)$ and has simple
poles at $s=0$ and $s=2$, the residues of which form an irreducible
$n$-dimensional representation of $su(2)$,
\item[\bf{N3}]
$\displaystyle{T_i(s)=-T_i^\dagger(s),\qquad T_i(s)=T_i^t(2-s). }$
\end{itemize}
Upon defining
\begin{align*}
A(\zeta)&=T_1+iT_2-2iT_3\zeta+(T_1-iT_2)\zeta\sp2\\
M(\zeta)&=-iT_3+(T_1-iT_2)\zeta \intertext{we find that Nahm's
equation is equivalent to the Lax equation}
\frac{dT_i}{ds}&=\frac{1}{2}\sum_{j,k=1}^3\epsilon_{ijk}[T_j,T_k]
\Longleftrightarrow
[\dfrac{d}{ds}+M,A]=0.
\end{align*}
Here $\zeta$ is a spectral parameter. Following from the Lax
equation we have the invariance of the spectral curve
\begin{equation}\label{spectcurve}
\hat{\mathcal{C}}:\ 0=P(\eta,\zeta):=\det(\eta 1_n+A(\zeta))
\end{equation}
where
\begin{equation}\label{spectcurveP}
P(\eta,\zeta)=\eta\sp{n}+a_1(\zeta)\eta\sp{n-1}+\ldots+a_n(\zeta),\qquad
\deg a_r(\zeta)\le2r.\end{equation} As with any spectral curve
presented in the form (\ref{spectcurve}) one should always ask
where $\hat{\mathcal{C}}$ lies. Typically the spectral curve lies
in a surface, $\hat{\mathcal{C}}\subset\mathcal{S}$, and
properties of the surface are closely allied with the integrable
system encoded by the Lax equation. (For example, for suitable
surfaces, the separation of variables may be described by
$\mathop{\rm Hilb}\nolimits\sp{[N]}(\mathcal{S})$, the Hilbert scheme of points on
$\mathcal{S}$.) For the case at hand
\begin{equation*}
{\hat{\mathcal{C}}
\subset
T\mathbb{P}\sp1:=\mathcal{S},\qquad (\eta,\zeta)\rightarrow
\eta\frac{d}{d\zeta}\in T\mathbb{P}\sp1,
\end{equation*}
and monopoles admit a minitwistor description: the curve
$\hat{\mathcal{C}}$ corresponds to those lines in $\mathbb{R}\sp3$
which admit normalizable solutions of an appropriate scattering
problem in both directions \cite{hitchin82, hitchin83}. This latter
description makes clear that $\hat{\mathcal{C}}$ comes equipped with
an antiholomorphic involution or real structure coming from the
reversal of orientation of lines
$$(\eta,\zeta)\rightarrow(-\bar\eta/{\bar\zeta}\sp2,-1/{\bar\zeta}).$$
This means the coefficients of (\ref{spectcurveP}) are such that
\begin{equation}\label{reality}
{a_r(\zeta)=(-1)\sp{r}\zeta\sp{2r}\overline{a_r(-1/{\overline\zeta}\,)}}
\end{equation}
and so each may be expressed in terms of $2r+1$ (real) parameters
$$ a_r(\zeta)= \chi_r\, \left[\prod_{l=1}\sp{r}\left(
\frac{\overline{\alpha}_{r,l}}{\alpha_{r,l}}\right)\sp{1/2}\right]
\prod_{k=1}\sp{r}(\zeta-\alpha_{r,k})(\zeta+\frac{1}{\overline{\alpha}_{r,k}}),\qquad
\alpha_{r,k}\in \mathbb{C},\ \chi_r\in\mathbb{R}. $$ We remark that a
real structure constrains the form of the period matrix of a curve
and that while in general there may be between $0$ and $\hat g+1$
ovals of fixed points of an antiholomorphic involution ($\hat g$
being the genus of $\hat{\mathcal{C}}$) for the case at hand there
are no fixed points. For the monopole spectral curve
(\ref{spectcurveP}) we have (generically) $\hat g=(n-1)\sp2$.
Although in many situations the solution of the integrable system
encoded by a Lax pair (with spectral parameter) only depends on
intrinsic properties of the spectral curve the monopole physical
setting means that extrinsic properties of our curve in
$T\mathbb{P}\sp1$ are relevant here. Spatial symmetries act on the
monopole spectral curve via fractional linear transformations.
Although a general M\"obius transformation does not change the
period matrix of a curve $\hat{\mathcal{C}}$ only the subgroup
$PSU(2)< PSL(2,\mathbb{C})$ preserves the reality properties
necessary for a monopole spectral curve. These reality conditions
are an extrinsic feature of the curve (encoding the space-time
aspect of the problem) whereas the intrinsic properties of the curve
are invariant under birational transformations or the full M\"obius
group. Such extrinsic aspects are not a part of the usual integrable
system story. Thus $SO(3)$ spatial rotations induce an action on
$T{\mathbb P}\sp1$ via $PSU(2)$: if $\begin{pmatrix} p&q\\-\bar q&\bar
p\end{pmatrix}\in PSU(2)$, ($ |p|\sp2+|q|\sp2=1$) then
\begin{equation}\label{pstransf}
\zeta\rightarrow\tilde\zeta:=\dfrac{\bar p\, \zeta-\bar q}{q\,
\zeta+p}, \qquad \eta\rightarrow\tilde\eta:= \dfrac{\eta}{(q\,
\zeta+p)\sp2}
\end{equation}
corresponds to a rotation by $\theta$ around ${\mathbf n}\in S\sp2$
where $$ n_1\sin{(\theta/2)}=\mathop{\rm Im}\nolimits q,\ n_2\sin{(\theta/2)}=-\mathop{\rm Re}\nolimits q,\
n_3\sin{(\theta/2)}=\mathop{\rm Im}\nolimits p,\
\cos{(\theta/2)}=-\mathop{\rm Re}\nolimits p.
$$
This $SO(3)$ action commutes with the standard real structure on
$T{\mathbb P}\sp1$. The action on the spectral curve may be expressed as
\begin{equation}
P(\tilde\eta,\tilde\zeta)=\dfrac{{\tilde P}(\eta,\zeta)}{(q\,
\zeta+p)\sp{2n}}, \qquad {\tilde
P}(\eta,\zeta)=\eta\sp{n}+\sum_{r=1}\sp{n}\eta\sp{n-r}\,{\tilde
a}_r(\zeta),
\end{equation}
where in terms of the parameterization above
$$
a_r(\zeta)\rightarrow \frac{{\tilde a}_r(\zeta)}{(q\,
\zeta+p)\sp{2r}}\equiv \frac{{\tilde\chi}_r}{(q\, \zeta+p)\sp{2r}}
\left[\prod_{l=1}\sp{r}\left(
\frac{\overline{\tilde\alpha}_l}{\tilde\alpha_l}\right)\sp{1/2}\right]
\prod_{k=1}\sp{r}(\zeta-{\tilde\alpha}_k)(\zeta+\frac{1}{\overline{\tilde\alpha}_k})
$$
with
$$\alpha_k\rightarrow {\tilde\alpha}_k \equiv\frac{p\alpha_k+{\bar
q}}{{\bar p}-\alpha_k q},\qquad \chi_r \rightarrow
{\tilde\chi}_r\equiv \chi_r \prod_{k=1}\sp{r}\left[\frac{({\bar
p}-\alpha_k q)({ p}-{\bar\alpha_k}{\bar q})({\bar\alpha_k}{\bar
p}+q)(\alpha_k{ p}+ {\bar
q})}{{\alpha_k}{\bar\alpha_k}}\right]\sp{1/2}.$$ In particular the
form of the curve does not change under a rotation: that is, if
$a_r=0$ then so also ${\tilde a}_r=0$.
Hitchin, Manton and Murray \cite{hmm95} showed how curves
invariant under finite subgroups of $SO(3)$ or their binary covers
yield symmetric monopoles. Suppose we have a symmetry; the
spectral curve $0=P(\eta,\zeta)$ is transformed to the same curve,
$0=P(\tilde\eta,\tilde\zeta)= {{\tilde P}(\eta,\zeta)}/{(q\,
\zeta+p)\sp{2n}}$. Then $P(\eta,\zeta)={\tilde P}(\eta,\zeta)$, or
equivalently $a_r(\zeta)={\tilde a}_r(\zeta)$. Relevant for us is
the example of cyclically symmetric monopoles. Let
$\omega=\exp(2\pi i/n)$. A rotation of order $n$ is then given by
$\bar p=\omega\sp{1/2}$, $q=0$ which yields
$$\phi:\ (\eta,\zeta)\rightarrow (\omega\eta,\omega\zeta).$$
Correspondingly $\eta\sp{i}\zeta\sp{j}$ is invariant for $i+j\equiv
0 \mod n$ and the spectral curve
$$\eta\sp{n}+a_1 \eta\sp{n-1}\zeta+a_2
\eta\sp{n-2}\zeta\sp2+\ldots+a_n\zeta\sp{n}+\beta\zeta\sp{2n}+\gamma=0$$
is invariant under the cyclic group $\texttt{C}_n$ generated by this
rotation. Imposing the reality conditions (\ref{reality}) and
centering the monopole (setting $a_1=0$) then gives us the spectral
curve in the form \begin{equation}\label{cycliccurve}\eta\sp{n}+a_2
\eta\sp{n-2}\zeta\sp2+\ldots+a_n\zeta\sp{n}+\beta\zeta\sp{2n}+(-1)\sp{n}\bar\beta=0,\qquad
a_i\in\mathbf{R} \end{equation} and by an overall rotation we may
choose $\beta$ real.
Now the $\texttt{C}_n$-invariant curve $\hat{\mathcal{C}}$
(\ref{cycliccurve}) of genus $\hat g=(n-1)^2$ is an $n$-fold
unbranched cover of a genus $g=n-1$ curve $\mathcal{C}$. The
Riemann-Hurwitz theorem yields the relation $\hat g =n(g-1)+1$.
Introduce the rational invariants $x=\eta/\zeta$, $
\nu=\zeta\sp{n}\beta$, then
$$x\sp{n}+a_2 x\sp{n-2}+\ldots
+a_n+\nu+\frac{(-1)\sp{n}|\beta|\sp2}{\nu}=0$$ and upon setting
$y=\nu-{(-1)\sp{n}|\beta|\sp2}/{\nu}$ we obtain the curve
\begin{equation}\label{Todacurve}
y^2=(x\sp{n}+a_2 x\sp{n-2}+\ldots +a_n)\sp2-4(-1)\sp{n}|\beta|\sp2.
\end{equation}
This curve is the spectral curve of $su(n)$ affine Toda theory in
standard hyperelliptic form.
For future reference we note that the $n$-points $\hat\infty_j$
above the point $\zeta=\infty$ project to one of the infinite
points, $\infty_+$, of the curve (\ref{Todacurve}), while the
$n$-points above the point $\zeta=0$ project to the other infinite
point. At $\hat\infty_j$ we have ${\eta}/{\zeta}\sim\rho_j \zeta$ as
$ \zeta\sim\hat{\infty}_j$, with $\rho_j=\beta\sp{1/n}\exp(2\pi
i[j+1/2]/n)$.
\noindent{{\bf{{The $n=2$ example:}}} The reality conditions for
$n=2$ and $a_2(\zeta)=\beta \zeta^4+\gamma\zeta^2+\delta$ means
that $\delta=\bar\beta$ and $\gamma=\bar\gamma$ and
(\ref{cycliccurve}) becomes
$$\eta^2+\beta\zeta^4+\gamma\zeta^2+\bar\beta=0.$$ This is an
elliptic curve. If $\beta=|\beta|e\sp{2i\theta}$ let $U=\zeta
e\sp{i\theta}$ and $V=i\eta e\sp{i\theta}/|\beta|\sp{1/2}$ and
this may be rewritten as
\begin{equation}\label{examplen2a}
V^2=U^4+t\,U^2+1, \qquad
t=\gamma/|\beta|.
\end{equation}
For irreducibility $t\ne2$. Now the curve (\ref{Todacurve})
becomes (with $Y=y/\sqrt{\gamma^2-4|\beta|\sp2}$)
\begin{equation}\label{examplen2b}
Y^2=x^4+t'\,x^2+1, \qquad t'=\frac{2t}{\sqrt{t^2-4}}.
\end{equation}
These two curves (\ref{examplen2a}, \ref{examplen2b}) are
$2$-isogenous: if we quotient the former curve under the
involution $(U,V)\rightarrow(-U,-V)$ we obtain the latter.
\section{The Sutcliffe Ansatz}
Some years ago Sutcliffe \cite{sut96} introduced the following
ansatz for cyclically symmetric monopoles. Let
\begin{align}
T_1+iT_2&=\begin{pmatrix} 0&e\sp{(q_1-q_2)/2}&0&\ldots&0\\
0&0&e\sp{(q_2-q_3)/2}&\ldots&0\\
\vdots&&&\ddots&\vdots\\
0&0&0&\ldots&e\sp{(q_{n-1}-q_n)/2}\\
e\sp{(q_n-q_1)/2}&0&0&\ldots&0
\end{pmatrix}\\
T_1-iT_2&=-\begin{pmatrix}0&0&\ldots&0&e\sp{(q_n-q_1)/2}\\
e\sp{(q_1-q_2)/2}&0&\ldots&0&0\\
0&e\sp{(q_2-q_3)/2}&\ldots&0&0\\
\vdots&&\ddots&&\vdots\\
0&0&\ldots&e\sp{(q_{n-1}-q_n)/2}&0
\end{pmatrix}\\
T_3&=-\frac{i}{2}\begin{pmatrix} p_1&0&\ldots&0\\
0&p_2&\ldots&0\\
\vdots&&\ddots&\vdots\\
0&0&\ldots&p_n
\end{pmatrix}
\end{align}
where $p_i$, $q_i$ are real. Then $T_i(s)=-T_i^\dagger(s)$ and
Nahm's equations yield
\begin{align*}
&\frac{d}{ds}\left(T_1+iT_2\right)=i[T_3,T_1+iT_2]&\Rightarrow&
\begin{cases}p_1-p_2=\dot q_1-\dot q_2,\\
\qquad\vdots\\ p_n-p_1=\dot q_n-\dot q_1.
\end{cases}\\
&\frac{d}{ds}T_3=[T_1,T_2]=\frac{i}2[T_1+iT_2,T_1-iT_2]&\Rightarrow&
\begin{cases}
\dot p_1=-e\sp{q_1- q_2}+e\sp{q_n-q_1},\\
\qquad\vdots \\ \dot p_n=-e\sp{q_n- q_1}+e\sp{q_{n-1}-q_n}.
\end{cases}
\end{align*}
These equations then follow from the equations of motion of the
affine Toda Hamiltonian
\begin{equation}\label{affToda}
H=\frac12\left(p_1\sp2+\ldots+p_n\sp2\right)-\left[e\sp{q_1-
q_2}+e\sp{q_2- q_3}+\ldots+e\sp{q_{n}- q_1} \right] .
\end{equation}
Sutcliffe's observation is that particular solutions of these
equations will then yield cyclically invariant monopoles. In fact
the monopole Lax operator $A(\zeta)$ here is essentially the usual
Toda Lax operator and $$\frac12{\rm Tr} A(\zeta)\sp2=\zeta\sp2\, H.$$
The spectral curve of the affine Toda system is then
(\ref{Todacurve}) upon restricting the center of mass motion
$\sum_i p_i=0=\sum_i q_i$. The constant $\beta$ may be related to
the coefficient of the scaling element when the Toda equations are
expressed in terms of the affine algebra $\widehat{\text{sl}}_n$.
In fact we may strengthen Sutcliffe's ansatz substantially. At this
stage we only have that solutions of the Toda equations will yield
some solutions of the Nahm equations with cyclic symmetry. First we
will show that any $\texttt{C}_n$ invariant solution of Nahm's
equations (for charge $n$ $su(2)$ monopoles) are given by solutions
of the affine Toda equations. Then we will very concretely relate
the solutions.
We have that $G\subset SO(3)$ acts on triples
${\bf{t}}=(T_1,T_2,T_3)\in\mathbb{R}\sp3\otimes SL(n,\mathbb{C})$
via the natural action on $\mathbb{R}\sp3$ and conjugation on
$SL(n,\mathbb{C})$. This natural action may be identified with the
$SU(2)$ action on $\mathcal{O}(2)$ given above. If $g'\in SO(3)$ and
$g=\rho(g')$ is its image in $SL(n,\mathbb{C})$ then we have
\begin{align*}g'\circ&\left[
\eta+(T_1+iT_2)-2iT_3\zeta+(T_1-iT_2)\zeta\sp2\right]\\&=
\omega\left[
\eta+\omega\sp{-1}g(T_1+iT_2)g\sp{-1}-2igT_3g\sp{-1}\zeta+ \omega
g(T_1-iT_2)g\sp{-1}\zeta\sp2\right] .\end{align*} Thus invariance of
the spectral curve gives
\begin{align*}
g(T_1+iT_2)g\sp{-1}&=\omega(T_1+iT_2), \\ gT_3g\sp{-1}&=T_3, \\
g(T_1-iT_2)g\sp{-1}&=\omega\sp{-1}(T_1-iT_2).
\end{align*}
Now Hitchin, Manton and Murray \cite{hmm95} have described how the
$SO(3)$ action on $SL(n,\mathbb{C})$ decomposes as the direct sum
$\underline{2n-1}\oplus
\underline{2n-3}\oplus\ldots\oplus\underline{5}\oplus\underline{3}$
where $\underline{2k-1}$ denotes the $SO(3)$ irreducible
representation of dimension $2k-1$. We may identify $SO(3)$ and
its image in $SL(n,\mathbb{C})$ and because this decomposition has
$\mathop{\rm rank}\nolimits SL(n,\mathbb{C})=n-1$ summands then, by a theorem of
Kostant \cite{kostant}, the Lie algebra of this $SO(3)$ is a
principal three-dimensional subalgebra. By conjugation we may
express our generator $g'$ of
$\texttt{C}_n$ as $g'=\exp\left[\dfrac{2\pi}{n}\begin{pmatrix}0&1&0\\
-1&0&0\\0&0&0\end{pmatrix}\right]$ and then
$g=\rho(g')=\exp\left[\frac{2\pi}{n}H\right]$ where $H$ is
semi-simple and the generator of the principal three-dimensional
algebra's Cartan subalgebra. Kostant described the action of such
elements on arbitrary semi-simple Lie algebras and their roots.
For the case at hand we have that $g$ is equivalent to
$\mathop{\rm Diag}\nolimits(\omega\sp{n-1},\ldots,\omega,1)$ and that
$$g E_{ij}g\sp{-1}=\omega\sp{j-i}\,E_{ij}.$$
Therefore at this stage we know that for a cyclically invariant
monopole we may write
$$T_1+iT_2=\sum_{\alpha\in\hat \Delta}e\sp{(\alpha,\tilde
q)/2}\,E_\alpha,\qquad T_3=-\frac{i}2\sum_j \tilde p_j\,H_j$$
where in principle $\tilde q_i$, $\tilde p_i\in\mathbb{C}$, and
$\alpha\in\hat \Delta$ are the simple roots together with minus
the highest root. (The sum over $H_i$ may be taken as either the
Cartan subalgebra of $SL(n,\mathbb{C})$ or, by reinstating the
center of mass, the Cartan subalgebra of $GL(n,\mathbb{C})$.) The
Sutcliffe ansatz follows if the $\tilde q_i$ and $\tilde p_i$ may
be chosen real. Now by an $SU(n)$ transformation
$\mathop{\rm Diag}\nolimits(e\sp{i\theta_1},e\sp{i\theta_2},\ldots, e\sp{i\theta_n})$
(where $\sum_i\theta_i=0$) together with an overall $SO(3)$
rotation the reality of $\tilde q_i$ may be achieved. The reality
of $\tilde p_i$ follows upon imposing $T_i(s)=-T_i\sp\dagger(s)$
which also fixes $T_1-iT_2$. At this stage we have established the
following.
\begin{theorem} Any cyclically symmetric monopole is gauge equivalent
to Nahm data given by Sutcliffe's ansatz, and so obtained from the
affine Toda equations.
\end{theorem}
\section{Flows and Solutions}
The relation between the Nahm data and the affine Toda system is
much closer than simply that they yield the same equations of
motion. Let $\hat{\mathcal{C}}$ denote the genus $(n-1)^2$ spectral
curve of the monopole and $\mathcal{C}$ denote the genus $n-1$
spectral curve of the Toda theory. We have already noted that
$\mathcal{C}=\hat{\mathcal{C}}/\texttt{C}_n$ and the natural
projection $\pi:\hat{\mathcal{C}}\rightarrow\mathcal{C}$ is an
$n$-fold unbranched cover. The solution of an integrable system is
typically expressed in terms of the straight line motion on the
Jacobian of the system's spectral curve. Such a line is determined
both by its direction and a point on the Jacobian. We shall now show
that both the direction and point relevant for monopole solutions
are obtained as pull-backs of Toda data.
First we recall that meromorphic differentials describe flows, and
that a meromorphic differential on a Riemann surface is uniquely
specified by its singular parts together with some normalisation
conditions. If
${\left\{\hat{\mathfrak{a}}_i,\hat{\mathfrak{b}}_i\right\}_{i=1}\sp{\hat
g}}$ form a canonical basis for
$H_1(\hat{\mathcal{C}},\mathbb{Z})$,
$$\hat{\mathfrak{a}}_i\cap\hat{\mathfrak{b}}_j
=-\hat{\mathfrak{b}}_j\cap\hat{\mathfrak{a}}_i=\delta_{ij},$$
then one such normalisation condition is that the
$\hat{\mathfrak{a}}$-periods of the meromorphic differential vanish.
(Thus the freedom to add to the meromorphic differential a
holomorphic differential without changing its singular part is
eliminated.) In what follows we denote by
${\left\{\mathfrak{a}_i,\mathfrak{b}_i\right\}_{i=1}\sp{g}}$ a
similar canonical basis for $H_1(\mathcal{C},\mathbb{Z})$.
For the monopole the Lax operator
$A(\zeta
$ has poles at $\zeta=\infty$. If we denote $\hat{\infty}_j$ to be
the $n$ points on the spectral curve above $\zeta=\infty$ (and these
may be assumed distinct) then we find that ${\eta}/{\zeta}=\rho_j
\zeta$ as $ \zeta\sim\hat{\infty}_j$. Consequently in terms of a
local coordinate
$t$ at $\hat{\infty}_j$, $\zeta=1/{t}$, then
$$d\left(\frac{\eta}{\zeta}\right)=\left(-\frac{\rho_j}{t\sp2}+O(1)\right)dt.
$$
Thus on the monopole spectral curve we may uniquely define a
meromorphic differential by the pole behaviour at $\hat{\infty}_j$
and normalization
\begin{align*}
\gamma_\infty&=\left(\frac{\rho_j}{t\sp2}+O(1)\right)dt, \qquad
0=\oint_{\hat{\mathfrak{a}_i}}\gamma_\infty.
\end{align*}
The vector of $\mathbf{\hat{\mathfrak{b}}}$-periods,
$$
\boldsymbol{\widehat U} =\frac{1}{2i\pi}
\oint_{\mathbf{\hat{\mathfrak{b}}}} \gamma_\infty,$$ known as the
Ercolani-Sinha vector \cite{es89}, determines the direction of the
monopole flow on $\mathop{\rm Jac}\nolimits(\hat{\mathcal{C}})$. This vector is in fact
constrained. Let us first recall Hitchin's conditions on a monopole
spectral curve, equivalent to the Nahm data already given. These are
\begin{itemize}\item[\bf{H1}] Reality conditions
$\displaystyle{
a_r(\zeta)=(-1)^r\zeta^{2r}\overline{a_r(-1/{\overline{\zeta}}\,)}}$
\item[\bf{H2}] Let $L^{\lambda}$ denote the holomorphic line bundle on
$T{\mathbb P}\sp1$ defined by the transition function
$g_{01}=\rm{exp}(-\lambda\eta/\zeta)$
and let
$L^{\lambda}(m)\equiv
L^{\lambda}\otimes\pi\sp*\mathcal{O}(m)$ be similarly defined in
terms of the transition function
$g_{01}=\zeta^m\exp{(-\lambda\eta/\zeta)}$. Then $L^2$ is trivial on
$\hat{\mathcal{C}}$ and $L\sp1(n-1)$ is real.
\item[\bf{H3}] $H^0(\hat{\mathcal{C}},L^{\lambda}(n-2))=0$ for
$\lambda\in(0,2)$
\end{itemize}
We have already seen the reality conditions. Here the triviality of
$L^2$ means that there exists a nowhere-vanishing holomorphic
section. The following are equivalent \cite{es89, hmr99}:
\begin{enumerate} \item $L\sp2$ is trivial on $\hat{\mathcal{C}}$.
\item $2\boldsymbol{\widehat U}\in \Lambda\Longleftrightarrow$ $
\boldsymbol{\widehat
U}=\frac{1}{2\pi\imath}\left(\oint_{\hat{\mathfrak{b}}_1}\gamma_{\infty},
\ldots,\oint_{\hat{\mathfrak{b}}_{\hat
g}}\gamma_{\infty}\right)\sp{T}= \frac12
\boldsymbol{n}+\frac12\hat{\tau}\boldsymbol{m} . $
\item There exists a 1-cycle
$\widehat{\mathfrak{es}}=\boldsymbol{n}\cdot{\hat{\mathfrak{a}}}+
\boldsymbol{m}\cdot{\hat{\mathfrak{b}}}$ such that for every
holomorphic differential
$$\Omega=\dfrac{\beta_0\eta^{n-2}+\beta_1(\zeta)\eta^{n
-3}+\ldots+\beta_{n-2}(\zeta)}{{\partial\mathcal{P}}/{\partial
\eta}}\,d\zeta, \qquad
\oint\limits_{\widehat{\mathfrak{es}}}\Omega=-2\beta_0. $$
\end{enumerate}
Here $\hat\tau$ is the period matrix of $\hat{\mathcal{C}}$ and
$\Lambda$ is the associated period lattice of the curve. Thus
$\boldsymbol{\widehat{U}}$ is constrained to be a half-period. These
are known as the Ercolani-Sinha constraints and they impose $\hat
g$ \emph{transcendental constraints} on the curve yielding
$$\sum_{j=2}\sp{n}(2j+1)-\hat{g}=(n+3)(n-1)-(n-1)^2=4(n-1)$$
degrees of freedom.
We now turn to consider the behaviour of the Ercolani-Sinha vector
under a symmetry. Clearly our group acting the curve leads to an
action on divisors and consequently on the Jacobian. We now show
that the Ercolani-Sinha vector describing the flow is fixed under
the symmetry. This means the vector may be obtained from the
pull-back of a vector on the Jacobian of the quotient (Toda) curve.
Suppose we have a symmetry
$$0=P(\eta,\zeta)=P(\tilde\eta,\tilde\zeta)= \dfrac{{\tilde
P}(\eta,\zeta)}{(q\, \zeta+p)\sp{2n}}.$$ In particular
\begin{equation}
\partial_{\tilde\eta}P(\tilde\eta,\tilde\zeta)=(q\, \zeta+p)\sp{2}\partial_\eta
P(\tilde\eta,\tilde\zeta)=\frac{\partial_\eta {\tilde
P}(\eta,\zeta)}{(q\, \zeta+p)\sp{2n-2}}= \frac{\partial_\eta
{P}(\eta,\zeta)}{(q\, \zeta+p)\sp{2n-2}}.
\end{equation}
Using
$$d\tilde\zeta=\frac{d\zeta}{(q\, \zeta+p)\sp{2}}$$
we see then that
$$\frac{\tilde\zeta\sp{r}\tilde\eta\sp{s}d\tilde\zeta}{\partial_{\tilde\eta}P(\tilde\eta,\tilde\zeta)}=
\frac{(\bar p \zeta-\bar q)\sp{r}(q\,
\zeta+p)\sp{2n-4-r-2s}\eta\sp{s}d\zeta}{
\partial_\eta { P}(\eta,\zeta)}.$$
Bringing these together
\begin{lemma}The differential $\hat\omega_{r,s}=\dfrac{\zeta\sp{r}\eta\sp{s}d\zeta}{
\partial_\eta { P}(\eta,\zeta)}$ is invariant under the rotation
(\ref{pstransf}) if and only if
$$\zeta\sp{r}=(\bar p \zeta-\bar q)\sp{r}(q\,
\zeta+p)\sp{2n-4-r-2s}.$$ This always has a solution, the
holomorphic differential
$$\hat \omega=\dfrac{\eta\sp{n-2}d\zeta}{
\partial_\eta { P}(\eta,\zeta)}.$$
\end{lemma}
For the particular case of interest here, for rotations given by
$q=0$, $|p|\sp2=1$, then
\begin{equation}\label{phionhol}
\phi\sp\ast\left( \dfrac{\zeta\sp{r}\eta\sp{s}d\zeta}{
\partial_\eta { P}(\eta,\zeta)} \right)=\omega\sp{r+s+2}\,
\dfrac{\zeta\sp{r}\eta\sp{s}d\zeta}{
\partial_\eta { P}(\eta,\zeta)}
\end{equation}
and we also have solutions for each $s$ ($0\le s\le n-2$) and
$r=n-2-s$. These give us $g=n-1$ $\texttt{C}_n$-invariant
holomorphic differentials which are pullbacks of the holomorphic
differentials on $\mathcal{C}$. We remark also that the symmetry
always fixes the subspaces $\sum_r \mu_r \omega_{r,s}$ for fixed
$s$. Thus on the space of holomorphic differentials
$\{\hat\omega_I\}_{I=1}\sp{\hat g -1}\cup\{\hat\omega_{0,n-2}\}$
(for appropriate $I=(r,s)$ whose order does not matter) we have
\begin{equation}\label{symL}
\phi\sp\ast(\hat\omega_{1},\ldots,\hat\omega_{\hat g -1},\hat\omega_{0,n-2})=
(\hat\omega_{1},\ldots,\hat\omega_{\hat g -1},\hat\omega_{0,n-2})
\begin{pmatrix}\ast&\ast&0\\
\ast&\ast&0\\
0&0&1\end{pmatrix}:= (\hat\omega_{1},\ldots,\hat\omega_{\hat g -1},\hat\omega_{0,n-2})L
\end{equation}
where $L$ is a $\hat{g}\times \hat{g}$ complex matrix. As $L^n=1$,
the matrix is both invertible and diagonalizable.
With $\{\hat{\mathfrak{a}}_i,\hat{\mathfrak{b}}_i\}$ the canonical
homology basis introduced earlier and $\{\hat{u}_j\}$ a basis of
holomorphic differentials for our Riemann surface
$\hat{\mathcal{C}}$ we have the matrix of periods
\begin{equation}\begin{pmatrix}\oint_{\hat{\mathfrak{a}}_i}\hat{u}_j\\
\oint_{\hat{\mathfrak{b}}_i}\hat{u}_j\end{pmatrix}=
\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}=\begin{pmatrix}1\\
\hat{\tau}\end{pmatrix}\hat{\mathcal{A}}
\end{equation}
with $\hat{\tau}=\hat{\mathcal{B}}\hat{\mathcal{A}}\sp{-1}$ the
period matrix. If $\sigma$ is any automorphism of
$\hat{\mathcal{C}}$ then $\sigma$ acts on
$H_1(\hat{\mathcal{C}},\mathbb{Z})$ and the holomorphic
differentials by
$$\sigma_\ast\begin{pmatrix}\hat{\mathfrak{a}}_i\\
\hat{\mathfrak{b}}_i\end{pmatrix}=\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}\begin{pmatrix}\hat{\mathfrak{a}}_i\\
\hat{\mathfrak{b}}_i\end{pmatrix},\qquad
\sigma\sp\ast\hat{u}_j=\hat{u}_k
L\sp{k}_j,$$ where $\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}\in Sp(2\hat{g},\mathbb{Z})$ and $L\in
GL(\hat{g},\mathbb{C})$. Then from
\begin{equation*}\oint_{\sigma_\ast\gamma}\hat{u}=\oint_\gamma\sigma\sp\ast\hat{u}
\end{equation*}
we obtain
\begin{align}
\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}&=\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}L .
\end{align}
With the ordering of holomorphic differentials of (\ref{symL}) the
second of the equivalent conditions for the Ercolani-Sinha vector
says there exist integral vectors $\mathbf{n}$, $\mathbf{m}$ such
that
\begin{equation}\label{escondab}
(\mathbf{n},\mathbf{m})\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}=-2(0,\ldots,0,1).
\end{equation}
Now suppose $\sigma$ corresponds to a symmetry coming from a
rotation. Then the form of $L$ in (\ref{symL}) gives
\begin{align*}
(\mathbf{n},\mathbf{m})\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}=-2(0,\ldots,0,1)=-2(0,\ldots,0,1).L=
(\mathbf{n},\mathbf{m})\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}.L
=(\mathbf{n},\mathbf{m})\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}
\end{align*}
and so
$$\left( (\mathbf{n},\mathbf{m})-(\mathbf{n},\mathbf{m})\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}\right)\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}=0.$$ As the rows of the lattice generated by $\begin{pmatrix}\hat{\mathcal{A}}\\
\hat{\mathcal{B}}\end{pmatrix}$ are independent over $\mathbb{Z}$ we
therefore have that
$$(\mathbf{n},\mathbf{m})=(\mathbf{n},\mathbf{m})\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}$$ for all symplectic matrices $\begin{pmatrix}{A}&B\\
C&D\end{pmatrix}$ representing the symmetries coming from spatial
rotations. In particular $(\mathbf{n},\mathbf{m})$ is invariant
under the group of symmetries. Therefore the Ercolani-Sinha vector
is invariant and so as an element of the Jacobian, this will reduce
to a vector of the Jacobian of the quotient curve. Viewing this
vector as a divisor on the curve it projects to a divisor on the
quotient curve. Thus we have established
\begin{theorem} The Ercolani-Sinha vector is invariant under the
group of symmetries of the spectral curve arising from rotations
(\ref{pstransf}),
\begin{equation}
\boldsymbol{\widehat U}=\pi\sp\ast(\boldsymbol{ U}),\qquad
\boldsymbol{ U}\in\mathop{\rm Jac}\nolimits(\mathcal{C}).
\end{equation}
\end{theorem}
For the cyclic symmetry under consideration we have from
\begin{align}dy&=n\left(\nu+\frac{(-1)\sp{n}|\beta|\sp2}{\nu}\right)\frac{d\zeta}{\zeta}=
-n(x\sp{n}+a_2 x\sp{n-2}+\ldots +a_n)\frac{d\zeta}{\zeta}, \nonumber \\
\partial_\eta { P}(\eta,\zeta)&=\zeta\sp{n-1}\partial_x
(x\sp{n}+a_2 x\sp{n-2}+\ldots +a_n), \nonumber \intertext{that}
\dfrac{\zeta\sp{n-2-s}\eta\sp{s}d\zeta}{
\partial_\eta { P}(\eta,\zeta)}&=
\pi\sp\ast\left(-\frac{1}{n}\,\frac{x\sp{s}dx}{y}\right).\label{pullbackhyp}
\end{align}
Thus each of the invariant differentials (for $0\le s\le n-2$)
reduce to hyperelliptic differentials.
\section{The base point}
In the construction of monopoles there is a distinguished point
$\widetilde{\boldsymbol{K}}\in\mathop{\rm Jac}\nolimits(\hat{\mathcal{C}})$ that Hitchin
uses to identify degree $\hat{g}-1$ line bundles with
$\mathop{\rm Jac}\nolimits(\hat{\mathcal{C}})$. For $n\ge3$ this point is a singular
point of the theta divisor, $ \widetilde{\boldsymbol{K}}\in
\Theta_{\rm singular}$ \cite{bren06}. If we denote the Abel map by
$$\mathcal{A}_{\hat{Q}}(\hat{P})=\int_{\hat{Q}}\sp{\hat{P}}\hat{u}_i$$
then \begin{equation}\label{tildeK}\widetilde{\boldsymbol{K}}=
\boldsymbol{\hat{K}}_{\hat{Q}}+\mathcal{A}_{\hat{Q}}\left((n-2)
\sum_{k=1}\sp{n}\hat{\infty}_k\right).\end{equation} Here
$\boldsymbol{\hat{K}}_{\hat{Q}}$ is the vector of Riemann
constants for the curve $\hat{\mathcal{C}}$. If
$\mathcal{K}_{\hat{\mathcal{C}}}$ is the canonical divisor of the
curve then
$\mathcal{A}_{\hat{Q}}(\mathcal{K}_{\hat{\mathcal{C}}})=-2\boldsymbol{\hat{K}}_{\hat{Q}}$.
The righthand side of (\ref{tildeK}) is in fact independent of the
base point $\hat{Q}$ in its definition.
The point $\widetilde{\boldsymbol{K}}$ is the base point of the
linear motion in the Jacobian referred to earlier and we shall now
relate this to a point in the Jacobian of the Toda spectral curve
$\mathcal{C}$. Let $\mathcal{A}_Q(\mathcal{K}_\mathcal{C})=-2{\boldsymbol{K}}_Q$ be
the corresponding quantities for the curve $\mathcal{C}$ with basis
of holomorphic differentials $\{ u_a\}$. We first relate $\pi\sp\ast
{\boldsymbol{K}}_Q$ and ${\hat {\boldsymbol{K}}}_{\hat Q}$ where $\pi({\hat Q})=Q$ is some
preimage of $Q$. Let our symmetry be
$\phi:\hat{\mathcal{C}}\rightarrow \hat{\mathcal{C}}$,
$\phi\sp{n}=1$, and observe that (with $\pi({\hat P})=P$, $\pi({\hat
Q})=Q$)
\begin{align*}\pi\sp\ast(\mathcal{A}_Q(P))&=\pi\sp\ast\left(\int_Q\sp{P}u \right)
=\sum_{s=0}\sp{n-1}\int_{\phi\sp{s}(\hat Q)}\sp{\phi\sp{s}(\hat
P)}\hat u =\sum_{s=0}\sp{n-1}\left[\mathcal{A}_{\hat
Q}\left(\phi\sp{s}(\hat P)\right)-\mathcal{A}_{\hat
Q}\left(\phi\sp{s}(\hat Q)\right)\right].
\end{align*}
(This is actually independent of the base-point chosen for the Abel
map, so well-defined.) Now if $\sum_{\alpha=1}\sp{2g-2}P_\alpha$ is
a canonical divisor for $\mathcal{C}$ then
$\sum_{\alpha=1}\sp{2g-2}\sum_{s=0}\sp{n-1}\phi\sp{s}(\hat
P_\alpha)$ is a canonical divisor for $\hat{\mathcal{C}}$. Thus
\begin{align*}
\pi\sp\ast(-2{\boldsymbol{K}}_Q)&=\pi\sp\ast\left(\mathcal{A}_Q(\mathcal{K}_\mathcal{C})
\right)\\
&=\pi\sp\ast\left(\sum_{\alpha=1}\sp{2g-2}\int_Q\sp{P_\alpha}u
\right)\\
&=\mathcal{A}_{\hat Q}(\hat{\mathcal{K}}_\mathcal{\hat{C}})-2(g-1)
\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat Q}\left(\phi\sp{s}(\hat
Q)\right)\\
&=-2{\hat {\boldsymbol{K}}}_{\hat Q}-2(g-1) \sum_{s=0}\sp{n-1}\mathcal{A}_{\hat
Q}\left(\phi\sp{s}(\hat Q)\right).
\end{align*}
Therefore
\begin{equation}\label{pullbackK}
\pi\sp\ast({\boldsymbol{K}}_Q)={\hat {\boldsymbol{K}}}_{\hat Q}+(g-1)
\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat Q}\left(\phi\sp{s}(\hat
Q)\right)+\hat e,
\end{equation}
where $2\hat e\in \Lambda$ is a half-period. This expression may be
rewritten as
\begin{align*}
\pi\sp\ast({\boldsymbol{K}}_Q)&={\hat {\boldsymbol{K}}}_{\hat Q}+(g-1)
\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat Q}\left(\phi\sp{s}(\hat
Q)\right)+\hat e\\
&=\left[{\hat {\boldsymbol{K}}}_{\hat Q}+(\hat g-1)\mathcal{A}_{\hat Q}(\hat
P)\right]-(\hat g-1)\mathcal{A}_{\hat Q}(\hat P)+(g-1)
\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat Q}\left(\phi\sp{s}(\hat
Q)\right)+\hat e\\
&={\hat {\boldsymbol{K}}}_{\hat P}-n( g-1)\mathcal{A}_{\hat Q}(\hat P)+(g-1)
\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat Q}\left(\phi\sp{s}(\hat
Q)\right)+\hat e\\
&={\hat {\boldsymbol{K}}}_{\hat P}+(g-1)\sum_{s=0}\sp{n-1}\mathcal{A}_{\hat
P}\left(\phi\sp{s}(\hat Q)\right)+\hat e,
\end{align*}
showing the left-hand side is independent of the choice of
base-point for the Abel map.
Comparison of (\ref{tildeK}) and (\ref{pullbackK}) now shows that
\begin{equation}\label{tildeKpb}\widetilde{\boldsymbol{K}}=
\pi\sp\ast({\boldsymbol{K}}_{\infty_+})-\hat e\end{equation} where
$\pi(\hat{\infty}_k)=\infty_+$ as noted earlier. Now the half-period
$\hat e$ can be identified and is of the form $\hat
e=\pi\sp\ast(e)$. The actual identification depends on an homology
choice and will be given in the next section, but for the moment we
simply note the form
\begin{equation}\label{tildeKpbf}\widetilde{\boldsymbol{K}}=
\pi\sp\ast({\boldsymbol{K}}_{\infty_+}-e).\end{equation}
\section{Fay-Accola factorization}
The standard reconstruction of solutions for an integrable system with spectral
curve $\hat{\mathcal{C}}$ proceeds by constructing the Baker-Akhiezer functions
for this curve.
These may be calculated in terms of theta functions for the curve and for our
present purposes we may focus on the theta function
$\theta(\lambda\widehat{\boldsymbol{U}}-
\widetilde{\boldsymbol{K}}\,|\,\hat\tau)$. This describes a flow on the Jacobian of
$\hat{\mathcal{C}}$ in the direction of the Ercolani-Sinha vector $\widehat{\boldsymbol{U}}$
with base point $\widetilde{\boldsymbol{K}}$. We have observed that we have
a cyclic unramified covering $\pi:\hat{\mathcal{C}}\rightarrow\mathcal{C}$
of the affine Toda spectral curve by the monopole spectral curve.
The map $\pi$ leads to a map $\pi\sp\ast:
\text{Jac}({\mathcal{C}})\rightarrow\text{Jac}(\hat{\mathcal{C}})$
which may be lifted to $\pi\sp\ast:{\mathbb C}\sp{g}\rightarrow
{\mathbb C}\sp{\hat g}$. Further we have established that
$$\lambda\widehat{\boldsymbol{U}}-
\widetilde{\boldsymbol{K}}=
\pi\sp\ast(\lambda \boldsymbol{U} -{\boldsymbol{K}}_{\infty_+}+e).
$$
We now are in a position to make use of a remarkable factorization theorem due to
Accola and Fay \cite{acc71, fay73} and also observed by Mumford.
When $\hat z=\pi\sp\ast z$ the theta
functions on $\hat{\mathcal{C}}$ and ${\mathcal{C}}$ are related
by this factorization theorem,
\begin{theorem}[\bf Fay-Accola] \label{fayaccola}
With respect to the ordered canonical homology bases $\{
\hat{\mathfrak{a}}_i\sp{c},\hat{\mathfrak{b}}_i\sp{c}\}$ described
below and for arbitrary $\boldsymbol{ z}=\in \mathbb{C}^g$ we have
\begin{equation}
\frac{\theta[\hat e](\pi\sp\ast\boldsymbol{ z};\hat{\tau}\sp{c})}
{\prod_{k=0}^{n-1}\theta\left[\begin{matrix}0&0&\dots&0
\\ \frac{k}{n}&0&\dots&0 \end{matrix}\right]\left(\boldsymbol{ z};\tau\sp{c
\right)}
=c_0(\widehat{\tau}\sp{c
) \label{fafactora}
\end{equation}
is a non-zero modular constant
$c_0(\hat{\tau}\sp{c
)
$ independent of
$\boldsymbol{ z}$. Here $\hat{\tau}\sp{c}$ is the $\mathfrak{a}$-normalized
period matrix for the curve $\hat{\mathcal{C}}$ in this homology basis and
$$\hat e=\left[\begin{matrix}0&0&\dots&0
\\ \frac{n-1}{2}&0&\dots&0 \end{matrix}\right]=
\pi\sp\ast\left(e\right)=
\pi\sp\ast\left( \left[\begin{matrix}0&0&\dots&0
\\ \frac{n-1}{2n}&0&\dots&0 \end{matrix}\right]\right)
.$$
\end{theorem}
The significance of this theorem for our setting is that it means
we can reduce the construction of solutions to that of quantities
purely in terms of the hyperelliptic affine Toda spectral curve.
The theorem is expressed in terms of a particular choice of homology
basis which is well adapted to the symmetry at hand. In terms of the
conformal automorphism
$\phi:\hat{\mathcal{C}}\rightarrow\hat{\mathcal{C}}$ of
$\hat{\mathcal{C}}$ that generates the group
$\texttt{C}_n=\{\phi\sp{s}\,|\,0\le s\le n-1\}$ of cover
transformations of $\hat{\mathcal{C}}$ and the projection
$\pi:\hat{\mathcal{C}}\rightarrow\mathcal{C}$ there exists a basis
$\{\hat{\mathfrak{a}}_0\sp{c},\hat {\mathfrak{b}}_0\sp{c},
\hat{\mathfrak{a}}_1\sp{c},\hat
{\mathfrak{b}}_1\sp{c},\ldots,\hat
{\mathfrak{a}}_{\hat g-1}\sp{c},\hat {\mathfrak{b}}_{\hat
g-1}\sp{c}\}$ of homology cycles for $\hat{\mathcal{C}}$ and
$\{{\mathfrak{a}}_0\sp{c},{\mathfrak{b}}_0\sp{c},{\mathfrak{a}}_1\sp{c},
{\mathfrak{b}}_1\sp{c},\ldots,{\mathfrak{a}}_{g-1}\sp{c},{\mathfrak{b}}_{g-1}\sp{c}\}$
for $\mathcal{C}$ such that ( for $1\le j\le g-1,\ 0\le s\le n$)
\begin{align*}
\pi(\hat
{\mathfrak{a}}_0\sp{c})&={\mathfrak{a}}_0\sp{c},&\pi(\hat
{\mathfrak{a}}_{j+s(g-1)}\sp{c})&={\mathfrak{a}}_j\sp{c},&
\pi(\hat {\mathfrak{b}}_0\sp{c})&=n\,
{\mathfrak{b}}_0\sp{c},&\pi(\hat
{\mathfrak{b}}_{j+s(g-1)}\sp{c})&={\mathfrak{b}}_j\sp{c},&\\
\phi\sp{s}(\hat {\mathfrak{a}}_0\sp{c})&\sim {\hat
{\mathfrak{a}}_0\sp{c}},&\phi\sp{s}(\hat
{\mathfrak{a}}_j\sp{c})&=\hat
{\mathfrak{a}}_{j+s(g-1)}\sp{c},&
\phi\sp{s}(\hat {\mathfrak{b}}_0)&= {\hat
{\mathfrak{b}}_0\sp{c}},&\phi\sp{s}(\hat
{\mathfrak{b}}_j\sp{c})&=\hat {\mathfrak{b}}_{j+s(g-1)}\sp{c}.&
\end{align*}
Here $\phi\sp{s}(\hat {\mathfrak{a}}_0)$ is homologous to $\hat
{\mathfrak{a}}_0$. If $\hat v_i$ are the
$\hat{\mathfrak{a}}$-normalized differentials for
$\hat{\mathcal{C}}$, then
$$\delta_{i,j+s(g-1)}=\int_{\hat {\mathfrak{a}}_{j+s(g-1)}}\hat v_i=
\int_{\phi\sp{s}(\hat {\mathfrak{a}}_j)}\hat v_i=\int_{\hat
{\mathfrak{a}}_j}(\phi\sp{s})\sp{\ast}\hat v_i=\int_{\hat
{\mathfrak{a}}_j}\hat v_{i-s(g-1)},$$ and we find that
\begin{equation}(\phi\sp{s})\sp{\ast}\hat v_0= \hat
v_0,\qquad(\phi\sp{s})\sp{\ast}\hat v_i=\hat v_{i-s(g-1)}.
\label{gpdiff} \end{equation} If $v_i$ are the normalized
differentials for $\mathcal{C}$, then
$$
\delta_{ij}=\int_{{\mathfrak{a}}_j}v_i=\int_{\pi(\hat
{\mathfrak{a}}_{j+s(g-1)})}v_i
=\int_{{\mathfrak{a}}_{j+s(g-1)}}\pi\sp\ast (v_i)$$ shows that
$$\pi\sp\ast (v_i)=\hat v_i+(\phi)\sp{\ast}\hat v_i+\ldots
+(\phi\sp{p-1})\sp{\ast}\hat v_i$$ and similarly that
$$\pi\sp\ast (v_0)=\hat v_0.$$
We may use the characters of $\texttt{C}_n$ to construct the
remaining linearly independent differentials on
$\hat{\mathcal{C}}$.
From (\ref{gpdiff}) we have an action of $\texttt{C}_n$ on
$\text{Jac}(\hat{\mathcal{C}})$ which lifts to an automorphism of
${\mathbb C}\sp{\hat g}$ by
\begin{equation}\label{symmjac}
\phi\sp{s}(\hat z)=(\hat z_0,\hat z_{1-s(g-1)},\ldots,\hat
z_{g-1-s(g-1)},\ldots,\hat z_{1+(p-s-1)(g-1)},\ldots,\hat
z_{g-1+(p-s-1)(g-1)})
\end{equation}
Now (\ref{symmjac}) together with the invariance of the
Ercolani-Sinha vector mans that in this cyclic homology basis we
have
\begin{equation}\label{escyclic}
(\mathbf{n},\mathbf{m})=(r_0,\mathbf{r},\ldots,\mathbf{r},
s_0,\mathbf{s},\ldots,\mathbf{s})\end{equation}
where the vectors
$\mathbf{r}= (r_1,\ldots,r_{g-1})$ and similarly $\mathbf{s}$ are
each repeated $n$ times. We also have
\begin{align}
\pi_\ast(\widehat{\mathfrak{es}} )&
= r_0\mathfrak{a}_0+n
\mathbf{r}\cdot\boldsymbol{ \mathfrak{a}}+ns_0\mathfrak{b}_0+ n
\mathbf{s}\cdot\boldsymbol{ \mathfrak{b}} . \label{projhomes}
\end{align}
With the choices above (things are different for
$\hat{\mathfrak{b}}$-normalization) we may lift the map
$\pi\sp\ast: \mathop{\rm Jac}\nolimits(\mathcal{C})\rightarrow \mathop{\rm Jac}\nolimits(\hat{\mathcal{C}})$
to $\pi\sp\ast:{\mathbb C}\sp{g}\rightarrow {\mathbb C}\sp{\hat
g}$,
$$
\pi\sp\ast(z)= \pi\sp\ast(z_0,z_1,\ldots,z_{g-1})=
(n\,z_0,z_1,\ldots,z_{g-1},\ldots,z_1,\ldots,z_{g-1})=\hat z.$$ With
this homology basis the period matrices for the two curves are
related by the block form
$$
\hat\tau\sp{c}=\left(
\begin{array}{ccccc}
n\, \tau_{00}\sp{c}&\tau_{0j}\sp{c}&\tau_{0j}\sp{c}&\ldots&\tau_{0j}\sp{c}\\
\tau_{j0}\sp{c}&\mathcal{M}&\mathcal{M}\sp{(1)}&&\mathcal{M}\sp{(n-1)}\\
\vdots&\\
\tau_{j0}\sp{c}&\mathcal{M}\sp{(1)}&&&\mathcal{M}
\end{array}
\right)
$$
where $\mathcal{M}\sp{(s)}=\int_{\phi\sp{-s}(\hat
{\mathfrak{b}}_j)}\hat v_i$. The $(r,s)$ block here has entry
$\mathcal{M}\sp{s-r}$ and
$(\mathcal{M}\sp{(s-r)})\sp{T}=\mathcal{M}\sp{(r-s)}$ by the
bilinear identity. Then
$\tau\sp{c}_{ij}=\sum_{s=0}\sp{n-1}\mathcal{M}\sp{(s)}_{ij}$. The
case $n=3$ is instructive, for here the $n-2$ block matrices are
just numbers and we have
\begin{equation}\hat{\tau}\sp{c
=\left( \begin{array}{cccc} a&b&b&b\\
b&c&d&d\\
b&d&c&d\\
b&d&d&c
\end{array}\right),\qquad
\tau\sp{c
=\left( \begin{array}{cc} \frac13 a&b\\
b&c+2d
\end{array}\right) .\end{equation}
The point to note is that although the period matrix for
$\hat{\mathcal{C}}$ involves integrations of differentials that do
not reduce to hyperelliptic integrals, the combination of terms
appearing in the reduction can be expressed in terms of
hyperelliptic integrals. This is a definite simplification.
Further the $\Theta$ function defined by $\hat\tau\sp{c}$ has the
symmetries
$$\Theta(\hat z|\hat\tau\sp{c})=\Theta(\phi\sp{s}(\hat z)|\hat\tau\sp{c})$$
for all $\hat z\in{\mathbb C}\sp{\hat g}$. In particular, the
$\Theta$ divisor is fixed under $\texttt{C}_n$.
If we are to reduce the construction of cyclic monopoles to a
problem involving only hyperelliptic quantities we must describe
the Ercolani-Sinha constaints in the context of the curve
$\mathcal{C}$.
\begin{theorem} \label{ESHYP} The Ercolani-Sinha constraint on the curve
$\hat{\mathcal{C}}$ yields the constraint
\begin{equation}
-2(0,\ldots,0,1)=(r_0,n\mathbf{r},n s_0,
n\mathbf{s})\begin{pmatrix}\mathcal{A}\\
\mathcal{B}\end{pmatrix}
\end{equation}
on the curve $\mathcal{C}$ with respect to the differentials
$u_s=-{x\sp{s}dx}/{(ny)}$ ($s=0,\ldots,n-2$).
\end{theorem}
\begin{proof}
The invariance of the Ercolani-Sinha vector means that
$\phi\sp\ast(\widehat{\mathfrak{es}})=\widehat{\mathfrak{es}}$.
Thus
$$
\int_{\widehat{\mathfrak{es}} }\hat\omega_{r,s}=
\int_{\phi\sp\ast(\widehat{\mathfrak{es}})}\hat\omega_{r,s}=
\int_{\widehat{\mathfrak{es}} }\phi\sp\ast\hat\omega_{r,s}=
\omega\sp{r+s+2}\int_{\widehat{\mathfrak{es}} }\hat\omega_{r,s},
$$
where we have used (\ref{phionhol}). Thus the integral of any
noninvariant differential around the cycle
$\widehat{\mathfrak{es}}$ must vanish, while from
(\ref{pullbackhyp}) and the Ercolani-Sinha condition we have that
$$-2\,\delta_{s,n-2}=\int_{\widehat{\mathfrak{es}} }
\pi\sp\ast\left(-\frac{1}{n}\,\frac{x\sp{s}dx}{y}\right)=
\int_{\pi_\ast(\widehat{\mathfrak{es}} )}
-\frac{1}{n}\,\frac{x\sp{s}dx}{y}.
$$
The theorem then follows upon using (\ref{projhomes}).
\end{proof}
In actual calculations it is convenient to use the unnormalized
differentials $\hat\omega_{r,s}$ and ${x\sp{s}dx}/{(ny)}$ rather
than Fay's normalized differentials $\hat v_i$. An alternate proof
of Theorem \ref{ESHYP} via Poincar\'e's reducibility theorem is
given in the Appendix, which provides further useful relations
amongst the periods of the two curves.
\section{Discussion}
In this paper we have shown that any cyclically symmetric monopole
is gauge equivalent to Nahm data obtained via Sutcliffe's ansatz
from the affine Toda equations. Further, the data needed to
reconstruct the monopole, the Ercolani-Sinha vector and base point
for linear flow on the Jacobian, may also be obtained from data on
the affine Toda equation's hyperelliptic spectral curve
$\mathcal{C}$. A theorem of Fay and Accola then enables us to
express the theta functions for the monopole spectral curve in
terms of the theta functions for the curve $\mathcal{C}$. Finally
the transcendental constraints on the monopole's spectral curve
can be recast as transcendental constraints for the hyperelliptic
curve $\mathcal{C}$ (Theorem \ref{ESHYP}). At this stage then the
construction of cyclically symmetric monopoles has been reduced to
one entirely in terms of hyperelliptic curves. Although analogues
of both the transcendental constraints still exist this is a
significant simplification. We note that the structure of the
theta divisor is better understood in the hyperelliptic setting
\cite{vanh95} and the hyperelliptic integrals are somewhat simpler
than the general integrals appearing in the Ercolani-Sinha
constraint for the full monopole curve.
Other approaches to constructing monopoles are known. In
particular \cite{hmm95} describe cyclically symmetric monopoles
within the rational map approach (see also \cite[\S8.8]{ms04}).
These authors show that the rational map for monopoles with
$\texttt{C}_n$ invariance about the $x_3$-axis takes the form
$$R(z)=\frac{\mu z\sp{l}}{z\sp{n}-\nu}$$
where $0\le l\le n-1$. The complex quantity $\nu $ determines
$\mu$ when the monopoles are strongly centred. Here
$\nu=(-1)\sp{n-1}\bar\beta$ of equation (\ref{cycliccurve}). The
moduli space $\mathcal{M}_n\sp{l}$ is a 4-dimensional totally
geodesic submanifold of the full moduli space. It is interesting
that both the rational map description and the description we have
presented lead to extra discrete parameters ($l$ in the case of
rational maps, and $k$ in \ref{fafactora}). The connection, if
any, between these will be pursued elsewhere \cite{bde10}.
Clearly the ansatz for monopoles extends to other algebras. If we
construct the spectral curve from the $D_n$ Toda system using the
$2n$ dimensional representation we find a spectral curve
$\hat{\mathcal{C}}$ of the form
$$\eta^{2n} + a_1 \eta^{2n-2} \zeta^2 + a_2\eta^{2n-4} \zeta^4+
\ldots+ a_n\zeta^{2n}+\alpha\eta^2 (\frac1{w} + \zeta^{4n-4}w)
=0.$$
Letting $x=\eta/\zeta$ the curve (upon dividing by $\zeta\sp{2n}$)
becomes
$$x^{2n} + a_1 x^{2n-2} + a_2 x^{2n-4}+
\ldots+ a_n+\alpha x^2(\frac1{w\zeta^{2n-2}} + \zeta^{2n-2}w)
=0.$$
and so we get with $\nu=\alpha w\zeta^{2n-2}$
$$P_n(x^2)+x^2(\nu+\frac{\alpha\sp2}{\nu})
=0$$
leading to a hyperelliptic curve $\tilde{\mathcal{C}}$
$$y^2=P_n(x^2)^2-4\alpha\sp2 x^4.$$
This curve has cyclic symmetry $\texttt{C}_{2n-2}$ from the
appearance of $\zeta^{2n-2}$ and $\texttt{C}_2$ due to the
appearance of $x^2$. The genus of $\hat{\mathcal{C}}$ is
$(2n-1)^2-2n$. The genus of $\tilde{\mathcal{C}}$ is $2n-1$.
Finally $\tilde{\mathcal{C}}$ covers a genus $n-1$ curve
${\mathcal{C}}$
$$y^2=P_n(u)^2-4\alpha\sp2 u^2.$$
Here we expect the Toda motion to lie in the Prym of this
covering, but the general theory warrants further study.
\section*{Acknowledgements}
I have benefited from many discussions with Antonella D'Avanzo,
Victor Enolskii and Timothy P. Northover. The results presented
here were described at the MISGAM supported meeting ``From
Integrable Structures to Topological Strings and Back'', Trieste
2008, and the Lorentz Center meeting ``Integrable Systems in
Quantum Theory'', Leiden 2008. I am grateful to the organisers of
these meetings for providing such a stimulating and pleasant
environment.
|
\section{Introduction}
Correlated equilibria are a natural generalization of Nash equilibria introduced by Aumann \cite{a:scrs}. They are defined to be joint probability distributions over the players' strategy spaces, such that if each player receives a private recommendation sampled according to the distribution, no player has an incentive to deviate unilaterally from playing his recommended strategy. In finite games the set of correlated equilibria is a compact convex polytope, and therefore seemingly much simpler than the set of Nash equilibria, which can be essentially any algebraic variety \cite{d:une}. Even in the simple case of two-player finite games, the set of Nash equilibria is a union of finitely many polytopes: seemingly more complicated than the set of correlated equilibria.
Nonetheless we will see that there are two-player zero-sum games in which the set of correlated equilibria has many more extreme points than the set of Nash equilibria has. This behavior does not seem to be pathological in any way: it occurs in very simple finite games and the simplest of infinite games. We take this as evidence that this complexity is likely to be quite common.
\paragraph{Contributions}
\begin{itemize}
\item We give a family of examples of two-player zero-sum finite games in which the set of Nash equilibria has polynomially many extreme points (Section~\ref{sec:nash}), while the set of correlated equilibria has factorially many extreme points (Section~\ref{sec:corr}).
For bimatrix games, this shows that while extreme Nash equilibria are a subset of the extreme correlated equilibria (see Related Literature below), enumerating all the extreme correlated equilibria is in general a bad way of computing all the extreme Nash equilibria. In particular, it would be faster to enumerate all subsets of the strategy spaces (there are ``only'' exponentially many) and check whether each was the support of a Nash equilibrium.
\item We give a related example of a continuous game with strategy sets equal to $[-1,1]$ and bilinear utility functions. This game is just the mixed extension of matching pennies, but we show that it has extreme correlated equilibria with arbitrarily large finite support (Proposition~\ref{prop:mpextfinsuppcorr}) and also with infinite support (Proposition~\ref{prop:mpextcorr}). This is in contrast to the extreme Nash equilibria, which always have uniformly bounded finite support in zero-sum games with polynomial utilities \cite{karlin:tig}.
Once the existence of Nash equilibria in continuous games has been established \cite{glicksberg:cg}, it is straightforward to show that polynomial games\footnote{For the purposes of this paper a \emph{polynomial game} is one in which the strategy spaces are compact intervals and the utility functions are polynomials in the players' strategies.} admit Nash equilibria with finite support \cite{karlin:tig, sop:slrcg}. There are more elementary ways of showing that continuous games have correlated equilibria \cite{hs:ece}, but to the authors' knowledge there is no proof that polynomial games have finitely supported correlated equilibria which does not rely on the existence of Nash equilibria. This example shows that the plausible-sounding proof idea that all extreme correlated equilibria are finitely supported simply isn't true.
\item Comparing Proposition~\ref{prop:momentmeasure} with this example shows that in general there is no finite-dimensional description of the set of correlated equilibria of a zero-sum polynomial game. That is to say, one cannot check if a measure is a correlated equilibrium merely by examining finitely many generalized moments (parameters such as mean, covariance, etc.\ -- any compactly supported distribution can be specified by countably many such parameters). Such a description for the Nash equilibria has been known for over fifty years \cite{karlin:tig}.
Intuitively, the reason for this difference is that being a correlated equilibrium is a statement about conditional distributions, and these are too delicate to be controlled by finitely many moments. This example confirms the intuition.
Experience from finite games suggests that correlated equilibria should be easier to compute than Nash equilibria. While there are computational methods which converge asymptotically to correlated equilibria of polynomial games \cite{spo:cecgcc}, the only exact algorithm the authors are aware of consists of computing a Nash equilibrium by quantifier elimination (extremely slow), which is possible because of the finite-dimensional description. In particular, no provably efficient method for computing correlated equilibria of polynomial games exactly or approximately is known. The lack of a finite-dimensional description of the problem seems to be an important part of what makes it difficult.
\end{itemize}
\paragraph{Related Literature}
The geometry of Nash and correlated equilibria has been studied extensively. Therefore we only mention work below if it is directly connected to ours and we do not attempt to be exhaustive.
The result most closely related to the present paper states that in two-player finite games, extreme Nash equilibria (viewed as product distributions) are a subset of the extreme correlated equilibria. Cripps \cite{c:ecnetpg} and Evangelista and Raghavan \cite{er:nce} proved this independently. This result shows that it makes sense to compare the number of extreme Nash and correlated equilibria. It also raises the natural question of whether all extreme Nash equilibria could be enumerated efficiently by enumerating the extreme correlated equilibria. We show that there can be many more extreme correlated equilibria than extreme Nash equilibria, answering this question in the negative.
In a similar vein, Nau et al.\ \cite{nch:ognece} show that for non-trivial finite games with any number of players, the Nash equilibria lie on the boundary of the correlated equilibrium polytope. With three or more players, the Nash equilibria need not be extreme correlated equilibria. For example consider the three-player poker game analyzed by Nash in \cite{nash:ncg} which has rational payoffs, hence rational extreme correlated equilibria, but whose unique Nash equilibrium uses irrational probabilities.
Separable games, a generalization of polynomial games, were first studied around the $1950$'s by Dresher, Karlin, and Shapley in papers such as \cite{dks:pg}, \cite{dk:scgfp}, and \cite{ks:gms}, which were later combined in Karlin's book \cite{karlin:tig}. Their work focuses on the zero-sum case, which contains some of the key ideas for the nonzero-sum case. In particular, they show how to replace the infinite-dimensional mixed strategy spaces (sets of probability distributions over compact metric spaces) with finite-dimensional moment spaces. Carath\'{e}odory's theorem \cite{bno:convex} then applies to show that finitely-supported Nash equilibria exist.
There are many similarities between separable games and finite games whose payoff matrices satisfy low-rank conditions. Lipton et al.\ \cite{lmm:plgss} consider two-player finite games and provide bounds on the cardinality of the support of extreme Nash equilibrium strategies in terms of the ranks of the payoff matrices. The main technical tool here is again Carath\'{e}odory's theorem.
Germano and Lugosi show that in finite games with three or more players there exist correlated equilibria with smaller support than one might expect for Nash equilibria \cite{gl:essce}. The proof is geometrical; it essentially views correlated equilibria as living in a subspace of low codimension and it too uses Carath\'{e}odory's theorem \cite{bno:convex}.
The bounds on the support of equilibria in finite and separable games of the previous three paragraphs are all synthesized in \cite{s:mastersthesis}; the portion on Nash equilibria has appeared in \cite{sop:slrcg}. The general idea is that simple payoffs (low-rank matrices, low-degree polynomials, etc.) lead to simple Nash equilibria (small support), and those in turn lead to simple correlated equilibria (small support again).
To produce upper bounds on the minimal support of correlated equilibria which depend only on the rank of the payoff matrices and not on the size of the strategy sets, this work does not bound the support of all extreme correlated equilibria, but rather only those whose support is contained inside a Nash equilibrium of small support, which must exist. Similar results hold for polynomial games with, for example, degree used in place of rank (the notions of degree and rank are generalized in \cite{s:mastersthesis} and \cite{sop:slrcg}).
This work left open the question of whether all extreme correlated equilibria have support size which can be bounded in terms of the rank of the payoff matrices, independently of the size of the strategy sets. Here we show that this is not the case, because our examples have payoffs which are of rank $1$ and extreme correlated equilibria of arbitrarily large, even infinite, support.
Correlated equilibria without finite support have been defined and studied by several authors. An important example of this line of research is the paper by Hart and Schmeidler \cite{hs:ece}. The definition of correlated equilibria presented in \cite{hs:ece} is convenient for proving some theoretical results (they focus on existence) but not usually for computation.
The authors of the present paper have developed several equivalent characterizations of correlated equilibria in continuous games which are more suitable for computation \cite{spo:cecgcc}. One of these forms the basis for the analysis in Section~\ref{sec:corr} below. Other such characterizations lead to algorithms for approximating correlated equilibria of continuous games \cite{spo:cecgcc}.
\paragraph{Outline}
The remainder of this paper is organized as follows. Section $2$ introduces the examples to be studied. The two types of example are closely related -- the finite game examples are just restrictions of the strategy spaces in the infinite game example to fixed finite sets. This allows us to analyze both examples on equal footing. In Section $3$ we define and compute the extreme Nash equilibria of these examples, counting them in the finite game example. Then we define and analyze the extreme correlated equilibria in Section $4$. This analysis is somewhat long and at times technical, so we present a detailed roadmap before beginning. We close with Section $5$, where we outline directions for future work.
\section{Description of the examples}
\label{sec:examples}
First we fix notation. When $S$ is a topological space, $\Delta(S)$ will denote the set of Borel probability measures on $S$ and $\Delta^*(S)$ the set of finite Borel measures on $S$. In particular $\Delta(S)$ is the set of measures in $\Delta^*(S)$ with unit mass. If $S$ is finite it will be given the discrete topology by default so $\Delta(S)$ is a simplex and $\Delta^*(S)$ is an orthant in $\mathbb{R}^{\lvert S \rvert}$. We abuse notation and write the measure of a singleton $\{p\}$ as $\mu(p)$ rather than $\mu(\{p\})$. For any $p\in S$, define $\delta_p\in \Delta(S)$ to be the measure which assigns unit mass to the point $p$. Let $I = [-1,1]\subset\mathbb{R}$.
We will focus on two related examples, one with finite strategy sets and one with infinite strategy sets. We will develop them in parallel by analyzing arbitrary games satisfying the following condition. The condition does not have any game theoretic content; it was merely chosen for simplicity and the results to which it leads.
\begin{assumption}
\label{assump:mp}
The game is a zero-sum strategic form game with two players, called $X$ and $Y$. The strategy sets $C_X$ and $C_Y$ are compact subsets of $I = [-1,1]$, each of which contains at least one positive element and at least one negative element. Player $X$ chooses a strategy $x\in C_X$ and player $Y$ chooses $y\in C_Y$. The utility functions\footnote{By inspection of the utilities we can see that for any $C_X$ and $C_Y$ with at least two points, the rank of this game in the sense of \cite{sop:slrcg} is $(1,1)$ (and in fact also in the stronger sense of Theorem $3.3$ of that paper). The notion of the rank of a game is related to the rank of the payoff matrices and will not play a significant role in this paper; we merely wish to note that under this definition of complexity of payoffs, the games we consider are extremely simple.} are $u_X(x,y) = xy = -u_Y(x,y)$.
\end{assumption}
\savecounter{exfin}
\begin{example}
\label{ex:finite}
Fix an integer $n>0$. Let $C_X$ and $C_Y$ each have $2n$ elements, $n$ of which are positive and $n$ of which are negative. If we take $n = 1$ and $C_X = C_Y = \{-1,1\}$ then we recover the matching pennies game, as shown in Table~\ref{tab:matchingpennies}.
\end{example}
\begin{table}
\centering
\begin{tabular}{|c|cc|}
\hline
$(u_X,u_Y)$ & $x = -1$ & $x = 1$ \\
\hline
$y = -1$ & $(1,-1)$ & $(-1,1)$\\
$y = 1$ & $(-1,1)$ & $(1,-1)$\\
\hline
\end{tabular}
\caption{Utilities for matching pennies}
\label{tab:matchingpennies}
\end{table}
\savecounter{exinf}
\begin{example}
\label{ex:infinite}
Let $C_X = C_Y = [-1,1]$. Then the game is essentially the mixed extension of matching pennies. That is to say, suppose two players play matching pennies and choose their strategies independently, playing $1$ with probabilities $p\in [0,1]$ and $q\in [0,1]$. Define the utilities for the mixed extension to be the expected utilities under this random choice of strategies. Letting $x = 2p - 1$ and $y = 2q - 1$, the utility to the first player is $xy$ and the utility to the second player is $-xy$. Therefore this example is the mixed extension of matching pennies, up to an affine scaling of the strategies.
\end{example}
Usually one looks at pure equilibria of the mixed extension of a game; these are exactly the mixed equilibria of the original game. We will instead be looking at mixed Nash equilibria and correlated equilibria of the mixed extension itself, a game with a continuum of actions. The relationship between correlated equilibria of the mixed extension and those of the original game is much more complicated than the corresponding relationship for mixed Nash equilibria. This drives the results of the paper.
\section{Extreme Nash equilibria}
\label{sec:nash}
We now characterize and count the extreme points of the sets of Nash equilibria in games satisfying Assumption~\ref{assump:mp}. Since the games are zero-sum, the set of Nash equilibria can be viewed as a Cartesian product of two (weak*) compact convex sets, the sets of maximin and minimax strategies \cite{glicksberg:cg}. The Krein-Milman theorem completely characterizes such sets by their extreme points \cite{r:fa}, explaining our focus on extreme points throughout.
We define Nash equilibria in two-player games, which will be sufficient for our purposes, as well as the standard notions of extreme point and extreme ray from convex analysis.
\begin{definition}A \textbf{Nash equilibrium} is a pair $(\sigma,\tau)\in \Delta(C_X)\times\Delta(C_Y)$ such that $u_X(x,\tau)\leq u_X(\sigma,\tau)$ for all $x\in C_X$ and $u_Y(\sigma,y)\leq u_Y(\sigma,\tau)$ for all $y\in C_Y$ (where we extend utilities by expectation in the usual fashion $u_X(x,\tau) = \int u_X(x,y)\,d\tau(y)$, etc.).
\end{definition}
In other words, a Nash equilibrium is a strategy pair in which each player is playing a best reply to his opponent's strategy.
\begin{definition}
A point $x$ in a (usually convex) subset $K$ of a real vector space is an \textbf{extreme point} if $x = \lambda y + (1-\lambda)z$ for $y,z\in K$ and $\lambda\in (0,1)$ implies $x = y = z$.
\end{definition}
The related notion of extreme ray will not be used until the next section, but we record it here for comparison.
\begin{definition}
A convex set $K$ such that $x\in K$ and $\lambda\geq 0$ implies $\lambda x \in K$ is called a \textbf{convex cone}. A point $x\neq 0$ is an \textbf{extreme ray} of the convex cone $K$ if $x = y + z$ and $y,z\in K$ implies that $y$ or $z$ is a scalar multiple of $x$.
\end{definition}
The Nash equilibria of games satisfying Assumption~\ref{assump:mp} take the following particularly simple form.
\begin{proposition}
\label{prop:nashzeromean}
A pair $(\sigma,\tau)\in\Delta(C_X)\times\Delta(C_Y)$ is a Nash equilibrium of a game satisfying Assumption~\ref{assump:mp} if and only if $\int x\,d\sigma(x) = \int y\,d\tau(y) = 0$.
\end{proposition}
\begin{proof}
If $\int x\,d\sigma(x) = 0$ then $u_Y(\sigma, y) = 0$ for all $y\in C_Y$, so any $\tau\in\Delta(C_Y)$ is a best response to $\sigma$. If $\int y\,d\tau(y) = 0$ as well then $\sigma$ is also a best response to $\tau$, so $(\sigma,\tau)$ is a Nash equilibrium.
Suppose for a contradiction that there exists a Nash equilibrium $(\sigma,\tau)$ such that $\int x\,d\sigma(x) > 0$; the other cases are similar. Player $y$ must play a best response, so $\int y\,d\tau(y) < 0$, which is possible by assumption. Player $x$ plays a best response to that, so $\int x\,d\sigma(x) < 0$, a contradiction.
\end{proof}
We introduce the notion of extreme Nash equilibrium in two-player zero-sum games. For an extension of this definition to two-player finite games and a proof that extreme Nash equilibria are always extreme points of the set of correlated equilibria in this setting, see \cite{c:ecnetpg} or \cite{er:nce}.
\begin{definition}
In a two-player zero-sum game, \textbf{maximin} and \textbf{minimax} strategies are those mixed strategies for player $X$ and $Y$, respectively, which appear in a Nash equilibrium. A Nash equilibrium of a zero-sum game is called \textbf{extreme} if $\sigma$ and $\tau$ are extreme points of the maximin and minimax sets, respectively.
\end{definition}
Applying Proposition~\ref{prop:nashzeromean} to this definition, we can characterize the extreme Nash equilibria of games satisfying Assumption~\ref{assump:mp}.
\begin{proposition}
\label{prop:extremenash}
Consider a game satisfying Assumption~\ref{assump:mp}. A pair $(\sigma,\tau)\in\Delta(C_X)\times\Delta(C_Y)$ is an extreme Nash equilibrium if and only if $\sigma$ and $\tau$ are each either $\delta_0$ or of the form $\alpha\delta_u + \beta\delta_v$ where $u<0$, $v, \alpha,\beta>0$, $\alpha+\beta = 1$, and $\alpha u + \beta v = 0$.
\end{proposition}
\begin{proof}
By Proposition~\ref{prop:nashzeromean} we must show that these distributions are the extreme points of the set of probability distributions having zero mean. Since $\delta_0$ is an extreme point of the set of probability distributions, it must be an extreme point of the subset which has zero mean. To see that $\alpha\delta_u + \beta\delta_v$ is also an extreme point, suppose we could write it as a convex combination of two other probability distributions with zero mean. The condition that both be positive measures implies that both must be of the form $\alpha'\delta_u + \beta'\delta_v$. But $\alpha$ and $\beta$ as specified above are the unique coefficients which make this be a probability measure with zero mean. Therefore $\alpha' = \alpha$ and $\beta' = \beta$, so $\alpha\delta_u + \beta\delta_v$ cannot be written as a nontrivial convex combination of probability distributions with zero mean, i.e., it is an extreme point.
Suppose $\sigma$ were an extreme point which was not of one of these types. Then $\sigma$ could not be supported on one or two points, so either $[0,1]$ or $[-1,0)$ could be partitioned into two sets of positive measure. We will only treat the first case; the second is similar. Let $[0,1] = A\cup B$ where $A\cap B = \emptyset$ and $\sigma(A),\sigma(B)>0$. Since $\sigma$ has zero mean we must have $\sigma([-1,0)) > 0$ as well.
For a set $D$ we define the restriction measure $\sigma\vert_D$ by $\sigma\vert_D(C) = \sigma(D\cap C)$ for all $C$. Then $\sigma = \sigma\vert_A + \sigma\vert_B + \sigma\vert_{[-1,0)}$. Let $a = \int_A x\,d\sigma(x)$, $b = \int_B x\,d\sigma(x)$, and $c = \int_{[-1,0)}x\,d\sigma(x)$. Since $\sigma([-1,0))>0$ and $x$ is less than zero everywhere on $[-1,0)$, we must have $c < 0$ and similarly $a,b\geq 0$. By assumption $a + b + c = 0$. Therefore we can write:
\begin{equation*}
\sigma = \left(\sigma\vert_A + \frac{a}{\lvert c \rvert} \sigma\vert_{[-1,0)}\right) + \left(\sigma\vert_B + \frac{b}{\lvert c \rvert}\sigma\vert_{[-1,0)}\right)
\end{equation*}
Being an extreme point of the set of probability measures with zero mean, $\sigma$ must be an extreme ray of the set of positive measures with first moment equal to zero. But this means that we cannot write $\sigma = \sigma_1 + \sigma_2$ where the $\sigma_i$ are positive measures with zero first moment unless $\sigma_i$ is a multiple of $\sigma$. Neither of the measures in parentheses above is a multiple of $\sigma$, so we have a contradiction.
\end{proof}
We illustrate this proposition on both examples introduced in Section~\ref{sec:examples}.
\usesavedcounter{exfin}
\begin{example}[cont'd]
In this case neither $C_X$ nor $C_Y$ contains zero, so the only extreme Nash equilibria are those in which $\sigma$ and $\tau$ are of the form $\alpha\delta_u + \beta\delta_v$ for $u<0$ and $v>0$. For any choice of $u$ and $v$ simple algebra gives unique $\alpha$ and $\beta$ satisfying the conditions of Proposition~\ref{prop:extremenash}. There are $n$ possible choices for each of $u$ and $v$ for each of the two players, so there are $n^4$ extreme Nash equilibria.
\end{example}
\setcounter{theorem}{\value{tempthm}
\usesavedcounter{exinf}
\begin{example}[cont'd]
Since $C_X = C_Y = [-1,1]$, there are infinitely many extreme Nash equilibria in this case. However, they are all finitely supported and the size of the support of each player's strategy is always either one or two. Furthermore the condition that $(\sigma,\tau)$ be a Nash equilibrium is equivalent to both having zero mean. This illustrates the general facts that in games with polynomial utility functions the Nash equilibrium conditions only involve finitely many moments of $\sigma$ and $\tau$ (in this case, only the mean) and the extreme Nash equilibria (when defined, say for zero-sum games) have uniformly bounded support \cite{karlin:tig}.
\end{example}
\setcounter{theorem}{\value{tempthm}
\section{Extreme correlated equilibria}
\label{sec:corr}
In this section we will show that even in finite games, the number of extreme correlated equilibria can be much larger than the number of extreme Nash equilibria. It makes sense to compare these because all extreme Nash equilibria of a two-player game, viewed as product distributions, are automatically extreme correlated equilibria \cite{c:ecnetpg,er:nce}.
In the case of polynomial games we will show that there can be extreme correlated equilibria with arbitrarily large finite support and with infinite support. This implies that the set of correlated equilibria cannot be characterized in terms of finitely many joint moments.
\paragraph{Roadmap}The analysis proceeds in several steps which will be technical at times, so we start with an outline of what follows.
\begin{itemize}
\item We begin by defining correlated equilibria in games satisfying Assumption~\ref{assump:mp} using a characterization from \cite{spo:cecgcc}.
\item Proposition~\ref{prop:mpcorreqchar} shows that this characterization can be simplified because of our choice of utility functions.
\item We use this characterization to construct a family of finitely supported extreme correlated equilibria in Proposition~\ref{prop:mpextfinsuppcorr}.
\item Then we note that all extreme correlated equilibria of the games in Example~\ref{ex:finite} are of this form, so this allows us to count the extreme correlated equilibria and determine their asymptotic rate of growth as the number of pure strategies grows.
\item Next we introduce some ideas from ergodic theory. With these tools in hand, we construct in Proposition~\ref{prop:mpextcorr} a large family of extreme correlated equilibria without finite support for the game in Example~\ref{ex:infinite}.
\item Finally we show that if a set can be represented by finitely many moments then all its extreme points have uniformly bounded finite support. This shows that the set of correlated equilibria of the game in Example~\ref{ex:infinite} cannot be represented by finitely many moments and completes the analysis.
\end{itemize}
Having completed the roadmap, we are ready to begin. Correlated equilibria are meant to capture the notion of a joint distribution of private recommendations to the two players such that neither player can expect to improve his payoff by deviating unilaterally from his recommendation. For finitely supported probability distributions and games satisfying Assumption~\ref{assump:mp}, this can be written as per the standard definition (see \cite{myerson:gt} or \cite{ft:gt}):
\begin{definition}
\label{def:ce}A finitely supported probability distribution $\mu\in\Delta(C_X\times C_Y)$ is a \textbf{correlated equilibrium} of a game satisfying Assumption~\ref{assump:mp} if
\[
\sum_{y\in C_Y} \mu(x,y)[xy - x'y]\geq 0
\]
for all $x,x'\in C_X$ and
\[
\sum_{x\in C_X}\mu(x,y)[xy'-xy]\geq 0
\]
for all $y,y'\in C_Y$ (note that these sums are finite by the assumption on $\mu$).
\end{definition}
The standard definition extending this notion to arbitrary (not necessarily finitely supported) distributions is given in \cite{hs:ece}. This definition is difficult to compute with, so we will use the following equivalent characterization.
\begin{proposition}[\cite{spo:cecgcc}]
\label{prop:correqchar1}A probability distribution $\mu\in\Delta(C_X\times C_Y)$ is a correlated equilibrium of a game satisfying Assumption~\ref{assump:mp} if and only if
\begin{equation*}
\int_{A\times I} (xy-x'y)\,d\mu(x,y)\geq 0\text{\ \ \ and\ \ \ }\int_{I\times A}(xy-xy')\,d\mu(x,y)\leq 0
\end{equation*}
for all $x'\in C_X$, $y'\in C_Y$, and measurable $A\subseteq I$.
\end{proposition}
\begin{proof}
When $\mu$ is finitely supported this is clearly equivalent to Definition~\ref{def:ce}. The general case is part $(1)$ of Corollary $2.14$ in \cite{spo:cecgcc} with the present utilities substituted in.
\end{proof}
Note that these conditions are homogeneous (that is, invariant under positive scaling) in $\mu$. The only condition on $\mu$ that is not homogeneous is the probability measure condition $\mu(I\times I) = 1$. We will often ignore this condition to avoid having to normalize every expression, referring to a measure $\mu\in\Delta^*(C_X\times C_Y)$ satisfying the conditions of the proposition as a correlated equilibrium.
\begin{definition}When we need to distinguish these notions, we will refer to a measure $\mu\in\Delta^*(C_X\times C_Y)$ satisfying the conditions of Proposition~\ref{prop:correqchar1} as a \textbf{homogeneous correlated equilibrium} and a measure $\mu\in\Delta(C_X\times C_Y)$ satisfying the conditions as a \textbf{proper correlated equilibrium}. In the context of homogeneous correlated equilibria the term \textbf{extreme} will refer to extreme rays; for proper correlated equilibria it will refer to extreme points.
\end{definition}
When $\mu\neq 0$ is a homogenous correlated equilibrium, $\frac{1}{\mu(I\times I)}\mu$ is a proper correlated equilibrium. The set of homogenous correlated equilibria is a convex cone. The extreme rays of this cone are exactly those measures which are positive multiples of the extreme points of the set of proper correlated equilibria.
The following proposition characterizes correlated equilibria of games satisfying Assumption~\ref{assump:mp} and is analogous to Proposition~\ref{prop:nashzeromean} for Nash equilibria. Note how the Nash equilibrium measures were characterized in terms of their moments but the correlated equilibria are not. Whereas the Nash equilibria are pairs of mixed strategies with zero mean for each player, condition (\ref{item:mpxint}) of this proposition says that the correlated equilibria are joint distributions such that regardless of each player's own recommendation, the conditional mean of his opponent's recommended strategy is zero.
\begin{proposition}
\label{prop:mpcorreqchar}
For a game satisfying Assumption~\ref{assump:mp} and a measure $\mu\in\Delta^*(C_X\times C_Y)$ such that $xy\neq 0$ $\mu$-a.e., the following are equivalent:
\begin{enumerate}
\item \label{item:mpcorreq}$\mu$ is a correlated equilibrium;
\item \label{item:mpxyint}
\begin{equation*}
\kappa_x(A) := \int_{A\times I} xy\,d\mu(x,y)\text{\ \ \ and\ \ \ }\kappa_y(A) := \int_{I\times A}xy\,d\mu(x,y)
\end{equation*}
are both the zero measure, i.e., equal zero for all measurable $A\subseteq I$;
\item \label{item:mpxint}
\begin{equation*}
\lambda_x(A) := \int_{A\times I} y\,d\mu(x,y)\text{\ \ \ and\ \ \ }\lambda_y(A) := \int_{I\times A} x\,d\mu(x,y)
\end{equation*}
are both the zero measure.
\end{enumerate}
\end{proposition}
\begin{proof}
(\ref{item:mpcorreq} $\Rightarrow$~\ref{item:mpxyint}) We will consider only $\kappa_x$; $\kappa_y$ is similar. The conditions of Proposition~\ref{prop:correqchar1} with $A = I$ imply that
\begin{equation*}
x'\int_{I\times I}y\,d\mu(x,y) \leq \int_{I\times I}xy\,d\mu(x,y)\leq y'\int_{I\times I}x\,d\mu(x,y)
\end{equation*}
for all $x'\in C_X, y'\in C_Y$. By assumption it is possible to choose $x'$ and $y'$ either positive or negative, so $\int_{I\times I}xy\,d\mu(x,y) = 0$. A similar argument with any $A$ implies that $\int_{A\times I}xy\,d\mu(x,y)\geq 0$. Therefore we have
\begin{equation*}
0 = \int_{I\times I}xy\,d\mu(x,y) = \int_{A\times I}xy\,d\mu(x,y) + \int_{(I\setminus A)\times I}xy\,d\mu(x,y)\geq 0 + 0 = 0
\end{equation*}
for all $A$, so the inequality must be tight and we get $\int_{A\times I} xy\,d\mu(x,y) = 0$ for all $A$.
(\ref{item:mpxyint} $\Leftrightarrow$~\ref{item:mpxint}) By definition $d\kappa_x = x\,d\lambda_x$ and by assumption $\lambda_x(0) = 0$. If one of these measures is zero then so is the other, and respectively with $y$ in place of $x$.
(\ref{item:mpxyint} \& \ref{item:mpxint} $\Rightarrow$~\ref{item:mpcorreq}) The integrals in Proposition~\ref{prop:correqchar1} vanish.
\end{proof}
\begin{proposition}
\label{prop:mpextfinsuppcorr}
Fix a game satisfying Assumption~\ref{assump:mp}. Let $k > 0$ be even and $x_1,\ldots, x_{2k}$ and $y_1,\ldots,y_{2k}$ be such that:
\begin{enumerate}
\item $x_i\in C_X$ and $y_i\in C_Y$ are all nonzero;
\item the sequence $x_1,x_3,\ldots, x_{2k-1}$ has distinct elements and alternates in sign;
\item the sequence $y_1,y_3,\ldots,y_{2k-1}$ has distinct elements and alternates in sign;
\item $x_{2i} = x_{2i-1}$ and $y_{2i} = y_{2i+1}$ for all $i$ when subscripts are interpreted $\mod 2k$.
\end{enumerate}
Then $\mu = \sum_{i=1}^{2k} \frac{1}{\lvert x_i y_i\rvert}\delta_{(x_i,y_i)}$ is an extreme correlated equilibrium.
\end{proposition}
\begin{proof}
To show that $\mu$ is a correlated equilibrium define $d\kappa(x,y) = xy\,d\mu(x,y)$. Then $\kappa = \sum_{i=1}^{2k} \sign(x_i)\sign(y_i)\delta_{(x_i,y_i)}$. Defining the projection $\kappa_x$ as in Proposition~\ref{prop:mpcorreqchar}, we have
\begin{equation*}
\begin{split}
\kappa_x & = \sum_{i=1}^{2k} \sign(x_i)\sign(y_i)\delta_{x_i} = \sum_{i=1}^k \sign(x_{2i})\left(\sign(y_{2i}) + \sign(y_{2i-1})\right)\delta_{x_{2i}} \\ &= \sum_{i=1}^k\sign(x_{2i})(0)\delta_{x_{2i}} = 0,
\end{split}
\end{equation*}
because $x_{2i} = x_{2i-1}$ and $y_{2i}$ differs in sign from $y_{2i-1}$ by assumption. The same argument shows that $\kappa_y = 0$, so $\mu$ is a correlated equilibrium.
To see that $\mu$ is extreme, suppose $\mu = \mu' + \mu''$ where $\mu'$ and $\mu''$ are correlated equilibria. Clearly $\mu' = \sum_{i=1}^{2k}\alpha_i\delta_{(x_i,y_i)}$ for some $\alpha_i\geq 0$. Define $d\kappa' = xy\,d\mu'(x,y)$, so $\kappa' = \sum_{i=1}^{2k}\alpha_ix_iy_i\delta_{(x_i,y_i)}$. By assumption
\begin{equation*}
\kappa'_x = \sum_{i=1}^{k}x_{2i}\left(\alpha_{2i-1}y_{2i-1} + \alpha_{2i}y_{2i}\right)\delta_{x_{2i}}
\end{equation*}
is the zero measure. Since the $x_{2i}$ are distinct and nonzero we must have $\alpha_{2i-1}y_{2i-1} + \alpha_{2i}y_{2i} = 0$ for all $i$. Similarly since $\kappa'_y = 0$ we have $\alpha_{2i+1}x_{2i+1} + \alpha_{2i}x_{2i} = 0$ for all $i$ (with subscripts interpreted $\mod 2k$).
The $x_i$ and $y_i$ are all nonzero, so fixing one $\alpha_i$ fixes all the others by these equations. That is to say, these equations have a unique solution up to multiplication by a scalar, so $\mu'$ is a positive scalar multiple of $\mu$. But the splitting $\mu = \mu' + \mu''$ was arbitrary, so $\mu$ is extreme.
\end{proof}
An argument along the lines of the proof of Proposition~\ref{prop:mpextfinsuppcorr} shows that any finitely supported correlated equilibrium $\mu$ whose support does not contain any points with $x=0$ or $y=0$ can be written as $\mu = \mu' + \mu''$ where $\mu'\neq 0$ is a correlated equilibrium and $\mu''\neq 0$ is a correlated equilibrium of the form studied in Proposition~\ref{prop:mpextfinsuppcorr}. Therefore a finitely supported $\mu$ cannot be extreme unless it is of this form.
\usesavedcounter{exfin}
\begin{example}[cont'd]
For some examples of the supports of extreme correlated equilibria of games of this type, see Figures~\ref{fig:exfin1} and~\ref{fig:exfin2}.
\begin{figure
\begin{center}
\includegraphics[width = 0.4\textwidth]{ex1}
\caption{The support of an extreme correlated equilibrium. In the notation of Proposition~\ref{prop:mpextfinsuppcorr}, $k=2$, $x_1 = 0.4$, $x_3 = -0.6$, $y_1 = 0.2$, and $y_3 = -0.8$.}
\label{fig:exfin1}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[width=0.4\textwidth]{ex2}
\caption{The support of another extreme correlated equilibrium. In the notation of Proposition~\ref{prop:mpextfinsuppcorr}, $k=4$, $x_1 = 0.4$, $x_3 = -0.4$, $x_5 = 0.6$, $x_7 = -0.6$, $y_1 = 0.6$, $y_3 = -0.4$, $y_5 = 0.4$, and $y_7 = -0.6$.}
\label{fig:exfin2}
\end{center}
\end{figure}
To count the number of extreme correlated equilibria of this game we must count the number of essentially different sequences of $x_i$ and $y_i$ of the type mentioned in Proposition~\ref{prop:mpextfinsuppcorr}. Fix $k$ and let $k = 2r$ where $1\leq r \leq n$. Note that cyclically shifting the sequences of $x_i$'s and $y_i$'s by two does not change $\mu$, nor does reversing the sequence. Therefore we can assume without loss of generality that $x_1,y_1 > 0$. We then have $n$ possible choices for $x_1,y_1,x_3,\text{ and }y_3$, $n-1$ possible choices for $x_5,x_7,y_5,\text{ and }y_7$, etc., for a total of $\left(\frac{n!}{(n-r)!}\right)^4$ possible choices of the $x_i$ and $y_i$. These will always be essentially different (i.e., give rise to different $\mu$) unless we cyclically permute the sequences of $x_i$ and $y_i$ by some multiple of four, in which case the resulting sequence is essentially the same. The number of such cyclic permutations is $r$. Therefore the total number of extreme correlated equilibria is
\begin{equation*}
e(n) = \sum_{r = 1}^n \frac{1}{r}\left(\frac{n!}{(n-r)!}\right)^4.
\end{equation*}
We will see that $e(n) = \Theta\left(\frac{1}{n}(n!)^4\right)$. That is to say, $e(n)$ is asymptotically upper and lower bounded by a constant times $\frac{1}{n}(n!)^4$. The expression $\frac{1}{n}(n!)^4$ is just the final term in the summation for $e(n)$, so the lower bound is clear. Define
\begin{equation*}
f(n) = \frac{e(n)}{\frac{1}{n}(n!)^4} = \sum_{s = 0}^{n-1} \frac{n}{n-s}\cdot\frac{1}{(s!)^4}.
\end{equation*}
Then $f(n)\geq 1$ for all $n$. We will now show that $f(n)$ is also bounded above. Intuitively this is not surprising since the terms in the summation for $f(n)$ die off extremely quickly as $s$ grows.
For all $1\leq s < n - 1$ we have that the ratio of term $s+1$ in the summation to term $s$ is:
\begin{equation*}
\frac{\frac{n}{n-s-1}\cdot\frac{1}{((s+1)!)^4}}{\frac{n}{n-s}\cdot\frac{1}{(s!)^4}} = \frac{n-s}{n-s-1}\cdot\frac{1}{(s+1)^4} \leq \frac{1}{8},
\end{equation*}
so for $n > 1$ we can bound the sum by a geometric series:
\begin{equation*}
f(n) - 1 = \sum_{s=1}^{n-1} \frac{n}{n-s}\cdot\frac{1}{(s!)^4} \leq \frac{n}{n-1}\sum_{t=0}^{\infty}\frac{1}{8^t} = \frac{8n}{7(n-1)}\leq \frac{16}{7}.
\end{equation*}
Therefore $1\leq f(n)\leq\frac{23}{7}$ for all $n$, so $e(n) = \Theta\left(\frac{1}{n}(n!)^4\right)$ as claimed. Comparing this to the results of the previous section in which we saw that the number of extreme Nash equilibria of this game is $n^4$, we see that in this case there is a super-exponential separation between the number of extreme Nash and the number of extreme correlated equilibria. This implies, for example, that computing all extreme correlated equilibria is not an efficient method for computing all extreme Nash equilibria, even though all extreme Nash equilibria are extreme correlated equilibria and recognizing whether an extreme correlated equilibrium is an extreme Nash equilibrium is easy. There are simply too many extreme correlated equilibria.
\end{example}
\setcounter{theorem}{\value{tempthm}
Next we will prove a more abstract version of Proposition~\ref{prop:mpextfinsuppcorr} which includes certain extreme points which are not finitely supported. Before doing so we need a brief digression to ergodic theory. The first definition is the standard definition of compatibility between a measure and a transformation on a space. The second definition expresses one notion of what it means for a transformation to ``mix up'' a space -- in this case that the space cannot be partitioned into two sets of positive measure which do not interact under the transformation. Then we state the main ergodic theorem and a corollary which we will apply to exhibit extreme correlated equilibria of games satisfying Assumption~\ref{assump:mp}.
\begin{definition}
Given a measure $\mu\in\Delta^*(S)$ on a space $S$, a measurable function $g: S\rightarrow S$ is called \textbf{($\mu$-)measure preserving} if $\mu(g^{-1}(A)) = \mu(A)$ for all measurable $A\subseteq S$. Note that if $g$ is invertible (in the measure theoretic sense that an almost everywhere inverse exists), then this is equivalent to the condition that $\mu(g(A)) = \mu(A)$ for all $A$.
\end{definition}
\begin{definition}
Given a measure $\mu\in\Delta^*(S)$, a $\mu$-measure preserving transformation $g$ is called \textbf{ergodic} if $\mu(A\bigtriangleup g^{-1}(A))=0$ implies $\mu(A) = 0$ or $\mu(A) = \mu(S)$, where $A\bigtriangleup B$ denotes the symmetric difference $(A\setminus B) \cup (B\setminus A)$.
\end{definition}
\begin{example}
Fix a finite set $S$ and a function $g: S\rightarrow S$. Let $\mu$ be counting measure on $S$. Then $g$ is measure preserving if and only if it is a permutation. In this case a set $T$ satisfies $\mu(g^{-1}(T)\bigtriangleup T) = 0$ if and only if $g^{-1}(T) = T$ if and only if $T$ is a union of cycles of $g$. Therefore $g$ is ergodic if and only if it consists of a single cycle.
\end{example}
\begin{example}
\label{ex:ergrot}
Fix $\alpha\in\mathbb{R}$. Let $S = [0,1)$ and let $\mu$ be Lebesgue measure on $S$. Define $g: S\rightarrow S$ by $g(x) = (x+\alpha)\mod 1 = (x+\alpha)-\lfloor x+\alpha\rfloor$. Then $g$ is $\mu$-measure preserving because Lebesgue measure is translation invariant. It can be shown that $g$ is ergodic if and only if $\alpha$ is irrational. For a proof and more examples, see \cite{silva:iet}.
\end{example}
The following is one of the core theorems of ergodic theory. We will only use it to prove the corollary which follows, so it need not be read in detail. The proof can be found in any text on ergodic theory, e.g.\ \cite{silva:iet}.
\begin{theorem}[Birkhoff's ergodic theorem]
Fix a probability measure $\mu$ and a $\mu$-measure preserving transformation $g$. Then for any $f\in \mathcal{L}^1(\mu)$:
\begin{itemize}
\item $\tilde{f}(x) = \lim_{n\rightarrow\infty} \frac{1}{n}\sum_{k=0}^{n-1}f(g^k(x))$ exists $\mu$-almost everywhere,
\item $\tilde{f}\in\mathcal{L}^1(\mu)$,
\item $\int \tilde{f}\,d\mu = \int f\,d\mu$,
\item $\tilde{f}(g(x)) = \tilde{f}(x)$ $\mu$-almost everywhere, and
\item if $g$ is ergodic then $\tilde{f}(x) = \int f\,d\mu$ $\mu$-almost everywhere.
\end{itemize}
\end{theorem}
\begin{corollary}
\label{cor:uniquemeasure}
Suppose $\mu$ and $\nu$ are probability measures such that $\nu$ is absolutely continuous with respect to $\mu$. If a transformation $g$ preserves both $\mu$ and $\nu$ and $g$ is ergodic with respect to $\mu$, then $\nu = \mu$.
\end{corollary}
\begin{proof}Fix any measurable set $A$. Let $f$ be the indicator function for $A$, i.e.\ the function equal to unity on $A$ and zero elsewhere. Applying Birkhoff's ergodic theorem to $f$ and $\mu$ yields $\tilde{f}(x) = \mu(A)$ $\mu$-almost everywhere. Since $\nu$ is absolutely continuous with respect to $\mu$, $\tilde{f}(x) = \mu(A)$ $\nu$-almost everywhere also. If we now apply Birkhoff's ergodic theorem to $\nu$ we get:
\begin{equation*}
\nu(A) = \int f \,d\nu = \int \tilde{f}\,d\nu = \int \mu(A)\,d\nu = \mu(A).\qedhere
\end{equation*}
\end{proof}
We now construct a family of extreme correlated equilibria.
\begin{proposition}
\label{prop:mpextcorr}
Fix measures $\nu_1, \nu_2, \nu_3$, and $\nu_4\in\Delta^*((0,1])$ and maps $f_i: (0,1]\rightarrow (0,1]$ such that $\nu_{i+1} = \nu_i\circ f_i^{-1}$ (interpreting subscripts $\mod 4$). The portion of the measure $\mu$ in the $i^{\text{th}}$ quadrant of $I\times I$ will be constructed in terms of $f_i$ and $\nu_i$. Define $j_i:(0,1]\rightarrow I\times I$ by $j_1(x) = (x,f_1(x))$, $j_2(x) = (-f_2(x),x)$, $j_3(x) = (-x,-f_3(x))$, and $j_4(x) = (f_4(x),-x)$. Let $\lvert\kappa\rvert = \sum_{i=1}^4 \nu_i\circ j_i^{-1}$. If Assumption~\ref{assump:mp} is satisfied, $\supp\lvert\kappa\rvert\subseteq C_X\times C_Y$, and $\frac{1}{\lvert xy\rvert}\in\mathcal{L}^1(\lvert\kappa\rvert)$ then $d\mu = \frac{1}{\lvert xy\rvert}\,d\lvert\kappa\rvert$ is a correlated equilibrium.
By assumption
$f_4\circ f_3\circ f_2\circ f_1:(0,1]\rightarrow (0,1]$ is $\nu_1$-measure preserving. If it is also ergodic with respect to $\nu_1$, then $\mu$ is extreme.
\end{proposition}
\begin{proof}
First we must show that $\mu$ is a correlated equilibrium. It is a finite measure by the assumption $\frac{1}{\lvert xy\rvert}\in\mathcal{L}^1(\lvert\kappa\rvert)$ and $xy\neq 0$ $\mu$-a.e.\ by definition. Define $g: I\times I \rightarrow I\times I$ as follows.
\begin{equation*}
g(x,y) = \begin{cases}
j_1(x) & \text{if }x>0, y<0 \\
j_2(y) & \text{if }x>0, y>0 \\
j_3(-x) & \text{if }x<0,y>0 \\
j_4(-y) & \text{if }x<0,y<0 \\
\text{arbitrary} & \text{otherwise}
\end{cases}
\end{equation*}
The function $g$ is $\lvert\kappa\rvert$-measure preserving. To see this fix any measurable set $B\subseteq (0,1]\times(0,1]$. Let $A = j_1^{-1}(B)$. Then $\lvert\kappa\rvert(B) = \lvert\kappa\rvert(A\times (0,1]) = \nu_1(A)$ by definition of $\lvert\kappa\rvert$. But $g^{-1}(B) = g^{-1}(A\times(0,1]) = A\times [-1,0)$, so
\begin{equation*}
\lvert\kappa\rvert(g^{-1}(B)) = \lvert\kappa\rvert(A\times [-1,0)) = \nu_4(j_4^{-1}(A\times [-1,0))) = \nu_4(f_4^{-1}(A)) = \nu_1(A) = \lvert\kappa\rvert(B).
\end{equation*}
Therefore $g$ is measure preserving for subsets of $(0,1]\times (0,1]$. The arguments for the other quadrants are similar and since $g$ maps each quadrant into a different quadrant, $g$ is measure preserving on its entire domain.
Define the signed measure $\kappa$ by $d\kappa = xy \,d\mu = \sign(x)\sign(y)\,d\lvert\kappa\rvert$. We have seen that $\lvert\kappa\rvert(A\times (0,1]) = \lvert\kappa\rvert(A\times [-1,0))$, so $\kappa(A\times (0,1]) = -\kappa(A\times [-1,0))$. Since $\kappa(A\times \{0\}) = 0$, we have $\kappa(A\times I) = 0$, or using the terminology of Proposition~\ref{prop:mpcorreqchar}, $\kappa_x(A) = 0$. A similar argument implies $\kappa_x(A) = 0$ if $A\subseteq [-1,0)$. Clearly $\kappa_x(0) = 0$ by definition of $\kappa_x$, so $\kappa_x$ is the zero measure. In the same way we can show that $\kappa_y$ is the zero measure, so $\mu$ is a correlated equilibrium by Proposition~\ref{prop:mpcorreqchar}.
Now we will show via several steps that $\mu$ is extreme. Write $\mu = \mu_1 + \mu_2$ where the $\mu_i$ are nonzero correlated equilibria. Since these are all positive measures, the $\mu_i$ are absolutely continuous with respect to $\mu$.
Define $d\lvert\kappa_i\rvert = \lvert xy\rvert \,d\mu_i$ and $d\kappa_i = xy\,d\mu_i$
Next we show that $g$ is $\lvert\kappa_i\rvert$-measure preserving. We will demonstrate this fact for $B\subseteq (0,1]\times (0,1]$. As above, we define $A = j_1^{-1}(B)$. Then $\lvert\kappa_i\rvert(B) = \lvert\kappa_i\rvert(A\times (0,1])$ since $(A\times (0,1]) \bigtriangleup B$ has $\lvert\kappa\rvert$ measure zero and $\lvert\kappa_i\rvert$ is absolutely continuous with respect to $\lvert\kappa\rvert$. Furthermore, $\lvert\kappa_i\rvert(g^{-1}(B)) = \lvert\kappa_i\rvert(A\times [-1,0))$. But $\mu_i$ is a correlated equilibrium so $\kappa_i(A\times (0,1]) = -\kappa_i(A\times [-1,0))$. Hence $\lvert\kappa_i\rvert(g^{-1}(B)) = \lvert\kappa_i\rvert(A\times [-1,0)) = \lvert\kappa_i\rvert(A\times (0,1]) = \lvert\kappa_i\rvert(B)$. Again, the proof is the same for $B$ contained in other quadrants, so $g$ is $\lvert\kappa_i\rvert$-measure preserving.
For the second-to-last step we prove that $g$ is ergodic with respect to $\lvert\kappa\rvert$. Suppose $B\subseteq I\times I$ is such that $\lvert\kappa\rvert(g^{-1}(B)\bigtriangleup B) = 0$. Let $Q_i$ be the intersection of $B$ with the $i^{\text{th}}$ quadrant. Then $\lvert\kappa\rvert(g^{-1}(Q_{i+1})\bigtriangleup Q_i) = 0$, so $\lvert\kappa\rvert(g^{-4}(Q_1) \bigtriangleup Q_1) = 0$. Let $A = j_1^{-1}(Q_1)$. Then $\lvert\kappa\rvert(g^{-4}(Q_1)\bigtriangleup Q_1) = \nu_1((f_4\circ f_3 \circ f_2\circ f_1)^{-1}(A) \bigtriangleup A) = 0$. By assumption the map $f_4\circ f_3\circ f_2\circ f_1$ is ergodic, so $\nu_1(A) = 0$ or $\nu_1(A) = \nu_1((0,1]) = \lvert\kappa\rvert((0,1]\times(0,1])$. Therefore $\lvert\kappa\rvert(Q_1) = \nu_1(A) = 0$ or $\lvert\kappa\rvert(Q_1) = \lvert\kappa\rvert((0,1]\times(0,1])$. In either case since $g$ is $\lvert\kappa\rvert$-measure preserving we get $\lvert\kappa\rvert(Q_i) = \lvert\kappa\rvert(Q_1)$ for all $i$. Therefore $\lvert\kappa\rvert(B) = 0$ or $\lvert\kappa\rvert(B) = \lvert\kappa\rvert(I\times I)$, so $g$ is ergodic with respect to $\lvert\kappa\rvert$.
Normalizing $\lvert\kappa\rvert$ and $\lvert\kappa_i\rvert$ to be probability measures, we can apply Corollary~\ref{cor:uniquemeasure} to obtain $\lvert\kappa_i\rvert = \frac{\lvert\kappa_i\rvert(I\times I)}{\lvert\kappa\rvert(I\times I)}\lvert\kappa\rvert$. By definition the set on which $\lvert xy \rvert$ is zero has $\mu$ measure zero. Therefore
\begin{equation*}
d\mu_i = \frac{1}{\lvert xy \rvert}\,d\lvert\kappa_i\rvert = \frac{\lvert\kappa_i\rvert(I\times I)}{\lvert\kappa\rvert(I\times I)}\frac{1}{\lvert xy \rvert}\,d\lvert\kappa\rvert = \frac{\lvert\kappa_i\rvert(I\times I)}{\lvert\kappa\rvert(I\times I)}\,d\mu,
\end{equation*}
so $\mu_i = \frac{\lvert\kappa_i\rvert(I\times I)}{\lvert\kappa\rvert(I\times I)} \mu$ and $\mu$ is extreme.
\end{proof}
Above we have constructed $\mu$ and $g$ so that $g$ maps the quadrants counter-clockwise -- quadrant $1$ to quadrant $2$, etc. However, the same argument would go through if $g$ mapped the quadrants clockwise.
To view Proposition~\ref{prop:mpextfinsuppcorr} as a special case of Proposition~\ref{prop:mpextcorr}, let each $\nu_i$ be a uniform probability measure over a finite subset of $(0,1]$. The function $g$ is defined by $g(x_i,y_i) = (x_{i+1},y_{i+1})$ and the $f_i$ are defined to be compatible with this. The map $f_4\circ f_3\circ f_2\circ f_1$ is a permutation on the support of $\nu_1$, which is precisely the positive values of $x_i$. By construction this permutation consists of a single cycle, hence it is ergodic.
\usesavedcounter{exinf}
\begin{example}[cont'd]
We can combine Example~\ref{ex:ergrot} and Proposition~\ref{prop:mpextcorr} to exhibit extreme points of the set of correlated equilibria for this game which are not finitely supported. Let $0 < a < b < 1$. Let $\nu_i$ be Lebesgue measure on $[a,b)$ for all $i$. Fix $\alpha$ such that $\frac{\alpha}{b-a}$ is irrational. Define $f_1: [a,b)\rightarrow [a,b)$ by $f(x) = (x-a+\alpha\mod (b-a)) + a$. This is just an affinely scaled version of Example~\ref{ex:ergrot} so $f_1$ is $\nu_i$-measure preserving and ergodic. Define $f_1$ on $(0,1]\setminus [a,b)$ arbitrarily, because that is a set of measure zero. Let $f_2,f_3,f_4: (0,1]\rightarrow (0,1]$ be the identity. These data satisfy all the assumptions of Proposition~\ref{prop:mpextcorr}. In particular, since $0<a<b<1$, $xy$ is bounded away from zero on the support of $\lvert\kappa\rvert$. Therefore $\frac{1}{\lvert xy \rvert}\in\mathcal{L}^1(\lvert\kappa\rvert)$. Since $\nu_i$ is not finitely supported, $\mu$ is an extreme correlated equilibrium which is not finitely supported. The support of $\mu$ is shown in Figure~\ref{fig:exinf} with parameters $a = 0.2$, $b = 0.8$, and $\alpha = \frac{1}{\sqrt{5}}$.
\setcounter{theorem}{\value{tempthm}
\begin{figure
\begin{center}
\includegraphics[width = 0.4\textwidth]{ex3}
\caption{The support set of an extreme correlated equilibrium which is not finitely supported. Extremality of this equilibrium depends sensitively on the choices of endpoints for the line segments. In this case there are segments connecting: $(0.2,-0.2)$ to $(0.8, -0.8)$; $(-0.2,-0.2)$ to $(-0.8,-0.8)$; $(-0.2,0.2)$ to $(-0.8,0.8)$; $\left(0.2,0.2+\frac{1}{\sqrt{5}}\right)$ to $\left(0.8-\frac{1}{\sqrt{5}},0.8\right)$; and $\left(0.8-\frac{1}{\sqrt{5}},0.2\right)$ to $\left(0.8,0.2+\frac{1}{\sqrt{5}}\right)$.}
\label{fig:exinf}
\end{center}
\end{figure}
\end{example}
\setcounter{theorem}{\value{tempthm}
\begin{definition}Given a compact Hausdorff space $K$ we say that a set of measures $\mathcal{M}\subseteq\Delta^*(K)$ is \textbf{describable by moments} if there exists an integer $d$, bounded Borel measurable maps $g_1,\ldots,g_d: K\rightarrow\mathbb{R}$, and a set $M\subseteq\mathbb{R}^d$ such that a measure $\mu$ is in $\mathcal{M}$ if and only if $\left(\int g_1\,d\mu,\ldots,\int g_d\,d\mu\right)\in M$.
\end{definition}
The results of \cite{karlin:tig} show that the maximin and minimax strategy sets of a two-player zero-sum polynomial game can always be described by moments. Introducing a similar notion for $n$-tuples of measures, the set of Nash equilibria can always be described by moments in any polynomial game \cite{sop:slrcg}. However, combining this example with the following proposition we see that the set of correlated equilibria of a polynomial game cannot in general be described by moments.
This is important because the finite-dimensional representation in terms of moments is the primary tool for computing and characterizing Nash equilibria of polynomial games. One is therefore naturally drawn to try to find such a representation for the set of correlated equilibria. The example and this proposition show that no such representation exists in general.
\begin{proposition}
\label{prop:momentmeasure}
Let $\mathcal{M}\subseteq\Delta^*(K)$ be a set of measures describable by moments. Then all extreme points of $\mathcal{M}$ have finite support and this support is uniformly bounded by $d$, where $d$ is the integer associated with the description of $\mathcal{M}$ by moments.
\end{proposition}
\begin{proof}
Let $g_1,\ldots,g_d: K\to\mathbb{R}$ be the maps describing $\mathcal{M}$. Suppose there exists a measure $\mu\in\mathcal{M}$ which is extreme and supported on more than $d$ points, so we can partition the domain of $\mu$ into $d+1$ sets $B_1,\ldots,B_{d+1}$ of positive measure. For $c = (c_1,\ldots, c_{d+1})\in\mathbb{R}^{d+1}_{\geq 0}$, define $\mu_c = \sum_{i = 1}^{d+1} c_i \mu\vert_{B_i}$. The map $c\mapsto \mu_c$ is injective. Define
\begin{equation*}
K = \left\{c\in\mathbb{R}^{d+1}_{\geq 0} \,\bigg\vert\, \int g_i\,d\mu_c=\int g_i\,d\mu\text{ for }i =1,\ldots, d\right\},
\end{equation*}
so $(1,1,\ldots, 1)\in K$. Linearity of integration implies that the nonempty set $K$ is the intersection of an affine space of dimension at least one with the positive orthant. By Carath\'{e}odory's theorem (or equivalently the statement that a feasible linear program has a basic feasible solution), the extreme points of $K$ each have at most $d$ nonzero entries \cite{bno:convex}. Thus $(1,1,\ldots, 1)$ is not an extreme point of $K$, so we can write $(1,1,\ldots, 1) = \lambda c + (1 - \lambda) c'$ for $0<\lambda < 1$ and $(1,1,\ldots, 1)\neq c,c'\in K$. Therefore $\mu = \mu_{(1,1\ldots, 1)} = \lambda \mu_c + (1-\lambda)\mu_{c'}$ is not extreme.
\end{proof}
\section{Future work}
These results leave several open questions. If we define a moment map to be any map of the form $\pi\mapsto \left(\int f_1\,d\pi,\int f_2\,d\pi,\ldots,\int f_k\,d\pi\right)$ for bounded Borel measurable $f_i$, then we have shown that the set of correlated equilibria is not the inverse image of any set under any moment map. On the other hand, since moment maps are linear and weak* continuous, we know that the image of the set of correlated equilibria under any moment map is convex and compact.
Supposing the utilities and the $f_i$ are polynomials, is there anything more we can say about this image? In particular, is it semialgebraic (i.e.,\ describable in terms of finitely many polynomial inequalities)? If so, can we compute these inequalities or a solution thereof efficiently for given utilities? A sequence of easily computed outer bounds to this image is presented in \cite{spo:cecgcc}. Can we compute nonempty inner bounds?
\section*{Acknowledgements}
The first author would like to thank Prof. Cesar E. Silva for many discussions about ergodic theory, and in particular for the simple proof of Corollary~\ref{cor:uniquemeasure} using Birkhoff's ergodic theorem.
\bibliographystyle{plain}
|
\section{Introduction}
Black holes are the most striking predictions of Einstein's general theory of relativity of gravitation. A possible identification of black holes as thermodynamical systems has unanimously evolved over the years.
The study of black hole thermodynamics is therefore quite important. However, if one analyses properties like thermodynamical stability, phase transition, black hole evaporation etc. the semi-classical results in the canonical ensemble are not much encouraging\cite{BHT}. This is because the specific heat is always negative and therefore in a canonical ensemble it is not possible to attain a stable equilibrium with the surroundings. However, results in a microcanonical ensemble indicate the possibility of stability \cite{BHT,davies}.
The aim of the present paper is to make a thorough analysis of the possibility of phase transition and stability in black holes, within the framework of a canonical ensemble. As a specific example we consider the Kerr black hole which provides a striking difference from the usual spherically symmetric solutions of Einstein's equation. The main point is the appearance of a work term like $\Omega dJ$, where $\Omega$ and $J$ are the angular velocity and angular momentum, respectively, in the first law of black hole thermodynamics. As shown later, this is the crucial term for classifying phase transition. As we have already hinted in the previous paragraph it is necessary to go beyond the standard semiclassical approximation to discuss phase transition. The results for various thermodynamic variables, for the Kerr black hole, beyond the semiclassical approximation were recently obtained by two of us \cite{Modak}. In particular, the entropy was shown to receive logarithmic corrections. Such corrections have some interesting applications in braneworld cosmology \cite{Lidsey:2002ah}. In this picture the FRW equations get modified leading to a different early/late time cosmology. Also, it has been suggested that \cite{Lidsey:2002ah} in the presence of logarithmic corrections, small black holes can be stable even in a flat back ground. Our calculations in this paper actually quantify such suggestions by systematically analysing phase transition and aspects of stability. Although the logarithmic corrections to be used here were obtained in \cite{Modak} the utility of such terms in the context of phase transition were neither stated nor analyzed in \cite{Modak}. This is the new input in this paper.
Using the corrected expressions of different thermodynamic entities derived in \cite{Modak} we show that there exists a phase transition through which a black hole can switch over to a stable phase from an initial unstable phase. The continuity of the corrected entropy for any value of the black hole mass rules out the existence of a first order phase transition. However we find that there is a discontinuity in specific heat for a critical value of the mass. At this mass the black hole temperature is maximum. Then it sharply decreases finally attaining the absolute zero for a vanishing mass. Therefore this indicates a transition from an unstable phase to a stable phase. Having found this possibility of phase transition, we then analyze the nature of this phase transition. We first derive two Ehrenfest like equations for the rotating Kerr black hole. If this phase transition is second order in nature the Ehrenfest's equations must be satisfied. A numerical analysis reveals that the first Ehrenfest's equation is satisfied for a long range of angular velocity but the second Ehrenfest's equation is violated. Hence a second order phase transition is also ruled out.
The final possibility is the occurrence of a glassy phase transition. It is found from experiments that in a liquid to glass phase transition, the specific heat ($C_p$), compressibility ($k$) and thermal expansivity ($\alpha$) {\it i.e.} the second order derivatives of the Gibbs free energy, make a smeared jump. For this type of transition the Prigogine-Defay (PD) ratio \cite{jackle}
measures the deviation from the second Ehrenfest's equation. For a second order phase transition $\Pi=1$ exactly but for a glassy phase transition $\Pi$ varies from 2 to 5.
In the present paper we have shown that the second derivatives of free energy suffer a smeared discontinuity at the phase transition point and the PD ratio as found by us for the Kerr black hole is nearly 3. This enables us to conclude that the transition of the black hole from an unstable to a stable phase is a glassy phase transition.
In this paper the coefficient of the correction term to the semi-classical entropy is taken to be $\frac{1}{90}$ which was obtained in \cite{Modak}. To check the robustness of our conclusion it is desirable to take other normalisations. However, here we are faced with a lack of data. This is because, apart from our work \cite{Modak}, there is no other computation which explicitly gives the coefficient of the logarithmic term for the Kerr black hole. We therefore choose arbitrarily some values for this normalisation. Specifically we have taken this factor to be $\frac{1}{2}$ (which is also less than 1, like $\frac{1}{90}$) and $\frac{3}{2}$ (which is greater than 1) as well as 1. In all cases we observe a glassy phase transition with the same PD ratio. This establishes the soundness of our conclusion.
The novelty of this paper is not only confined to its results but also to the approach. The issues related to the thermodynamic stability of a black hole have never been explored by going beyond the semi-classical regime. As we have shown it is essential to go beyond this approximation to study thermodynamic stability through a phase transition. Of course there is a wide literature on the Hawking-Page \cite{Hawkpage} phase transition which however occurs only in a black hole of AdS space \cite{group}. In this paper, on the other hand we discuss phase transition and thermodynamic stability in an asymptotically flat Kerr black hole.
The paper is organised in the following manner. In section 2 we study the Kerr black hole as a thermodynamical system, staying within the semi-classical regime. In the next section we go beyond the semi-classical approximation and discover a phase transition for the Kerr black hole. Section 4 is devoted to analyse the nature of this phase transition. Here we show that it is a glassy phase transition which makes the Kerr black hole thermodynamically stable. We give our conclusions in section 5. There are two appendices presented at the end of the paper. In appendix 1 we derive two Ehrenfest's equations considering the Kerr black hole as a thermodynamical system and in appendix 2 basic results are compiled to prove the robustness of our conclusions for different normalisations of the correction (logarithmic) term.
\section{Kerr black hole as a thermodynamical system; a semi-classical study}
The Kerr black hole is an axisymmetric, non-static solution of the Einstein equation having two Killing vectors. It has two conserved parameters mass $(M)$ and angular momentum $(J)$ corresponding to those vectors. The location of outer/event ($r_+$) and inner/Cauchy ($r_-$) horizons are given by
\begin{equation}
r_{\pm}=M\pm \sqrt{M^{2}-a^{2}},
\label{equation 4}
\end{equation}
where
\begin{equation}
a=\frac{J}{M}~~(a\neq0)
\label{equation 4a}
\end{equation}
is the angular momentum per unit mass. The extremal condition (where the inner and the outer horizon coincide) for the Kerr black hole is given by $M^2=\frac{J^2}{M^2}$. The angular velocity and area of the event horizon are defined as \cite{Carrol}
\begin{eqnarray}
\Omega_{H}= \frac{a}{r^2_+ + a^2}.
\label{equation 5}
\end{eqnarray}
and
\begin{eqnarray}
A= 4\pi (r^2_+ + a^2)
\label{equation 6}
\end{eqnarray}
respectively. The semi-classical Hawking temperature and entropy of the rotating Kerr black hole are given by \cite{Carrol}
\begin{eqnarray}
T=\frac{\hbar\kappa}{2\pi}=\frac{\hbar}{2\pi}\frac{\sqrt{M^2-a^2}}{(r^2_+ +a^2)}.
\label{equation 7}
\end{eqnarray}
and
\begin{equation}
S=\frac{A}{4\hbar}=\frac{\pi(r_+^2+a^2)}{\hbar}.
\label{equation 8}
\end{equation}
respectively. Note that for the extremal Kerr black hole the semi-classical Hawking temperature vanishes. To calculate the semi-classical specific heat it is desirable to first make the appropriate identifications. Comparison between the work terms of the first law of thermodynamics and first law of black hole thermodynamics,
\begin{eqnarray}
dE= TdS - PdV,\label{equation 2}\\
dM=TdS+\Omega_H dJ,\label{equation 3a}
\end{eqnarray}
yields the following map
\begin{eqnarray}
P \rightarrow -\Omega_H,~~V\rightarrow J.
\label{equation 9}
\end{eqnarray}
This is a well known analogy where one can treat $-\Omega_H dJ$ as the external work done on the black hole \cite{Carrol}.
The semi-classical specific heat for the Kerr black hole can now be written, following its standard thermodynamic definition, as,
\begin{eqnarray}
C_{\Omega_H}=T\left(\frac{dS}{dT}\right)_{\Omega_H}=\left(\frac{dM}{dT}\right)_{\Omega_H}-\Omega_{H}\left(\frac{dJ}{dT}\right)_{\Omega_H}
\label{equation 10}
\end{eqnarray}
where, in the second equality, (\ref{equation 3a}) has been used. For convenience, instead of working with the variables ($M,\Omega_H,J$), we take ($M,\Omega_H,a$) as the basic variables and express $\left(\frac{dJ}{dT}\right)_{\Omega_H}$ as
\begin{eqnarray}
\left(\frac{dJ}{dT}\right)_{\Omega_H}=a\left(\frac{dM}{dT}\right)_{\Omega_H}+M\left(\frac{da}{dT}\right)_{\Omega_H}
\label{equation 11}
\end{eqnarray}
where (\ref{equation 4a}) has been used. From (\ref{equation 10}) and (\ref{equation 11}) we find
\begin{eqnarray}
C_{\Omega_H}=(1-a\Omega_H)\left(\frac{dM}{dT}\right)_{\Omega_H}-M\Omega_{H}\left(\frac{da}{dT}\right)_{\Omega_H}.
\label{equation 12}
\end{eqnarray}
To calculate $C_{\Omega_H}$, we first recast the semi-classical Hawking temperature (\ref{equation 7}), by eliminating $r_+$ using (\ref{equation 5}),
\begin{eqnarray}
T=\frac{\hbar\Omega_H\sqrt{M^2-a^2}}{2\pi a}=T(M, \Omega_H, a)
\label{equation 13}
\end{eqnarray}
Likewise from (\ref{equation 4}) and (\ref{equation 5}), we find,
\begin{eqnarray}
a=\frac{4M^2\Omega_H}{4M^2\Omega_H^2+1}\label{equation 14}
\end{eqnarray}
Exploiting this, one of the variables ($M,a$) can be replaced in favour of the others and thus (\ref{equation 13}) can be written in either of the following ways {\footnote{From now on by $a,T,M$ and $\Omega_H$ we mean $\frac{a}{\sqrt\hbar},\frac{T}{\sqrt\hbar},\frac{M}{\sqrt\hbar}$ and $\sqrt\hbar\Omega_H$ respectively. With this convention, $\hbar$ no longer appears explicitly in any of the equations.}},
\begin{eqnarray}
T=T(M,\Omega_H)=\frac{1}{8\pi M}(1-4M^2\Omega_H^2),\label{equation 16}\\
T=T(\Omega_H,a)=\frac{(1-2a\Omega_H)}{4\pi}\sqrt\frac{\Omega_H}{a(1-a\Omega_H)}.\label{equation 17}
\end{eqnarray}
Now using these two equations together with (\ref{equation 14}) we find the semi-classical specific heat (\ref{equation 12}) as
\begin{eqnarray}
C_{\Omega_H}=C_{\Omega_H}(M,\Omega_H)= \frac{-8\pi M^2(1-4M^2\Omega_H^2)}{(1+4M^2\Omega_H^2)^3}.
\label{equation 18}
\end{eqnarray}
In the above equation we took $M$ and $\Omega_H$ as the two independent variables for the purpose of graphical analysis. We shall stick to this convention for other cases also. Using (\ref{equation 4}) and (\ref{equation 14}), the semi-classical entropy (\ref{equation 8}) can be expressed in terms of $M, \Omega_H$ as
\begin{eqnarray}
S=2\pi M\left(M+\sqrt{\frac{M^2(1-4M^2\Omega_H^2)^2}{(1+4M^2\Omega_H^2)^2}}\right).
\label{equation 18a}
\end{eqnarray}
For the non-extremal region ($T>0$), it follows from (\ref{equation 16}) that $(1-4M^2\Omega_H^2)>0$ and therefore one has the following expression for the semi-classical entropy for Kerr black hole,
\begin{eqnarray}
S=S(M,\Omega_H)= \frac{4\pi M^2}{(1+4M^2\Omega_H^2)}.
\label{equation 19}
\end{eqnarray}
Following the thermodynamical definition Gibbs free energy, for the Kerr black hole is defined as,
\begin{eqnarray}
G=M-\Omega_HJ-TS.
\label{equation 19a}
\end{eqnarray}
For Kerr black hole it is found to be,
\begin{eqnarray}
G=\frac{M}{1+4M^2\Omega_H^2}-\frac{1}{4}(1-4M^2\Omega_H^2)\left(M+\sqrt{\frac{(M-4M^3\Omega_H^2)^2}{(1+4M^2\Omega_H^2)^2}}\right).
\label{equation 19aa}
\end{eqnarray}
\begin{figure}[ht]
\centering
\includegraphics[angle=0,width=15cm,height=12cm]{figure1.ps}
\caption[]{\it{Semi-classical Hawking temperature ($T$), entropy ($S$), specific heat ($C_{\Omega_H}$) and Gibbs free energy ($G$) vs. mass ($M$) plot: In all curves $\Omega_H=10^3$.}}
\label{figure 1}
\end{figure}
In figure 1 we plot the semi-classical Hawking temperature (\ref{equation 16}), entropy (\ref{equation 18a}), specific heat (\ref{equation 18}) and Gibb's free energy (\ref{equation 19a}) with respect to $M$, keeping the other parameter $\Omega_H$ fixed at a particular value $\Omega_H=10^3$. The temperature diverges as the mass of the black hole tends to zero. The point at which $T=0$ is the extremal case ($M^2=\frac{J^2}{M^2}$) where the Hawking temperature vanishes. From (\ref{equation 16}) this yields $M^2=\frac{1}{4\Omega_H^2}=25\times10^{-8}$. Hence $M=0.0005$ which is also reflected in the $T-M$ graph (figure 1(a)). Beyond this point, the temperature acquires a negative value ($M^2\leq \frac{J^2}{M^2}$) and the curve does not have any physical significance. The entire non-extremal region where $T$ is positive, corresponding to $M^2\geq \frac{J^2}{M^2}$, is relevant. In this non-extremal region specific heat is negative (figure 1(c)) for all values of the mass and thus renders the system unstable. The semi-classical entropy (figure 1(b)) and Gibbs free energy (figure 1(d)) are continuous for the entire non-extremal region. The absence of discontinuity in $S$ and $C_{\Omega_H}$ suggests that there is no phase transition. Thus at the semi-classical level, the system remains unstable all throughout. Later we shall repeat this analysis by considering correction to the semi-classical Hawking temperature and entropy of the Kerr black hole which will reveal some dramatic changes.
\section{Kerr black hole as a thermodynamical system; beyond semi-classical approximation}
Now we want to investigate the situation when a correction to the semi-classical approximation is taken into account. In a recent work involving two of us \cite{Modak} the black hole temperature and entropy for Kerr black hole were found to be
\begin{eqnarray}
\tilde T=T\left(1+\frac{\sqrt\hbar}{180\pi Mr_+}+{\cal O}(\hbar^2)\right)^{-1},
\label{equation 20}
\end{eqnarray}
and
\begin{eqnarray}
\tilde S=\frac{A}{4\hbar}+\frac{1}{90}\log{\frac{A}{4\hbar}}+{\cal O}(\hbar),
\label{equation 21}
\end{eqnarray}
where the usual semi-classical expressions for $A$ and $T$ were given in (\ref{equation 6}) and (\ref{equation 7}) respectively. We remark that the WKB type methods adopted in \cite{Modak} were also successfully employed \cite{Modak1} in other black hole geometries reproducing well known results found by other approaches \cite{OTHER}.
With these structures, the first law of black hole thermodynamics for the Kerr black hole is now given by \cite{Modak}
\begin{eqnarray}
dM= \tilde T d\tilde S+\Omega_HdJ.
\label{equation 24}
\end{eqnarray}
In the literature there exist several approaches which find the logarithmic correction to the semiclassical Bekenstein-Hawking entropy for various black holes\cite{OTHER}. However, the coefficient of the correction term is not identical in these approaches. Also as far as we are aware, there exists no other paper apart from \cite{Modak} which fixes the coefficient of the logarithmic correction to the semiclassical entropy for the Kerr black hole.
In order to calculate the specific heat we follow the method discussed in the previous section and write $\tilde T$ as a function of ($M,\Omega_H$) and ($a,\Omega_H$) in the following way:
\begin{eqnarray}
\tilde T=\tilde T(M,\Omega_H)=T(M,\Omega_H)\left(1+\frac{1+4M^2\Omega_H^2}{360\pi M^2}\right)^{-1},
\label{equation 22a}\\
\tilde T=\tilde T(a,\Omega_H)=T(a,\Omega_H)\left(1+\frac{\Omega_H}{90\pi a}\right)^{-1},
\label{equation 22b}
\end{eqnarray}
where $T$ is just the semi-classical expression defined in (\ref{equation 16}) and (\ref{equation 17}). Using (\ref{equation 14}), the corrected entropy (\ref{equation 21}) can be written as a function of $M,\Omega_H$ in the following way
\begin{eqnarray}
\tilde S=S+\frac{1}{90}\log{S}
\label{equation 23}
\end{eqnarray}
where the semiclassical entropy $S$ is defined in (\ref{equation 18a}). Following the semi-classical analysis, the specific heat is defined as
\begin{eqnarray}
\tilde C_{\Omega_H}=\tilde T\left(\frac{d\tilde S}{d\tilde T}\right)_{\Omega_H}=(1-a\Omega_H)\left(\frac{dM}{d\tilde T}\right)_{\Omega_H}-M\Omega_{H}\left(\frac{da}{d\tilde T}\right)_{\Omega_H}.
\label{equation 25a}
\end{eqnarray}
It is now straightforward to derive the corrected specific heat for the Kerr black hole by using (\ref{equation 22a}), (\ref{equation 22b}) and (\ref{equation 25a}). This finally leads to
\begin{eqnarray}
\tilde C_{\Omega_H}=-\frac{(1+4M^2\Omega_H^2+360\pi M^2)^2(1-4M^2\Omega_H^2)}{45(1+4M^2\Omega_H^2)^2(16M^4\Omega_H^4+1440\pi M^4\Omega_H^2+16M^2\Omega_H^2+360\pi M^2-1)}.
\label{equation 25b}
\end{eqnarray}
Gibbs free energy, in the presence of quantum correction, is found to be
\begin{eqnarray}
&& \tilde G = \frac{M}{1+4M^2\Omega_H^2}+\nonumber\\
&& \frac{M(4M^2\Omega_H^2-1)\left(180\pi M(M+\sqrt{\frac{(M-4M^3\Omega_H^2)^2}{(1+4M^2\Omega_H^2)^2}})\right)+\log\left(2M\pi(M+\sqrt{\frac{(M-4M^3\Omega_H^2)^2}{(1+4M^2\Omega_H^2)^2}})\right)}{2+8M^2(\Omega_H^2+90\pi)}~~~
\label{equation 23a}
\end{eqnarray}
Making use of equations (\ref{equation 22a}), (\ref{equation 23}) and (\ref{equation 25b}), we plot $\tilde T$, $\tilde S$, $\tilde C_{\Omega_H}$ and $\tilde G$ with mass $M$ for the previously fixed value of $\Omega_H=10^3$ (figure 2). Let us discuss the physical significance of the different plots one by one.
\begin{figure}[h]
\centering
\includegraphics[angle=0,width=15cm,height=12cm]{figure2.ps}
\caption[]{\it{Corrected Hawking temperature ($\tilde T$), entropy ($\tilde S$), specific heat ($\tilde C_{\Omega_H}$) and Gibbs free energy ($\tilde G$) vs. mass ($M$) plot: In all curves $\Omega_H=10^3$.}}
\label{figure 2}
\end{figure}
Figure 2(a) shows that the corrected Hawking temperature is not diverging when the mass of the black hole goes to zero, rather it vanishes. This behaviour was absent in the semi-classical case. The extremality condition is still satisfied for $M=0.0005$ since at this point the corrected temperature vanishes. There is a critical mass ($M_{\textrm{crit}}=0.0002425$) for which the temperature is maximum given by $T_{\textrm{max}}=T_{\textrm{crit}}=0.006756$. From figure 2(c) one can see that specific heat $C_{\Omega_H}$ now suffers a discontinuity at $M_{\textrm{crit}}$. At this point $\tilde C_{\Omega_H}$ becomes positive from the initial negative value. This point can be elaborated in the following manner. During the black hole evaporation as the mass decreases temperature of the black hole initially increases. This is shown in the region phase 1 of figure 2(c). At $M=M_{\textrm{crit}}$ a phase transition is taking place where the specific heat switches to a positive value. Consequently the black hole cools down during further evaporation and finally at zero mass the temperature vanishes. This behaviour is found in phase 2 region of the graph 2(c). The corrected entropy ($\tilde S$) in figure 2(b) and corrected free energy ($\tilde G$) in figure 2(d), are continuous for the entire region.
Note that we cannot write the corrected temperature (\ref{equation 20}) as $\tilde T=T\left(1-\frac{\sqrt\hbar}{180\pi Mr_+}+{\cal O}(\hbar^2)\right)$ since at the phase transition point, $\frac{\sqrt \hbar}{180\pi Mr_+}\sim 10^5$. As this is considerably greater than one, a naive simplification of (\ref{equation 20}) by series expansion is not possible.
So far it seems reasonable from the plots that there is a phase transition which eventually leads to the thermodynamic stability of the Kerr black hole. In the next section we shall make a systematic study to identify the type of phase transition.
\section{Nature of the black hole phase transition}
In the previous section we found a signature of phase transition with the corrected versions of Hawking temperature and entropy of Kerr black hole. This phase transition is not a first order phase transition as there is no discontinuity in the corrected entropy for the entire physical region. Instead, the discontinuity in specific heat suggests that there may be a second order phase transition. If that is the case then one naturally expects that there must be some analog of Ehrenfest's equations in black hole thermodynamics which must be satisfied. Till now Ehrenfest's relations were not studied in the context of black hole thermodynamics. Hence we shall systematically derive (see appendix) the Ehrenfest's equations and then check the validity of these relations afterwards. These (first and second) Ehrenfest's equations are given by (\ref{eren1}) and (\ref{eren2}) respectively. It is noteworthy that, since in a second order phase transition both $S$ and $J$ remain constant, left hand sides of these two equations have the same value. In the next subsection we shall check the validity of these two equations in the context of black hole thermodynamics.
Let us now convince ourselves that there is indeed a discontinuity present in the behaviour of corrected specific heat $C_{\Omega_H}$ (\ref{corsp}), analog of volume expansion coefficient $\alpha$ (\ref{volexp}) and compressibility $k_{T}$ (\ref{compr}). As far as $C_{\Omega_H}$ is concerned, its values at two different phases are given by $C_{\Omega_{H_1}}=0$ and $C_{\Omega_{H_2}}=0.0222$, so that $C_{\Omega_{H_2}}-C_{\Omega_{H_1}}=0.0222$ (figure 2(c)). For others, the expressions for $\alpha$ and $k_T$ in terms of $M$ and $\Omega_H$ are derived and appropriate graphs are plotted.
Comparing (\ref{volexp}) and (\ref{equation 11}) we find
\begin{equation}
J\alpha=a\left(\frac{\partial M}{\partial T}\right)_{\Omega_H}+M\left(\frac{\partial a}{\partial T}\right)_{\Omega_H}
\label{equation 69}
\end{equation}
Using the expressions of the corrected Hawking temperature from (\ref{equation 22a}) and (\ref{equation 22b}) and substituting $a$ from (\ref{equation 14}) this equation is simplified to
\begin{equation}
J\alpha=-\frac{4M^2\Omega_H(1+4M^2(90\pi+\Omega_H^2))^2(4M^2\Omega_H^2+3)}{45(1+4M^2\Omega_H^2)^2[16M^4(90\pi\Omega_H^2+\Omega_H^4)+8M^2(2\Omega_H^2+45\pi)-1]}.
\label{equation 57}
\end{equation}
The plot of $J\alpha$ with the mass ($M$) for our chosen constant angular velocity $\Omega_H=10^3$ of Kerr black hole is shown in figure 3a.
\begin{figure}[ht]
\centering
\includegraphics[angle=0, width=15cm, height=7cm]{figure3.ps}
\caption[]{\it{$J\alpha$, $Jk_T$ vs. mass ($M$) plot for $\Omega_H=10^3$.}}
\label{figure 3}
\end{figure}
There is a discontinuity present in this plot exactly at the same mass value where the specific heat also suffered a discontinuity. The values of $J\alpha$ in different phases are given by $J{\alpha_1}=-0.000022$ and $J{\alpha_2}=0$.
To express $k_{T}$ in terms of $M$ and $\Omega_H$ we proceed as follows. Using (\ref{equation 4a}) we re-express (\ref{compr}) in the following way,
\begin{eqnarray}
Jk_T=\left(\frac{\partial J}{\partial\Omega_H}\right)_{\tilde T}=\left(\frac{\partial (Ma)}{\partial\Omega_H}\right)_{\tilde T}\nonumber\\
=M\left(\frac{\partial a}{\partial \Omega_H}\right)_{\tilde T}+a\left(\frac{\partial M}{\partial \Omega_H}\right)_{\tilde T}.
\label{equation 63}
\end{eqnarray}
Using the property of partial differentiation,
\begin{eqnarray}
\left(\frac{\partial a}{\partial\Omega_H}\right)_{\tilde T}\left(\frac{\partial \Omega_H}{\partial\tilde T}\right)_{a}\left(\frac{\partial\tilde T}{\partial a}\right)_{\Omega_H}=-1;~~\left(\frac{\partial M}{\partial\Omega_H}\right)_{\tilde T}\left(\frac{\partial \Omega_H}{\partial\tilde T}\right)_{M}\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}=-1
\label{equation 63a}
\end{eqnarray}
we obtain the following two equalities
\begin{eqnarray}
\left(\frac{\partial a}{\partial\Omega_H}\right)_{\tilde T}=-\frac{\left(\frac{\partial \tilde T}{\partial\Omega_H}\right)_{a}}{\left(\frac{\partial\tilde T}{\partial a}\right)_{\Omega_H}};~~\left(\frac{\partial M}{\partial\Omega_H}\right)_{\tilde T}=-\frac{\left(\frac{\partial\tilde T}{\partial\Omega_H}\right)_{M}}{\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}}.
\label{equationn 64}
\end{eqnarray}
These two equations can easily be simplified by using (\ref{equation 22a}) and (\ref{equation 22b}). Substituting their values in (\ref{equation 63}) and using (\ref{equation 14}) to eliminate $a$, we finally obtain,
\begin{equation}
Jk_{T}=-\frac{4[M^3+4M^5(7\Omega_H^2-90\pi)+16M^7(540\pi\Omega_H^2+7\Omega_H^4)+64M^9(270\pi\Omega_H^4+\Omega_H^6)]}{(1+4M^2\Omega_H^2)^2[16M^4(\Omega_H^4+90\pi\Omega_H^2)+8M^2(2\Omega_H^2+45\pi)-1]}.
\label{equation 66}
\end{equation}
This is plotted in figure 3(b) for constant $\Omega_H=10^3$. The curve shows a discontinuity at the same critical mass value ($M_{{\textrm{crit}}}=0.0002425$). This plot gives $Jk_{T_{1}}=-4.1\times10^{-10}$ and $Jk_{T_2}=0$.
The above analysis strongly suggests the occurrence of phase transition. All the relevant physical parameters ($C_{\Omega_H}$, $J\alpha$ and $Jk_{T}$) appearing in the Ehrenfest's equations show discontinuity at the same critical value of mass. To conclusively argue in favour of a phase transition, we explicitly check the validity of the Ehrenfest's equations. In the next subsection we shall address this issue.
\subsection{Validity of Ehrenfest's equations}
Let us consider the first Ehrenfest's equation (\ref{eren1}). Using the chain rule of partial differentiation we have {\footnote {Under the change of variables $u=u(x,y)$ and $v=v(x,y)$ for a function $f=f(x,y)$, $\left(\frac{\partial f}{\partial x}\right)_y=\left(\frac{\partial f}{\partial u}\right)_v\left(\frac{\partial u}{\partial x}\right)_y+\left(\frac{\partial f}{\partial v}\right)_u\left(\frac{\partial v}{\partial x}\right)_y$. In the special case when $u=x$, $\left(\frac{\partial f}{\partial x}\right)_y=\left(\frac{\partial f}{\partial u}\right)_v+\left(\frac{\partial f}{\partial v}\right)_u\left(\frac{\partial v}{\partial x}\right)_y$.}}
\begin{equation}
\left(\frac{\partial\tilde T}{\partial\Omega_H}\right)_{\tilde S}=\left(\frac{\partial {\tilde T}}{\partial \Omega_H}\right)_{a}+\left(\frac{\partial {\tilde T}}{\partial a}\right)_{\Omega_H}\left(\frac{\partial a}{\partial\Omega_{H}}\right)_{\tilde S}
\label{equation 53}
\end{equation}
The derivatives of the corrected temperature (r.h.s) are computed from (\ref{equation 22b}). To calculate $\left(\frac{\partial a}{\partial\Omega_H}\right)_{\tilde S}$, we first derive an expression for the corrected entropy. On using (\ref{equation 5}) and (\ref{equation 23}) this is written as,
\begin{equation}
\tilde S=\tilde S(a,\Omega_H)=\frac{\pi a}{\Omega_H}+\frac{1}{90}\log{\frac{\pi a}{\Omega_H}}.
\label{equation 54}
\end{equation}
Exploiting the rule of partial differentiation, one obtains
\begin{eqnarray}
\left(\frac{\partial a}{\partial \Omega_H}\right)_{\tilde S}=-\frac{\left(\frac{\partial{\tilde S}}{\partial\Omega_H}\right)_{a}}{\left(\frac{\partial {\tilde S}}{\partial a}\right)_{\Omega_H}}.
\label{equation 54a}
\end{eqnarray}
From (\ref{equation 54}) and (\ref{equation 54a}) we find $\left(\frac{\partial a}{\partial\Omega_{H}}\right)_{\tilde S}=\frac{a}{\Omega_H}$. Substituting this in (\ref{equation 53}) and using (\ref{equation 22b}) yields,
\begin{equation}
\left(\frac{\partial\Omega_H}{\partial\tilde T}\right)_{\tilde S}=\frac{2(90\pi a+\Omega_H)(1-a\Omega_H)^2}{45a\Omega_H(2a\Omega_H-3)}\sqrt{\frac{\Omega_H}{a(1-a\Omega_H)}}.
\label{equation 55}
\end{equation}
This is the left hand side of the first Ehrenfest's equation. Eliminating $a$ in terms of $M,\Omega_H$ by using (\ref{equation 14}), we find the cherished expression,
\begin{equation}
-\left(\frac{\partial\Omega_H}{\partial\tilde T}\right)_{\tilde S}=\frac{(1+4M^2\Omega_H^2)(1+4M^2(\Omega_H^2+90\pi))}{180M^3\Omega_H(16M^4\Omega_H^4+16M^2\Omega_H^2+3)}.
\label{equation 55a}
\end{equation}
For $\Omega_H=10^3$ and $M=M_{\textrm{crit}}=0.0002425$, we get the value of the left hand side of the first Ehrenfest's equation (\ref{eren1}) that follows directly from (\ref{equation 55a}),
\begin{equation}
-\left(\frac{\partial\Omega_H}{\partial\tilde T}\right)_{\tilde S}=1.487\times10^5
\label{equation 58a}
\end{equation}
The right hand side of (\ref{eren1}) is now calculated by using results dictated earlier from different graphs ({\it i.e.} figure 2(a), 2(c), 3(a)). From figure 2(a) we find, at $M=M_{\textrm{crit}}$ the critical temperature $\tilde T=T_{\textrm{crit}}=0.006756$. Using these values in the right hand side of (\ref{eren1}), we get
\begin{eqnarray}
\frac{{\tilde C}_{\Omega_{H_2}}-{\tilde C}_{\Omega{_{H_1}}}}{{\tilde T}_{\textrm{crit}}(J\alpha_2-J\alpha_1)}=\frac{(0.222-0)}{0.006756\times0.000022}=1.494\times10^5
\label{equation 58}
\end{eqnarray}
The remarkable agreement between (\ref{equation 58a}) and (\ref{equation 58}) clearly shows the validity of the first Ehrenfest's equation for the Kerr black hole.
Let us now take the second Ehrenfest's equation (\ref{eren2}). The left hand side of (\ref{eren2}) can be calculated in the following manner. First we simplify (\ref{equation 14}) to find $(1-a\Omega_H)=\frac{J}{4M^3\Omega_H}$. Substituting this in (\ref{equation 22b}) and replacing $a$ from (\ref{equation 4a}) we write the angular momentum in the following way
\begin{equation}
J=J(M,\Omega_H,\tilde T)=\frac{M\Omega_H(45M-\tilde T)}{90(\pi \tilde T+M\Omega_H^2)}.
\label{equation 58b}
\end{equation}
For constant value of $J$ ({\it i.e.} $dJ=0$) we have
\begin{equation}
\left(\frac{\partial J}{\partial M}\right)_{\Omega_H,\tilde T}dM+\left(\frac{\partial J}{\partial\Omega_H}\right)_{M,\tilde T}d\Omega_H+\left(\frac{\partial J}{\partial \tilde T}\right)_{M,\Omega_H}d\tilde T=0.
\label{equation 59}
\end{equation}
Since $\tilde T$, $M$ and $\Omega_H$ are connected by (\ref{equation 22a}), we take only the variables $M$ and $\Omega_H$ as independent. Then $d\tilde T$ can be written as
\begin{equation}
d\tilde T=\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}dM+\left(\frac{\partial\tilde T}{\partial\Omega_H}\right)_{M}d\Omega_H.
\label{equation 60}
\end{equation}
Now multiplying (\ref{equation 59}) by $\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}$ and (\ref{equation 60}) by $\left(\frac{\partial J}{\partial M}\right)_{\Omega_H,\tilde T}$ and subtracting one from the other we arrive at the result
\begin{equation}
\left(\frac{\partial\Omega_H}{\partial\tilde T}\right)_J=\frac{\left(\frac{\partial J}{\partial\tilde T}\right)_{M,\Omega_H}\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}+\left(\frac{\partial J}{\partial M}\right)_{\tilde T,\Omega_H}}{\left(\frac{\partial J}{\partial M}\right)_{\Omega_H,\tilde T}\left(\frac{\partial\tilde T}{\partial\Omega_H}\right)_{M}-\left(\frac{\partial J}{\partial\Omega_H}\right)_{M,\tilde T}\left(\frac{\partial\tilde T}{\partial M}\right)_{\Omega_H}}
\label{equation 61}
\end{equation}
Derivatives of $J$ and $\tilde T$, obtained from (\ref{equation 58b}) and (\ref{equation 22a}) respectively, are substituted in (\ref{equation 61}) to finally yield the left hand side of the second Ehrenfest's equation (\ref{eren2}) as
\begin{equation}
-\left(\frac{\partial\Omega_H}{\partial\tilde T}\right)_J=\frac{\Omega_H(4M^2\Omega_H^2+3)(1+4M^2\Omega_H^2+360M^2\Omega_H^2)^2}{45M[1-4M^2(90\pi-7\Omega_H^2)+16M^4(540\pi\Omega_H^2+7\Omega_H^4)+64M^6(270\pi\Omega_H^4+\Omega_H^6)]} \label{equation 62}
\end{equation}
For $\Omega_H=10^3$ and $M=M_{\textrm{crit}}=0.0002425$ we find directly from (\ref{equation 62}),
\begin{equation}
-\left(\frac{\partial\Omega_H}{\partial T}\right)_J=1.4875\times10^5.
\label{equation 68}
\end{equation}
To calculate the right hand side of (\ref{eren2}) we employ the graphical results, as stated earlier. This yields,
\begin{equation}
\frac{(J\alpha_2-J\alpha_1)}{(Jk_{T_2}-Jk_{T_1})}=\frac{(2.2\times10^{-5})}{(4.1\times 10^{-10})}=5.37\times10^4.
\label{equation 67}
\end{equation}
The disagreement between (\ref{equation 68}) and (\ref{equation 67}) spoils the possibility of a second order phase transition in black hole thermodynamics. In the following section we shall argue that this phase transition is very special and usually found in glass forming liquids when they are supercooled to form glasses.
\subsection{Glassy phase transition and Prigogine-Defay ratio for Kerr black hole}
When some liquids are supercooled they solidify without crystallisation and form glasses, most common of which is the silicate glass. Though the underlying properties of materials which favour a glassy phase transition have not been completely understood \cite{rawson}, it is known that the rate of cooling is an important factor. Metals and alloys which were previously known as non glass forming substances are now found to form glasses under very fast cooling ($10^5$ K/s)\cite{giessen}. This is almost $10^7$ times faster than the cooling rate to form normal glass. In spite of the fact that in the glassy phase transition the specific heat, expansion coefficient and isothermal compressibility suffer discontinuity at the critical temperature, one cannot identify this as second order phase transition for the following reason. The critical temperature and the width of the transformation region for the glassy transition depend on the cooling rate\cite{jackle}. As a consequence the second Ehrenfest's equation (\ref{eren2}) is violated for this type of phase change.
For a normal second order phase transition, first order derivatives of Gibb's free energy are continuous at the phase transition point and hence one can equate the left hand sides of two Ehrenfest's equations to get the result $\Pi=1$, where $\Pi$, the Prigogine-Defay ratio \cite{jackle}, is defined by
\begin{eqnarray}
\Pi=\frac{\Delta C_p\Delta k}{TV(\Delta \alpha)^2}
\label{equation 3b}
\end{eqnarray}
In the case of black hole phase transition, from (\ref{equation 58a}) and (\ref{equation 68}) we see that left hand sides of (\ref{eren1}) and (\ref{eren2}) are also identical. Since in the previous subsection we discussed about the violation of second Ehrenfest's equation, we rectify it by putting a correction term ($\Pi$) in the right hand side of (\ref{eren2}) to get
\begin{eqnarray}
-\left(\frac{\partial\Omega_{H}}{\partial T}\right)_{J}=\Pi\frac{\alpha_{2}-\alpha_{1}}{k_{T_2}-k_{T_1}}
\label{equation 69z}
\end{eqnarray}
Equating (\ref{equation 69z}) and (\ref{equation 58a}) we get the Prigogine-Defay ratio (which is the black hole analogue of (\ref{equation 3b}))
\begin{equation}
\Pi=\frac{\Delta C_{\Omega_H}\Delta k_{T}}{TJ(\Delta\alpha)^2}.
\label{equation 70}
\end{equation}
For the Kerr black hole an explicit calculation gives $\Pi=\frac{0.0222\times 4.1\times10^{-10}}{0.006756\times{0.000022}^2}=2.78$. This is strikingly well placed in the known limit $2\leq\Pi\leq 5$ for glassy phase transitions in liquids and polymers \cite{gupta}.
Our studies further reveal that for values of $\Omega_H$ ($\geq 10^2$), the critical point ($M_{\textrm{crit}}$ and $\tilde T_{\textrm{crit}}$) is shifted but the value of PD ratio remains same. For small values of $\Omega_H$ ($<10$), the change in $Jk_{T}$ between two phases is extremely small and it is very difficult to check the validity of Ehrenfest's equations. When $\Omega_H=10$ both of the Ehrenfest's equations are violated and only further studies can reveal the peculiarity of this point. We thus infer that the glassy phase transition occurs for $\Omega_H \gtrsim 10$.
One may wonder whether the results are dependent on the normalisation of the logarithmic term appearing in (\ref{equation 21}). While it is difficult to answer this question in complete generality, nevertheless, we provide some arguments that establish the robustness of our conclusions. The first point to note is that, contrary to other black holes (like Schwarzschild and Reissner-Nordstrom etc) where different approaches yield different normalisations \cite{OTHER}, explicit results for the Kerr metric are only given in \cite{Modak}. However that does not prevent us from taking other normalisations and testing the inferences. Specifically, if the normalisation of the logarithmic term in (\ref{equation 21}) is taken as $\frac{1}{2}$ (instead of $\frac{1}{90}$) we again get a PD ratio close to 3, signifying a similar type of glassy phase transition. Likewise, if we consider other normalisations like $\frac{3}{2}$ (which, contrary to the earlier $\frac{1}{90}$ or $\frac{1}{2}$, is greater than one) or 1, our basic result remains unaffected. For details, see appendix 2. On the other hand, a negative normalisation of the logarithmic correction term modifies the corrected Hawking temperature in such a way that it is found to be negative in the entire non-extremal region which is clearly unphysical. So our conclusion is strictly valid for a positive normalisation of the correction term.
\section{Conclusions}
In spite of considerable research in black hole physics, issues related to thermodynamical stability are often overlooked. This is because in most of the cases black holes are studied within the semi-classical regime where it is always thermodynamically unstable. This implies that the possibility of phase transition and stability can be studied by going beyond the semi-classical approximation. Also, it is necessary to have the work term ($PdV$-type) in the law of black hole thermodynamics to properly identify and classify a phase transition. Both these criteria are satisfied by considering the corrected expressions for the thermodynamic variables in the Kerr black hole. These expressions were taken from an earlier work done by two of us \cite{Modak}. This reference, however, did not consider at all the issues related to phase transition.
Here we showed that inclusion of correction beyond the semi-classical approximation makes the Kerr black hole stable via a phase transition. We discovered that this phase transition cannot be classified either as first or second order. This type of phase transition has a close analogy with liquid to glass transition and is known as a glassy phase transition. In particular, the Prigogine-Defay ratio found for the Kerr black hole fits within the bound obtained from experimental results for a glassy phase transition. Furthermore, we showed the robustness of our result by choosing three more arbitrary positive coefficients of the correction (logarithmic) term to the semi-classical entropy for the Kerr black hole. In all these cases a glassy phase transition occurred with almost the same PD ratio. This ratio, close to 3, was well within the bounds (2 to 5) quoted in the literature \cite{gupta} for glassy phase transitions in liquids and polymers. We feel that this work can open up a new field of research where the deep connection between black hole physics and thermodynamics (statistical mechanics) can be further analyzed.
\section*{Appendix 1}
\subsection*{Derivation of Ehrenfest's equations for Kerr black hole}
Using the first law of black hole thermodynamics (\ref{equation 3a}) and the definition of Gibb's function (\ref{equation 19a}), the differential form of Gibbs free energy is written as,
\begin{eqnarray}
dG=-Jd\Omega_H-SdT\label{dg}
\end{eqnarray}
From this equation, black hole entropy and the angular momentum can be written as the derivatives of the Gibbs free energy as,
\begin{eqnarray}
&&S=-\left(\frac{\partial G}{\partial T}\right)_{\Omega_H}\\
&&J= -\left(\frac{\partial G}{\partial \Omega_H}\right)_{T}
\end{eqnarray}
Since, by definition, the Gibbs free energy (\ref{equation 19a}) is a state function, $dG$ is an exact differential and hence we get the following Maxwell relation from (\ref{dg})
\begin{eqnarray}
\left(\frac{\partial J}{\partial T}\right)_{\Omega_H}=\left(\frac{\partial S}{\partial \Omega_H}\right)_{T}\label{maxwellrelation}
\end{eqnarray}
In a second order phase transition Gibbs free energy and its first order derivatives {\it i.e.} entropy and angular momentum, are all continuous. So at a phase transition point, characterized by some temperature $T$ and an angular velocity $\Omega_H$
\begin{eqnarray}
&&G_1=G_2\label{freeenergyequality}\\
&&S_1=S_2 \label{entropyequality}\\
&&J_1=J_2\label{angularmomentumequality}
\end{eqnarray}
where the subscripts 1 and 2 denote the values of the different physical quantities in the two phases.
Let us now consider the equality (\ref{entropyequality}). If the temperature and angular velocity are increased infinitesimally to $T+dT$ and $\Omega_H+d\Omega_H$ then
\begin{eqnarray}
S_1+dS_1=S_2+dS_2\label{2ndentropyequality}
\end{eqnarray}
From (\ref{entropyequality}) and (\ref{2ndentropyequality}) we find
\begin{eqnarray}
dS_1=dS_2\label{entropychange}
\end{eqnarray}
Taking $S$ as a function of $T$ and $\Omega_H$
\begin{eqnarray}
S=S(T,\Omega_H)
\end{eqnarray}
we write the infinitesimal change in entropy as
\begin{eqnarray}
dS=\left(\frac{\partial S}{\partial T}\right)_{\Omega_H}dT+\left(\frac{\partial S}{\partial \Omega_H}\right)_{T}d\Omega_H
\end{eqnarray}
Using (\ref{maxwellrelation}), the above equation takes the form,
\begin{eqnarray}
dS&=\frac{C_{\Omega_H}}{T}dT+J\alpha d\Omega_H
\end{eqnarray}
where
\begin{equation}
\alpha=\frac{1}{J}\left(\frac{\partial J}{\partial T}\right)_{\Omega_H}
\label{volexp}
\end{equation}
is the coefficient of change in angular momentum and
\begin{equation}
C_{\Omega_H}=T\left(\frac{\partial S}{\partial T}\right)_{\Omega_H}.
\label{corsp}
\end{equation}
Since $J$ is same in both phases,
\begin{eqnarray}
&&dS_1=\frac{C_{\Omega_{H_1}}}{T}dT+J\alpha_1 d\Omega_H \label{ds1}\\
&&dS_2=\frac{C_{\Omega_{H_2}}}{T}dT+J\alpha_2 d\Omega_H \label{ds2}
\end{eqnarray}
Now use of the condition (\ref{entropychange}) requires the equality of (\ref{ds1}) and (\ref{ds2}),
\begin{eqnarray}
\frac{C_{\Omega_{H_1}}}{T}dT+J\alpha_1 d\Omega_H=\frac{C_{\Omega_{H_2}}}{T}dT+J\alpha_2 d\Omega_H.
\end{eqnarray}
This gives the first Ehrenfest's equation
\begin{eqnarray}
-\left(\frac{\partial\Omega_H}{\partial T}\right)_{S}=\frac{C_{\Omega_{H_2}}-C_{\Omega{_{H_1}}}}{TJ(\alpha_2-\alpha_1)}
\label{eren1}
\end{eqnarray}
We now take the constancy of angular momentum (\ref{angularmomentumequality}) which, under infinitesimal change of temperature and angular velocity, gives
\begin{eqnarray}
dJ_1=dJ_2.\label{angularmomentumchange}
\end{eqnarray}
Taking angular momentum as a function of $T$ and $\Omega_H$,
\begin{eqnarray}
J=J(T,\Omega_H)
\end{eqnarray}
we get the differential relation
\begin{eqnarray}
dJ&=&\left(\frac{\partial J}{\partial T}\right)_{\Omega_H}dT+\left(\frac{\partial J}{\partial \Omega_H}\right)_Td\Omega_H\\
&=&J\alpha dT-Jk_Td\Omega_H
\end{eqnarray}
where
\begin{equation}
k_T=\frac{1}{J}\left(\frac{\partial J}{\partial {\Omega_H}}\right)_T
\label{compr}
\end{equation}
is the analog of compressibility. Since $dJ$ is same for the two phases, we get the second Ehrenfest's equation by mimicking the previous steps used to derive (\ref{eren1})
\begin{eqnarray}
-\left(\frac{\partial\Omega_{H}}{\partial T}\right)_{J}=\frac{\alpha_{2}-\alpha_{1}}{k_{T_2}-k_{T_1}}.
\label{eren2}
\end{eqnarray}
The two Ehrenfest's equations (\ref{eren1}) and (\ref{eren2}) which must be satisfied for a second order phase transition play the same role as the Clapeyron's equation for the first order phase transition.
\section*{Appendix 2}
\subsection*{Summary of results for distinct normalisations of the logarithmic term in (\ref{equation 21})}
Here we summarise our findings for different normalisations of the logarithmic term in (\ref{equation 21}). For instance, if the normalisation is taken as $\frac{1}{2}$, the corrected entropy and temperature are given by,
\begin{eqnarray}
\tilde S=\frac{A}{4\hbar}+\frac{1}{2}\log{\frac{A}{4\hbar}}+{\cal O}(\hbar),
\label{diffs1}
\end{eqnarray}
and
\begin{eqnarray}
\tilde T=T\left(1+\frac{\sqrt\hbar}{4\pi Mr_+}+{\cal O}(\hbar^2)\right)^{-1}.
\label{difft1}
\end{eqnarray}
After carrying out the analysis with these corrected expressions one would find the following results for two Ehrenfest's equations, which are the analogues of (\ref{equation 58a}) and (\ref{equation 58}), obtained with the earlier normalisation ($\frac{1}{90}$),
\begin{eqnarray}
-\left(\frac{\partial\Omega_H}{\partial T}\right)_{S}=6.6934\times10^{6}\nonumber\\
\frac{C_{\Omega_{H_2}}-C_{\Omega{_{H_1}}}}{TJ(\alpha_2-\alpha_1)}=7.0109\times 10^{6}.
\label{diff1eren1}
\end{eqnarray}
and
\begin{eqnarray}
-\left(\frac{\partial\Omega_{H}}{\partial T}\right)_{J}=6.68077\times10^{6}\nonumber\\
\frac{\alpha_{2}-\alpha_{1}}{k_{T_2}-k_{T_1}}=2.2500\times 10^{6}.
\label{diff1eren2}
\end{eqnarray}
The above pair of equations (\ref{diff1eren2}) are the analogues of (\ref{equation 68}) and (\ref{equation 67}). This clearly shows that while the first Ehrenfest's equation is closely satisfied the second one is violated in such a way that the PD ratio (\ref{equation 70}) following from (\ref{diff1eren2}) becomes $\Pi=\frac{6.68077\times10^{6}}{2.2500\times10^{6}}=2.96$.
Similarly, taking the normalisation as $\frac{3}{2}$ of the logarithmic term in (\ref{equation 21}), we get
\begin{eqnarray}
\tilde S=\frac{A}{4\hbar}+\frac{3}{2}\log{\frac{A}{4\hbar}}+{\cal O}(\hbar),
\label{diffs2}
\end{eqnarray}
\begin{eqnarray}
\tilde T=T\left(1+\frac{3\sqrt\hbar}{4\pi Mr_+}+{\cal O}(\hbar^2)\right)^{-1}.
\label{difft2}
\end{eqnarray}
Correspondingly, the results for the two Ehrenfest's equations are given by
\begin{eqnarray}
-\left(\frac{\partial\Omega_H}{\partial T}\right)_{S}=2.00802\times10^{7}\nonumber\\
\frac{C_{\Omega_{H_2}}-C_{\Omega{_{H_1}}}}{TJ(\alpha_2-\alpha_1)}=2.14084\times 10^{7}.
\label{diff2eren1}
\end{eqnarray}
and
\begin{eqnarray}
-\left(\frac{\partial\Omega_{H}}{\partial T}\right)_{J}=2.00423\times10^{7}\nonumber\\
\frac{\alpha_{2}-\alpha_{1}}{k_{T_2}-k_{T_1}}=6.82923\times 10^{6}.
\label{diff2eren2}
\end{eqnarray}
These results finally give a PD ratio $\Pi=2.93$ for the above choice of the normalisation.
Finally, if we set the normalisation as unity, the corrected temperature and entropy are given by,
\begin{eqnarray}
\tilde S=\frac{A}{4\hbar}+\log{\frac{A}{4\hbar}}+{\cal O}(\hbar),
\label{diffs3}
\end{eqnarray}
\begin{eqnarray}
\tilde T=T\left(1+\frac{\sqrt\hbar}{2\pi Mr_+}+{\cal O}(\hbar^2)\right)^{-1},
\label{difft3}
\end{eqnarray}
leading to the Ehrenfest's equations,
\begin{eqnarray}
-\left(\frac{\partial\Omega_H}{\partial T}\right)_{S}=1.33868\times10^{7}\nonumber\\
\frac{C_{\Omega_{H_2}}-C_{\Omega{_{H_1}}}}{TJ(\alpha_2-\alpha_1)}=1.36735\times 10^{7}.
\label{diff3eren1}
\end{eqnarray}
and
\begin{eqnarray}
-\left(\frac{\partial\Omega_{H}}{\partial T}\right)_{J}=1.33615\times10^{7}\nonumber\\
\frac{\alpha_{2}-\alpha_{1}}{k_{T_2}-k_{T_1}}=4.63415\times 10^{6}.
\label{diff3eren2}
\end{eqnarray}
Once again one finds a PD ratio $\Pi=2.88$ which closely matches with other cases.
\section*{Acknowledgement}
One of the authors (S.K.M) thanks the Council of Scientific and Industrial Research (C.S.I.R), Government of India, for financial support.
|
\section{Introduction}
In recent years, the analysis of the equal-time cross-correlation matrix for a variety of multivariate data sets such as financial data, \cite{Laloux_1999,Plerou_1999,Laloux_2000,Plerou_2000,Gopikrishnan_2000,Utsuki_2004,Bouchaud_book,Wilcox_2004,Sharifi_2004,Conlon_2007,Conlon_2008,Podobnik_2008}, electroencephalographic (EEG) recordings, \cite{Schindler_2007,Schindler_2007_b}, magnetoencephalographic (MEG) recordings, \cite{Kwapien_2000}, and others, has been studied extensively. Other authors have investigated the relationship between stock price changes and liquidity or trading volume, \cite{Ying_1966, Karpoff_1987,LeBaron_1999}. In particular, Random Matrix Theory, (RMT), has been applied to filter the relevant information from the statistical fluctuations, inherent in empirical cross-correlation matrices, for various financial time series, \cite{Laloux_1999,Plerou_1999,Laloux_2000,Plerou_2000,Gopikrishnan_2000,Utsuki_2004,Bouchaud_book,Wilcox_2004,Sharifi_2004,Conlon_2007}. By comparing the eigenvalue spectrum of the correlation matrix to the analytical results, obtained for random matrix ensembles, significant deviations from RMT eigenvalue predictions provide genuine information about the correlation structure of the system. This information has been used to reduce the difference between predicted and realised risk of different portfolios.
Several authors have suggested recently that some real correlation information may be hidden in the RMT defined random part of the eigenvalue spectrum. A technique, involving the use of power mapping to identify and estimate the noise in financial correlation matrices, \cite{Guhr_2003}, allows the suppression of those eigenvalues, associated with the noise, in order to reveal different correlation structures buried underneath. The relationship, between the eigenvalue density $c$ of the true correlation matrix, and that of the empirical correlation matrix $C$, was derived to show that correlations can be measured in the random part of the spectrum, \cite{Burda_2004_a, Burda_2004_b}. A Kolmogorov test was applied to demonstrate that the bulk of the spectrum is not in the Wishart RMT class, \cite{Malevergne}, while the existence of factors, such as an overall market effect, firm size and industry type, is due to collective influence of the assets. More evidence that the RMT fit is not perfect was provided, \cite{Kwapien_2006}, where it was shown that the dispersion of ``noise'' eigenvalues is inflated, indicating that the bulk of the eigenvalue spectrum contains correlations masked by measurement noise.
The behaviour of the largest eigenvalue of a cross-correlation matrix for small windows of time, has been studied, \cite{Drozdz_2000}, for the DAX and Dow Jones Industrial average Indices (DJIA). Evidence of a time-dependence between `drawdowns' (`draw-ups') and an increase (decrease) in the largest eigenvalue was obtained, resulting in an increase of the \textit{information entropy}\footnote{In information theory, the Shannon entropy or information entropy is a measure of the uncertainty associated with a random variable.}of the system. Similar techniques were used, \cite{Drozdz_2001}, to investigate the dynamics between the stocks of two different markets (DAX and DJIA). In this case, two distinct eigenvalues of the cross-correlation matrix emerged, corresponding to each of the markets. By adjusting for time-zone delays, the two eigenvalues were then shown to coincide, implying that one market leads the dynamics in the other.
Equal-time cross-correlation matrices have been used, \cite{Muller_2005}, to characterise dynamical changes in nonstationary multivariate time-series. It was specifically noted that, for increased synchronisation of $k$ series within an $M-$dimensional multivariate time series, a repulsion between eigenstates of the correlation matrix results, in which $k$ levels participate. Through the use of artifically-created time series with pre-defined correlation dynamics, it was demonstrated that there exist situations, where the relative change in eigenvalues from the lower edge of the spectrum is greater than that for the large eigenvalues, implying that information drawn from the smaller eigenvalues is highly relevant.
The technique was subsequently applied to the eigenvalue spectrum of the equal time cross-correlation matrix of multivariate Epileptic Seizure time series and information on the correlation dynamics was found to be visible in \emph{both} the lower and upper eigenstates. Further studies of the equal-time correlation matrix of EEG signals, \cite{Schindler_2007}, investigated temporal dynamics of focal onset epileptic seizures\footnote{A partial or focal onset seizure occurs when the discharge starts in one area of the brain and then spreads over other areas.} and showed that zero-lag correlations between multichannel EEG signals tend to decrease during the first half of a seizure and increase gradually before the seizure ends. A further extension (to the case of \textit{Status Epilepticus}, \cite{Schindler_2007_b}), used the equal-time correlation matrix to assess neuronal synchronisation prior to seizure termination.
Examples have demonstrated, \cite{Muller_2006_b}, that information about cross correlations can be found in the RMT bulk of eigenvalues and that the information extracted at the \emph{lower} edge is statistically \emph{more significant} than that extracted from the larger eigenvalues. The authors introduced a method of unfolding the eigenvalue level density, through the normalisation of each of the level distances by its ensemble average, and used this to calculate the corresponding individual nearest-neighbour distance. Those parts of the spectrum, dominated by noise, could be distinguished from those containing information about correlations. Application of this technique to multichannel EEG data showed the smallest eigenvalues to be more sensitive to detection of subtle changes in the brain dynamics than the largest.
In this paper, we examine eigenvalue dynamics of the cross-correlation matrix for multivariate financial data with a view to characterising market behaviour. Methods are reviewed in Section~\ref{methods}, the data described in Section~\ref{Data} and in Section~\ref{Eresults} we look at the results obtained both for the empirical correlation matrix and the model correlation matrices described.
\section{Methods}
\label{methods}
\subsection{Empirical Dynamics}
\label{Emp_Dyn}
The equal-time cross-correlation matrix, between time series of equity returns, is calculated using a sliding window where the number of assets, $N$, is smaller than the window size $T$. Given returns $G_{i} \left(t\right)$, $i = 1, \ldots,N$, of a collection of equities, we define a normalised return, within each window, in order to standardise the different equity volatilities. We normalise $G_{i}$ with respect to its standard deviation $\sigma_{i}$ as follows:
\begin{equation}
g_{i}\left(t\right) = \frac{G_{i} \left(t\right) - \widehat{G_{i} \left(t\right)}}{\sigma_{i}}
\end{equation}
Where $\sigma_{i}$ is the standard deviation of $G_{i}$ for assets $i = 1, \ldots,N$ and $\widehat{G_{i}}$ is the time average of $G_{i}$ over a time window of size $T$.
Then, the equal-time cross-correlation matrix is expressed in terms of $g_{i}\left(t\right)$
\begin{equation}
C_{ij} \equiv \left\langle g_{i}\left(t\right) g_{j}\left(t\right) \right\rangle
\label{cross_corr}
\end{equation}
The elements of $C_{ij}$ are limited to the domain $-1 \leq C_{ij} \leq 1$, where $C_{ij} = 1$ defines perfect positive correlation, $C_{ij} = -1$ corresponds to perfect negative correlation and $C_{ij} = 0$ corresponds to no correlation. In matrix notation, the correlation matrix can be expressed as
\begin{equation}
\mathbf{C} = \frac{1}{T} \mathbf{GG}^{\tau}
\end{equation}
Where $\bf{G}$ is an $N\times T$ matrix with elements $g_{it}$.
The eigenvalues $\mathbf{\lambda}_{i}$ and eigenvectors $\mathbf{\hat{v}}_{i}$ of the correlation matrix $\mathbf{C}$ are found from the following
\begin{equation}
\mathbf{C} \mathbf{\hat{v}}_{i} = \mathbf{\lambda}_{i} \mathbf{\hat{v}}_{i}
\end{equation}
The eigenvalues are then ordered according to size, such that $\mathbf{\lambda}_{1} \leq \mathbf{\lambda}_{2}\leq \ldots \leq \mathbf{\lambda}_{N}$. The sum of the diagonal elements of a matrix, (the Trace), must always remain constant under linear transformation. Thus, the sum of the eigenvalues must always equal the Trace of the original correlation matrix. Hence, if some eigenvalues increase then others must decrease, to compensate, and vice versa (\textit{Eigenvalue Repulsion}).
There are two limiting cases for the distribution of the eigenvalues \cite{Schindler_2007,Muller_2005}. When all of the time series are perfectly correlated, $C_{i} \approx 1$, the largest eigenvalue is maximised with a value equal to $N$, while for time series consisting of random numbers with average correlation $C_{i} \approx 0$, the corresponding eigenvalues are distributed around $1$, (where any deviation is due to spurious random correlations).
For cases between these two extremes, the eigenvalues at the lower end of the spectrum can be much smaller than $\lambda_{max}$. To study the dynamics of each of the eigenvalues using a sliding window, we normalise each eigenvalue in time using
\begin{equation}
{\tilde{\lambda}}_{i}(t) = \frac{\left(\mathbf{\lambda}_{i}(t) - {\widehat{\lambda_{i}(\tau)}}\right)}{\sigma^{\lambda_{i}\left(\tau\right)}}
\label{normalise}
\end{equation}
where $\mathbf{\widehat{\lambda}}$ and $\sigma^{\lambda}$ are the mean and standard deviation of the eigenvalue $i$ over a particular reference period, $\tau$. This normalisation allows us to visually compare eigenvalues at both ends of the spectrum, even if their magnitudes are significantly different. The reference period, used to calculate mean and standard deviation of the eigenvalue spectrum, can be chosen to be a low volatility sub-period, (which helps to enhance the visibility of high volatility periods), or the full time period studied.
\subsection{One-factor Model}
\label{OneFModel}
In the \emph{one-factor model} of stock returns, we assume a \emph{global correlation} with the cross-correlation between all stocks the same, $\rho_{0}$. The spectrum of the associated correlation matrix consists of only two values, a large eigenvalue of order $(N-1)\rho_{0} + 1$, associated with the market, and an $(N-1)-$fold degenerate eigenvalue of size $1-\rho_{0}<1$. Any deviation from these values is due to the finite length of time series used to calculate the correlations. In the limit $N\rightarrow\infty$ (even for \emph{small correlation}, i.e. $\rho\rightarrow0$) a large eigenvalue appears, which is associated with the eigenvector $v_{1} = \left(\frac{1}{\sqrt{N}}\right)\left(1,1,1\ldots1\right)$, and which dominates the correlation structure of the system.
\subsection{`Market plus sectors' model}
\label{MultiFModel}
To expand the above to a `market plus sectors' model, we perturb a number of pairs $N$ of the correlations $\rho_{0} + \rho_{n}$, where $-1-\rho_{0}\leq\rho_{n}\leq1-\rho_{0}$. Additionally, we impose a constraint $\displaystyle\sum_{N}\rho_n = 0$, ensuring that the average correlation of the system remains equal to $\rho_{0}$. These perturbations allow us to introduce \textit{groups of stocks} with similar correlations, (corresponding to Market Sectors).
Using the correlation matrix from the ``one-factor model'' and the ``market plus sectors model'', we can construct correlated time series using the Cholesky decomposition $A$ of a correlation matrix $C = {AA}^{\tau}$. We can then generate finite correlated time series of length $T$,
\begin{equation}
x_{it} = \sum_{j} A_{ij} y_{jt} \quad \quad t = 1,\dots,T
\label{chol}
\end{equation}
where $y_{jt}$ is a random Gaussian variable with mean zero and variance $1$ at time $t$. Using Eqn.~\ref{cross_corr} we can then construct a correlation matrix using the simulated time series. The finite size of the time series introduces `noise' into the system and hence empirical correlations will vary from sample to sample. This `noise' could be reduced through the use of longer simulated time series or through averaging over a large number of time series.
In order to compare the eigenvectors from each of the Model Correlation matrices to that constructed from the equity returns time series, we use the Inverse Participation Ratio (IPR) \cite{Plerou_2000,Noh}. The IPR allows quantification of the number of components that participate significantly in each eigenvector and tells us more about the level and nature of deviation from RMT. The IPR of the eigenvector $u^{k}$ is given by
$ I^{k} \equiv \sum^{N}_{l = 1} \left( u^{k}_{l}\right)^{4} $ and allows us to compute the inverse of the number of eigenvector components that contribute significantly to each eigenvector.
\section{Data}
\label{Data}
In order to study the dynamics of the empirical correlation matrix over time, we analyse two different data sets. The first data set comprises the $384$ equities of the Standard \& Poors (S\&P) 500 where full price data is available from January $1996$ to August $2007$ resulting in 2938 daily returns. The S\&P 500 is an index consisting of 500 large capitalisation equities, which are predominantly from the US. In order to demonstrate that our results are not market specific, however, we examine a second data set, made up of the 49 equities of the Dow Jones Euro Stoxx 50 where full price data is available from January $2001$ to August $2007$ resulting in $1619$ daily returns. The Dow Jones Euro Stoxx 50 is a stock index of Eurozone equities designed to provide a blue-chip representation of supersector leaders in the Eurozone.
\section{Results}
\label{Eresults}
We analyse the eigenvalue dynamics of the correlation matrix of a subset of 100 S\&P equities, chosen randomly, using a sliding window of 200 days. This subsector was chosen such that $Q = \frac{T}{N} = 2$, thus ensuring that the data would be close to non-stationary in each sliding window. Figure \ref{S&Peigs}(a) shows broadly similar sample dynamics from the $5^{th}$, $15^{th}$ and $25^{th}$ largest eigenvalues over each of these sliding windows. The sum of the 80 smallest eigenvalues are shown in Figure \ref{S&Peigs}(b), while the dynamics of the largest eigenvalue is displayed in Figure \ref{S&Peigs}(c). The repulsion between the largest eigenvalue and the small eigenvalues is evident here, (comparing \ref{S&Peigs}(b) and \ref{S&Peigs}(c)), with the dynamics of the small eigenvalues contrary to those of the largest eigenvalue. As noted earlier (Section~\ref{Emp_Dyn}), this is a consequence of the fact that the trace of the correlation matrix must remain constant under transformations and any change in the largest eigenvalue must be reflected by a change in one or more of the other eigenvalues. Similar results were obtained for different subsets of the S\&P and also for the members of the Dow Jones Euro Stoxx 50.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=100mm,width=130mm]{SandPEigenvalues.eps}
\caption{Time Evolution of (a) Three small eigenvalues (b) Sum of the 80 smallest eigenvalues (c) The largest eigenvalue}
\label{S&Peigs}
\end{center}
\end{figure}
\subsection{Normalised Eigenvalue Dynamics}
Using normalised eigenvalues as described above, (Eqn. \ref{normalise}), we performed a number of experiments to investigate the dynamics of a set of small eigenvalues versus the largest eigenvalue. The various experiments are described below:
\begin{enumerate}
\item
As in Section~\ref{Eresults}, the dynamics for the same subset of 100 equities are analysed using a sliding window of 200 days. The normalisation is carried out using the mean and standard deviation of each of the eigenvalues over the entire time-period. Figure \ref{S&PeigsNorm}(a) shows the value of the S\&P index from $1997$ to mid$-2007$.
The normalised largest eigenvalue is shown in Figure \ref{S&PeigsNorm}(b) together with the average of the 80 normalised small eigenvalues. The compensatory dynamics mentioned earlier are shown more clearly here, with the largest and average of the smallest 80 eigenvalues having opposite movements. The normalised eigenvalues for the entire eigenvalue spectrum are shown in Figure \ref{S&PeigsNorm}(c), where the colour indicates the number of standard deviations from the time average for each of the eigenvalues over time. As shown, there is very little to differentiate the dynamics of the $80-90$ or so smallest eigenvalues. In contrast, the behaviour of the largest eigenvalue is clearly opposite to that of the smaller eigenvalues. However, from the $90^{th}$ and subsequent eigenvalue there is a marked change in the behaviour, (Figure \ref{S&PeigsNorm}(d)), and the eigenvalue dynamics are distinctly different. This may correspond to the area outside the ``Random Bulk'' in RMT. Similar to \cite{Drozdz_2000,Drozdz_2001}, we also find evidence of an increase/decrease in the largest eigenvalue with respect to `drawdowns'/`draw-ups'. Additionally, we find the highlighted \textit{compensatory dynamics} of the small eigenvalues. These results were tested for various time windows and normalisation periods, with smaller windows and normalisation periods found to capture and emphasise additional features.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=120mm,width=130mm]{SandPEigenvaluesNormalised.eps}
\caption{(a) S\&P Index (b) Normalised Largest Eigenvalue vs. Average of 80 smallest normalised eigenvalues (c) All Normalised Eigenvalues (d) Largest 12 Normalised Eigenvalues}
\label{S&PeigsNorm}
\end{center}
\end{figure}
\item
To demonstrate the above result for a \emph{different level of granularity}, we chose $50$ equities randomly with a time window of $500$ days, giving $Q = \frac{T}{N} = 10$. The results obtained, (Figure \ref{S&Peigs50}), are in keeping with those for $Q = 2$ earlier, with a broad-band increase (decrease) of the $40$ smallest eigenvalues concurrent to a decrease (increase) of the largest eigenvalue.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=70mm,width=130mm]{SandPEigenvaluesLargeQ.eps}
\caption{(a) Normalised Largest Eigenvalue vs. Average of 40 smallest normalised eigenvalues (b) All Normalised Eigenvalues}
\label{S&Peigs50}
\end{center}
\end{figure}
\item
The previous examples used random subsets of the S\&P universe in order to keep $Q= \frac{T}{N}$ as large as possible. To demonstrate that the above results were not sampling artifacts, we also looked at the full sample of 384 equities, (that survived the entire $11$ year period), with a time window of $500$ days ($Q = 1.30$). The results, as shown in Figure \ref{S&Peigs384}, are similar to those above, with the majority of the small eigenvalues compensating for changes in the large eigenvalue. As indicated previously, however, there is a small band of large eigenvalues, for which behaviour is different to that of both the band of small eigenvalues and the largest eigenvalue.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=70mm,width=130mm]{SandPEigenvaluesAllStocks.eps}
\caption{(a) Normalised Largest Eigenvalue vs. Average of 325 smallest normalised eigenvalues (b) All Normalised Eigenvalues}
\label{S&Peigs384}
\end{center}
\end{figure}
\item
All examples discussed so far have focused on the universe of equities from the S\&P $500$ that have survived since $1997$. To ensure that the results obtained were not exclusive to the S\&P $500$, we also applied the same technique to the $49$ equities of the EuroStoxx 50 index that survived from January $2001$ to August $2007$. The sliding window used was $200$ days, such that $Q = 4.082$. The results, (Figure \ref{SX5E}), were again similar to before, with a wide band of small eigenvalues ``responding to'' movements in the largest eigenvalues. In this case, the band of deviating large eigenvalues (ie. those which correspond to the area outside the ``Random Bulk'' in RMT), (Figure \ref{SX5E}(d)), is not as marked as in the previous example. This effectively implies that equities in this index are dominated by ``the Market''.
\end{enumerate}
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=120mm,width=130mm]{SX5EEigenvalues.eps}
\caption{(a) EuroStoxx 50 Index, Jan 2001 - Aug 2007 (b) Normalised Largest Eigenvalue vs. Average of 40 smallest normalised eigenvalues for EuroStoxx 50 (c) All Normalised Eigenvalues (d) The 9 Largest Normalised Eigenvalues}
\label{SX5E}
\end{center}
\end{figure}
\subsection{Model Correlation Matrix}
\label{corr_model}
The results, described, demonstrate that the time dependent dynamics of the small eigenvalues of the correlation matrix of stock returns move counter to those of the largest eigenvalue. Here, we show how a simple one-factor model, Section~\ref{OneFModel}, of the correlation structure reproduces much of this behaviour. Furthermore, we show how additional features can be captured by including perturbations in this model, essentially a \textit{``market plus sectors''} model, Section~\ref{MultiFModel}, \cite{Malevergne,Noh,Papp}.
In order to compare the empirical results, Section~\ref{Eresults}, to those of the single factor model, we first constructed a correlation matrix where each non-diagonal element was equal to the average correlation of the empirical matrix in each sliding window. We then calculated the eigenvalues of this matrix over each sliding window and normalised these as before, (Section \ref{methods}). The dynamics of the largest eigenvalue for the single-factor model are displayed in Figure \ref{modelEuroStoxx}(a) for the EuroStoxx 50 index using a sliding window of $200$ days. As can be seen, the main features of the dynamics are in agreement with those of Figure~\ref{SX5E} for the empirical data. The dynamics of the remaining eigenvalues, shown in figure \ref{modelEuroStoxx}(c), are found to be within a narrow range, with any change in time due to compensation for change in the largest eigenvalue.
For the `market plus sectors' model, we introducted perturbations with two groups of $5$ stocks having correlation $\rho_{0}-0.15$ and $\rho_{0}+0.15$, with the average correlation at each time window remaining the same.
The largest eigenvalue dynamics for the `market plus sectors' model is shown in \ref{modelEuroStoxx}(b), with the dynamics almost identical to those for the largest eigenvalue of the `one-factor' model, (any differences are due to random fluctuations induced in the Cholesky decomposition, Section~\ref{MultiFModel}). However, looking at the remaining eigenvalues, \ref{modelEuroStoxx}(d), we see that a number are found to be deviate significantly from the bulk. These deviations are due to the additional sectorial information included in the market plus sectors model. This agrees with previous results, \cite{Plerou_2000, Utsuki_2004}, where small eigenvalues were found to deviate from the random bulk, in addition to large deviating eigenvalues containing sectorial information.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=95mm,width=135mm]{CompareSingleSectorModel.eps}
\caption{(a) Largest Eigenavalue, one factor model (b) Largest Eigenavalue, Market plus sectors model (c) Eigenvalues 1-48, one factor model (d) Eigenvalues 1-48, Market plus sectors model}
\label{modelEuroStoxx}
\end{center}
\end{figure}
To examine properties of the eigenvector components themselves, we use the Inverse Participation Ratio (IPR). Figure~\ref{IPR}(a), displays the IPR found using the empirical data from the Eurostoxx $50$. This has similar characteristics to those found for different indices, \cite{Plerou_2000}, with the IPR for the largest eigenvalue much smaller than the mean, a large IPR corresponding to sectorial information in the $2^{nd}$ or $3^{rd}$ largest eigenvalues and the IPR for the small eigenvalues raised. For the single factor model, we created a \emph{synthetic} correlation matrix using Eqn. (\ref{chol}), with average correlation $(0.204)$ equal to that of the Euro Stoxx 50 over the time period studied. As shown in Figure~\ref{IPR}(b), the IPR retains some of the features found for empirical data, \cite{Plerou_2000,Noh}, with that corresponding to the largest eigenvector having a much smaller value than the mean. This corresponds to an eigenvector to which many stocks contribute, (effectively the market eigenvector), \cite{Plerou_2000,Noh}.
In an attempt to include additional empirical features, such as the band of deviating large eigenvalues between the bulk and the largest eigenvalue, we also considered a perturbation, with two groups of stocks having correlation $\rho_{0}-0.15$ and $\rho_{0}+0.15$, and the same average correlation. In this case, Figure~\ref{IPR}(b), additional features are found, with larger IPR for both smallest and second largest eigenvalue. This agrees with \cite{Plerou_2000} where, for empirical data, the group structure resulted in a number of small and large eigenvalues with larger IPR than that of the bulk of eigenvalues. These large eigenvalues were shown, \cite{Plerou_2000,Conlon_2007}, to be associated with correlation information related to the group structure.
\begin{figure}[htbp!]
\begin{center}
\includegraphics[height=70mm,width=130mm]{IPRsim.eps}
\caption{(a) IPR Empirical Data (b) IPR Simulated Data}
\label{IPR}
\end{center}
\end{figure}
\subsection{Drawdown Analysis}
As described, \cite{Drozdz_2000}, and discussed above, drawdowns (periods of large negative returns) tend to reflect an increase of one eigenstate of the cross-correlation matrix, with the opposite occurring during drawups. In this section, we attempt to characterise the market according to the relative size of both the largest and small eigenvalues using stocks of the Dow Jones Euro Stoxx 50.
Returns, correlation matrix and eigenvalue spectrum time series for overlapping windows of $200$ days were calculated and normalised using the mean and standard deviation over the entire series, (Eqn.~\ref{normalise}). Representing normalised eigenvalues in terms of standard deviation units (SDU) allows partitioning according to the magnitude of the eigenvalues. Table~\ref{evalueReturns} shows the average return of the index, (during periods when the largest eigenvalue is $\pm1$ SDU), for both the largest eigenvalue and the average of the normalised $40$ smallest eigenvalues.
The results, Table~\ref{evalueReturns}, demonstrate that when the largest eigenvalue is $>1$ SDU, the average index return over a $200$ day period is found to be $-16.80\%$. When it is small ($<-1$ SDU), the average index return is $18.46\%$. Hence, the largest eigenvalue can be used to characterise markets, with the eigenvalue peaking during periods of negative returns (Drawdowns) and bottoming out when the market is rising (Drawup). For the average of the $40$ smallest eigenvalues, the partition also reflected drawdowns and drawups, but with opposite signs. This indicates that information about the correlation dynamics of financial time series is visible in both the lower and upper eigenstates, in agreement with the results found by \cite{Schindler_2007,Muller_2005} for both synthetic data and EEG seizure data.
\begin{table*}[htbp!]
\centering
\begin{tabular}{c c c c}
& \emph{Eigenvalues} & \emph{No. Std} & \emph{Index Return} \\
\hline
\hline
& & $<$-1 & 18.46\% \\[0ex]
& \raisebox{1.5ex}{\emph{Large}} & $>$1 & -16.80\% \\[0ex]
& & $<$-1 & -23.90\% \\[0ex]
& \raisebox{1.5ex}{\emph{Average 40 Smallest}} & $>$1 & 19.32\% \\[0ex]
\end{tabular}
\caption{Drawdown/Drawup analysis, average index returns for eigenvalue partitions in SDU.}
\label{evalueReturns}
\end{table*}
\section{Conclusions}
The correlation structure of multivariate financial time series was studied by investigation of the eigenvalue spectrum of the equal-time cross-correlation matrix. By filtering the correlation matrix through the use of a sliding window we have examined behaviour of the largest eigenvalue over time. As shown, Figures (\ref{S&PeigsNorm} -~\ref{SX5E}), the largest eigenvalue moves counter to that of a band of small eigenvalues, due to \textit{eigenvalue repulsion}. A decrease in the largest eigenvalue, with a corresponding increase in the small eigenvalues, corresponds to a redistribution of the correlation structure across more dimensions of the vector space spanned by the correlation matrix. Hence, additional eigenvalues are needed to explain the correlation structure in the data. Conversely, when the correlation structure is dominated by a smaller number of factors (eg. the ``single-factor model'' of equity returns), the number of eigenvalues needed to describe the correlation structure in the data is reduced. In the context of previous work, \cite{Drozdz_2000,Drozdz_2001}, this means that fewer eigenvalues are needed to describe the correlation structure of `drawdowns' than that of `draw-ups'.
By introducing a simple `one-factor model' of the correlation in the system (Section~\ref{corr_model}), we were able to reproduce the main results of the empirical study. The compensatory dynamics, described, were clearly seen for a correlation matrix with all elements equal to the average of the empirical correlation matrix. The model was then adapted, by the addition of pertubations to the correlations, with the average correlation remaining unchanged. This `markets plus sectors' type model was able to reproduce additional features of the empirical correlation matrix with eigenvalues deviating from below and above the bulk. The Inverse Participation Ratio of the ``markets plus sectors'' model was also shown to have group characteristics typically associated with known Industrial Sectors, with a larger than average value for the smallest eigenvalue and for the second largest eigenvalue. Through a partition of the time-normalised eigenvalues, it was then shown quantitatively that the largest eigenvalue is greatest/smallest during drawdowns/drawups, and vice versa for the small eigenvalues.
Suggested future work includes a more detailed study of the relationship between the direction of the market and magnitude of the eigenvalues of the correlation matrix. Studying the multiscaled correlation dynamics over \emph{different granularities} may shed some light on the different collective behaviour of traders with different strategies and time horizons. Additional analysis of high frequency data may also be useful in the characterisation of correlation dynamics, especially prior to market crashes. Study of the possible relationship between the dynamics of the small eigenvalues and additional correlation information which, according to some authors, \cite{Guhr_2003,Burda_2004_a,Burda_2004_b,Malevergne,Kwapien_2006,Muller_2005,Muller_2006_b}, may be hidden in the part of the eigenvalue spectrum normally classifed as noise could be achieved through analysis of the relative dynamics of the small and large eigenvalues at times of extreme volatility, (such as during market crashes).
|
\section{\protect\bigskip Introduction}
The experimental evidence that Lorentz symmetry is preserved for effective
four-dimensional theories is overwhelming. In curved space-time this Lorentz
symmetry is realized as a local symmetry of the tangent manifold \cite%
{utiyama} \cite{Kibble}. Moreover, to incorporate spinors in general
relativity, we are forced to consider this local symmetry because there are
no spinor representations of the diffeomorphism group. Usually the dimension
of the tangent space is taken to be equal to the dimension of the curved
manifold and then the Lorentz symmetry is simply a manifestation of the
equivalence principle, which is valid in torsion-free theories. General
relativity could then be formulated as a gauge theory of the Lorentz group
where the gauge fields are the spin-connection. In reality one can search
for all possible tangent groups in $d$-dimensional space-time \cite{Wein}.
In this paper we will investigate whether it is possible to have a larger
group of symmetry in the tangent space and still unambiguously reproduce
general relativity. We will show in section 2, that this is indeed possible
by taking the tangent space to be real with de Sitter group symmetry. The de
Sitter gauge invariant action which is linear in curvature is shown to be
identical to Einstein gravity, provided that metricity condition is imposed
on the spin and affine connections. In section 3 we consider matter
interactions of gravity with the de Sitter group as the tangent group. We
then, in section 4, consider a complex tangent space and show that the
relevant symmetry in this case is the unitary symmetry. The resultant theory
is the Einstein-Strauss theory. Section 5 is the conclusion. An appendix
treats the special limit of Poincare symmetry, and examines the relation of
our new formalism in three dimensions with Witten's formulation of
Chern-Simons gravity.
\section{Gravity with de Sitter tangent group}
Let us begin with a $d$-dimensional manifold and assume that at every point
of this manifold there is a real $N$-dimensional tangent space spanned by
linearly independent vectors $\mathbf{v}_{A}$, where $A=1,2...N.$ Assuming
that $d\leq N$, the coordinate basis vectors $\mathbf{e}_{\alpha }\equiv
\partial /\partial x^{\alpha },$ where $\alpha =1,2...d,$ span $d$%
-dimensional space. Next we define the scalar product in the tangent space
and take the vectors $\mathbf{v}_{A}$ to be orthonormal\footnote{%
We use the notation and methods of Misner, Thorne and Wheeler (\cite{MTW}),
in particular Chapters 9 and 10.}
\begin{equation}
\mathbf{v}_{A}\cdot \mathbf{v}_{B}=\eta _{AB}. \label{1}
\end{equation}%
where $\eta _{AB}$ is Minkowski matrix. The Lorentz transformations
\begin{equation}
\mathbf{\tilde{v}}_{A}=\Lambda _{A}^{\hspace{0.05in}B}\mathbf{v}_{B},\text{
\ \ \ \ \ \ }\Lambda _{A}^{\hspace{0.05in}C}\eta _{CD}\Lambda _{A}^{\hspace{%
0.05in}D}=\eta _{AB}\text{\ } \label{1a}
\end{equation}%
preserve the orthogonality of the vielbein, $\mathbf{\tilde{v}}_{A}\cdot
\mathbf{\tilde{v}}_{B}=\eta _{AB}.$ The scalar product of coordinate basis
vectors then induces the metric in $d$-dimensional manifold%
\begin{equation}
\mathbf{e}_{\alpha }\cdot \mathbf{e}_{\beta }=g_{\alpha \beta }(x^{\gamma }).
\label{2}
\end{equation}%
Expanding $\mathbf{e}_{\alpha }$ in $\mathbf{v}_{A}$-basis%
\begin{equation}
\mathbf{e}_{\alpha }=e_{\alpha }^{B}\mathbf{v}_{B}, \label{3}
\end{equation}%
and substituting in (\ref{2}) we obtain the following expression for the
metric $g_{\alpha \beta }$
\begin{equation}
g_{\alpha \beta }=e_{\alpha }^{A}e_{\beta }^{B}\eta _{AB}, \label{4}
\end{equation}%
in terms of components. Tangent space indices are raised and lowered with
the Minkowski metric, thus%
\begin{equation}
e_{A\alpha }=\eta _{AB}e_{\alpha }^{B}=\left( \mathbf{v}_{A}\cdot \mathbf{e}%
_{\alpha }\right) , \label{4a}
\end{equation}%
and $\eta ^{AB}$ is inverse to Minkowski matrix $\eta _{AB}.$ Next we
consider parallel transport on the manifold relating vectors in
\textquotedblleft nearby\textquotedblright\ tangent spaces. The affine and
spin connections determining the rules for parallel transport of the
coordinate basis vectors and vielbein are defined via%
\begin{equation}
\mathbf{\nabla }_{\mathbf{e}_{\beta }}\mathbf{e}_{\alpha }\equiv \mathbf{%
\nabla }_{\beta }\mathbf{e}_{\alpha }=\Gamma _{\alpha \beta }^{\nu }\mathbf{e%
}_{\nu },\ \ \mathbf{\nabla }_{\beta }\mathbf{v}_{A}=-\omega _{\beta
A}^{\quad \hspace{0.03in}B}\mathbf{v}_{B}, \label{5a}
\end{equation}%
where $\mathbf{\nabla }_{\beta }$ is the derivative defining the rate of
change of vectors along a basis vector $\mathbf{e}_{\beta }$. When applied
to a scalar function $f$ \ this derivative acts as a partial derivative with
respect to the appropriate coordinates, that is, $\mathbf{\nabla }_{\beta
}f=\partial f/\partial x^{\beta }$. Notice that $\eta _{AB}$ and $g_{\alpha
\beta }$ as defined in (\ref{1}) and (\ref{2}) are the sets of scalar
functions and, hence, $\mathbf{\nabla }_{\beta }\eta _{AB}=0,$ $\mathbf{%
\nabla }_{\gamma }g_{\alpha \beta }=\partial g_{\alpha \beta }/\partial
x^{\gamma }\equiv \partial _{\gamma }g_{\alpha \beta }$.
Given $\eta _{AB},$ $g_{\alpha \beta }$ and $e_{\alpha }^{A}$ we derive the
consistency (metricity) conditions for the connections by taking derivative
of equations (\ref{1}), (\ref{2}) and (\ref{4a}). In particular, we obtain%
\begin{equation}
\left( \mathbf{\nabla }_{\alpha }\mathbf{v}_{A}\right) \cdot \mathbf{v}_{B}+%
\mathbf{v}_{A}\cdot \left( \mathbf{\nabla }_{\alpha }\mathbf{v}_{B}\right)
=-\omega _{\alpha AB}^{\quad }-\omega _{\alpha BA}^{\quad }=\mathbf{\nabla }%
_{\alpha }\eta _{AB}=0, \label{6}
\end{equation}%
that is, the spin connection should be antisymmetric with respect to tangent
space indices, $\omega _{\alpha AB}^{\quad }=-\omega _{\alpha BA}^{\quad }.$
Applying the derivative $\mathbf{\nabla }_{\gamma }$ to (\ref{2}) gives%
\begin{equation}
\Gamma _{\alpha \gamma }^{\nu }g_{\nu \beta }+\Gamma _{\beta \gamma }^{\nu
}g_{\alpha \nu }=\partial _{\gamma }g_{\alpha \beta } \label{7}
\end{equation}%
Assuming that torsion is absent, $\Gamma _{\alpha \beta }^{\nu }=\Gamma
_{\beta \alpha }^{\nu }$, these equations are solved unambiguously, giving
the well known result%
\begin{equation}
\Gamma _{\alpha \beta }^{\gamma }=\frac{1}{2}g^{\gamma \sigma }\left(
g_{\alpha \sigma ,\beta }+g_{\sigma \beta ,\alpha }-g_{\alpha \beta ,\sigma
}\right) , \label{8}
\end{equation}%
where $g^{\gamma \sigma }$ is inverse to $g_{\alpha \beta },$ that is, $%
g^{\alpha \sigma }g_{\sigma \beta }=\delta _{\beta }^{\alpha }.$ We would
like to stress that affine connections are determined unambiguously
irrespective of the group of tangent space. Finally, from (\ref{4a}) we
obtain%
\begin{equation}
\partial _{\beta }e_{A\alpha }=-\omega _{\beta A}^{\quad \hspace{0.03in}%
B}e_{B\alpha }+\Gamma _{\alpha \beta }^{\nu }e_{A\nu }. \label{9}
\end{equation}%
Let us find when these equations can unambiguously be solved for $\omega
_{\beta B}^{\quad \ A}$ in terms of the soldering form $e_{\alpha }^{B}$ and
metric $g_{\alpha \beta }.$ The total number of components of $e_{\alpha
}^{B}$ is $Nd$. Given a metric $g_{\alpha \beta },$ whose derivatives
determine $\Gamma _{\alpha \beta }^{\nu }$ via (\ref{8}), and hence impose $%
\frac{1}{2}d^{2}\left( d+1\right) $ constraints on $\partial _{\beta
}e_{A\alpha }$, leaves us with $d\left( Nd-\frac{1}{2}d\left( d+1\right)
\right) $ \textit{independent} equations (\ref{9}) to determine $\frac{1}{2}%
dN\left( N-1\right) $ antisymmetric spin connections $\omega _{\beta
AB}^{\quad }.$ Note that for any $N$ and $d$ \ the number of equations can
never exceed the number of independent $\omega _{\beta AB}^{\quad }$ to be
determined, and hence for any dimension of tangent space the system of
equations is not overdetermined. However, the spin connection is
unambiguously determined only if the number of equations is equal to the
number of its unknown components:%
\begin{equation*}
d\left( Nd-\frac{1}{2}d\left( d+1\right) \right) =\frac{1}{2}dN\left(
N-1\right) .
\end{equation*}%
The only solutions of this equation are $N=d$ \ and $N=d+1.$ The first case
is well known and thus we shall concentrate on the second case which
corresponds to the larger symmetry group $SO(1,d)$ of the tangent space. In
the case of a four-dimensional manifold the tangent space is five
dimensional. The metric in 5d tangent space can then be taken either to be $%
\eta _{AB}=\mathrm{diag}\left( 1,-1,-1,-1,-1\right) $ or $\eta _{AB}=\mathrm{%
diag}\left( 1,1,-1,-1,-1\right) $. In the first case the gauge group is 5d
Lorentz group $SO(1,4)$ which is also the group of symmetry of 4d de Sitter
space (de Sitter group), while in the second case the group is $SO(2,3)$
(the group of symmetry of 4d anti de Sitter space). For definiteness and
from \ here on, we consider these cases only. Note that although the
consistency equations do not lead to any contradiction for an arbitrary
dimension of tangent space the connections are entirely determined by the
soldering form only if $N=d$ or $N=d+1.$ Otherwise the spin connection is
not unambiguously determined by the fundamental soldering form and the
theory is not well defined.
In order to construct gauge invariant Lagrangians we need to define%
\begin{equation}
e_{A}^{\alpha }=g^{\alpha \gamma }e_{\gamma A}=g^{\alpha \gamma }\eta
_{AB}e_{\gamma }^{B}. \label{10}
\end{equation}%
Rewritten in terms of $e_{A}^{\alpha },$ equation (\ref{9}) becomes
\begin{equation}
\partial _{\beta }e_{A}^{\alpha }=-\omega _{\beta A}^{\quad \hspace{0.03in}%
B}e_{B}^{\alpha }-\Gamma _{\nu \beta }^{\alpha }e_{A}^{\nu }. \label{11}
\end{equation}%
The soldering form $e_{A}^{\alpha }$ is inverse to $e_{\beta }^{B}$ only if
the of dimension of the tangent space and the dimension of the manifold
match. In case of a de Sitter tangent group contraction over tangent space
indices gives%
\begin{equation}
e_{A}^{\alpha }e_{\beta }^{A}=g^{\alpha \gamma }\eta _{AB}e_{\gamma
}^{B}e_{\beta }^{A}=g^{\alpha \gamma }g_{\gamma \beta }=\delta _{\beta
}^{\alpha }, \label{12}
\end{equation}%
however, contraction over space-time indices gives
\begin{equation}
e_{A}^{\alpha }e_{\alpha }^{B}\neq \delta _{B}^{A}.
\end{equation}%
To prove this, let us introduce the unit vector $\mathbf{n}$ orthogonal to
all $\mathbf{e}_{\alpha },$ that is, $\mathbf{n}\cdot \mathbf{e}_{\alpha }=0$
and $\mathbf{n}\cdot \mathbf{n=}\varepsilon ,$ where $\varepsilon =-1$ or $%
+1 $ for de Sitter and anti de Sitter groups correspondingly. The vectors $%
\mathbf{n}$ and $\mathbf{e}_{\alpha }$ form a complete basis in tangent
space and therefore
\begin{equation}
\mathbf{v}_{A}=v_{A}^{\alpha }\mathbf{e}_{\alpha }+n_{A}\mathbf{n.}
\label{13}
\end{equation}%
Taking into account (\ref{4a}) we have%
\begin{equation}
v_{A}^{\alpha }=g^{\alpha \gamma }\left( \mathbf{v}_{A}\cdot \mathbf{e}%
_{\gamma }\right) =g^{\alpha \gamma }\eta _{AB}e_{\gamma }^{B}=e_{A}^{\alpha
}, \label{14}
\end{equation}%
that is, the soldering form $e_{A}^{\alpha }$ coincides with the coefficient
$v_{A}^{\alpha }$ in expansion (\ref{13}). Taking this into account one gets
\begin{equation}
\eta _{AB}=\mathbf{v}_{A}\cdot \mathbf{v}_{B}=v_{A}^{\alpha }v_{B}^{\beta
}g_{\alpha \beta }+\varepsilon n_{A}n_{B}=e_{A}^{\alpha }e_{\alpha
B}+\varepsilon n_{A}n_{B}, \label{15}
\end{equation}%
or after rasing the tangent space index we obtain%
\begin{equation}
e_{A}^{\alpha }e_{\alpha }^{B}=\delta _{A}^{B}-\varepsilon n_{A}n^{B}\equiv
P_{B}^{A} \label{16}
\end{equation}%
where $P_{B}^{A}$ is a projection operator: $P_{C}^{A}P_{B}^{C}=P_{B}^{A}.$
The components $n_{A}$ satisfy the following relations%
\begin{equation}
n^{A}e_{A}^{\alpha }=0,\text{ \ }n_{A}n^{A}=\varepsilon . \label{16a}
\end{equation}%
To prove this let us note that it follows from (\ref{13}) that $\mathbf{v}%
_{A}\cdot \mathbf{n}=\varepsilon n_{A}.$ Substituting here the expansion
\begin{equation}
\mathbf{n}=\tilde{n}^{B}\mathbf{v}_{B}, \label{16b}
\end{equation}%
we infer that $\tilde{n}^{B}=\varepsilon n^{B}$ and hence%
\begin{equation}
\mathbf{n}=\varepsilon n^{B}\mathbf{v}_{B}=\varepsilon \left(
n^{B}e_{B}^{\alpha }\mathbf{e}_{\alpha }+n^{B}n_{B}\mathbf{n}\right) ,
\label{16c}
\end{equation}%
from which (\ref{16a}) immediately follows.
In vielbein formalism the soldering form $e_{A}^{\alpha }$ is a fundamental
quantity and the group of symmetry under which the theory is required to be
invariant is the group of \textit{local} Lorentz transformations (\ref{1a}),
where $\Lambda _{A}^{\quad B}=\Lambda _{A}^{\quad B}\left( x\right) .$ Under
Lorentz transformation we have%
\begin{equation}
\mathbf{\tilde{v}}_{A}=\Lambda _{A}^{\hspace{0.05in}B}\mathbf{v}_{B}=\Lambda
_{A}^{\hspace{0.05in}B}\left( e_{B}^{\alpha }\mathbf{e}_{\alpha }+n_{B}%
\mathbf{n}\right) =\tilde{e}_{A}^{\alpha }\mathbf{e}_{\alpha }+\tilde{n}_{A}%
\mathbf{n,} \label{17}
\end{equation}%
and hence%
\begin{equation}
e_{A}^{\alpha }\rightarrow \tilde{e}_{A}^{\alpha }=\Lambda _{A}^{\hspace{%
0.05in}B}e_{B}^{\alpha } \label{18}
\end{equation}%
The transformation law for the spin connection follows from its definition:%
\begin{equation*}
\tilde{\omega}_{\beta A}^{\quad \,B}\mathbf{\tilde{v}}_{B}=-\mathbf{\nabla }%
_{\beta }\mathbf{\tilde{v}}_{A}
\end{equation*}%
Substituting $\mathbf{\tilde{v}}_{B}=\Lambda _{A}^{\quad C}\mathbf{v}_{C}$
and taking into account (\ref{5a}) we infer that
\begin{equation}
\omega _{\mu A}^{\quad B}\rightarrow \tilde{\omega}_{\mu A}^{\quad \hspace{%
0.03in}B}=\left( \Lambda \omega _{\mu }\Lambda ^{-1}\right)
_{A}^{\,B}+\left( \Lambda \partial _{\mu }\Lambda ^{-1}\right) _{A}^{\,B},
\label{20}
\end{equation}%
where $\Lambda $ and $\Lambda ^{-1}$ are the matrices corresponding to
Lorentz transformation and its inverse. Up to this point, we have considered
only vector representations of the Lorentz group. In general,
\begin{equation}
\Lambda =\exp \left( \lambda ^{AB}J_{AB}\right)
\end{equation}%
where $J_{AB}$ are corresponding generators of the Lie algebra which satisfy
the commutation relations%
\begin{equation}
\left[ J_{AB},J_{CD}\right] =\frac{1}{2}\left( \eta _{BC}J_{AD}-\eta
_{AC}J_{BD}-\eta _{BD}J_{AC}+\eta _{AD}J_{BC}\right)
\end{equation}%
Consider spinors $\psi $ which transforms according to
\begin{equation}
\psi \rightarrow \exp \left( \frac{1}{4}\lambda ^{AB}\Gamma _{AB}\right)
\psi , \label{21}
\end{equation}%
where$\ \Gamma _{AB}=\frac{1}{2}\left( \Gamma _{A}\Gamma _{B}-\Gamma
_{B}\Gamma _{A}\right) $ are generators of the Lie algebra in the spinor
representation and $\Gamma _{A}$ are $d+1$ Dirac matrices satisfying%
\begin{equation}
\left\{ \Gamma ^{A},\Gamma ^{B}\right\} =2\eta ^{AB},\quad \Gamma ^{\dagger
A}=\Gamma ^{0}\Gamma ^{A}\Gamma ^{0}. \label{22}
\end{equation}%
We note that the signature of $\eta ^{AB}$ does not play any significant
role in the derivations that follow, and thus our results holds equally well
for both de Sitter and anti de Sitter tangent groups. The Dirac action
\begin{equation}
\dint d^{4}x\sqrt{g}\,\overline{\psi }i\Gamma ^{C}e_{C}^{\alpha }D_{\alpha
}\psi , \label{23}
\end{equation}%
where \
\begin{equation}
D_{\alpha }\equiv \partial _{\alpha }+\frac{1}{4}\omega _{\alpha }^{\hspace{%
0.03in}AB}\Gamma _{AB}, \label{24}
\end{equation}%
is invariant under gauge transformations (\ref{18}), (\ref{20}) and (\ref{21}%
). This action is real, thanks to the metricity conditions (\ref{11}).
Next one constructs the curvature of the connection $D_{\mu }$ defined by%
\begin{equation}
\left[ D_{\mu },D_{\nu }\right] =\frac{1}{4}R_{\mu \nu }^{\hspace{0.05in}%
\hspace{0.05in}AB}\Gamma _{AB}, \label{25}
\end{equation}%
where%
\begin{equation}
R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}\left( \omega \right)
=\partial _{\mu }\omega _{\nu }^{\,\,\,AB}-\partial _{\nu }\omega _{\mu
}^{\,\,\,AB}+\omega _{\mu }^{\,\,\,AC}\omega _{\nu C}^{\quad B}-\omega _{\nu
}^{\,\,\,AC}\omega _{\mu C}^{\quad B}. \label{26}
\end{equation}%
This curvature$\ $transforms as
\begin{equation}
\left( R_{\mu \nu }\right) _{A}^{\hspace{0.05in}B}\rightarrow \left( \Lambda
R\Lambda ^{\,-1}\right) _{A}^{\hspace{0.05in}B}, \label{27}
\end{equation}%
and hence
\begin{equation}
\,R\left( \omega \right) =\,e_{A}^{\mu }R_{\mu \nu }^{\hspace{0.05in}\hspace{%
0.05in}AB}\left( \omega \right) e_{B}^{\nu }, \label{28}
\end{equation}%
is invariant under local gauge transformations. The gauge invariant action
is then given by
\begin{equation}
S=-\frac{1}{2\kappa ^{2}}\dint d^{4}x\sqrt{g}R\left( \omega \right)
\label{29}
\end{equation}%
Although this action appears to depend on the non-diagonal $e_{A}^{\mu }$,
it is a function of $g_{\mu \nu }$ only.
To prove this we first find how the tangent space covariant derivative acts
on the components of a vector $\mathbf{l=}l^{C}\mathbf{v}_{C}.$ Using spinor
representation for the vector we have%
\begin{equation}
D_{\nu }\left( l^{D}\Gamma _{D}\right) =\partial _{\nu }l^{D}\Gamma _{D}+%
\frac{1}{4}\omega _{\nu }^{\hspace{0.05in}BC}\left[ \Gamma _{BC},\Gamma _{D}%
\right] l^{D}. \label{30}
\end{equation}%
Taking into account the commutation relation $\left[ \Gamma _{BC},\Gamma _{D}%
\right] =2\left( \eta _{CD}\Gamma _{B}-\eta _{BD}\Gamma _{C}\right) $ one
gets%
\begin{equation}
D_{\nu }\left( l^{D}\Gamma _{D}\right) =\left( \partial _{\nu }l^{D}+\omega
_{\nu \hspace{0.05in}C}^{\hspace{0.05in}D}l^{C}\right) \Gamma _{D},
\label{31}
\end{equation}%
and hence we deduce%
\begin{equation}
D_{\nu }l^{D}=\partial _{\nu }l^{D}+\omega _{\nu \hspace{0.05in}C}^{\hspace{%
0.05in}D}l^{C}. \label{32}
\end{equation}%
In particular, it follows that
\begin{equation}
D_{\nu }e^{\rho A}=\partial _{\nu }e^{\rho A}+\omega _{\nu \hspace{0.05in}%
B}^{\hspace{0.05in}A}e^{\rho B}, \label{33}
\end{equation}%
which in turn implies that
\begin{equation}
\left[ D_{\mu },D_{\nu }\right] e^{\rho A}=R_{\mu \nu }^{\hspace{0.03in}%
\hspace{0.05in}AB}\left( \omega \right) e_{B}^{\rho }. \label{34}
\end{equation}%
On the other hand, using metricity condition (\ref{11}), we have%
\begin{equation}
D_{\nu }e^{\rho A}=-\Gamma _{\nu \sigma }^{\rho }e^{\sigma A}, \label{35}
\end{equation}%
and therefore
\begin{align}
D_{\mu }\left( D_{\nu }e^{\rho A}\right) & =-D_{\mu }\left( \Gamma _{\nu
\sigma }^{\rho }e^{\sigma A}\right) =-\left( \partial _{\mu }\Gamma _{\nu
\sigma }^{\rho }\right) e^{\sigma A}-\Gamma _{\nu \sigma }^{\rho }\left(
D_{\mu }e^{\sigma A}\right) \notag \\
& =-\partial _{\mu }\Gamma _{\nu \sigma }^{\rho }e^{\sigma A}+\Gamma _{\nu
\sigma }^{\rho }\Gamma _{\mu \kappa }^{\sigma }e^{\kappa A}. \label{36}
\end{align}%
Taking the commutator one gets
\begin{align}
\left[ D_{\mu },D_{\nu }\right] e^{\rho A}& =-\left( \partial _{\mu }\Gamma
_{\nu \sigma }^{\rho }-\partial _{\nu }\Gamma _{\mu \sigma }^{\rho }+\Gamma
_{\mu \kappa }^{\rho }\Gamma _{\nu \sigma }^{\kappa }-\Gamma _{\nu \kappa
}^{\rho }\Gamma _{\mu \sigma }^{\kappa }\right) e^{\sigma A} \notag \\
& =-R_{\,\,\,\sigma \mu \nu }^{\rho }\left( \Gamma \right) e^{\sigma A}.
\label{37}
\end{align}%
Comparing this result with (\ref{34}) we arrive at the identity
\begin{equation}
R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}\left( \omega \right)
e_{B}^{\rho }=-R_{\,\,\,\sigma \mu \nu }^{\rho }\left( \Gamma \right)
e^{\sigma A}, \label{38}
\end{equation}%
which in turn leads to
\begin{align}
\,R\left( \omega \right) & =e_{A}^{\mu }R_{\mu \nu }^{\hspace{0.05in}\hspace{%
0.05in}AB}\left( \omega \right) e_{B}^{\nu }=-R_{\,\,\,\sigma \mu \nu }^{\nu
}\left( \Gamma \right) e^{\sigma A}e_{A}^{\mu } \notag \\
& =R_{\,\,\,\sigma \nu \mu }^{\nu }\left( \Gamma \right) g^{\sigma \mu
}=R\left( \Gamma \right) . \label{39}
\end{align}%
This completes the proof that the gauge invariant action (\ref{29}) is
equivalent to Einstein action and involves only those combinations of $%
e_{A}^{\mu }$ which reduce to the metric $g_{\mu \nu }$. The remaining $%
\frac{1}{2}d\left( d+1\right) $ independent combinations of $e_{A}^{\mu }$
components represent the $\frac{1}{2}d\left( d+1\right) $ gauge degrees of
freedom associated with $SO(1,d)$. Thus, we conclude that it is possible to
formulate Einstein gravity as a gauge invariant theory with the tangent
group being de Sitter or anti de Sitter.
We would like to stress that in proving identity (\ref{39}) we never (and
could not) assume that the soldering form $e_{A}^{\mu }$ has an inverse and,
moreover, this result is valid for an arbitrary dimension of tangent space.
However, as it was noticed above the theory is well defined only if $N=d$ or
$N=d+1.$ We could also consider a gauge invariant action involving higher
order curvature invariants. One can show that even in this case the action
depends only on the metric $g_{\mu \nu }.$ To give an example consider all
possible terms which are of second order in curvature
\begin{equation}
R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}R_{\rho \sigma }^{\hspace{%
0.05in}\hspace{0.05in}CD}\left( c_{1}\ e_{A}^{\mu }e_{B}^{\nu }e_{C}^{\rho
}e_{D}^{\sigma }+c_{2}\ e_{A}^{\mu }e_{C}^{\nu }e_{D}^{\rho }e_{B}^{\sigma
}+c_{3}\ e_{C}^{\mu }e_{D}^{\nu }e_{A}^{\rho }e_{B}^{\sigma }\right) ,\
\end{equation}%
because other terms are related to these three by symmetry. The first term
is identical to $R^{2}\left( \Gamma \right) $, while for the second term we
have
\begin{equation}
R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}\left( \omega \right)
e_{A}^{\mu }e_{B}^{\sigma }R_{\rho \sigma }^{\hspace{0.05in}\hspace{0.05in}%
CD}\left( \omega \right) e_{C}^{\nu }e_{D}^{\rho }=g^{\mu \kappa
}R_{\,\,\,\kappa \mu \nu }^{\sigma }\left( \Gamma \right) g^{\nu \lambda
}R_{\,\,\,\lambda \rho \sigma }^{\rho }\left( \Gamma \right) .
\end{equation}%
after using the identity (\ref{38}) twice. Similarly, the third term gives
\begin{equation}
R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}\left( \omega \right)
e_{A}^{\rho }e_{B}^{\sigma }R_{\rho \sigma }^{\hspace{0.05in}\hspace{0.05in}%
CD}\left( \omega \right) e_{C}^{\mu }e_{D}^{\nu }=g^{\kappa \rho
}R_{\,\,\,\kappa \mu \nu }^{\sigma }\left( \Gamma \right) g^{\mu \lambda
}R_{\,\,\,\lambda \rho \sigma }^{\nu }\left( \Gamma \right) ,
\end{equation}%
which proves that the most general action which is second order in
spin-connection curvature is identical to the one that depends on
affine-connection curvature.
\section{Matter couplings}
We have seen that gravity is insensitive to the gauge group of the tangent
space. In this section we will show that, to the contrary, matter
\textquotedblleft feels\textquotedblright\ the tangent space group. Let us
consider the matter couplings in the case of de Sitter tangent group. In
this case the fundamental spinors, vectors and tensors are defined as
representations of the 5d Lorentz group of tangent space, and their
Lagrangians must be invariant with respect to de Sitter symmetry. In
vierbein formulation of gravity, we can exchange space-time tensors with
Lorentz tensors. This is no longer valid for de Sitter tangent group because
in this case the vielbein $e_{A}^{\mu }$ is not invertible and, for example,
a vector in the tangent space is not equivalent to a space-time vector. In
fact as we will show now the 5d de Sitter vector is equivalent to 4d space
time vector and real space time scalar. Therefore, de Sitter tangent space
\textquotedblleft unifies\textquotedblright\ 4d vectors and scalars.
Let us consider a 5d vector $\mathbf{H,}$ which can be expanded in terms of
components as (see (\ref{13}), (\ref{14})):
\begin{equation}
\mathbf{H}=H^{A}\mathbf{v}_{A}=H^{A}e_{A}^{\alpha }\mathbf{e}_{\alpha
}+H^{A}n_{A}\mathbf{n=}H^{\alpha }\mathbf{e}_{\alpha }+\phi \mathbf{n,}
\label{40a}
\end{equation}%
where
\begin{equation}
H^{\alpha }=H^{A}e_{A}^{\alpha },\text{ \ }\phi =H^{A}n_{A}, \label{41a}
\end{equation}%
are the components of a 4d vector and a scalar, respectively. Multiplying
the first equation by $e_{\alpha }^{B}$ and taking into account (\ref{16})
we derive
\begin{equation}
H^{B}=H^{\alpha }e_{\alpha }^{B}+\varepsilon \phi n^{B}; \label{42a}
\end{equation}%
since $e_{\alpha }^{B}n_{B}=0$ and $n^{A}n_{A}=\varepsilon $ (see (\ref{16a}%
)) it follows from here that
\begin{equation}
H^{B}H_{B}=g_{\alpha \beta }H^{\alpha }H^{\beta }+\varepsilon \phi ^{2}.
\label{43aa}
\end{equation}%
Let us construct the curvature of $H_{A}$%
\begin{equation}
F_{AB}=D_{A}H_{B}-D_{B}H_{A}, \label{44a}
\end{equation}%
where $D_{A}\equiv e_{A}^{\alpha }D_{\alpha }$ and $D_{\alpha }$ is
covariant derivative with respect to tangent space vector indices (see (\ref%
{32})); therefore, the components with only space time indices are scalars
with respect to this derivative, for example, $D_{\alpha }H^{\beta
}=\partial _{\alpha }H^{\beta }.$ Taking this into account and using
decomposition (\ref{42a}) we find%
\begin{equation}
D_{A}H^{B}=e_{A}^{\beta }e_{\alpha }^{B}\partial _{\beta }H^{\alpha
}+e_{A}^{\beta }H^{\alpha }D_{\beta }e_{\alpha }^{B}+\varepsilon
e_{A}^{\beta }n^{B}\partial _{\beta }\phi +\varepsilon e_{A}^{\beta }\phi
D_{\beta }n^{B}. \label{45a}
\end{equation}%
The last term here is equal to zero. In fact, using the definition (\ref{5a}%
) we have
\begin{equation}
\partial _{\beta }n_{A}=\varepsilon \mathbf{\nabla }_{\beta }\left( \mathbf{v%
}_{A}\cdot \mathbf{n}\right) =-\omega _{\beta A}^{\quad \
B}n_{B}+\varepsilon \mathbf{v}_{A}\cdot \mathbf{\nabla }_{\beta }\mathbf{n,}
\label{46aaa}
\end{equation}%
and hence $D_{\beta }n_{A}=-\varepsilon \mathbf{v}_{A}\cdot \mathbf{\nabla }%
_{\beta }\mathbf{n.}$ In turn, one can immediately conclude from $\mathbf{%
\nabla }_{\beta }\left( \mathbf{e}_{\alpha }\cdot \mathbf{n}\right) =0$ and $%
\mathbf{\nabla }_{\beta }\left( \mathbf{n}\cdot \mathbf{n}\right) =0$ that $%
\mathbf{\nabla }_{\beta }\mathbf{n=}0$ and therefore $D_{\beta }n_{A}=0.$
Using metricity condition (\ref{35}) to express $D_{\beta }e_{\alpha }^{B}$
in terms of $\Gamma _{\nu \sigma }^{\rho }$ and interchanging indices we
then find%
\begin{equation}
F_{AB}=e_{A}^{\beta }e_{B}^{\alpha }\left( \partial _{\beta }H_{\alpha
}-\partial _{\alpha }H_{\beta }\right) +\varepsilon \left( e_{A}^{\beta
}n_{B}-e_{B}^{\beta }n_{A}\right) \partial _{\beta }\phi . \label{47a}
\end{equation}%
Note that $F_{AB}$ is invariant under the $U(1)$ gauge transformation%
\begin{equation}
H_{A}\rightarrow H_{A}+e_{A}^{\alpha }\partial _{\alpha }\Lambda ,
\label{48a}
\end{equation}%
which in terms of the space time components become $H_{\alpha }\rightarrow
H_{\alpha }+\partial _{\alpha }\Lambda ,$ $\phi \rightarrow \phi .$ Squaring
(\ref{47a}) we will find the gauge invariant Lagrangian density for the
massless vector field
\begin{equation}
L=-\frac{1}{4}F_{AB}F^{AB}=-\frac{1}{4}F_{\alpha \beta }F^{\alpha \beta }-%
\frac{1}{2}\varepsilon \partial _{\alpha }\phi \partial ^{\alpha }\phi ,
\label{49a}
\end{equation}%
where
\begin{equation}
F_{\alpha \beta }=\partial _{\alpha }H_{\beta }-\partial _{\beta }H_{\alpha
}. \label{50a}
\end{equation}%
Notice that we get the correct sign for the kinetic energy of the scalar
field $\phi $ only in the case of de Sitter group ($\varepsilon =-1$) while
for anti de Sitter group $\varepsilon =1$ we get a ghost$.$ We deduce that
the formulation of gravity where the tangent group is $SO(1,d)$ instead of $%
SO(1,d-1)$ unifies spins zero and spin one in one vector field. If we add to
the Lagrangian the term (\ref{43aa}) both fields acquire the same mass.
We now turn to spinors. Because they should respect 5d tangent Lorentz group
it is well known that neither Majorana or Weyl conditions can be imposed on
them \cite{Scherk}. Thus the spinors $\psi $ must be Dirac spinors. The
Dirac action in this case is
\begin{equation*}
\dint \sqrt{g}d^{4}x\left( i\overline{\psi }\Gamma ^{A}D_{A}\psi -i\overline{%
D_{A}\psi }\Gamma ^{A}\psi \right)
\end{equation*}%
The spinors do feel the full $SO(1,4)$ local symmetry. This seems to be a
very strong constraint as it implies that chiral spinors cannot exist if the
tangent group is $SO(1,4).$ This is similar to the situation in case of
supersymmetry in five dimensions \cite{Cremmer}, \cite{CN}, or $N=2$
supersymmetry. There, it was shown that it is possible to generalize the
Majorana condition by taking a doublet of spinors \cite{Scherk}. The
conclusion we must draw is then that the $SO(1,4)$ tangent group implies
that spinors must be treated in the same way as in $N=2$ supersymmetry. To
couple the spinors to vectors, some gauge symmetry must be introduced. As an
example, let us assume the existence of a $U(1)$ gauge symmetry. In this
case the covariant derivative $D_{A}\psi $ becomes
\begin{equation}
D_{A}\psi =\left( e_{A}^{\mu }\left( \partial _{\mu }+\frac{1}{4}\omega
_{\mu }^{\hspace{0.03in}AB}\Gamma _{AB}\right) \ +iH_{A}\right) \psi \ ,
\end{equation}%
which shows that the spinors exist in a unified interactions with both a
scalar and a vector field, as was seen in the decomposition of the vector $%
H_{A}$ into a vector $H_{\mu }$ and a scalar $\phi .$
\section{Complex gravity and unitary $U(1,d-1)$ tangent group}
As a tangent space one can also consider a complex vector space with
Hermitian scalar product satisfying
\begin{equation}
\left( \mathbf{v},\mathbf{u}\right) =\left( \mathbf{u},\mathbf{v}\right)
^{\ast },\text{ \ }\left( \mathbf{v},\alpha \mathbf{u}\right) =\alpha \left(
\mathbf{v},\mathbf{u}\right) , \label{40}
\end{equation}%
where $\alpha $ is a complex number. It follows from here that $\left(
\alpha \mathbf{v},\mathbf{u}\right) =\alpha ^{\ast }\left( \mathbf{v},%
\mathbf{u}\right) .$ As before let us introduce in this space the
orthonormal basis $\mathbf{v}_{A}$ $(A=1,...N)$:
\begin{equation}
\left( \mathbf{v}_{A},\mathbf{v}_{B}\right) =\eta _{AB}. \label{41}
\end{equation}%
The condition of orthogonality is preserved under $U(1,N-1)$ transformations%
\begin{equation}
\mathbf{\tilde{v}}_{A}=U_{A}^{\hspace{0.05in}C}\mathbf{v}_{C},\ \ \ \ \ \
U_{A}^{\hspace{0.05in}C}\eta _{CD}\left( U_{A}^{\hspace{0.05in}D}\right)
^{\ast }=\eta _{AB}. \label{43}
\end{equation}%
For generality let us first consider the complex coordinate basis vectors $%
\mathbf{e}_{\alpha }$ $(\alpha =1,...d)$ in $d$-dimensional manifold and
show that in this case we obtain the Hermitian theory of gravity as
formulated by Einstein and Strauss \cite{Ein}, \cite{ES}. Later on we will
show that this theory can be consistently truncated to General Relativity
while preserving the unitary structure of the tangent space.
Assuming that $N\geq d$ \ we can expand the coordinate basis vectors in
terms of vielbein vectors, $\mathbf{e}_{\alpha }=e_{\alpha }^{A}\mathbf{v}%
_{A},$ and then the metric on the manifold can be expressed as%
\begin{equation}
g_{\alpha \beta }\equiv \left( \mathbf{e}_{\alpha },\mathbf{e}_{\beta
}\right) =e_{\alpha }^{A}e_{\beta }^{B\ast }\eta _{AB}. \label{44}
\end{equation}%
This metric is Hermitian%
\begin{equation*}
g_{\alpha \beta }=\left( \mathbf{e}_{\alpha },\mathbf{e}_{\beta }\right)
=\left( \mathbf{e}_{\beta },\mathbf{e}_{\alpha }\right) ^{\ast }=g_{\beta
\alpha }^{\ast }.
\end{equation*}%
In the case under consideration the affine and spin connections are defined
exactly as in (\ref{5a}). Taking derivative of (\ref{44}) and using
definition in (\ref{5a}) we obtain%
\begin{equation}
\partial _{\gamma }g_{\alpha \beta }=\left( \mathbf{\nabla }_{\gamma }%
\mathbf{e}_{\alpha },\mathbf{e}_{\beta }\right) +\left( \mathbf{e}_{\alpha },%
\mathbf{\nabla }_{\gamma }\mathbf{e}_{\beta }\right) =\Gamma _{\alpha \gamma
}^{\nu \ast }g_{\nu \beta }+\Gamma _{\beta \gamma }^{\nu }g_{\alpha \nu }.
\label{45}
\end{equation}%
These $d^{\,3}$ equations can be solved unambiguously for $\Gamma _{\kappa
\rho }^{\mu }$ in terms of metric $g_{\alpha \beta }$ only if we impose the
hermiticity condition%
\begin{equation}
\Gamma _{\rho \mu }^{\nu \ast }=\Gamma _{\mu \rho }^{\nu }, \label{46}
\end{equation}%
which leaves us with $d^{3}$ components to be determined. Unlike the real
case equations (\ref{46}) can be solved only perturbatively. They were first
imposed by Einstein in his formulation of Hermitian gravity which he
referred to as the "$+-$" condition \cite{Ein}, \cite{ES}, \cite{DD}.
Similar to (\ref{6}) we derive a condition on spin connection%
\begin{equation}
\omega _{\alpha A}^{\quad C}\eta _{CB}=-\left( \omega _{\alpha B}^{\quad
C}\right) ^{\ast }\eta _{CA}, \label{46a}
\end{equation}%
which leaves $N^{2}d$ independent components. Taking derivative of $\left(
\mathbf{v}_{A},\mathbf{e}_{\alpha }\right) =e_{\alpha }^{B}\eta _{AB}$ we
derive the following metricity conditions
\begin{equation}
\partial _{\gamma }e_{\alpha }^{A}=\omega _{\gamma B}^{\quad \,A}e_{\alpha
}^{B}+\Gamma _{\alpha \gamma }^{\nu }e_{\nu }^{A}. \label{47}
\end{equation}%
Taking into account that $d^{3}$ equations (\ref{45}) determine $\Gamma
_{\beta \gamma }^{\nu }$ through $\partial _{\gamma }e_{\alpha }^{A}$ we are
left with $2Nd^{2}-d^{3}$ equations to find $N^{2}d$ independent components
of $\omega _{\alpha A}^{\quad C}.$ The number of equations match the number
of unknown components only if $N=d,$ that is, when dimension of complex
tangent space coincides with the dimension of the manifold. Hence the gauge
group of the tangent space can be only $U(1,d-1)$ \cite{AHC}. In this case
we can define the soldering form $e_{B}^{\beta },$ which is inverse to $%
e_{\alpha }^{A}:$%
\begin{equation}
e_{B}^{\alpha }e_{\alpha }^{A}=\delta _{B}^{A},\text{ \ \ }e_{A}^{\alpha
}e_{\beta }^{A}=\delta _{\beta }^{\alpha }.\text{\ } \label{48}
\end{equation}%
The metric with upper indices is then given by%
\begin{equation}
g^{\mu \nu }=e_{A}^{\mu }e_{B}^{\nu \ast }\eta ^{AB}, \label{49}
\end{equation}%
and it is inverse to $g_{\alpha \beta }$
\begin{equation}
g_{\alpha \nu }g^{\beta \nu }=\delta _{\alpha }^{\beta }\neq g_{\alpha \nu
}g^{\nu \beta }. \label{50}
\end{equation}%
Similar to (\ref{34}) the curvature of the connection $\omega _{\mu
A}^{\quad B}$ can be defined as
\begin{align}
\left[ D_{\mu },D_{\nu }\right] e_{A}^{\sigma }& \equiv R_{\mu \nu A}^{\quad
\hspace{0.03in}\hspace{0.03in}B}\left( \omega \right) e_{B}^{\sigma } \notag
\\
& =\left( \partial _{\mu }\omega _{\nu A}^{\quad B}-\partial _{\nu }\omega
_{\mu A}^{\quad B}+\omega _{\mu A}^{\quad C}\omega _{\nu C}^{\quad B}-\omega
_{\nu A}^{\quad C}\omega _{\mu C}^{\quad B}\right) e_{B}^{\sigma }.
\label{52}
\end{align}%
On the other hand, using the metricity condition, we have
\begin{align}
\left[ D_{\mu },D_{\nu }\right] e_{A}^{\sigma }& =-\left( \partial _{\mu
}\Gamma _{\rho \nu }^{\sigma }-\partial _{\nu }\Gamma _{\rho \mu }^{\sigma
}+\Gamma _{\kappa \mu }^{\sigma }\Gamma _{\rho \nu }^{\kappa }-\Gamma
_{\kappa \nu }^{\sigma }\Gamma _{\rho \mu }^{\kappa }\right) e_{A}^{\sigma }
\notag \\
& \equiv -R_{\hspace{0.03in}\hspace{0.03in}\rho \mu \nu }^{\sigma }\left(
\Gamma \right) e_{A}^{\rho }, \label{53}
\end{align}%
and it follows from here that%
\begin{equation}
R_{\hspace{0.03in}\hspace{0.03in}\rho \mu \nu }^{\sigma }\left( \Gamma
\right) =-e_{\rho }^{A}R_{\mu \nu A}^{\quad \,\,\,\,B}\left( \omega \right)
e_{B}^{\sigma }. \label{54}
\end{equation}%
In particular, the scalar curvature
\begin{align}
R\left( \omega \right) & =\eta ^{AC}e_{C}^{\mu \ast }R_{\mu \nu A}^{\quad
\,\,\,\,B}\left( \omega \right) e_{B}^{\nu }=-\eta ^{AC}e_{C}^{\mu \ast
}R_{\,\,\,\rho \mu \nu }^{\nu }\left( \Gamma \right) e_{A}^{\rho } \notag \\
& =g^{\rho \mu }R_{\,\,\,\rho \nu \mu }^{\nu }\left( \Gamma \right) =R\left(
\Gamma \right) , \label{55}
\end{align}%
is $U\left( 1,d-1\right) $ gauge invariant. The scalar curvature is real,%
\begin{equation}
R^{\ast }\left( \omega \right) =R\left( \omega \right) . \label{56}
\end{equation}%
\ To prove this we first note the identity%
\begin{equation}
\left( R_{\mu \nu A}^{\quad \,\,\,B}\left( \omega \right) \right) ^{\ast
}=-R_{\mu \nu C}^{\quad \,\,\,\,D}\left( \omega \right) \eta ^{CB}\eta _{DA},
\label{57}
\end{equation}%
which follows from equation (\ref{52}) taking into account (\ref{46a}).
Using this relation together with (\ref{54}) we obtain
\begin{align}
\left( R_{\hspace{0.03in}\hspace{0.03in}\rho \mu \nu }^{\sigma }\left(
\Gamma \right) \right) ^{\ast }& =-e_{\rho }^{A\ast }\left( R_{\mu \nu
A}^{\quad \,\,\,B}\left( \omega \right) \right) ^{\ast }e_{B}^{\sigma \ast
}=e_{\rho }^{A\ast }R_{\mu \nu C}^{\quad \,\,\,D}\left( \omega \right) \eta
^{CB}\eta _{DA}e_{B}^{\sigma \ast } \notag \\
& =-\eta ^{CB}e_{C}^{\kappa }e_{B}^{\sigma \ast }R_{\,\,\,\kappa \mu \nu
}^{\lambda }\left( \Gamma \right) \eta _{DA}e_{\lambda }^{D}e_{\rho }^{A\ast
}=-g^{\kappa \sigma }R_{\,\,\,\kappa \mu \nu }^{\lambda }\left( \Gamma
\right) g_{\lambda \rho }. \label{58}
\end{align}%
It follows from here that the tensor
\begin{equation}
R_{\rho \kappa \mu \nu }\left( \Gamma \right) =R_{\,\,\,\kappa \mu \nu
}^{\lambda }\left( \Gamma \right) g_{\lambda \rho }, \label{59}
\end{equation}%
is antihermitian with respect to exchange of first two indices
\begin{equation}
\left( R_{\kappa \rho \mu \nu }\left( \Gamma \right) \right) ^{\ast
}=-R_{\rho \kappa \mu \nu }\left( \Gamma \right) , \label{60}
\end{equation}%
and it is antisymmetric with respect to exchange of the last two indices
(see (\ref{53}). Taking this into account we have%
\begin{equation}
R^{\ast }\left( \Gamma \right) =\left( g^{\rho \mu }g^{\nu \sigma }R_{\sigma
\rho \nu \mu }\right) ^{\ast }=g^{\mu \rho }g^{\sigma \nu }R_{\rho \sigma
\mu \nu }=R\left( \Gamma \right) , \label{61}
\end{equation}%
and because $R\left( \omega \right) =R\left( \Gamma \right) ,$ this
completes the proof of reality of gauge invariant scalar curvature.
The identity (\ref{60}) was not noticed by Einstein and this forced him to
construct Hermitian combinations of the curvature tensor. As we see this is
not necessary because one can use instead the real scalar curvature as
Lagrangian density.
If we write the connection as
\begin{equation}
\omega _{\mu A}^{\quad B}=\bar{\omega}_{\mu A}^{\quad B}+\frac{1}{d}\hat{%
\omega}_{\mu }\delta _{A}^{B}, \label{62}
\end{equation}%
where
\begin{equation}
\bar{\omega}_{\mu A}^{\quad A}=0,\text{ \ }\hat{\omega}_{\mu }=\omega _{\mu
A}^{\quad A}, \label{63}
\end{equation}%
the curvature splits into two pieces
\begin{equation}
R_{\mu \nu A}^{\quad \,\,\,B}\left( \omega \right) =R_{\mu \nu A}^{\quad
\,\,B}\left( \bar{\omega}\right) +\frac{1}{d}R_{\mu \nu C}^{\quad
\,\,\,C}\left( \hat{\omega}\right) \delta _{A}^{B}, \label{64}
\end{equation}%
where
\begin{align}
R_{\mu \nu A}^{\quad \,\,\,B}\left( \bar{\omega}\right) & =\left( \partial
_{\mu }\bar{\omega}_{\nu A}^{\quad B}-\partial _{\nu }\bar{\omega}_{\mu
A}^{\quad B}+\bar{\omega}_{\mu A}^{\quad C}\bar{\omega}_{\nu C}^{\ \ \ \ B}-%
\bar{\omega}_{\nu A}^{\quad C}\bar{\omega}_{\mu C}^{\ \ \ \ B}\right) ,
\notag \\
R_{\mu \nu C}^{\quad \,\,\,C}\left( \omega \right) & =\partial _{\mu }\hat{%
\omega}_{\nu }-\partial _{\nu }\hat{\omega}_{\mu }. \label{65}
\end{align}%
It follows from here that
\begin{equation}
R\left( \omega \right) =\eta ^{AC}e_{C}^{\mu \ast }R_{\mu \nu A}^{\quad
\,\,\,B}\left( \bar{\omega}\right) e_{B}^{\nu }+\frac{1}{d}g^{\nu \mu
}R_{\mu \nu C}^{\quad \,\,\,C}\left( \hat{\omega}\right) =R\left( \bar{\omega%
}\right) +\frac{1}{d}\tilde{R}\left( \hat{\omega}\right) , \label{66}
\end{equation}%
where $\tilde{R}=g^{\nu \mu }R_{\mu \nu A}^{\quad \,\,\,A}$ is another
scalar curvature invariant. \ Therefore it can be added to the action with
an arbitrary coefficient leading to the following most general gauge
invariant first order action%
\begin{equation}
S=\dint d^{4}x\left\vert \det e_{\mu }^{A}\right\vert \left( \alpha R\left(
\bar{\omega}\right) +\beta \tilde{R}\left( \hat{\omega}\right) \right) .
\label{67}
\end{equation}%
It must be stressed that we are using here a second order formalism where
the field $\omega _{\mu A}^{\quad B}$ is determined by the metricity
condition and not by the field equations. The best strategy to analyze this
action is to solve for $\omega _{\mu A}^{\quad B}$ in a perturbative
expansion in terms of $e_{\mu }^{A}.$
We can understand the above results by noting that the gauge invariant
action allows to use the gauge invariance to reduce the independent
components of $e_{\alpha }^{A}$ to those of $g_{\alpha \beta }.$ In other
words we expect that because of $U(1,d-1)$\ gauge invariance, the action
depends only on the metric%
\begin{equation*}
g_{\alpha \beta }=e_{\alpha }^{A}e_{\beta }^{B\ast }\eta _{AB}\equiv
G_{\alpha \beta }+iB_{\alpha \beta }.
\end{equation*}%
This theory was considered before using a first order formalism where the
spin-connection was determined from the equations of motion \cite{AHC}%
\textit{. }This is possible only when the action depends quadratically on
the spin-connection. However, the $U(1)$ part $\hat{\omega}$ of the $%
U(1,d-1) $ connection being abelian, appears linearly. This then imposes a
constraint on the antisymmetric part of the metric%
\begin{equation}
\partial _{\alpha }\left( \left\vert \det e_{\mu }^{A}\right\vert B^{\alpha
\beta }\right) =0,
\end{equation}%
which thus remains undetermined \cite{AHC}. This is to be contrasted with
the second order formalism where all spin-connections are determined from
the metricity condition. \textit{\ }
We arrive to an interesting case by requiring that the metric $g_{\alpha
\beta }$ to be real. This is equivalent to truncating the $B_{\alpha \beta }$
field. Let
\begin{equation}
e_{\alpha }^{A}=e_{\alpha \left( 0\right) }^{A}+ie_{\alpha \left( 1\right)
}^{A},
\end{equation}%
so that
\begin{eqnarray}
G_{\alpha \beta } &=&\left( e_{\alpha \left( 0\right) }^{A}e_{\beta \left(
0\right) }^{B}+e_{\alpha \left( 1\right) }^{A}e_{\beta \left( 1\right)
}^{B}\right) \ \eta _{AB}, \\
B_{\alpha \beta } &=&\left( e_{\alpha \left( 1\right) }^{A}e_{\beta \left(
0\right) }^{B}-e_{\alpha \left( 0\right) }^{A}e_{\beta \left( 1\right)
}^{B}\right) \ \eta _{AB},
\end{eqnarray}%
Truncating $B_{\alpha \beta }$ gives $\frac{1}{2}d\left( d-1\right) $
constraints on the $2d^{2}$ (real) fields $e_{\alpha \left( 0\right) }^{A}$
and $e_{\alpha \left( 1\right) }^{A}$. In this case the affine connection is
also real and its $\frac{1}{2}d^{2}\left( d+1\right) $ components are
Christoffel connection for the metric $G_{\alpha \beta }.$ The remaining%
\begin{equation*}
2d^{3}-\frac{1}{2}dd\left( d-1\right) -\frac{1}{2}d^{2}\left( d+1\right)
=d^{3}
\end{equation*}%
independent equations (\ref{47}) are then enough to unambiguously determine $%
d^{3}$ components of $\omega _{\mu A}^{\quad B}.$ This implies that it is
possible to enlarge the tangent group to become $U(1,d-1)$ and still obtain
the Einstein gravity without any modification. The coupling to matter will,
however, feel the tangent group $U(1,d-1)$.
\textbf{Matter coupling.} When the tangent group is $U(1,3)$ then from the
previous discussion it should be clear that neither the Majorana nor the
Weyl condition could be imposed, except if a doublet of spinors is taken.
Thus, as with the $SO(1,4)$ case we must take a Dirac spinor, or a doublet
of Majorana or Weyl spinors, again as in the $N=2$ supersymmetric case. We
note the isomorphism of the algebras
\begin{equation}
U\left( 1,3\right) \sim SO(1,5)\times SO(1,1).
\end{equation}%
It is easy to see that $U\left( 1,3\right) $ has ten compact generators and
six non-compact generators, while $SO(1,5)$ has ten compact generators and
five non-compact generators and $SO(1,1)$ has one non-compact generator.
Thus spinors in the case of unitary tangent group will exhibit conformal
local symmetry.
Gravity has a universal coupling to matter. One way to classify the fields
is according to their behavior under the diffeomorphism group, or
equivalently under the tangent Lorentz group. A complex scalar field has the
following couplings%
\begin{equation}
\dint d^{4}x\sqrt{\det g}g^{\mu \nu }\partial _{\mu }\phi \partial _{\nu
}\phi ^{\ast }.
\end{equation}%
For a massless vector it can be easily seen that the action can be written
in terms of a complex space-time vector $H_{\mu }$\ with the action
\begin{equation}
\ \dint d^{4}x\sqrt{\det g}g^{\mu \rho }g^{\nu \sigma }F_{\mu \nu }F_{\rho
\sigma }^{\ast }.
\end{equation}%
Similarly we can treat the case of fields which are in the vector
representations of the gauge group. The fermions have more complicated
couplings. First, a Dirac spinor has the $U(1,d-1)$ transformation%
\begin{equation}
\psi \rightarrow e^{i\lambda _{B}^{\hspace{0.03in}A}\Gamma _{A}\Gamma
^{B}}\psi ,
\end{equation}%
where $\Gamma ^{A}$ and $\Gamma _{A}$ satisfy the relations
\begin{equation*}
\left\{ \Gamma ^{A},\Gamma ^{B}\right\} =0,\quad \left\{ \Gamma _{A},\Gamma
_{B}\right\} =0,\quad \left\{ \Gamma ^{A},\Gamma _{B}\right\} =\delta
_{B}^{A},
\end{equation*}%
and thus $\Gamma _{A}\Gamma ^{B}$ are the generators of $U(1,d-1).$ We can
define the Hermitian Dirac matrices
\begin{align*}
\gamma ^{\mu }& =e_{A}^{\mu }\Gamma ^{A}+e^{\mu A}\Gamma _{A}, \\
\left\{ \gamma ^{\mu },\gamma ^{\nu }\right\} & =g^{\mu \nu }+g^{\nu \mu },
\end{align*}%
The covariant derivative is given by
\begin{equation*}
D_{\mu }\psi =\partial _{\mu }\psi +\omega _{\mu B}^{\quad A}\Gamma
_{A}\Gamma ^{B}\psi ,
\end{equation*}%
Hermitian Dirac action is then
\begin{equation}
\dint d^{4}x\left\vert \det e_{\mu }^{A}\right\vert \overline{\psi }\gamma
^{\mu }D_{\mu }\psi .
\end{equation}%
Therefore, Dirac spinors do couple to both the symmetric and antisymmetric
components of the Hermitian metric.
\section{Conclusions}
We have shown that Einstein gravity exhibits universality when formulated as
a gauge theory of tangent space group. Besides of the well known natural
case when the tangent space has the same dimension as the manifold, we
discovered two other possibilities for General Relativity to be reproduced
and the theory still remains unambiguous. Namely, we have shown that in the
four dimensional case the tangent space can be five dimensional and possess
(anti) de Sitter group of symmetry. This group is important when we
incorporate matter couplings to the gravitational field. As an example, we
have shown that de Sitter tangent space group allows us to \textquotedblleft
unify\textquotedblright\ 4d vectors and scalars which become components of
the same five dimensional vector in tangent space. Even more dramatic are
the consequences of the tangent space symmetry group on fermions. They
become fundamentally five dimensional and neither Majorana nor Weyl
conditions could be imposed on them. This situations is similar to $N=2$
supersymmetry where we are forced to generalize the Majorana condition by
taking a doublet of spinors. We also would like to note that if we impose an
extra $U\left( 1\right) $ local symmetry in the tangent space then the
spinors would exist in a unified interaction with both scalar and vector
fields.
Another interesting possibility arise when we consider complex tangent space
of the same dimension as the manifold. In this case the group of symmetry is
the unitary group. This gives rise generically to the theory of Hermitian
gravity, where the basic fields are the symmetric and antisymmetric
components of the metric, which coincide with the basic fields appearing in
effective open string field theory. It is interesting that this theory can
be consistently truncated to Einstein gravity, while still preserving the
unitary group of tangent space. In turn, this has interesting and nontrivial
consequences for the coupling to matter which should respect this symmetry.
In a forthcoming paper \cite{CM} we shall explore the implications of these
new formulations of gravity, especially in regard to the spontaneous
breakdown of these larger symmetries down to the $SO(1,d-1)$ symmetry.
\section{Appendix: The Poincare limit and 3d CS gravity}
In this appendix we examine the special case when the radius of the de
Sitter tangent group becomes infinite, which corresponds to Poincare
symmetry. Later we shall also investigate the correspondence with
Chern-Simons gravity in three dimensions which also have de Sitter or
Poincare symmetry \cite{AT}, \cite{Witten}.
The $SO(1,d)$ group generators satisfy the commutation relations%
\begin{equation}
\left[ J_{AB},J_{CD}\right] =-\frac{1}{2}\left( \eta _{AC}J_{BD}-\eta
_{BC}J_{AD}-\eta _{AD}J_{BC}+\eta _{BD}J_{AC}\right) .
\end{equation}%
Splitting the range of the index $A=a,\overline{d},$ \ where $a=0,1,\cdots
,d-1,$ and similarly for the other indices we get the usual $SO(1,d-1)$ for
the $J_{ab},$ while for $J_{a\overline{d}}\equiv RP_{a}$ we have
\begin{equation}
\left[ P_{a},P_{b}\right] =-\frac{1}{R^{2}}J_{ab}.
\end{equation}%
Thus, in the limit $R\rightarrow \infty $ the de Sitter tangent group
becomes the inhomogeneous Lorentz group, i.e. $ISO(1,d-1)$ also known as the
Poincare group. The covariant derivative
\begin{equation}
D_{\mu }=\partial _{\mu }+\omega _{\mu }^{\,\,AB}J_{AB},
\end{equation}%
implies that the field $\omega _{\mu }^{a\overline{d}}$ must be defined as $%
\omega _{\mu }^{a\overline{d}}\equiv \frac{1}{2R}b_{\mu }^{a}$ so that
\begin{equation}
D_{\mu }=\partial _{\mu }+\omega _{\mu }^{\,\,ab}J_{ab}+b_{\mu }^{\,a}P_{a},
\end{equation}%
is independent of the radius $R.$ The curvatures in terms of the redefined
fields are
\begin{align}
R_{\mu \nu }^{\quad ab}& =\partial _{\mu }\omega _{\nu }^{\,\,\,ab}-\partial
_{\nu }\omega _{\mu }^{\,\,\,ab}+\omega _{\mu }^{\,\,\,ac}\omega _{\nu
c}^{\quad b}-\omega _{\nu }^{\,\,\,ac}\omega _{\mu c}^{\quad b}-\frac{1}{%
4R^{2}}\left( b_{\mu }^{\,a}b_{\nu }^{\,b}-b_{\mu }^{\,b}b_{\nu
}^{\,a}\right) , \\
R_{\mu \nu }^{\quad a\overline{d}}& =\frac{1}{2R}\left( \partial _{\mu
}b_{\nu }^{\,a}-\partial _{\mu }b_{\nu }^{\,a}+\omega _{\mu
}^{\,\,\,ac}b_{\nu c}-\omega _{\nu }^{\,\,\,ac}b_{\mu c}\right) .
\end{align}%
The zero torsion condition on $e_{A}^{\mu }$ is consistent in the limit $%
R\rightarrow \infty $ if we define
\begin{equation}
e_{\overline{d}}^{\mu }\equiv \frac{1}{R}c^{\mu },
\end{equation}%
so that
\begin{equation}
\ \partial _{\mu }c^{\nu }-\frac{1}{2}b_{\mu }^{\,a}e_{a}^{\nu }+\Gamma
_{\rho \mu }^{\nu }c^{\rho }=0,
\end{equation}%
which allows us to calculate $b_{\mu }^{a}$ in terms of $c^{\mu }.$ The
field $\omega _{\mu }^{\quad ab}$ is solved from the condition
\begin{equation}
\ \partial _{\mu }e_{a}^{\nu }+\omega _{\mu a}^{\quad b}e_{b}^{\nu }+\frac{1%
}{2R^{2}}b_{\mu }^{a}c^{\nu }+\Gamma _{\rho \mu }^{\nu }e_{a}^{\rho }=0.
\end{equation}%
Writing the gravitational action in terms of the rescaled fields, we expand $%
e_{A}^{\mu }R_{\mu \nu }^{\hspace{0.05in}\hspace{0.05in}AB}\left( \omega
\right) e_{B}^{\nu }$ to get
\begin{align}
& e_{a}^{\mu }e_{b}^{\nu }\left( \partial _{\mu }\omega _{\nu }^{\hspace{%
0.05in}ab}-\partial _{\nu }\omega _{\mu }^{\,\hspace{0.05in}ab}+\omega _{\mu
}^{\,\,\,ac}\omega _{\nu c}^{\quad b}-\omega _{\nu }^{\,\,\,ac}\omega _{\mu
c}^{\quad b}-\frac{1}{4R^{2}}\ \left( b_{\mu }^{\,a}b_{\nu }^{\,b}-b_{\mu
}^{\,b}b_{\nu }^{\,a}\right) \right) \notag \\
& \ +\frac{1}{R^{2}}e_{a}^{\mu }c^{\nu }\left( \partial _{\mu }b_{\nu
}^{\,a}-\partial _{\nu }b_{\mu }^{\,a}+\omega _{\mu }^{\hspace{0.05in}%
ac}b_{\nu c}-\omega _{\nu }^{\hspace{0.05in}ac}b_{\mu c}\right) .
\end{align}%
Therefore it is clear that in the limit $R\rightarrow \infty $ the
connection $\omega _{\mu }^{\,\,ab}$ coincides with the $SO(1,d-1)$ Lorentz
connection and the action becomes identical to the Einstein-Hilbert action.
The fields $b_{\mu }^{\,a}$ and $c^{\mu }$ drop out of the action. Thus in
the limit of $ISO(1,d-1)$ the action is indistinguishable from the $%
SO(1,d-1) $ invariant action for gravity.
For matter couplings, especially for the vector $H_{A}$, the gauge
transformation is
\begin{equation*}
\delta H_{A}=\lambda _{AB}H^{B},\quad \lambda _{AB}=-\lambda _{BA}.
\end{equation*}%
Denoting $H_{\overline{d}}=\phi $ \ and $\lambda _{a\overline{d}}=\frac{1}{2R%
}\lambda _{a}$, the gauge transformations of $H_{a}$ and $\phi $ are
\begin{align*}
\delta H_{a}& =\lambda _{ab}H^{b}+\frac{1}{2R}\lambda _{a}\phi , \\
\delta \phi & =-\frac{1}{2R}\lambda _{a}H^{a}.
\end{align*}%
Thus, in the limit $R\rightarrow \infty $ the fields $H_{a}$ and $\phi $
remain in the action as spin one and spin zero fields, but they decouple in
the transformations and become independent.
When our $SO(1,d)$ gauge invariant gravitational action is taken in three
dimensions, it is natural to ask whether the action obtained is identical to
the Chern-Simons action which was also shown by Achucarro-Townsend \cite{AT}
and Witten \cite{Witten} to be equivalent to the Einstein action in three
dimensions, but with a cosmological constant. In the Chern-Simons
construction one uses only the gauge field $\omega _{\mu }^{AB}$ where the
CS action is
\begin{equation}
I_{\mathrm{CS}}=\frac{1}{2}\dint d^{3}x\epsilon ^{\mu \nu \rho }\epsilon
_{ABCD}\left( \omega _{\mu }^{\,\,AB}\partial _{\nu }\omega _{\rho
}^{\,\,CD}+\frac{2}{3}\omega _{\mu }^{\,\,AB}\omega _{\nu }^{\,\,CE}\omega
_{\rho E}^{\quad \,D}\right) .
\end{equation}%
Using the same decomposition for $\omega _{\mu }^{AB}$ as before, we get
\begin{equation}
\frac{1}{R}\dint d^{3}x\epsilon ^{\mu \nu \rho }\epsilon _{abc}b_{\mu
}^{a}\left( \partial _{\nu }\omega _{\rho }^{\,\,bc}+\omega _{\nu
}^{\,\,be}\omega _{\rho e}^{\quad c}-\frac{1}{12R^{2}}b_{\nu }^{\,b}b_{\rho
}^{\,\,c}\right) ,
\end{equation}%
which is the the first order formulation of the Einstein action plus a
cosmological constant, with the dreibein field $b_{\mu }^{\,a}$. The special
case with the $ISO(1,d-1)$ gauge group can be recovered by rescaling the
action by $R$ and then taking the limit $R\rightarrow \infty .$ In our
treatment, there is also the additional field $e_{A}^{\mu }$ which is not a
gauge field. The field $b_{\mu }^{\,a}$ is given by
\begin{equation*}
b_{\mu }^{\,a}=2e_{\nu }^{a}\nabla _{\mu }c^{\nu },
\end{equation*}%
where $e_{\nu }^{a}$ is the inverse of $e_{a}^{\nu }.$ Our action can be
expressed in terms of $e_{a}^{\nu }$ and a non-propagating field $c^{\mu }.$
Comparing the two formulations, we deduce that the field $b_{\mu }^{\,a}$
must be identified with $e_{\mu }^{a}$. Although $e_{\mu }^{a}$ is not a
gauge field, it can be shown, using the torsion constraint, that its
diffeomorphism transformation with parameters $\zeta ^{\mu }$ can yield the
same gauge transformation as $b_{\mu }^{\,a}$ with the gauge parameter $%
\lambda ^{a}=e_{\mu }^{a}\zeta ^{\mu }$ \cite{Witten}. It then clear that
although both formulations have the same gauge symmetry, they have different
field configurations. Moreover, the usual matter couplings in the CS
formulation are not possible because the dreibein $b_{\mu }^{\,a}$ is a
gauge field. Any direct coupling to matter breaks gauge invariance, except
for coupling to Wilson lines. In our case since $e_{A}^{\mu }$ is not a
gauge field, a gauge invariant metric can be easily formed $g^{\mu \nu
}=e_{A}^{\mu }e^{\nu A}$ and coupled to any form of matter desired.
\begin{acknowledgement}
The work of AHC is supported in part by the Alexander von Humboldt
Foundation and by the National Science Foundation 0854779. V.M. is supported
by TRR 33 \textquotedblleft The Dark Universe\textquotedblright\ and the
Cluster of Excellence EXC 153 \textquotedblleft Origin and Structure of the
Universe\textquotedblright .
\end{acknowledgement}
|
\section{Introduction}
Photometric redshifts have been used extensively in studies of galaxy
evolution, particularly for sources that are too faint for feasible
spectroscopic detection of absorption or emission lines. Photo-z's
are just beginning to find application to measurements of cosmic
structure, particularly in the determination of source redshifts for
weak gravitational lensing surveys \citep[e.g.][]{Combo17,Cosmos}.
Projects in development plan to
measure photometric redshifts for $\sim10^9$ galaxies, far more than
can be practically determined spectroscopically.
Yet the cosmological
measures that weak-lensing surveys have as goals are extremely
sensitive to biases or other miscalibrations of the photo-z scale:
biases as small as $\approx10^{-3}(1+z)$ will overwhelm the
statistical errors on cosmology in these experiments \citep{HTBJ}.
The photo-z technique has never been subject to such demands on its
accuracy, so we need to examine carefully the strategies for achieving
them.
A photo-z algorithm must somehow be taught to convert broadband
fluxes into redshifts. This can be done by fitting to theoretical or
empirical templates for galaxy spectra \citep[e.g.][]{Benitez}; or by completely empirical
machine-learning techniques, {\it e.g.} neural nets \citep{Annz}, that are trained
with a sample of galaxies of known spectroscopic redshift. There is a
difficulty, however, in that many imaging surveys intend to measure
photo-z's from galaxies that are too faint to obtain the spectroscopic
information needed to form templates or obtain spectroscopic training
redshifts. One must therefore take on some degree of faith the
assertion that a photo-z algorithm trained on brighter galaxies will
return sufficiently unbiased redshifts for galaxies beyond the
spectroscopic limit.
We investigate in this paper the sizes of biases that we might expect
to arise because of differences between the faint galaxy population
and the brighter spectroscopic population. In particular:
\begin{itemize}
\item Fainter galaxies will tend to be at higher redshifts, where
active galactic nuclei (AGN) are more frequent or brighter than at
low redshift. Furthermore, optically-based spectroscopy is
inefficient in the ``desert'' between $1.5\lesssim z \lesssim 3$,
where quasar activity peaks. We can ask, therefore, whether the
different AGN characteristics of the fainter galaxies could bias the
photo-z's trained on low-z galaxies.
\item For elliptical galaxies, there is a well-known color-magnitude
relation. As we move to lower luminosity galaxies in this red sequence, stellar populations traverse a path in the
age-metallicity plane. Stellar populations of
the luminous galaxies of different Hubble types that would typically be used for photo-z training will not cover the same locus in the age-metallicity plane.
Would photo-z's trained on bright galaxies produce biased
redshifts for fainter ellipticals? We investigate this by examining the metallicity dependence of photo-z results for old stellar populations when training on the bright Hubble sequence. There are additional complications from the age-metallicity degeneracy. Accordingly, we also estimate the age dependent behavior of photo-z biases.
\end{itemize}
In the Rumsfeld parlance, these are ``known unknowns'' of the faint
galaxy population: a photo-z algorithm might be designed which could
compensate for these two differences between faint and bright
galaxies. But we will use a photo-z algorithm that is ignorant of
these effects, so that they can serve as models for ``unknown
unknowns'' which might bias faint photo-z estimates. In this way we
might learn how safe or dangerous it is to extrapolate training sets
or templates to fainter magnitudes.
\section{Photo-z fitting technique and template sets}
Probing AGN contamination biases necessitates the preparation of an appropriate galaxy spectrum and the use of fitting software to determine the photometric redshift $z_p$. Contaminated galaxies are modeled by introducing a QSO spectrum as an impurity to the spectral energy distributions (SEDs) of galaxies. The {{\it Le~Phare}\/}\footnote{{\url{http://www.oamp.fr/people/arnouts/LE\_PHARE.html}}} software package is used to compute the photometric redshifts. {\it Le~Phare\/} compares theoretical and known source magnitudes, estimating $z_p$ via $\chi^2$ minimization \citep[{\it cf}.][]{steve1,steve2}.
{\it Le~Phare\/} fits to a magnitude catalog of template galaxies. In this case, the catalog consists of 4 known galaxies from \citet{CWW}[CWW] and 3 simulated galaxies from \citet{BC96}[BC96] as well as linear interpolations (via flux) of these. The CWW galaxies are early types, Sbc and Scd spirals, and irregulars, while the BC96 galaxies are $2$, $0.5$, and $.05$\,Gyr in age. Altogether, including interpolations, the catalog is composed of 72 models. We investigate two sets of filters. The first includes 4 Hubble Deep Field North (HDFN) filters (F300W, F450W, F606W, and F814W) and three near-infrared filters (Jbb, H, and K) included in the {\it Le~Phare\/} software package. The second set is the 5 Sloan Digital Sky Survey (SDSS) filter set (u, g, r, i, and z). {\it Le~Phare\/} compiles this catalog by calculating the magnitudes for each galaxy SED with each filter at redshifts from $0$ to $6$ in increments of $.04$. Note that the template galaxies contain no AGN signal.
A second input catalog describes the galaxies for which $z_p$ is to be estimated. We synthesize AGN-contaminated magnitudes by: making a QSO template spectrum; normalizing this template to the galaxy SEDs; contaminating the {\it Le~Phare\/} templates with a specified fraction of AGN flux; then calculating apparent magnitudes at various redshifts and AGN contaminations.
A template QSO spectrum over a wide wavelength span is needed in order to synthesize AGN contaminated SEDs. \citet{VB} and \citet{GHW} provide suitable QSO template spectra; the former covers a broad UV-NIR (800.5\AA - 8554.5\AA) wavelength range while the latter extends further into the NIR (5801\AA - 35095.4\AA). We scale the Glikman spectrum to match the Vanden Berk spectrum at 847~nm, where both are well measured, and then concatenate the two.
The QSO contamination is specified by the fraction $0\le h\le 0.2$ of the flux contributed by the QSO. Specifically, for a chosen galaxy template $g(\lambda)$ and our template QSO spectrum $q(\lambda)$ we first determine the normalization $N$ which yields
\begin{equation}
\int^{\lambda_2}_{\lambda_2} g(\lambda)\,d\lambda = N\int^{\lambda_2}_{\lambda_1} q(\lambda)\,d\lambda,
\end{equation}
where the integration bounds $\lambda_1 =80$~nm to $\lambda_2=3.5\,\mu$m span the range over which both spectra are known.
For a desired contamination $h$, the galaxy rest-frame SED is taken as
\begin{equation}
f(\lambda) = (1 - h)\,g(\lambda) + N\,h\,q(\lambda).
\end{equation}
Apparent magnitudes in the chosen filter sets are calculated using the standard formulae for redshifted sources.
We use the same QSO template spectrum for all levels of contamination. In reality, low-luminosity AGN will have different spectra than bright QSOs, {\it e.g.} weaker broad lines, however the effect of contamination on photo-z's that we derive will still be crudely correct.
The catalog of AGN-contaminated galaxies contains
7 galaxy types---the 4 CWW templates (of type E, Sbc, Scd, and Irr) and the 3 BC96 simulated star forming galaxies (denoted I2, I05, and I005 for their ages)---at redshifts in the range $0 \leq z \leq 3$ in steps of $.1$ for SDSS filters, and in steps of $.1$ from $0 \leq z \leq 1$ and steps of $.2$ for $1 \leq z \leq 3$ for the HDFN filters, with AGN contaminations in the range $0 \leq h \leq 0.2$ in $0.01$ increments.
We run {\it Le~Phare\/} to obtain estimates for $z_p$. The quantity of interest is the redshift bias between the nominal and photometric redshifts, to wit, $\Delta z = z_p - z$.
\section{Bias from AGN contamination}
All galaxy types using either filter set have a non-trivial bias $\Delta z$ induced by AGN contamination. We first examine how $\Delta z$ depends on contamination fraction $h$, at fixed $z$ and galaxy type. Figure~\ref{gbu} illustrates the range of behaviors found. The I2 galaxy shows a simple linear bias-contamination response with a well-defined $d(\Delta z)/dh$. In cases like the Scd galaxy at $z=0.7$, the linear trend gains steps or wiggles, which we believe is induced by the discrete steps in galaxy type that {\it Le~Phare\/} takes when finding a best-fit template. In this case we define $d(\Delta z)/dh$ by a least-squares fit through all the data points.
Finally there are cases like the high redshift E-type in which the best-fit photo-z becomes ``catastrophically'' incorrect above some contamination level. Here the AGN light has moved the galaxy toward some new part of the color manifold of the templates rather than inducing a small differential change. We define
$d(\Delta z)/dh$ by manually selecting a region of $h$ that is below the catastrophe and fitting a slope to this region. We essentially ignore the catastrophic failures because we are more interested in biases induced by gradual evolution in the AGN fraction. We will focus henceforth on the slope
$d(\Delta z)/dh$ in the vicinity of $h=0$.
\begin{figure}[t!]
\epsscale{.323}\plotone{f1a.eps}
\epsscale{.323}\plotone{f1b.eps}
\epsscale{.323}\plotone{f1c.eps}
\caption{\small Plots of $\Delta z$ vs. $h$ with various behaviors.}\label{gbu}
\end{figure}
\begin{figure}[b!]
\epsscale{.406}\plotone{f2a.eps}
\epsscale{.406}\plotone{f2b.eps}
\caption{\small The response $d(\Delta z)/dh$ of photo-z bias to AGN contamination is shown for two filter sets (HDFN on the left and SDSS on the right). The bias response is plotted vs redshift for different galaxy types. For clarity, in the plot we require $|d(\Delta z)/dh| < 2.5$ which cuts out some catastrophic errors.}\label{dzdp}
\end{figure}
Figure~\ref{dzdp} shows $d(\Delta z)/dh$ from the low-contamination domain in both filter sets. The different templates show a variety of behaviors, and some are erratic due to discreteness in the {\it Le~Phare\/} fitting, particularly with the SDSS filter set, which offers poor photo-z constraints in some domains due to lack of NIR coverage.
The important point, however, is that the bias responses $d(\Delta z)/dh$ are $\approx\pm0.5$ over most of the color and galaxy-type domain. {\em A 1\% contamination with AGN light causes a photo-z bias of $\Delta z\sim0.005$} if the photo-z algorithm is using uncontaminated templates. Application of photo-z's to cosmic-shear measurements require that biases be held below $\sim0.001$, which means that the mean AGN contamination of the target sample would need to be determined to $\ll1\%$ in order for the resultant photo-z biases to make insignificant impact on the inferred cosmology. We will not know the faint-galaxy AGN fraction to anywhere near this precision, so this would be a major problem.
Of course we are using a naive photo-z estimator: one could include a QSO template in the fitting process (as in the option {\it Le~Phare\/} offers, or as described in \cite{WLF}, which includes object classification), or attempt to construct an empirical training set of bright galaxies that spans the range of AGN contamination of the faint set. This would greatly reduce the photo-z bias in our simple simulated catalog. The behavior of real AGN is, however, much more complex than our simple single QSO template, and it is likely that the faint population has different AGN spectra than the bright galaxies that might serve as templates. We simply point out that photo-z biases can be very sensitive to AGN light, and it is risky to extrapolate AGN characteristics to a new population.
\section{Age-Metallicity Models}
We use stellar population models to estimate the redshift bias induced by a metallicity mismatch between target galaxies and {\it Le~Phare}'s\/ templates ({\it Le~Phare\/} uses the same extended CWW catalog used above). The galaxy evolution simulation code GALAXEV\footnote{\url{http://www2.iap.fr/users/charlot/bc2003/}} is used here to generate a magnitude catalog for galaxies with near-solar metallicity with a $z=0$ ages of 10\,Gyr. We use the standard model as described in \citet{BC03}[BC03]. Specifically, the galaxy models employed here use the STELIB and BaSeL spectral libraries, a simple stellar population (SSP) star formation rate, the Chabrier initial mass function, and the 1994 Padova evolutionary tracks (per recommendation of the code's authors).
GALAXEV produces a table of magnitudes and ages at various redshifts for a given galaxy model and filter set. The same SDSS and infrared supplemented HDFN filters from the AGN contamination study are used to calculate these magnitudes. GALAXEV produces colors for metallicities $Z/Z_{\odot}=0.2, 0.4, 1.0,$ and 2.5 (models m42, m52, m62, and m72 in GALAXEV). We also produce colors for metallicities 0.3, 0.7, and $1.75Z_{\odot}$ by interpolating between these four GALAXEV spectra.
We further employ GALAXEV to investigate biases due to age misestimates. We generate solar metallicity models with z = 0 ages of 11, 9, and 8 Gyr. Note that the fiducial model here is still the 10\,Gyr old solar metallicity galaxy, but in contrast to what is done in the metallicity study, we now hold the metallicity fixed and vary the age.
\section{Bias From Metallicity Indeterminacy}
Calculating the bias due to a small metallicity mismatch requires removing the fiducial bias present resulting from inherent differences between the GALAXEV spectra and {\it Le~Phare}'s\/ catalog of templates. At each redshift, we take an elliptical galaxy with solar metallicity as our benchmark. The solar bias, $\delta z_{\odot}$, is defined as the difference between the solar model's redshift and {\it Le~Phare}'s\/ estimate:~$\delta z_{\odot}=z_{\odot,\mbox{\tiny mod}}-z_{\odot,\mbox{\tiny phot}}$. For any particular metallicity, the bias due to metallicity mismatch is then $\Delta z = z_{\rm phot} - z_{\rm model} - \delta z_\odot$. Hence by definition we assign zero metallicity bias to a solar-metallicity population, for any chosen filter set, population age, and redshift.
Figure~\ref{mzb} illustrates the biases from metallicity offsets as a function of redshift between $z=0$ and $z=.7$. Five models are present in the plot, each is 10\,Gyr old at $z=0$ with metallicities 0.2--2.5$Z_\odot$. As the models get more metal rich, the redshift bias increases, that is, the photometric redshift gets larger than the model's redshift. In the redshift range that was studied, the super- and sub-solar models both had biases of up to $\pm .2$ for metallicities $\approx\pm0.5$~dex away from solar. This suggests that biases of $\approx0.001(1+z)$ would result from shifts of only $\sim0.003$~dex in the mean metallicity of an old population away from the metallicity of the training set. For comparison, current modelling of age and metallicity variation along the red sequence of SDSS galaxies suggests a $\sim0.3$~dex change in metallicity across 4 magnitudes of galaxy luminosity \citep{MB}. Therefore the biases in photo-z's of faint red-sequence galaxies can be highly significant if the photo-z's are trained exclusively on brighter red-sequence members.
\begin{figure}[!hb]
\epsscale{.406}\plotone{f3a.eps}
\epsscale{.406}\plotone{f3b.eps}
\caption{\small Metallicity-induced biases are shown from rest-frame to z=.7 for HDFN (left) and SDSS (right) filters. Each model's metallicity is shown on the plot in solar units. The dashed and dotted lines serve distinguish between the models in the crowded regions. The smoother behavior of the HDFN+IR plot is attributable to the supplementary IR bands---the HDFN data acquire jagged profiles when the analysis is done sans the IR bands. Indeed, these plots (not shown) resemble the SDSS plots; while the overall range of the bias decreases, the curves behave more erratically.}\label{mzb}
\end{figure}
\section{Bias from Age Indeterminacy}
Biases from age estimates errors are computed in the same manner as in the metallicity case. The intrinsic bias from the benchmark 10\,Gyr solar metallicity galaxy, $\delta z_{\sub{10\,Gyr}}$, must be subtracted from the total, $\delta z_{\sub{mod}}$, to arrive at the bias attributable to age mismatching: $\Delta z_{\sub{age}} = z_{\sub{mod}} - z_{\sub{p}} - \delta z_{\sub{10\,Gyr}} = \delta z_{\sub{mod}} - \delta z_{\sub{10\,Gyr}}$.
\begin{figure}[ht!]
\epsscale{.406}\plotone{f4a.eps}
\epsscale{.406}\plotone{f4b.eps}
\caption{\small Shown above are the age-induced biases for solar metallicity models with $0 \leq z \leq .7$ for the HDFN (left) and SDSS (right) filter sets. Model ages are distinguished by color. As with the metallicity-induced bias, a reanalysis of the HDFN data with NIR data excluded yields plots comparable to the SDSS.}\label{azb}
\end{figure}
Figure~\ref{azb} portrays the induced biases in both filter sets. Compared to the metallicity-induced biases, these show little structure outside of the feature near $z=.4$ for the SDSS filters (also found in the metallicity plots), perhaps suggesting that the photo-z's are more sensitive to metallicity offsets. For the HDFN filter set, the bias only becomes comparable at higher $z$ when the age offset is 2\,Gyr. In both studies, the HDFN sets exhibit a slightly wider range in biases which may mean that extra bands may be more confusing for {\it Le~Phare\/} if the templates and data are not initially well matched.
\section{Conclusion}
Photometric redshifts resemble the real redshifts only as closely as the training sets mimic the objects of interest. In the cases of AGN activity and metallicity variation, we find that even very small mismatches between the mean photometric target and the training set can induce photo-z biases large enough to corrupt significantly the validity of derived cosmological parameters. In particular, we find that a metallicity shift of $\sim0.003$~dex in an old population, or contamination of any galaxy spectrum with $\sim 0.2\%$ AGN flux, is sufficient to induce a $10^{-3}$ bias in photo-z. Miscalculating the age of a galaxy by 2\,Gyr may also result in similar a bias. In a real survey, we can expect differences far larger than this between the faint photometry survey targets and the bright spectroscopic targets used for photo-z training.
While simple models are used here to study redshift biases, the results indicate a need to develop training sets that encompass the full range of behavior of the photo-z target population. Our study uses a worst-case situation in that the photo-z algorithm is given no information on the physical effects underlying the bias (AGN or a luminosity-metallicity correlation). In reality the training sets can include galaxies with AGN of varying brightness and some galaxies of lower luminosity, which would ameliorate the biases we have found. However the extreme sensitivity of the photo-z to these effects suggests that it will be dangerous to extrapolate from the behavior of bright training-set galaxies to a target photo-z sample that extends to significantly higher redshift and lower luminosity, since even subtle differences in spectral behavior can lead to important biases. Furthermore we have investigated two known differences between the faint and bright populations, but there may be differences that have not yet been discovered because there are no detailed spectroscopic surveys of the galaxies fainter than 24th magnitude that will comprise the bulk of future weak-lensing surveys. These results highlight the desirability of training photo-z's with complete spectroscopic surveys to $\sim$25 mag, so that the training set is representative of the galaxies for which photo-z's are being obtained.
\begin{acknowledgments}
This work is supported by grant AST-0607667 from the NSF and DOE grant
DE-FG02-95ER40893. We are very grateful to Mariangela Bernardi and Gordon Richards for their advice and assistance. We also thank Stephanie Jouvel for instruction in the use of {\it Le~Phare}\/.
\end{acknowledgments}
|
\section{Introduction}
Soon after the observation of the Hall effect, Hall also observed
the anomalous Hall effect (AHE)~\cite{Hall1880_a,Hall1880_b} in
ferromagnetic metals, where the Hall resistance arises from the
spin-orbit coupling (SOC) between electric current and magnetic
moments, even in the absence of external magnetic field. Recent
progress on the mechanism of AHE have established a link between the
AHE and the topological nature of Hall current due to SOC by
adopting the Berry-phase
concepts~\cite{Nagaosa-AHE,Jungwirth,Fang2003_a,Fang2003_b}, in
close analogy to the intrinsic spin Hall effect~\cite{SHE1,SHE2}.
Given the experimental discovery of the quantum Hall~\cite{Klitzing}
and quantum spin Hall (QSH) effects~\cite{Bernevig2006, Konig2007},
it is natural to ask whether the AHE can also be quantized in a
topological magnetic insulator, without the external magnetic field
and the associated Landau levels.
A simple mechanism for the QAH insulator has been proposed in a
two-band model of a two dimensional (2D) magnetic
insulator~\cite{Qi2006}. In the limit of vanishing spin-orbit
coupling, and large enough exchange splitting, the majority spin
band is completely filled and minority spin band is empty. When the
exchange splitting is reduced, the two bands intersect each other,
leading to a band inversion, say near the $\Gamma$ point. The
degeneracy at the interaction region can be removed by turning on
the SOC, giving rise to an insulator state with a topologically
non-trivial band structure characterized by a finite Chern number
and chiral edge states characteristic of the QAH state (see figures
(1) and (2) of Ref.~\cite{Qi2006}). In alternative mechanisms of
realizing the quantum Hall state without an external magnetic field,
bond currents on a honeycomb lattice has been
proposed~\cite{Haldane}, and the localization of the band electron
could also give rise to a quantized plateau~\cite{QAHE-Nagaosa}.
However, these mechanisms seem harder to realize experimentally.
Therefore, the crucial ingredients for realizing an QAH state are: (1)
a ferromagnetically ordered 2D insulator which breaks the
time-reversal symmetry; (2) a band inversion transition with strong
SOC. In comparison with the QSH state, which satisfies the second
condition~\cite{Bernevig2006, Konig2007}, the first condition of the
ferromagnetic order in an insulating state is harder to realize.
However, the QSH insulators are usually good starting candidates for
finding QAH effect due to its proximity to the band inversion. Given
the QSH effect in HgTe quantum wells~\cite{Bernevig2006, Konig2007}
through a band inversion transition by varying the thickness of the
quantum well, the magnetically doped HgMnTe has been proposed as a
candidate for the QAH insulator~\cite{CXLiu2008}. Unfortunately, the
Mn moments do not order spontaneously in HgMnTe, and an additional,
small Zeeman field is required for the magnetic alignment of the Mn
moments.
Recently, tetradymite semiconductors Bi$_2$Te$_3$, Bi$_2$Se$_3$, and
Sb$_2$Te$_3$ have been theoretically predicted and experimentally
observed to be topological insulators with the bulk band gap as
large as 0.3eV in Bi$_2$Se$_3$~\cite{HJZhang2009, Xia2009,
Chen2009}. The electronic states close to the Fermi level are the
bonding and anti-bonding $p$ orbitals, and these states undergo a
band inversion at the $\Gamma$ point when the strength of the
spin-orbit coupling is increased~\cite{HJZhang2009}. When the
thickness is reduced, the three dimensional (3D) topological
insulator crosses over to a 2D topological insulator in an
oscillatory fashion, as a function of the thickness~\cite{2Dcross}.
In this work, we predict that this family of compounds doped with
proper transition metal elements (Cr or Fe) should give the QAH
state, the long sought-after quantum Hall state without any external
magnetic field. In sharp contrast to the ferromagnetic order in
conventional dilute-magnetic-semiconductors (DMS), where free
carriers are necessary to mediate the ferromagnetic couplings via
the RKKY mechanism~\cite{DMS,Zener,Dietl,Ohno}, the magnetic dopants in tetradymite
compounds naturally order ferromagnetically, mediated by the large
spin susceptibility (of the van Vleck type) of the host insulator
state. Due to the proximity to the band inversion transition, the
exchange splitting leads to a topologically non-trivial insulating
band structure with a finite Chern number. The recent experimental
progresses on this family of compounds have demonstrated that
well-controlled layer-by-layer MBE thin film growth can be
achieved~\cite{Kehui,Xue1,Xue2,Xue3}, and various transition metal
elements (such as Ti, V, Cr, Fe) can be substituted into the parent
compounds with observable ferromagnetism even above
100K~\cite{FM-BiSe_a,FM-BiSe_b,FM-BiSe_c}. By quantitative
first-principles calculations, we further show that such
ferromagnetic QAH insulating state can be also reached with Tc as
high as around 70K. We will start from simple physical pictures
based on model analysis, and then conclude by quantitative
first-principles calculations.
\section{The Van-Vleck paramagnetism in Bi$_2$Se$_3$ family compounds}
We first discuss minimal requirements for the ferromagnetic insulator
phase in a semiconductor system doped with dilute magnetic ions under the
assumption that the magnetic exchange among local moments is mediated by
the band electrons. The
whole system can then be divided into two sub-systems describing the
local moments and band electrons respectively, which are coupled by a
magnetic exchange term. Therefore if we only consider the spatially
homogeneous phase, the total free energy of the system can thus be
written as,
\[
F_{total}=\frac{1}{2}\chi_{L}^{-1}M_{L}^{2}+\frac{1}{2}
\chi_{e}^{-1}M_{e}^{2}-J_{ex}M_{L}M_{e}-\left(M_{L}+M_{e}\right)H\]
where $\chi_{L/e}$ is the spin susceptibility of the local
moments/electrons, $M_{L/e}$ denotes the magnetization for the local
moment and electron sub-system and $J_{ex}$ is the magnetic exchange
coupling between the local moments and electrons. The onset of the
ferromagnetic phase can be determined when the minimization
procedure of the above free energy gives a non-zero magnetization
without the external magnetic field $H$, which leads to
$J_{ex}^{2}-\chi_{L}^{-1}\chi_{e}^{-1}>0$, or equivalently
$\chi_{L}>\frac{1}{J_{ex}^{2}\chi_{e}}$. In the dilute limit, we can
neglect the direct coupling among the local moments, and then the
spin susceptibility of the local moment sub-system takes the
Curie-Weiss form, $\chi_{L}=x\mu^2_{J}/(3k_BT)$, where $x$ is the
concentration of the magnetic ions and $\mu_J$ is the magnetic
moment of a single magnetic ion. Therefore, in order to have non-zero
FM transition temperature, the above criterion requires a sizable
spin susceptibility of the electron sub-system $\chi_e$. In most of
the DMS systems, e.g. (Ga$_{1-x}$Mn$_x$)As, the electronic spin
susceptibility is negligible for the insulator phase and finite
carrier concentration is required to conduct the magnetic coupling
among local moments. As we shall discuss below, unlike the situation
in GaAs, sizable spin susceptibility even exists through the Van
Vleck paramagnetism for the insulator Bi$_2$Se$_3$, providing a new
mechanism for ordering the doped magnetic ions.
The non-zero paramagnetic spin susceptibility for a band insulator has
been known as the Van-Vleck paramagnetism\cite{Van_Vleck} for a long
time. For temperature much less than the band gap, the spin
susceptibility for a band insulator can be obtained by the second
order perturbation on the ground state, which can be written as,
\begin{equation}
\chi^{zz}_{e}=\sum_{E_{nk}<\mu;E_{mk}>\mu}4\mu_0\mu^2_B\frac{\left\langle
nk\right|\hat{S_{z}}\left|mk\right\rangle \left\langle
mk\right|\hat{S_{z}}\left|nk\right\rangle }{E_{mk}-E_{nk}}.
\end{equation}
where $\mu_0$ is the vacuum permeability and $\mu_B$ is the Bohr magneton.
It is then quite obvious that the van-Vleck paramagnetism is caused
by the mixing of the conduction and valence bands induced by the
spin operator in the presence of SOC. Such a mechanism is absent in
the GaAs system, because the gap in GaAs is between mostly $s$ and
$p$ bands and the matrix element of $\hat{S_{z}}$ is nearly zero
between conduction and valence bands. In the case of Bi$_2$Se$_3$
family, the situation is completely different: both the conduction
and valence bands are formed by the bonding and anti-bonding $p$
orbitals and the energy gap is opened by the spin-orbital coupling
with band inversion, therefore the matrix element of the spin
operator can be very large.
The Van-Vleck paramagnetism in Bi$_2$Se$_3$ family can be well
understood using the following effective $k\cdot p$ Hamiltonian
derived in ref.~\cite{HJZhang2009},
\begin{equation}
H_{3D}\left(\mathbf{k}\right)=\left[\begin{array}{cccc}
\mathit{\mathcal{M}}(\mathbf{k}) & A_{xy}k_{-} & 0 &
A_{z}k_{z}\\ A_{xy}k_{+} & \mathit{-\mathcal{M}}(\mathbf{k}) &
A_{z}k_{z} & 0\\ 0 & A_{z}k_{z} & \mathit{\mathcal{M}}(\mathbf{k})
& -A_{xy}k_{+}\\ A_{z}k_{z} & 0 & -A_{xy}k_{-} &
-\mathit{\mathcal{M}}(\mathbf{k})\end{array}\right]
+\epsilon_{0}\left(\mathbf{k}\right)
\label{BiSe-4band}
\end{equation}
with $k_{\pm}=k_{x}\pm ik_{y}$,
$\epsilon_{0}\left(\mathbf{k}\right)=C+D_{xy}
\left(k_{x}^{2}+k_{y}^{2}\right)+D_{z}k_{z}^{2}$,
$\mathit{\mathcal{M}}(\mathbf{k})=M_{0}+B_{xy}
\left(k_{x}^{2}+k_{y}^{2}\right)+B_{z}k_{z}^{2}$~\cite{Param}, in
the basis of $\left|P_{z}^{+},\uparrow\right\rangle$,
$\left|P_{z}^{-},\downarrow\right\rangle$,
$\left|P_{z}^{+},\downarrow\right\rangle$,
$\left|P_{z}^{-},\uparrow\right\rangle$. The $\pm$ in the basis
denote the even and odd parity states and $\downarrow\uparrow$
indicate the spin. As discussed before in Ref. ~\cite{HJZhang2009},
the four low energy states originate from the bonding and
anti-bonding combinations of Bi and Se $P_z$ orbitals, and the
parity is a good quantum number at the $\Gamma$ point due to the
spatial inversion symmetry. The topological properties of the
Bi$_2$Se$_3$ family can be well explained by the band inversion
between two sets of Kramer's degenerate states with opposite parity
at the $\Gamma$ point. Here we will show that the band inversion not
only gives the non-trivial topological index but also increases the
spin susceptibility dramatically. For simplicity, we focus on a 2D
sheet in the k-space with $k_{z}=0$ (the situation of general $k_z$
is qualitatively similar), where the above effective Hamiltonian
reduces to two diagonal blocks as\cite{Bernevig2006,2Dcross}
\begin{equation}
H_{2D}\left(\mathbf{k}\right)=\left[\begin{array}{cc}
h\left(k\right)
& 0\\ 0 & h^{*}\left(-k\right)\end{array}\right]
\label{H2D}
\end{equation}
with
$h\left(k\right)=\mathit{\mathcal{M}}(\mathbf{k})\sigma_{z}+A_{xy}\left(k_{x}\sigma_{x}+k_{y}\sigma_{y}\right)+\epsilon_{0}\left(\mathbf{k}\right)$.
The two eigen-states of $h\left(k\right)$ give the occupied and
unoccupied bands respectively, which can be denoted as
$\left|ck\right\rangle $and $\left|vk\right\rangle $. It is then
obvious that the amplitude of the matrix element $\left\langle
ck\right|\hat{S_{z}}\left|vk\right\rangle $ reaches the maximum
value when $\mathit{\mathcal{M}}(\mathbf{k})=0$. The above
condition can only be satisfied when $M_{0}B_{xy}<0$, which is
exactly the condition for the band inversion. In the normal phase
without inverted bands, the maximum value of $\left\langle
ck\right|\hat{S_{z}}\left|vk\right\rangle $ can not be reached and
the spin susceptibility is less pronounced compared to the system
with band inversion.
To further demonstrate this argument, we perform first-principles
calculations (see ref.~\cite{Method} for method) for the electronic
structures and the spin susceptibility of Bi$_2$Se$_3$ bulk. The
results as function of SOC strength are plotted in Fig. 1(a). When the
relative SOC strength $\lambda/\lambda_0$ is larger than 0.5, the
bands at the $\Gamma$ point are inverted, and the spin susceptibility
starts to increase appreciably around this point.
\section{Mean Field Theory for the Magnetic Topological Insulators}
In the following, we apply a tight binding version of the Zener
model\cite{DMS,Zener,Dietl} to calculate the possible Curie temperature at the mean field
level. Let's first assume that some magnetic local moments are
introduced iso-electronically into the Bi$_2$Se$_3$ compounds
without changing the insulating behavior. We show that ferromagnetic
ordering with $T_c$ as high as 70K can be achieved via the van-Vleck
paramagnetism, without requiring the existence of free carriers as
in conventional DMS systems. The total Hamiltonian of our system can
be written as
\begin{equation}
H_{total}=H_{LDA}+\sum_{I\mu\nu,\alpha=xyz}J_{eff}
\mathbf{S}_{I}^{\alpha}\cdot
\mathbf{s}_{\mu\nu}^{\alpha}c_{I,\mu}^{\dagger}c_{I,\nu}+h.c.\label{Ht}
\end{equation}
with
\[H_{LDA}=\sum_{ij,\mu\nu}t_{ij}^{\mu\nu}c_{i,\mu}^{\dagger}c_{j,\nu}\]
is the single particle Hamiltonian (obtained from first-principles
calculations) expressed in the basis of local Wannier
functions~\cite{Method}. Here $i$ and $j$ denote the sites, $I$ only
runs over all the unit cells containing magnetic ions and $\mu\nu$
are the orbital/spin indices. The coupling strength between the
local spin $\mathbf{S}_I$ and the band spin $\mathbf{s}_{\mu\nu}$ of
band electrons is described by the effective exchange parameter
$J_{eff}$.
Performing the mean field approximation and applying the virtual
crystal approximation (CPA) to average out the spatial inhomogeneity,
Eq.(\ref{Ht}) can be written as
\begin{equation}
H_{MF}=H_{e}+H_{L}-xJ_{eff}\mathbf{M}_{L}^{\alpha}
\cdot\mathbf{M}_{e}^{\alpha}
\end{equation}
where $\mathbf{M}_{L}^{\alpha}=\left\langle
\mathbf{S}_{I=0}^{\alpha}\right\rangle $,
$\mathbf{M}_{e}^{\alpha}=\left\langle
\mathbf{s}_{\mu\nu}^{\alpha}c_{0,\mu}^{\dagger}c_{0,\nu}\right\rangle
$ are the mean field magnetization for the local moments and band
electrons respectively, with
\[H_{e}=H_{LDA}+\sum_{i\mu\nu}^{\alpha=xyz}J_{eff}\mathbf{M}_{L}^{\alpha}\cdot
s_{\mu\nu}^{\alpha}c_{i,\mu}^{\dagger}c_{i,\nu}+h.c.\]
and
\[H_{L}=\sum_{I}^{\alpha=xyz}J_{eff}\mathbf{S}_{I}^{\alpha}\cdot
\mathbf{M}_{e}^{\alpha}\]
We choose the concentration of the magnetic ion to be $5\%$, the total
angular momentum of a single magnetic ion to be $J=3/2$ and the
effective exchange coupling $J_{eff}=2.0eV$, which are in the same
order with the quantitative first-principles calculations (as
described later). The above mean field equations have been solved
self-consistently, and the ordered moment as well as Curie temperature
can be calculated. In Fig.\ref{fig: bise_spin_sus_and_TC}(b), the
calculated magnetization versus temperature are plotted for both bulk
material and thin film system with 3, 4, 5 quintuple layers (QL)
thicknesses. The Curie temperature for the bulk system is around 70K,
which can be well compared with previous experimental
studies~\cite{FM-BiSe_a,FM-BiSe_b,FM-BiSe_c}. The ferromagnetic
ordering is anisotropic, and the Curie transition is of the Ising
type, with a finite transition temperature in 2D. Our mean field
calculation also indicates that for the thin film system, the Curie
temperature is only slightly reduced due to the finite size effect,
which give rise to the possible QAHE as discussed below.
\section{Electronic structures of Bi$_2$Se$_3$ doped with
transition metal elements}
In this section, we shall show, by parameter-free first principles
calculations, that the insulating magnetic ground state discussed
above can be indeed obtained by proper choice of magnetic dopants. It
has been suggested by experiments~\cite{FM-BiSe_a,FM-BiSe_b,FM-BiSe_c}
that the magnetic dopants, such as Ti, V, Cr and Fe, will mostly
substitute the Bi ions (or the Sb ions in the case of Sb$_2$Te$_3$),
we therefore concentrate on this situation in the following
discussions. Since the nominal valence of Bi (or Sb) ions are 3+, a
general rule is to find a transition metal element which may have a
stable 3+ chemical state, so that no free carriers are introduced by
this iso-electronic substitution. In addition, due to the coexistence
of orbital and spin degrees of freedom of magnetic dopants (due to the
partially filled $d$-shells), we need a mechanism to quench these
degrees of freedom and stabilize the insulating state. We shall show
that such conditions can be satisfied by the combination of crystal
field splitting and large Hund's rule coupling for Cr and Fe dopants.
We performed self-consistent first-principle
calculations~\cite{Method} for Bi$_2$Se$_3$ doped with various
transition metal elements, Ti, V, Cr and Fe. We use a large supercell
(of size 3$\times$3$\times$1) with one of the Bi atoms substituted by
a magnetic dopant. The crystal structure and the magnetic state are
fully optimized to obtain the ground state. The results shown in Fig.2
suggest that a insulating magnetic state is obtained for Cr or Fe
doping, while the states are metallic for Ti or V doping cases. To
understand the results, we first point out that, for all the cases,
the dopants are nearly in the 3+ valence state, and we always obtain
the high-spin state due to the large Hund's rule coupling of $3d$
transition metal ions. Thus the charge and spin degrees of freedom
are nearly frozen. This will directly lead to the insulating state of
Fe-doped samples, because the Fe$^{3+}$ has five $3d$ electrons,
favoring the $d^{5\uparrow}d^{0\downarrow}$ configuration in a
high-spin state, resulting in a gap between the majority and minority
spins. We now consider the orbitals for other dopants. The local
environment of dopants, which substitute the Bi sites, is a octahedral
formed by six nearest neighboring Se$^{2-}$ ions. Such a local crystal
field splits the $d$-shell into $t_{2g}$ and $e_g$ manifolds. This
splitting is large enough to stabilize the
$t_{2g}^{3\uparrow}e_{g}^{0\uparrow}t_{2g}^{0\downarrow}e_{g}^{0\downarrow}$
configuration of Cr$^{3+}$ ion, resulting in a gap between the
$t_{2g}$ and $e_g$ manifolds. For the case of Ti or V doping, even the
$t_{2g}$ manifold is partially occupied, leading to the metallic
state. It is important to note that, although the local density
approximation in the density functional theory may underestimate the
electron correlation effects, the inclusion of electron-electron
interaction $U$ (such as in LDA+$U$ method) should further enhanced
the gap.
The energy gain due to the spin polarization is about 0.9eV/Fe,
1.5eV/Cr, 0.7eV/V and 0.02eV/Ti, respectively, which are very large
numbers except for the case of Ti substitution. From the spin
splitting of $p$ orbitals of the band electrons, we can estimate the
effective exchange coupling $J_{eff}$ between the local moments and
the $p$ electrons~\cite{J_cal}. The estimated value is around 2.7eV
for Cr and 2.8eV for Fe in Bi$_2$Se$_3$. This exchange coupling is
comparable to that in GaMnAs~\cite{DMS,J_cal}.
\section{The quantized anomalous Hall effect}
Once the ferromagnetically ordered insulating state is achieved with
a reasonable $T_c$, here in this section, we show that QAH effect
can be realized for the 2D thin films. In the thick slab geometry,
the spatially-separated two pairs of surface states are well defined
for the top and bottom surfaces respectively. However, with the
reduction of the film thickness, quantum tunneling between the top
and bottom surfaces becomes more and more pronounced, giving rise to
finite mass term in the effective 2D model\cite{2Dcross}. In the
basis of $\left|+\uparrow\right\rangle $
,$\left|-\downarrow\right\rangle $ ,$\left|+\downarrow\right\rangle
$ ,$\left|-\uparrow\right\rangle $ , with $\pm$ referring to the
bonding and anti-bonding combinations of the top and bottom surface
states, and $\downarrow\uparrow$ indicating the spin, the low energy
effective model for the thin film can be written in the form of
$H_{2D}$ as in Eq. \ref{H2D}, with
$h(k)=\Delta_{k}\sigma_{z}+v_{F}\left(k_{x}\sigma_{y}-k_{y}\sigma_{x}\right)$,
$\Delta_{k}=\Delta_{0}+Bk_{\parallel}^{2}$, similar to the
Bernevig-Hughes-Zhang model describing the low energy physics in
HgTe/CdTe quantum well\cite{Bernevig2006}. When $\Delta_{0}B<0$,
band inversion occurs, and the system will be in the QSH phase if
this is the only band inversion between two subbands with opposite
parity, which is exactly the situation in HgTe/CdTe quantum well
system. The situation for Bi$_2$Se$_3$ film is somewhat different in
the sense that more pairs of subbands can get inverted at the
$\Gamma$ point, and the system may still remain in the topologically
trivial state if the band inversion occurs even number of
times~\cite{2Dcross}. Nevertheless, as we shall show below,
regardless of whether the 4-bands system is originally in the
topologically non-trivial phase or not, a strong enough exchange
field will induce QAH effect in this system.
The exchange coupling term induced by the finite magnetization in
the FM phase can be written as,
\begin{equation}
H_{exchange}=\left[\begin{array}{cc} g_{eff}M\sigma_{z} & 0\\ 0 &
-g_{eff}M\sigma_{z}\end{array}\right]\label{exchange}
\end{equation}
with $M$ denoting the strength of the exchange field and $g_{eff}$
being the effective g-factor of the surface states. Added to the 2D
model (\ref{H2D}), the exchange field increases the mass term of the
upper block and reduces it for the lower block, breaking the time
reversal symmetry. More importantly, a sufficiently large exchange
field can change the Chern number of one of the two blocks. As
schematically illustrated in Fig.3, if the 4-band system is originally
in the topologically trivial phase, the exchange field will induce a
band inversion in the upper block and push the two sub-bands in the
lower block even farther away from each other. Therefore the 2D model
with a negative mass in the upper block contributes $e^{2}/h$ for the
Hall conductance. On the other hand, if the system is originally in
the topologically non-trivial phase, both blocks have inverted band
structures. In this case, a sufficiently large exchange field can
increase the band inversion in the upper block and revert the band
inversion in the lower block. Again the negative mass in the upper
block contributes $e^{2}/h$ for the Hall conductance. These two
scenarios are illustrated in Fig.\ref{fig:chemata}.
To be realistic, we carried out quantitative calculations for
Bi$_2$Se$_3$ films, based on the density functional theory and
linear response formalism~\cite{Method}. A spacial uniform exchange
field is included to take into account the effect of magnetization
in the FM state at the mean field level. In Fig.\ref{fig:
bise_m_gap}(a), (b) and (c), we plot the lowest four sub-band levels
at $\Gamma$ point as a function of exchange field. The level
crossings (indicated by arrows) between the lowest conduction bands
(blue lines) and valence bands (red lines) are found in all the
three Bi$_2$Se$_3$ thin film systems with different layer thickness.
The corresponding Hall conductance can be obtained using the
following Kubo formula\cite{Nagaosa-AHE},
\begin{equation}
\sigma_{xy}={e^{2}}{\hbar}\sum_{nmk}\frac{Im\left[\left\langle
nk\right|\hat{v}_{x}\left|mk\right\rangle \left\langle
mk\right|\hat{v}_{y}\left|nk\right\rangle
\right]}{\left(E_{nk}-E_{mk}\right)^{2}}
\left(n_{f}(E_{nk})-n_{f}(E_{mk})\right)
\label{kubo_sxy}
\end{equation}
with $E_{nk}$ being the dispersion of the quantum well sub-bands and
$\hat{v}_{x/y}$ being the velocity operators. In the insulator case,
where the chemical potential is located inside the energy gap between
conduction and valence sub-bands, the Hall conductance is determined
by the first Chern number of the occupied bands and must be an exact
integer in the unit of $e^2/h$. The calculated Hall conductance for
Bi$_2$Se$_3$ thin film with three typical thickness are plotted in
figure \ref{fig: bise_m_gap} (d) as function of exchange field, where
a jump from "0" to "1" in the Hall conductance are observed at the
corresponding critical exchange field for the level crossing.
In Fig.\ref{fig:bise_hall_Platform}, we plot the sub-band dispersion
and corresponding Hall conductance as function of chemical potential
for 3 and 5 QL Bi$_{2}$Se$_3$ thin films with different exchange
fields. We can find that the change induced by the exchange field is
mainly in the area near $\Gamma$ point. For the QAH phases ((d) and
(h)), non-zero integer plateau is observed when the chemical
potential falls into the energy gap. The difference between the 3
and 5 QL cases suggests that the QAH plateau is much narrower for
thick film due to the narrower energy gap coming from the
intersection of top and bottom surface states.
In conclusion we have shown that magnetic dopants Cr and Fe form long
range magnetic order in the insulating state of the Bi$_2$Se$_3$
family of topological insulators. The ferromagnetic order lead to a
topologically non-trivial electronic structure with quantized Hall
conductance without any external magnetic field. In real samples, a
small amount of bulk carriers could always be present, however, if the
concentration is low enough, they will be localized by disorder in two
dimensions, and will not affect the precise quantization of the Hall
plateau. The edge states of our QAH state conducts charge current
without any dissipation in the absence of any external magnetic
field. This effect could enable a new generation of low power
electronic devices.
\begin{acknowledgments}
We acknowledge the valuable discussions with C. X. Liu,
X. L. Qi, Q. Niu, S. Q. Shen, and the supports from the NSF of
China, the 973 Program of China (No. 2007CB925000 and 2010CB923000),
and the International Science and Technology Cooperation Program of
China. SCZ is supported by the US NSF under grant numbers
DMR-0904264.
\end{acknowledgments}
|
\section{Introduction}
\noindent
One of the major open questions in modern cosmology is the origin of the matter-antimatter asymmetry in our Universe. The bayon asymmetry is nowadays known to be
$\eta_B\equiv (n_B-n_{\bar{B}})/n_\gamma =6.14 \times 10^{-10}(1.00 \pm 0.04)$
thanks to the measurement of the CMB anisotropies by WMAP \cite{Spergel:2003cb}. Sphalerons effects in baryogenesis and leptogenesis scenarios \cite{Buchmuller:2005eh} can equilibrate cosmic lepton and baryon asymmetries at the same level. Since the lepton asymmetry is only possible in the neutrino sector because of charge conservation, the observation of a non-zero neutrino degeneracy parameter $\xi$ can furnish important information to our understanding of the matter-antimatter asymmetry in the Universe.
In analogy with $\eta_B$ related to the baryon asymmetry, the total neutrino asymmetry $L_\nu=L_{\nu_e}+L_{\nu_\mu}+L_{\nu_\tau}$ can be quantified by the neutrino chemical potentials $\mu_{\nu_\alpha}$ ($\alpha \equiv e, \mu, \tau$) or, equivalently, the degeneracy parameters $\xi_{\nu_\alpha} \equiv \mu_{\nu_\alpha}/T_\nu$:
\begin{equation}\label{e:nuasym}
L_{\nu_{\alpha}} = \frac{n_{\nu_{\alpha}}-n_{\overline{\nu}_{\alpha}}}{n_{\gamma}}=\frac{\pi^2}{12 \zeta (3)} \left(\frac{T_{\nu_\alpha}}{T_\gamma} \right)^3 \left(\xi_{\nu_{\alpha}}+\frac{\xi^3_{\nu_{\alpha}}}{\pi^2}\right)
\end{equation}
where $n_{\nu_{\alpha}}$ ($n_{\overline{\nu}_{\alpha}}$) are the neutrinos (anti-neutrinos) occupation numbers and $\zeta (3) \simeq 1.202$.
Non-zero electron, muon and tau neutrino degeneracy parameters influence the abundance of light elements produced in
Big-Bang Nucleosynthesis (BBN) in two aspects. While all flavours influence the expansion rate of the Universe,
by modifying the
effective number of degrees of freedom, only $\xi_{\nu{e}}$ impacts the neutron/proton ratio, a key parameter for the $^{4}$He abundance. Indeed $^{4}$He, among all the light elements formed during BBN, is the most sensitive one to the neutrino
degeneracy parameters. Extensive work has been performed to extract information on the relic lepton asymmetries either
from Big-Bang Nucleosynthesis, as in e.g.
\cite{Wagoner:1966pv,Kang:1991xa,Esposito:2000hh,Barger:2003rt,Serpico:2005bc,Simha:2008mt}
or from the cosmic microwave background and large scale anisotropies, like in \cite{Lesgourgues:1999wu}.
Major advances have been performed in neutrino physics in the last ten years.
The change in neutrino flavour content due to oscillations is at present a well established phenomenon.
This implies that the neutrino flavour basis is related to the mass basis
\begin{equation}
\label{e:basis}
\psi_{{\nu}_{\alpha}} = \sum_i U_{\alpha i} \psi_i .
\end{equation}
where the unitary Maki-Nakagawa-Sakata-Pontecorvo (MNSP) matrix can be written as
a product of three matrices $U=T_{23}T_{13}T_{12}$
\begin{equation}
\label{e:U}
U = \left(\matrix{
1 & 0 & 0 \cr
0 & c_{23} & s_{23} \cr
0 & - s_{23} & c_{23} }\right)
\left(\matrix{
c_{13} & 0 & s_{13} e^{-i\delta}\cr
0 & 1 & 0 \cr
- s_{13} e^{i\delta} & 0& c_{13} }\right)
\left(\matrix{
c_{12} & s_{12} &0 \cr
- s_{12} & c_{12} & 0 \cr
0 & 0& 1 }\right) ,
\end{equation}
\noindent
with $c_{ij} = cos \theta_{ij}$ ($s_{ij} = sin \theta_{ij}$) and $\theta_{12},\theta_{23}$ and $\theta_{13}$ the three neutrino mixing angles.
These oscillation parameters have been well determined, except for the third neutrino mixing angle $\theta_{13}$ and a possible Dirac CP violating phase\footnote{Note that Majorana phases can also be present. They can influence the neutrinoless double-beta decay half-lives, while neutrino oscillations are not affected by such phases. For this reason they will not be considered here.}.
If $\theta_{13}$ is close to the Chooz limit, i.e. sin$^2 2\theta_{13}<0.02$, reactor experiments (Double-Chooz, RENO and Daya-Bay) should soon measure this angle \cite{Huber:2009cw}.
The two squared mass differences\footnote{Although the existence of sterile neutrinos is an attractive possibility, here we consider three active neutrino families, in agreement with the ensemble of experimental data.} have been measured with good precision \cite{Amsler:2008zzb}.
Since the sign of $\Delta m^2_{23}$ has not been determined yet, two mass hierarchies are possible: inverted ($\Delta m^2_{23}>0$) or normal ($\Delta m^2_{23}<0$).
This is known as the mass hierarchy problem.
The absolute neutrino mass scale is also still unknown, since neutrino oscillations are only sensitive to mass squared differences. The KATRIN experiment will soon reach the sub-eV sensitivity \cite{Osipowicz:2001sq} while important indirect limits on the sum of the neutrino masses are obtained using CMB and LSS data (see e.g. \cite{Lesgourgues:2006nd,Fogli:2006yq,Hannestad:2006zg}). Indeed, so far, only indirect effects of cosmological neutrinos have been observed. Their detection represents one of the major future challenges. An interesting possibility has been proposed recently in \cite{Cocco:2007za}, namely to exploit the capture on radioactive nuclei. This idea has been further investigated in \cite{Lazauskas:2007da,Blennow:2008fh}.
The observation of CP violation in the neutrino sector is a key open question. The breaking of the CP symmetry
can arise from the presence of a non-zero Dirac $\delta $ phase that renders the $U$ matrix complex Eq.(\ref{e:U}). Long-term expensive accelerator complex (super-beams, beta-beams or neutrino factories)
might be required to tackle this issue \cite{Volpe:2006in}.
Therefore it is important to explore complementary avenues and search for indirect effects. For example, recently we have explored possible
CP violation effects in core-collapse supernovae \cite{Balantekin:2007es}. We have shown that they can arise e.g. if $\nu_{\mu}$ and $\nu_{\tau}$ experience a different
refractive index in the medium -- due to loop corrections and/or physics beyond the standard model.
These results have been extended in presence of the neutrino-neutrino interaction in \cite{Gava:2008rp}.
Note that important developments are currently ongoing in the
study of neutrino propagation in dense media due to temporally evolving density profiles
\cite{Schirato:2002tg,Fogli:2004ff,Kneller:2007kg,Dasgupta:2005wn,Gava:2009pj,Galais:2009wi}, and
the neutrino-neutrino interaction, which introduces collective phenomena (see e.g.
\cite{Samuel:1993uw,Sigl:1992fn,Duan:2005cp,Hannestad:2006nj,Raffelt:2007xt}).
The importance of the latter contribution has been first pointed out in the early Universe context \cite{Dolgov:2002ab,Abazajian:2002qx}.
Several calculations have been performed of the neutrino degeneracy evolution at the BBN epoch including neutrino oscillations \cite{Bell:1998ds,Dolgov:2002ab,Abazajian:2002qx,Mangano:2005cc,Pastor:2008ti}. In \cite{Dolgov:2002ab} it is shown that neutrino oscillations tend to equilibrate the electron, muon and tau neutrino degeneracies. However how much flavour equilibration really holds is still unclear \cite{Pastor:2008ti}. Only in case of flavour equilibration the constraints on $\xi_{\nu_e}$ coming from the abundance of primordial Helium-4 can be translated to the other flavours. In such a case
the limit of $-0.044 < \xi < 0.070$ for all flavours \cite{Serpico:2005bc}, if the conservative Olive and Skillman analysis is used, with an uncertainty of
the order of $5 \%$ on $Y_p$ \cite{Olive:2004kq}. Other detailed analysis of the $^{4}$He fraction exist \cite{Izotov:2007ed,Coc:2003ce}. If the systematic uncertainties, inherent to the helium abundance measurements, are better understood a precision as low as $10^{-3}$ might be reached. Besides, future studies of gravitational lensing distortions on both the temperature and the CMB polarization might reach sensitivities, close to the BBN ones, on the helium fraction, and even at the level of 5.$~10^{-3}$ \cite{Kaplinghat:2003bh}.
In this paper we explore possible CP violating effects, coming from the Dirac phase, on the neutrino degeneracy parameters, at the time of Big-Bang Nucleosynthesis. First we demonstrate analytically the conditions under which there can be such effects. Then we determine numerically the neutrino degeneracies evolution
including, for the first time, a non-zero Dirac phase.
We solve the equations for the three flavour density matrix taking into account the vacuum oscillations, the coupling to the plasma, the neutrino-neutrino interaction and the collisions, using a damping approximation.
The manuscript is structured as follows. The theoretical formalism is shown in section 2. Section 3 presents our analytical results that define the conditions to have CP effects on the neutrino degeneracy parameters. Section 4 illustrates the numerical results and the potential modifications introduced by the Dirac phase as well as a discussion of the possible implications on the $^{4}$He fraction. Section 5 is the conclusion.
\section{The neutrino evolution equations}
\noindent
The neutrino evolution including oscillations can be determined by using the density matrix:
\begin{equation}
\label{cosmo1}
\rho_{{\nu}}(p,t)\equiv\left(\begin{array}{ccc}
\rho_{{\nu}_{ee}} & \rho_{{\nu}_{e\mu}} & \rho_{{\nu}_{e\tau}}\ \\
\rho_{{\nu}_{\mu e}} & \rho_{{\nu}_{\mu\mu}} & \rho_{{\nu}_{\mu\tau}} \\
\rho_{{\nu}_{\tau e}} & \rho_{{\nu}_{\tau\mu}} & \rho_{{\nu}_{\tau\tau}}
\end{array} \right)
\end{equation}
in three flavours. Each neutrino state is characterized by the momentum $p$ and the time $t$.
The transposed of $\rho_{{\nu}}(p,t)$, $\bar{\rho}_{{\nu}}(p,t)$, is taken to describe anti-neutrinos.
In an expanding universe, the equations of motion are \cite{Sigl:1992fn}:
\begin{equation}
i(\partial_t -Hp\partial_p)\rho_p= \left[ H_{tot},\rho_p \right] +C(\rho_p) ,
\label{CosmoEOM}
\end{equation}
where the explicit dependence on $t$ is not shown for simplicity, the subscript $p$ refers to the momentum dependence, and
$H_{tot}$ is the total Hamiltonian describing neutrino propagation in the medium.
As long as the expansion rate of the Universe is smaller than the collision rate among the relativistic species, collisions play an important role and drive the system towards equilibrium.
Such contributions, proportional to $G_F^2$ are included here through the collision term $C(\rho)$.
The cosmic expansion is taken into account through the $Hp\partial_p$ contribution, with $H=\dot{a}(t)/a(t)$.
$a(t)$ is the scale factor that is normalized such as $a \approx 1/T$ at high temperatures or early times. The Hubble constant $H$ is determined through the Friedmann equation $H = \sqrt{8 \pi G\rho/3}$ with $G$ being the gravitational
constant and $\rho$ the total energy density of the relativistic particles.
By using co-moving variables $x \equiv ma,~y \equiv pa$ ($m$ is an arbitrary mass scale
that we take equal to 1 MeV) Eq.(\ref{CosmoEOM}) becomes adimensional:
\begin{equation}
iHx \partial_x \rho_y = \left[H_{tot},\rho_y\right] +C(\rho_y) \label{Cosmo2EOM}
\end{equation}
The
total Hamiltonian describing neutrino propagation involves three contributions
\begin{equation}\label{Htot}
H_{tot} = H_{vac} + H_{mat} + H_{\nu \nu} = \frac{U M^2 U^{\dagger}}{2p}-\frac{8\sqrt{2}G_F p}{3 m_{W}^2}E+\sqrt2 G_F (\rho-\overline{\rho}),\
\end{equation}
where $G_F$ is the Fermi constant and $m_W$ the W boson mass.
Anti-neutrino evolution is described by the same Eq.(\ref{Htot}) for $\bar{\rho}_{{\nu}}(p,t)$, but with a minus sign for $H_{vac}$.
The first term is the vacuum oscillation contribution with $M^2=diag(m_1^2,m_2^2,m_3^2)$ and $U$ the MNSP matrix Eq.(\ref{e:U}).
The second contribution
is proportional to the energy densities $E$ of charged leptons (electrons, positrons and muons) in the plasma and corresponds to the refractive effects of the medium that neutrinos experience \cite{Sigl:1992fn}.
Note that the background potential arising due to asymmetries in charged leptons
is negligible in comparison with the other terms in the Hamiltonian \cite{Abazajian:2002qx}.
The $\sqrt2 G_F (\rho-\overline{\rho})$ contribution represents the neutrino-neutrino interactions and is meant to be integrated over the neutrino momenta. This non-linear term is responsible for synchronizing the neutrino ensemble, as discussed in \cite{Bell:1998ds,Dolgov:2002ab,Abazajian:2002qx},
similarly to what occurs in core-collapse supernovae \cite{Samuel:1993uw,Duan:2005cp,Gava:2008rp}.
Concerning the neutrino scattering with $e^{\pm}$, $\mu^{\pm}$ or among themselves, in principle one should consider the exact collision integral $I_{\nu_{\alpha}}$
\cite{Dolgov:1997mb,Pastor:2008ti} including all relevant two-body weak reactions
of the type $\nu_{\alpha}(1) +2 \longrightarrow 3 + 4$. Here we
follow \cite{Dolgov:2002ab,Mangano:2005cc,Pastor:2008ti} and use a damping prescription
of the form\footnote{Note that in \cite{Mangano:2005cc,Pastor:2008ti} the damping prescription is used only for the off-diagonal contributions.}
\begin{eqnarray}\label{e:coll}
C(\rho_{y,\alpha \beta}) & = & -D_{\alpha \beta} \rho_{y,\alpha \beta}\\ \nonumber
C(\rho_{y,\alpha \alpha}) & = & D_{\alpha \alpha}(f(y,\xi_{\alpha})-\rho_{y,\alpha \alpha})
\end{eqnarray}
with $\alpha,\beta =e, \mu, \tau$ and $\xi_{\alpha}$ being the equilibrium solution.
The coefficients are fixed at
$D_{\alpha \beta}=2(4 sin^4 \theta_W - 2sin^2\theta_W + 2)F_0$ ($\theta_W$ being the Weinberg angle) for
$\alpha=e$ and $\beta=e,\mu$ or $\tau$, while for all other cases we take
$D_{\alpha \beta}$ or $D_{\alpha\alpha}=2 (2 sin^4 \theta_W + 6)F_0$ with
$F_0$ as in \cite{Dolgov:2002ab}.
We consider the plasma to be in thermal\footnote{Corrections to the Fermi-Dirac distributions
have been calculated to be very small \cite{Dolgov:1997mb,Hannestad:1999fj,Mangano:2005cc}.} but not chemical equilibrium.
Therefore, before making the density matrix evolve, we consider
the neutrino occupation numbers given by Fermi-Dirac distributions $f(y,\xi_{\nu_i})$, characterized
by the temperature $T$ and the
chemical potentials $\xi$\footnote{Opposite
chemical potentials are considered for $\nu$ and $\bar{\nu}$.}:
\begin{equation}
\label{initialcond}
\rho_{{\nu}}(y,t=0)\equiv\left(\begin{array}{ccc}
f(y,\xi_{\nu_e}) & 0 & 0 \\
0 & f(y,\xi_{\nu_{\mu}}) & 0 \\
0 & 0 & f(y,\xi_{\nu_{\tau}})
\end{array} \right).
\end{equation}
Before showing the possible impact of the CP phase $\delta$ on $\xi$, we now try to get an analytical insight on the possible sources for these effects.
\section{Conditions for CP effects on $\xi$ : Analytical results}
\noindent
Let us now demonstrate under which conditions there can be CP violation effects coming from
the Dirac phase at the BBN epoch. To this end, we follow partly the procedure established in Refs.\cite{Balantekin:2007es} and \cite{Gava:2008rp} within the context of core-collapse supernovae.
For our purpose it is convenient to work in the $T_{23}$ basis, as shown in Ref \cite{Balantekin:2007es}, but applied to the density matrix Eq.(\ref{cosmo1})
\begin{equation}
\label{e:equatt23}
\tilde{\rho}_{{\nu,y}} = T_{23}^{\dagger} \rho_{{\nu,y}} T_{23}
\end{equation}
We also define the useful quantity $\tilde{\rho}_{y,S} = S^{\dagger}\tilde{\rho}_y S$ where the $\delta$ dependence is contained in the unitary diagonal matrix
\begin{equation}
\label{e:S}
S = \left(\matrix{
1 & 0 & 0 \cr
0 & 1 & 0 \cr
0 & 0 & e^{i\delta} }\right)
\end{equation}
\noindent
Since the MNSP matrix Eq.(\ref{e:U}) can be written as $U=T_{23}ST_{13}^0T_{12}$, we multiply Eq.(\ref{CosmoEOM}) by $T_{23}^{\dagger}$ ($T_{23}$) on the left (right) and get
\begin{equation}
iHx \partial_x \tilde{\rho}_{y,S} = \left[ \tilde{H}_{tot},\tilde{\rho}_{y,S} \right] +C(\tilde{\rho}_{y,S}) \label{e:HtotT23}
\end{equation}
with
\begin{equation}\label{e:Htot2T23}
\tilde{H}_{tot} = \frac{T^0_{13}T_{12} M^2 T^{\dagger}_{12}{T^0}^{\dagger}_{13}}{2y}-\frac{8\sqrt{2}G_F y}{3 m_{W}^2}S\tilde{E}S^{\dagger} \\ \newline
+ \sqrt2 G_F (\tilde{\rho}_{y,S} -\tilde{\overline{{\rho}}}_{y,S})
\end{equation}
where $\tilde{E}=T_{23}^{\dagger}~E~T_{23}$ and $E= diag(E_{ee},E_{\mu \mu},0)$. The quantities $E_{ee}$ and $E_{\mu\mu}$ are the energy densities associated with the electrons, positrons and $\mu^+,\mu^-$ respectively.
We now show the conditions under which : i) the initial conditions for $\tilde{\rho}_S$ and $\tilde{\rho}$ are the same, ii) the evolution equations for $\tilde{\rho}_S$ Eq.(\ref{e:HtotT23}) are the same as for $\tilde{\rho}$ Eq.(\ref{Cosmo2EOM}). If both i) and ii) are fullfilled then
using time discretization by mathematical induction one can show $\tilde{\rho}_S$ is equal to $\tilde{\rho}$ at all times, and therefore the density matrix does not depend on $\delta$ at any time. On the contrary, there can be CP violating effects on $\delta$ and on $\xi$.
Let us consider the initial conditions for $\tilde{\rho}_S$ and for each of the quantities on the r.h.s. of Eq.(\ref{e:HtotT23}) to identify the conditions under which $\tilde{\rho}_S=\tilde{\rho}(t=0)$, namely when condition i) is satisfied.
At $t=0$ the density matrix is
\begin{equation}\label{e:mattini}
\tilde{\rho}_{y,S} =\left(\begin{array}{ccc}
f(y,\xi_{\nu_e}) & 0 & 0 \\
0 & c^2_{23}f(y,\xi_{\nu_{\mu}}) + s^2_{23}f(y,\xi_{\nu_{\tau}}) & c_{23}s_{23}(f(y,\xi_{\tau}) -f(y,\xi_{\nu_{\mu}}) )e^{i\delta} \\
0 & c_{23}s_{23}(f(y,\xi_{\nu_{\tau}}) -f(y,\xi_{\nu_{\mu}}))e^{-i\delta} & c^2_{23}f(y,\xi_{\nu_{\tau}})+s^2_{23}f(y,\xi_{\nu_{\mu}})
\end{array} \right)
\end{equation}
This implies that, if the initial muon and tau neutrino degeneracy parameters differ, this engenders
a dependence on $\delta$ of the density matrix.
Concerning the different terms on the r.h.s. of Eq.(\ref{e:HtotT23}),
it is first shown in Ref.\cite{Balantekin:2007es} that the vacuum contribution to the Hamiltonian satisfies the factorization $\tilde{H}_{vac}(\delta=0)=S^{\dagger}\tilde{H}_{vac}(\delta)S$ and therefore has no $\delta$ dependence at any time.
The matter related term in Eq.(\ref{e:HtotT23}) is given by:
\begin{equation}\label{e:mattini}
S\tilde{E}S^{\dagger}(t=0) =\left(\begin{array}{ccc}
E_{ee} & 0 & 0 \\
0 & -s^2_{23}E_{\mu \mu} & c_{23}s_{23}E_{\mu \mu}e^{i\delta} \\
0 & c_{23}s_{23}E_{\mu \mu}e^{-i\delta} & -c^2_{23}E_{\mu \mu}
\end{array} \right)
\end{equation}
One can see that, if the presence of muons and anti-muons in the relativistic plasma is not neglected at this epoch of the Universe evolution, then this will introduce a source of CP-violation.
Note that it has been explicitly shown in Ref.\cite{Gava:2009gt} that a difference in the
$\nu_{\mu}$ and $\nu_{\tau}$ refractive indeces induces a dependence on $\delta$ in the $\nu_e$ channel.
However, the more the temperature goes down, the less the $E_{\mu \mu}$ term will be important.
Concerning the $\nu\nu$ interaction contribution at initial time, one has before integrating over the neutrino momenta :
\begin{equation}\label{e:nunu}
\tilde{\rho}_{S} -\tilde{{\bar{\rho}}}_{S} =
n_\gamma \left(\begin{array}{ccc}
L_{\nu_e} & 0 & 0 \\
0 & c^2_{23}L_{\nu_{\mu}}+s^2_{23}L_{\nu_{\tau}} & c_{23}s_{23}(L_{\nu_{\mu}}-L_{\nu_{\tau}})e^{i\delta} \\
0 & c_{23}s_{23}(L_{\nu_{\mu}}-L_{\nu_{\tau}})e^{-i\delta} & s^2_{23}L_{\nu_{\mu}}+c^2_{23}L_{\nu_{\tau}}
\end{array} \right)
\end{equation}
where $L_{\nu_i}$ is the $i$ flavour lepton asymmetry and $n_\gamma$ the photon number density.
Any lepton flavour asymmetry between muon and tau neutrinos introduces a CP dependence of the neutrino-neutrino interaction Hamiltonian.
Finally, one has that the collision term $C(\rho(p,t))$ has no $\delta$ dependence since
$\rho_{\nu_i \nu_i}=f(y,\xi_i)$ and the $\rho_{\nu_i \nu_j}=0$ for $i \neq j$ making such a term equal to zero in any basis initially.
Let us now show that condition ii) holds. As shown in \cite{Balantekin:2007es} the matter term is equal in both equations if and only if one can neglect the $E_{\mu \mu}$ contribution. Concerning the neutrino-neutrino interaction term, since the corresponding Hamiltonian has a linear dependence in the density matrix
$S^{\dagger}\tilde{H}_{\nu\nu}(\delta)S=\tilde{H}_{\nu\nu}(\delta=0)$ at all times, as demonstrated by mathematical induction in Ref.\cite{Gava:2008rp}, using the Liouville-Von Neumann equation, in the supernova context. Here this requires the muon and tau neutrino asymmetries being equal ($L_{\nu_{\mu}}=L_{\nu_{\tau}}$) at initial time. Finally the collision term, here treated in the damping approximation, is also a linear function of the density matrix, therefore its dependence is the same if $\tilde{\rho}_S$ and $\tilde{\rho}$ are used.
In conclusion, if $E_{\mu \mu}$ is negligeable and $L_{\nu_{\mu}}=L_{\nu_{\tau}}$ at $t=0$, then
the Hamiltonian that governs the evolution of the density matrix with and without a dependence of the Dirac phase is the same. As a consequence, the evolution of $\tilde{\rho}_S$ and $\tilde{\rho}$ is the same. This implies that, under such conditions,
$\rho_{\nu_e\nu_e}(\delta)=\rho_{\nu_e\nu_e}(\delta=0)$, and therefore
$\xi_{\nu_e}(\delta)=\xi_{\nu_e}(\delta=0)$ at all times.
Summarizing, we have demonstrated that possible CP violating effects can arise from the Dirac phase $\delta$
if, initially, there is a difference between the muon and tau neutrino occupation numbers and/or degeneracy parameters. If the (anti)muon contribution to the energy density is non-negligible this can also, in principle, engender
CP violating effects.
\section{Impact of the CP phase on $\xi_{\nu_e}$ : Numerical results}
\noindent
The goal here is to quantify possible CP effects on the neutrino degeneracies
just before Big-Bang nucleosynthesis, having in mind that this can e.g. potentially impact the Helium-4 fraction.
The numerical results we present are obtained by solving
Eqs.(\ref{Cosmo2EOM}-\ref{e:coll}) for the density matrix
and the initial conditions Eq.(\ref{initialcond}).
The neutrino degeneracies are then obtained using Eq.(\ref{e:nuasym}).
We will give the variations on $\xi_{\nu_e}$ induced by $\delta$ since this is the relevant quantity to quantify the effect on the helium-4 fraction.
In our calculations, the oscillation
parameters are fixed at the values
$\Delta m^2_{12}= 8 \times 10^{-5}$eV$^2$, sin$^2 2\theta_{12}=0.83$ and
$\Delta m^2_{23}= 3 \times 10^{-3}$eV$^2$, sin$^2 2\theta_{23}=1$ for
the solar and atmospheric differences of the squared mass differences and
mixings, respectively \cite{Amsler:2008zzb}. For the third still unknown neutrino mixing angle
$\theta_{13}$, we have taken either a large value close to the Chooz limit, namely
sin$^2 2\theta_{13}=0.19$ at 90 $\%$ C.L., or a small value of
sin$^2 2\theta_{13}=3 \times 10^{-4}$.
Figures \ref{fig1}-\ref{fig4} present the evolution of the neutrino degeneracy parameters as a function of the temperature. Figure \ref{fig1} shows the results obtained for $\xi$ with $\delta=0^{\circ}$ and $\delta=180^{\circ}$, without the inclusion of the neutrino-neutrino contribution to the total Hamiltonian Eq.(\ref{e:Htot2T23}). Only the off-diagonal contributions to the collision term are included.
One can see that the effect of $\delta$ are very small, the curves for a non-zero $\delta$ value being indistiguishable from those for $\delta=0^{\circ}$, as we have been verifying for different initial conditions.
For example, if at initial time $\xi_{\nu_e}=-0.3$, $\xi_{\nu_{\mu}}=0.3$ and $\xi_{\nu_{\tau}}=0$, the variation induced by the phase is
$\Delta \xi_{\nu_e} = \xi_{\nu_e}(\delta) - \xi_{\nu_e}(\delta=0)=4 \times 10^{-6}$, while for $\xi_{\nu_e}=-0.5$, $\xi_{\nu_{\mu}}=0.5$ and $\xi_{\nu_{\tau}}=0$, we obtain $\Delta \xi_{\nu_e} =10^{-5}$.
Figures \ref{fig2} and \ref{fig3}-\ref{fig4} show the CP effect on $\xi$ without and with the neutrino-neutrino contribution\footnote{The calculations including the neutrino-neutrino contribution have been performed following the procedure used in \cite{Dolgov:2002ab,Pastor}.}, respectively. Here both off-diagonal and diagonal collision terms are included\footnote{We use the approximation that the equilibrium $\xi$ are kept at the initial value \cite{Dolgov:2002ab}.}.
Results are shown with $\delta=0^{\circ}$ and $\delta=180^{\circ}$.
In Figure \ref{fig2}
the effect of the CP phase on $\xi_{\nu_e}$ is at the level of $10^{-3}$.
The degeneracy parameters are compatible with the bounds valid if total flavour equilibration is assumed. In Fig.\ref{fig3} the effect is $\Delta \xi_{\nu_e} = 6 \times 10^{-3}$. Note that, for such initial conditions, while
the value of $\xi_{\nu_e}$ at decoupling is compatible with the present bound from BBN;
the other degeneracy parameters are compatible with the bounds valid without assuming total flavour equilibration.
Figure \ref{fig4} presents $\Delta \xi_{\nu_e}$ for different initial conditions, showing that the variations induced by $\delta$ increase, if $\xi_{\nu_e}$ at the temperature of neutrino decoupling is larger.
In all the calculations performed we have found that the modifications produced by
the CP effects increase for a maximal phase and if the third neutrino mixing angle is large. Note that the CP effects arising from the presence of a non-zero $E_{\mu\mu}$
have been found to be completely negligeable.
The inclusion of the non-linear $\nu\nu$ contribution, in general, reduces the phase impact. This can be qualitatively understood. In fact such term synchronizes the neutrino ensemble and freezes neutrino flavour conversion. A similar reduction of CP effects on the neutrino oscillation probabilities has been found in the context of core-collapse supernovae, when the neutrino coupling to neutrinos is included (see Figure 3 of Ref.\cite{Gava:2008rp}).
Finally, our results show that the effect
depends on the diagonal collision terms. Clearly, a definite conclusion on the quantitative effects needs the inclusion of the exact collision integrals. This will be the object of future investigations.
\begin{figure}[t]
\centerline{\includegraphics[scale=0.3,angle=0]{fig1.eps}}
\vspace{.25cm}
\caption{Neutrino degeneracy parameters as a function of the temperature for $\delta=0^{\circ}$ and $\delta=180^{\circ}$ , at the BBN epoch. The initial conditions are set
at $\xi_{\nu_e}=0.1$, $\xi_{\nu_{\tau}}=-0.1$ and $\xi_{\nu_{\mu}}=0$. Results obtained solving Eqs. (\ref{Cosmo2EOM}-\ref{e:coll})
and the initial conditions Eq.(\ref{initialcond}) but
without including the neutrino-neutrino interaction and the diagonal collision terms. The third neutrino mixing angle is taken at the Chooz limit. Here the curves corresponding to a non-zero Dirac phase are indistinguishable from those with a zero value.}
\label{fig1}
\end{figure}
\begin{figure}[t]
\centerline{\includegraphics[scale=0.4,angle=0]{newR2.eps}}
\vspace{.25cm}
\caption{CP effects on the $\xi$ as a function of the temperature. The initial conditions here are
$\xi_{\nu_e}=\xi_{\nu_{\tau}}=0.$ and $\xi_{\nu_{\mu}}=-0.1$.
The calculations include collision terms but do not include the neutrino-neutrino interaction. The mixing angle $\theta_{13}$ is taken at the Chooz limit.}
\label{fig2}
\end{figure}
As far as the impact on the on $^{4}$He fraction is concerned,
an estimate of the CP effects can be obtained using the simple relation $\Delta Y_p=-0.2 \Delta \xi_{\nu_e}$ \cite{Kneller:2004jz}. In fact, a modification of the order of $\Delta \xi_e$ of a several $10^{-3}$, as we have found for some initial conditions, modifies $Y_p$ at most by about $10^{-3}$. This is within the uncertainty from BBN observations, although the present uncertainty on the $^{4}$He fraction might be narrowed down at the level of $Y_p < 0.005 $ in the future \cite{Kaplinghat:2003bh}.
\begin{figure}[t]
\centerline{\includegraphics[scale=0.3,angle=0]{nn2.eps}}
\vspace{.25cm}
\caption{Neutrino degeneracy parameters, as a function of the temperature, when the initial conditions are taken equal to $\xi_{\nu_e}=\xi_{\nu_{\mu}}=0$ and $\xi_{\nu_{\tau}}=0.5$.
The results correspond to $\xi_{\nu_e}$ for $\delta=180^{\circ}$ and $\delta=0^{\circ}$ (lower lines), to $\xi_{\nu_{\tau}}$ for $\delta=180^{\circ}$ (dashed) and $\delta=0^{\circ}$ (dot-dot-dashed) and to $\xi_{\nu_{\mu}}$ for $\delta=180^{\circ}$ (dotted) and $\delta=0^{\circ}$ (dot-dashed). The calculations include the vacuum oscillation and matter term, the neutrino-neutrino interaction and the collisions. }
\label{fig3}
\end{figure}
\begin{figure}[t]
\centerline{\includegraphics[scale=0.3,angle=0]{nn1.eps}}
\vspace{.25cm}
\caption{Neutrino degeneracy parameters $\xi_{\nu_e}$, as a function of the temperature,
for different values of the initial degeneracy parameters. The lines correspond to $\delta=180^{\circ}$ in comparison with the case $\delta=0^{\circ}$, for the initial values of $\xi_{\nu_{\tau}}=0.1$ (lower), $0.3$ (middle) and $0.5$ (upper curves). The values of $\xi_{\nu_e}$ and $\xi_{\nu_{\mu}}$ at $t=0$ are set to zero, for all cases.}
\label{fig4}
\end{figure}
\section{Conclusions}
\noindent
We have explored the impact of the Dirac CP phase of the MNSP matrix on the neutrino degeneracy
parameters at the Big-Bang nucleosynthesis epoch. First we have established analytically the conditions under which there can be possible CP effects coming from this phase and shown, in particular, that these are present if there is a difference between the initial muon and tau neutrino degeneracy parameters. To quantify such
effects we have numerically solved the evolution equation for the density matrix in three flavours, including mixings in
vacuum, coupling to matter, the $\nu\nu$ interaction and collisions (in the damping approximation).
We have found that, depending on the initial conditions
for the neutrino degeneracy parameters, which are an unknown,
modifications up to almost $1.\%$ and $0.1\%$ might be present on $\xi_{\nu_e}$ and $Y_p $ respectively, when the $\nu\nu$ interaction and the collision terms are included.
\vspace{.4cm}
\noindent
{\bf Acknowledgement}
\vspace{.1cm}
\noindent
We thank Sergio Pastor and Georg Raffelt for providing us with important information, Alain Coc, James Kneller and Julien Serreau for useful discussions.
|
\section{Introduction}
Let us consider an elliptic curve $E$, defined over the rationals
and written in short Weierstrass form
\begin{equation}\label{eq_short}
E: Y^2 = X^3 + AX + B, \quad A,B \in {\mathbb Z}.
\end{equation}
We will use the standard notations for:
\begin{itemize}
\item $\Delta = -16(4A^3+27B^2) \neq 0$, the discriminant of $E$;
\item $E({\mathbb Q})$, the finitely generated abelian group of rational points on $E$, and
\item ${\mathcal O}$, the identity element of $E({\mathbb Q})$.
\end{itemize}
Given $P \in E({\mathbb Q})$, we will also write as customary $[m]P$ for the point resulting after adding $m$ times $P$.
The problem of computing the torsion of $E({\mathbb Q})$ has been solved in a
lot of very efficient ways \cite{Cremona,Doud,GOT}, and most
computer packages (say \verb+Maple-Apecs+, \verb+PARI/GP+, \verb+Magma+ or \verb+Sage+)
calculate the torsion of curves with huge coefficients in very
few seconds. The major result which made this possible (along with
others, like the Nagell--Lutz Theorem (\cite{Nagell},\cite{Lutz}) or the embedding theorem for good
reduction primes (see, for example, \cite[VIII.7]{Silverman} or \cite[Chap. 5]{Husemoller})) was Mazur's Theorem
\cite{MazurIHES,Mazur} who listed the fifteen possible torsion
groups.
In the above papers, it is proved that the possible structures of the torsion group of $E({\mathbb Q})$ are
$$
{\mathbb Z} / n {\mathbb Z} \mbox{ for } n=2,\dots,10,12, \quad \mbox{ or } \quad {\mathbb Z} / 2 {\mathbb Z}\times {\mathbb Z} / 2n {\mathbb Z} \mbox{ for } n=1,\dots,4.
$$
Besides, the fifteen of them actually happen as torsion subgroups of elliptic curves. Notice that thanks to the above theorem, the possible prime orders for a torsion point defined over ${\mathbb Q}$ are $2,3,5$ or $7$.
Let $p$ be a prime number and let $E[p]$ be the group of points of order $p$ on $E(\overline{\mathbb Q})$, where $\overline{\mathbb Q}$ denotes an algebraic closure of ${\mathbb Q}$. The action of the absolute Galois group ${\rm G}_{\mathbb Q}={\rm Gal}(\overline{\mathbb Q}/{\mathbb Q})$ on $E[p]$ defines a mod $p$ Galois representation
$$
\rho_{E,p}:{\rm G}_{\mathbb Q}\rightarrow {\rm Aut}(E[p])\cong {\rm GL}_2(\mathbb{F}_p).
$$
Let ${\mathbb Q}(E[p])$ be the number field generated by the coordinates of the points of $E[p]$. Therefore, the Galois extension ${\mathbb Q}(E[p])/{\mathbb Q}$ has Galois group
$$
{\rm Gal}({\mathbb Q}(E[p])/{\mathbb Q})\cong \rho_{E,p}({\rm G}_{\mathbb Q}).
$$
The prime $p$ is called exceptional for $E$ if $\rho_{E,p}$ is not surjective. If $E$ has complex multiplication then any odd prime number is excepcional. On the other hand, if $E$ does not have complex multiplication then Serre \cite{Serre2} proved that $E$ has only finitely many exceptional primes.
Duke \cite{Duke} proved that {\em almost all} elliptic curves over ${\mathbb Q}$ have no exceptional primes. More precisely, given an elliptic curve $E$ in a short Weierstrass form as in (\ref{eq_short}), the height of the elliptic curve is defined as
$$
H(E) = \max(|A|^3,|B|^2).
$$
Let $M$ be a positive integer, and let $\mathcal{C}_H(M)$ be the set of elliptic curves $E$ with $H(E)\le M^6$. For any prime $p$ denote by ${\mathcal E}_p(M)$ the set of elliptic curves $E\in \mathcal{C}_H(M)$ such that $p$ is an excepcional prime for $E$, and by ${\mathcal E}(M)$ the union of ${\mathcal E}_p(M)$ for all primes. Actually in both sets the elliptic curves were considered up to ${\mathbb Q}$--isomorphisms. Duke then proved that
$$
\lim_{M \to \infty} \frac{|{\mathcal E}(M)|}{|\mathcal{C}_H(M)|} = 0.
$$
His proof is based on a version of the Chebotarev density theorem, and uses a two-dimensional large sieve inequality together with results of Deuring, Hurwitz and Masser-W\"ustholz.
Duke also conjectured the following fact, later proved by Grant \cite{Grant}
$$
\left|{\mathcal E}(M) \right| \sim c \sqrt{M}.
$$
Being a bit more precise, Grant showed that, in order to efficiently estimate $\left|
{\mathcal E}(M) \right|$, only ${\mathcal E}_2(M)$ and ${\mathcal E}_3(M)$ had to be actually taken
into account.
\
Now recall that there is a tight relationship between exceptional primes and torsion orders, because if there is a point of order $p$, then $p$ is an exceptional prime \cite{Serre2}. Our aim is then giving a down-to-earth proof of the fact that {\em almost all} elliptic curves over ${\mathbb Q}$ have trivial torsion, motivated by Duke's paper.
We will use in order to achieve this the characterization of torsion structures given in \cite{GT,GT2}, Mazur's Theorem \cite{MazurIHES,Mazur}; and a theorem by Schmidt \cite{Schmidt} on Thue inequalites. Note that we have used a different height notion, more naive in some sense, but nevertheless better suited for our purposes.
Let us change a bit the notation and let us call
$$
E_{(A,B)}: \; Y^2=X^3+AX+B
$$
and, provided $\Delta \neq 0$, we will denote by $E_{(A,B)}({\mathbb Q})[m]$ the group of
points $P \in E_{(A,B)}({\mathbb Q})$ such that $[m]P = {\mathcal O}$. Let us write as well
\begin{eqnarray*}
{\mathcal C}(M) &=& \{ (A,B) \in {\mathbb Z}^2 \; | \; \Delta = -16(4A^3+27B^2) \neq 0, \; \; |A|,|B| \leq M \}.\\
{\mathcal T}_p(M) &=& \{ (A,B) \in {\mathcal C}(M) \; | \; E_{(A,B)}({\mathbb Q})[p] \neq \{ {\mathcal O} \} \}. \\
{\mathcal T} (M) &=& \bigcup_{p \; \mbox{\scriptsize prime }} {\mathcal T}_p (M)
\end{eqnarray*}
Our version of Duke's result is then as follows.
\
\begin{theorem}\label{teorema} With the notations above,
$$
\lim_{M \to \infty} \frac{|{\mathcal T}(M)|}{|{\mathcal C}(M)| } = 0.
$$
\end{theorem}
\
The proof will lead to extremely coarse bounds for $|{\mathcal T}_p(M)|$ which will be proved unsatisfactory in view of experimental data, which we will display subsequently.
\section{Proof of Theorem \ref{teorema}.
Recall that the possible prime orders of a torsion point defined over ${\mathbb Q}$ are $2,3,5$ or $7$.
We will make extensive use of the parametrizations of curves with
a point of prescribed order given in \cite{GT,GT2,Kubert}. These
results have recently been proved useful in showing new properties
of the torsion subgroup (see, for instance \cite{BI,GGT,Ingram}).
First, note that, for a given $A$ with $|A|\leq M$ there are, at
most, two possible choices for $B$ such that $\Delta =0$ (and
hence, the corresponding curve $E_{(A,B)}$ is not an elliptic curve).
Therefore
$$
|{\mathcal C}(M)| \geq (2M+1)^2 - 2(2M+1) = 4M^2-1.
$$
Let us recall from \cite{GT} that a curve $E_{(A,B)}$ with
a point of order $2$ must verify that there exist $z_1,z_2 \in
{\mathbb Z}$ such that
$$
A = z_1-z_2^2, \quad\quad\quad B = z_1z_2.
$$
Therefore $z_1|B$ and for a chosen $z_1$, both $z_2$ and $A$ are determined. Hence, there is at most one pair in ${\mathcal T}_2(M)$ for every divisor of $B$.
We need now an estimate for the average order of
the function $d(x)$, the number of positive divisors of $x$. The simplest estimation is,
probably, the one that can be found in \cite{HW},
$$
d(1)+d(2)+...+d(x) \sim x \log(x).
$$
Therefore, as $M$ tends to infinity,
$$
\left| {\mathcal T}_2(M) \right| \leq \sum_{x=1}^M 2 d(x) + \sum_{x=1}^M 2 d(x) + 2M,
$$
taking into account that we need to consider both positive and negative divisors, the cases where $x \in \{-M,...,-1\}$ and the $2M$ curves with $B=0$. Hence $\left| {\mathcal T}_2(M) \right|\sim c_2 M\log(M)$, where we can, in fact, take $c_2 = 4$.
As for points of order $3$ we can find in \cite{GT} a similar
characterization (a bit more complicated this time) based on the
existence of $z_1,z_2 \in {\mathbb Z}$ such that
$$
A = 27z_1^4+6z_1z_2, \quad\quad\quad B = z_2^2-27z_1^6.
$$
Analogously $z_1|A$ and, once we fix such a divisor, $z_2$ is necessarily given by
$$
z_2 = \frac{A-27z_1^4}{6z_1},
$$
which implies that again there is at most one pair in ${\mathcal T}_3(M)$
for every divisor of $A$. Hence, as $M$ tends to infinity
$$
\left| {\mathcal T}_3(M) \right| \leq c_3 M\log(M),
$$
and again $c_3=4$ suits us.
Points of order $5$ and $7$ need a similar, yet slightly different
argument. From \cite{GT2} we know that if there is a point of order $5$
in $E_{(A,B)}({\mathbb Q})$, then there must exist $p,q \in
{\mathbb Z}$ verifying:
\begin{eqnarray*}
A &=& -27(q^4-12q^3p+14q^2p^2+12p^3q+p^4), \\
B &=& 54(p^2+q^2)(q^4-18q^3p+74q^2p^2+18p^3q+p^4).
\end{eqnarray*}
The first equation is an irreducible Thue equation, hence we can apply the
following result by Schmidt:
\vspace{.3cm}
\noindent {\bf Theorem (Schmidt \cite{Schmidt}).--}
Let $F(x,y)$ be an irreducible binary form of degree $r>3$, with integral
coefficients. Suppose that not more than $s+1$ coefficients are nonzero. Then the number of solutions of the inequality $|F(x,y)| \leq h$ is, a most,
$$
(rs)^{1/2} h^{2/r} \left( 1+\log^{1/r}(h) \right).
$$
\vspace{.3cm}
As for our interests are concerned, this gives a bound for the number of possible $(p,q)$ such that
$$
\left| -27(q^4-12q^3p+14q^2p^2+12p^3q+p^4) \right| \leq M.
$$
Hence, as every such solution determines at most one pair in ${\mathcal T}_5(M)$,
$$
\left| {\mathcal T}_5(M) \right| \leq 4 \sqrt{M} \left( 1+\log^{1/4}(M) \right).
$$
A similar result can be applied for points of order $7$. The
equations which must have a solution are now
\begin{eqnarray*}
A &=& -27k^4(p^2-pq+q^2)(q^6+5q^5p-10q^4p^2-15q^3p^3+\\
& & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad 30q^2p^4 -11qp^5+p^6), \\
B &=& 54k^6(p^{12}-18p^{11}q+117p^{10}q^2-354p^9q^3+570p^8q^4-486p^7q^5\\
& & \quad +273p^6q^6-222p^5q^7+174p^4q^8-46p^3q^9-15p^2q^{10}+6pq^{11}+q^{12}).
\end{eqnarray*}
either for $k=1$ or for $k=1/3$. Hence, using the polynomial defining $B$ and with a similar argument as above
$$
\left| {\mathcal T}_7(M) \right| \leq 24 \sqrt[6]{M} \left( 1 + \log^{1/12} (M) \right).
$$
Therefore, for all $p$ there is an absolut constant $c_p \in {\mathbb Z}_+$ such that
$$
\lim_{M \to \infty} \frac{\left| {\mathcal T}_p(M) \right|}{|C(M)| } \leq \lim_{M \to \infty} \frac{c_p M\log(M)}{4M^2-1}
= 0.
$$
This proves the theorem.
\vspace{.3cm}
\noindent {\bf Remark.--} It must be noted here that our arguments are counting pairs $(A,B)$. So,
in fact, isomorphic curves may appear as separated cases. Both Duke and Grant estimated isomorphism
classes (over ${\mathbb Q}$) rather than curves.
But this can also be achieved by the arguments above with a little extra work. We will show now that
these instances of isomorphic curves are actually negligible as for counting is concerned.
First note that if two curves $E_{(A,B)}$ and $E_{(A',B')}$ are isomorphic over ${\mathbb Q}$, there must be some
$u \in {\mathbb Q}$ such that $A=u^4A'$ and $B=u^6B'$. Hence, there exists some prime $l$ such that,
say, $l^4|A$ and $l^6|B$ (the case $l^4|A'$ and $l^6|B'$ is analogous). Let us write, for a fixed prime $l$
$$
P_n(M,l) = \left\{ x \in {\mathbb Z}_+ \; | \; 1\leq x \leq M, \;\; l^n|M \right\},
$$
and by $P_n(M)$ the union of $P_n(M,l)$, where $l$ run the set of prime divisors of $M$.
Then it is clear that
$$
\begin{array}{l}
\displaystyle \left| P_n(M^n) \right| \leq \displaystyle \sum_{l\leq M} \left| P_n(M^n,l) \right|
= \sum_{l\leq M} \left[ \frac{M^n}{l^n} \right] = \displaystyle \sum_{l\leq M} \left( \frac{M^n}{l^n} + O(1) \right)=\\
\quad \,\, = M^n \sum_{l\leq M} \left( \frac{1}{l^n}\right) + O(M) = \displaystyle M^n \sum_{l \; \mbox{\scriptsize prime}} \frac{1}{l^n} + O(M)
= M^n \mathcal P(n) + O(M),
\end{array}
$$
where $\mathcal P$ is the prime zeta function (see \cite{Froberg}, for instance). So, changing $M^n$ for $M$ we get
\begin{eqnarray*}
\left| P_4(M) \right| &\leq& \displaystyle P(4)M + O \left( \sqrt[4]{M} \right) \simeq 0.0769931M + O \left( \sqrt[4]{M} \right), \\
\left| P_6(M) \right| &\leq& \displaystyle P(6)M + O \left( \sqrt[6]{M} \right) \simeq 0.0170701M + O \left( \sqrt[6]{M} \right). \\
\end{eqnarray*}
Hence, if we are interested in curves up to ${\mathbb Q}$--isomorphism, our bounds for $|{\mathcal T}_p(M)|$ are still correct, while we should change
$$
|{\mathcal C}(M)| \geq 4M^2-1
$$
by
$$
|{\mathcal C}(M)| \geq \left( 4 - P(4)P(6) \right) M^2 + O\left( \sqrt[6]{M} \right)
$$
which obviously makes no difference in the result.
\
\begin{remark} {\rm While all of our boundings for $|{\mathcal T}_p(M)|$ are of the form $c_p M \log(M)$, computational data show that the actual number of curves on ${\mathcal T}_p(M)$ depends heavily on $p$, as one might predict after the estimation given by Grant \cite{Grant} for ${\mathcal E}_p(M)$, the set of elliptic curves $E\in \mathcal{C}_H(M)$ such that $p$ is an excepcional prime for $E$. In fact, a hands--on \verb+Magma+ program gave us the following output
$$
\begin{array}{l|rrrr}
M & \quad |{\mathcal T}_2(M)| & \quad |{\mathcal T}_3(M)| &
\quad |{\mathcal T}_5(M)| & \quad \quad |{\mathcal T}_7(M)| \\
\hline
10^4 \,\,\, & 204,220 & 507 & 1 & 1 \\
10^5 & 2,484,196 & 1,935 & 3 & 1 \\
10^6 & 29,430,050 & 5,873 & 11 & 4 \\
10^7 & \,\,\,340,334,782 & 18,387 & 24 & 5 \\
\end{array}
$$
These actual figures are quite smaller than the bounds obtained.
}\end{remark}
\subsection*{Acknowledgement}The first author was supported in part by grants MTM 2009-07291 (Ministerio de Educaci{\'o}n y Ciencia, Spain) and CCG08-UAM/ESP-3906 (Universidad Auton{\'o}ma de Madrid-Comunidad de Madrid, Spain). The second author was supported by grants FQM--218 and P08--FQM--03894 (Junta de Andaluc\'{\i}a) and MTM 2007--66929 (Ministerio de Educaci\'on y Ciencia, Spain).
|
\section{INTRODUCTION}
What is the best way to assess the influence of scientific publications of
individual scientists? Traditionally, this assessment has been based on the
number of publications of a scientist or the total number of citations
received. However, in any creative endeavor, such as physics research, the
total amount of output is not necessarily the right metric for productivity.
In fact, L. D. Landau himself\cite{LL} kept a list of physicists that were
ranked on a logarithmic scale of achievement.
Recently, Hirsch~\cite{H} introduced the $h$-index that attempts to capture
the overall impact of an individual's publication record researcher by a
single number. The total number of publications can be misleading because an
individual could simply publish a large number of worthless articles.
Conversely, the total number of citations could also be misleading because an
individual might publish a single highly-cited article in a hot but transient
subfield but then nothing else of scientific value. Such a citation record
may not be valuable as that of someone who steadily authors good publications
that are reasonably cited.
The idea underlying the $h$-index is that an equitable integral measure of
citation impact is provided by the value $h$, such that an individual has
published at least $h$ papers with at least $h$ citations. It is obvious
that the $h$-index of a prolific author of trivial publications and that of a
researcher with a single great publication will be much less than someone who
publishes good papers at a steady rate. Because of its obvious appeal, the
$h$-index has become a universally-used metric of overall citation impact.
As one example of the prominence of the $h$-index, it is immediately quoted
in Web of Science citation reports~\cite{ISI}. Moreover, the original idea
of the $h$-index has spawned various of efforts to make the $h$-index more
``fair''~\cite{fair} by correcting for some of the obvious biases that are
part of the citation record, such as many co-authors, self-citations, role of
thesis advisor, {\it etc.}
However, as noted by Hirsch in his original publication~\cite{H}, the
$h$-index of an individual should scale as the square-root of the total
number citations to this individual. This square-root scaling arises in the
most simple model of citations in which an individual publishes papers at a
constant rate and each publication is cited at a constant rate. As a result,
the total number of citations grows quadratically with time while the
$h$-index grows linearly with time, {\it i.e.} $\sqrt{c}$ scales linearly
with $h$. Here, we test this observation for a representative sample of 255
condensed-matter and statistical physics theorists in North America and
Europe.
\begin{figure}[ht]
\vspace*{0.cm}
\includegraphics*[width=0.45\textwidth]{hsq-vs-c-nonames}
\caption{Plot of $c$ versus $4h^2$ for the 255 individuals in the dataset.
The line $c=4h^2$ is shown dashed.}
\label{hsq-vs-c}
\end{figure}
The data was obtained by starting with the names of well-known
condensed-matter and statistical physics theorists and looking up their
citation record in the ISI Web of Science. By scanning at the author lists
of the top-cited publications of these initial authors, the initial list of
authors was extended to their main collaborators, and then to collaborators
of collaborators, {\it etc.} After about 250 people, it became difficult to
find new people or people who could be unambiguously resolved in the ISI
database with the limited knowledge of the author. Primarily because of
limited personal knowledge, the dataset also under-represents junior people.
Moreover, because the Boston University institutional subscription for ISI
extends only to citations after 1973, individuals who began publishing before
this year were excluded to avoid the use of incomplete citation data for
their publications. The data were gathered during a two-day period January
30--31, 2010 between updates of the science citation index database.
If $\sqrt{c}$ scales linearly with $h$, then a plot of these two quantities
should yield a straight line. Fig.~\ref{hsq-vs-c} illustrates this behavior
for all the individuals in the dataset. To highlight the outliers to the
linear behavior that will be discussed below, Fig.~\ref{hsq-vs-c} actually
shows $c$ versus $4h^2$. A linear least-squares fit to all the data of
$\sqrt{c}$ versus $2h$ gives a best fit value of the slope $s\equiv
\sqrt{c}/2h$ of $s\approx 1.045$. The data therefore suggest that $\sqrt{c}$
is essentially equivalent to the $h$-index, up to an overall factor that is
close to 2.
\begin{figure}[ht]
\vspace*{0.cm}
\includegraphics*[width=0.45\textwidth]{dist}
\caption{Plot of the probability density $P(s)$ that an individual is
characterized by a value $s=\sqrt{c}/2h$.}
\label{dist}
\end{figure}
As a further test of the linearity of the dependence of $h$ versus
$\sqrt{c}$, the quantity $s=\sqrt{c}/2h\,$ is computed for each individual in
the dataset of 255 physicists and the resulting distribution, $P(s)$, is
shown in Fig.~\ref{dist}. This distribution is fairly symmetric and most of
the data lies within the range $|s-1|<0.2$. The tightness of the range of
$s$ again suggests that the relation $\sqrt{c}=2h$ accounts for most of the
citation data.
The outliers in the distribution $P(s)$ with $s<1$ and with $s>1$ are
particularly interesting. In the scatter plot of $c$ versus $4h^2$ in
Fig.~\ref{hsq-vs-c}, consider first the outliers with $s<1$ --- data points
that lie below the diagonal. As illustrated in table~\ref{S}, the citation
patterns of best-cited publications for the individuals with the smallest ten
values of $s$ are remarkably similar even though the $h$ indices of this
group of researches ranges over a factor of more than two. In particular,
the difference in the number of citations of successive top-cited papers is
relatively small in all cases. For example, the ratio of the number of
citations to the top-cited and third-cited paper for each individual is in
the range 1.025--2.072.
For the twenty individuals with the largest values of $s$, the citations
patterns are also quite similar within this subpopulation. Almost all have
one (or a few) papers whose citations are a substantial factor larger than
their second-ranked paper. For example, the largest ratio between the number
of citations of the top-cited and third-cited paper is now 10.03. This wide
disparity arises because each individual in this subpopulation (co)-authored
one (or a few) famous publications whose citation frequency outstrips the
remaining publications. Among the individuals that (co)-author these famous
publications, there are three clearly-defined situations: (i) individuals
that wrote a ground-breaking publication on their own or were the driver of
publication with a junior co-author, (ii) those that collaborated with a more
senior author in a famous publication, and (iii) those whose famous
publication was a particularly timely or authoritative review article.
\begin{widetext}
\begin{longtable}
{|p{0.3in}|p{0.35in}|p{0.4in}|p{0.1in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|}
\endhead
\caption{List of the top-10 cited publications of the individuals with the
ten smallest values of $s=\sqrt{c}/2h$. The first three columns give the
$h$-index, the total number of citations $c$, and $s=\sqrt{c}/2h$. The
columns labeled $c_i$ for $i=1,2,\ldots, 10$ are the respective number of
citations of the 10
best-cited papers for each individual.}\label{S}\\
\hline $h$ & $c$ & $\sqrt{c}/2h$ && $c_1$& $c_2$& $c_3$& $c_4$& $c_5$& $c_6$& $c_7$& $c_8$& $c_9$& $c_{10}$\\ \hline
25 & 1510 & 0.777 && 84 & 81 & 62 & 48 & 46 & 43 & 42 & 39 & 37 & 36\\ \hline
39 & 3983 & 0.809 && 260 & 177 & 144 & 127 & 126 & 92 & 91 & 90 & 89 & 85\\ \hline
18 & 853 & 0.811 && 172 & 153 & 83 & 72 & 49 & 39 & 36 & 35 &33 &23 \\ \hline
27 & 1966 & 0.821 && 197 & 191 & 139 & 110 & 66 & 66 & 52 & 51 & 48 & 44\\ \hline
26 & 1854 & 0.828 && 83 & 81 & 81 & 72 & 70 & 68 & 63 & 56 & 55 & 52\\ \hline
28 & 2169 & 0.832 && 100 & 95 & 92 & 89 & 83 & 75 & 73 & 67 & 64 & 64\\ \hline
19 & 1002 & 0.833 && 68 & 66 & 64 & 56 & 51 & 51 & 50 & 43 & 42 & 39\\ \hline
26 & 1879 & 0.833 && 148 & 141 & 84 & 76 & 75 & 65 & 64 & 62 & 56 & 54\\ \hline
23 & 1480 & 0.836 && 94 & 64 & 64 & 62 & 58 & 51 & 49 & 47 & 46 & 42\\ \hline
54 & 8209 & 0.839 && 316 & 297 & 285 & 199 & 198 &198 &181 &177& 162& 153 \\ \hline
\end{longtable}
\newpage
\begin{longtable}
{|p{0.3in}|p{0.35in}|p{0.4in}|p{0.1in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|p{0.3in}|}
\endhead
\caption{List of the citation record of the individuals with the twenty
largest values of $s=\sqrt{c}/2h$; the data format is the same as
Table~\ref{S}. Italicized entries denote review articles.}\label{HR}\\
\hline $h$ & $c$ & $\sqrt{c}/2h$ && $c_1$& $c_2$& $c_3$& $c_4$& $c_5$& $c_6$& $c_7$& $c_8$& $c_9$& $c_{10}$\\ \hline
8 & 544 & 1.458 && 141 & 135 & 50 & 34 & 31 & 17 & 13 &13 & 8 &8 \\ \hline
11 & 1011 & 1.445 && 329 & 220 & 105 & 75 & 73 & 37 & 28 & 24 &24 &17 \\ \hline
20 & 3163 & 1.406 && 480 & 303 & 276 & 264 & 257 & 212 & 198 & 191 & 165 & 157\\ \hline
59 & 26937 & 1.391 && 2259 & 1830 & 1310 & 1220 & 784 &777 &606 &355 &54 &312 \\ \hline
44 & 13789 & 1.334 && 1824 & 1469 & 1393 & 1042 & 570 &560 &504 &480&327 &316 \\ \hline
17 & 2058 & 1.334 && 550 & 255 & 197 & 194 & 123 & 97 & 81 & 73 & 70 & 70\\ \hline
27 & 4903 & 1.297 && 2004 & 371 &316 &243 &157 &133 &114 &100 & 98 &97 \\ \hline
61 & 25003& 1.296 &&4461& \textit{3778} & 1444 & 1333 & 1176 & 1104 &1101 &835 &651 &400\\ \hline
43 & 12403 & 1.295 && 4148 & 1561 & 551 & 495 &452 &405 &399 &339 &217 &214 \\ \hline
40 & 10347 & 1.271 && 2118 & 2004 & 857 &433 &292 & 281 & 274 &238 &223 &221\\ \hline
38 & 9331 & 1.271 && 2721 & 828 & 530 &472 &466&451&324&271&205&178 \\ \hline
32 & 6537 & 1.263 && 1105 & 735 & 650 & 525 & 516 & 320 & 174& 154& 151& 138 \\ \hline
47 & 14090 & 1.263 && \textit{3232} & 815 & 699 & 620 &477&466&420&353&329&274\\ \hline
45 & 12347 & 1.235 && \textit{2357} & 765 & 641 & 563 &495 &462 &405 &377 &350 &322 \\ \hline
28 & 4660 & 1.219 && 2260 & 274 & 206 & 140 &116 &86&84&83&81&79 \\ \hline
19 & 2137 & 1.271 && 766 & 301 & 182 & 77 &74 &71 &61 &58 &43 &41 \\ \hline
61 & 21446 & 1.200 && 7014 & 1102 & 699 & 626 & 502 &427 &331&325&304&296 \\ \hline
15 & 1274 & 1.190 && 242 & 232 & 140 & 96 & 66 & 57& 48 & 41 & 34 & 33 \\ \hline
49 & 13582 & 1.189 && 3051 & 985 & 883 & 864 & 698 & 374 &349 &349&302 &241 \\ \hline
22 & 2732 & 1.188 && 569 & 343 & 271 & 192 & 165 & 98 & 96 & 90 & 72 & 63\\ \hline
39 & 8584 & 1.188 && 2260 & 980 & 658 & 451 & 296 & 289 & 269 & 149& 147&144 \\ \hline
22 & 2699 & 1.181 && 507 & 340 & 192 & 184 & 145 & 130 & 121 & 93 & 92 & 90\\ \hline
\end{longtable}
\end{widetext}
One basic conclusion from this study is that the square-root of the total
number of citations that an individual receives very nearly coincides with
twice his or her $h$-index. A still an open question is why should
$\sqrt{c}$ provide the same integrated measure of the breadth and depth of an
individual's citation record as the $h$-index itself.
A second conclusion is that it is possible to identify outstanding
researchers as the outliers above the diagonal in the scatter plot of
Fig.~\ref{hsq-vs-c}. While there are roughly the same number of points below
the diagonal as above the diagonal, the above-diagonal points with roughly
9000 citations or greater are visually prominent and correspond to
individuals with seminal publications. This simple characteristic appears to
provide a useful predictor of research excellence.
A final caveat: while the outliers discussed here correspond to researchers
with excellent publications to their credit, there are many examples of
excellent researchers that do not fit this outlier criterion. It is
important to be aware of the limitations of using citations alone, or some
function of the number of citations, as a measure of research excellence.
\acknowledgments{I gratefully acknowledge financial support from the US
National Science Foundation grant DMR0906504. I also thank S. Dorogovtsev
for initial correspondence that kindled my old interest in this subject and
J. E. Hirsch for friendly correspondences and advice.}
|
\section{Introduction}
Given a graph $H$, the {\it Ramsey number} $r(H)$ is defined to be the smallest
natural number $N$ such that, in any two-coloring of the edges of $K_N$, there
exists a monochromatic copy of $H$. That these numbers exist was first proven
by Ramsey \cite{R30} and rediscovered independently by Erd\H{o}s and Szekeres
\cite{ErSz}. Since their time, and particularly since the 1970s, Ramsey theory
has grown into one of the most active areas of research within combinatorics,
overlapping variously with graph theory, number theory, geometry and logic.
The most famous question in the field is that of estimating the Ramsey number
$r(t)$ of the complete graph $K_t$ on $t$ vertices. However, despite some small
improvements \cite{S75, Co}, the standard estimates, that $2^{t/2} \leq r(t)
\leq 2^{2t}$, have remained largely unchanged for over sixty years.
Unsurprisingly
then, the field has stretched in different directions. One such direction that
has become fundamental in its own right is that of looking at what happens to
the Ramsey number when we are dealing with various types of sparse graphs.
Another is that of determining induced Ramsey numbers, i.e., proving, for
any given $H$, that there is a graph $G$ such that any two-coloring of the
edges of $G$ contains an induced monochromatic copy of $H$. In this paper, we
present a unified approach which allows us to make improvements to two
classical questions in these areas.
In 1975, Burr and Erd\H{o}s \cite{BE75} posed the problem of showing that every
graph $H$ with $n$ vertices and maximum degree $\Delta$ satisfied $r(H) \leq
c(\Delta) n$,
where the constant $c(\Delta)$ depends only on $\Delta$. That this is indeed
the case was shown by Chv\'atal, R\"{o}dl, Szemer\'edi and Trotter
\cite{CRST83} in one of the earliest applications of Szemer\'edi's celebrated
regularity lemma \cite{Sz76}. Remarkably, this means that for graphs of fixed
maximum degree the Ramsey number only has a linear dependence on the number of
vertices. Unfortunately, because it uses the regularity lemma, the bounds that
the original method gives on $c(\Delta)$ are (and are necessarily \cite{G97})
of tower type in $\Delta$. More precisely, $c(\Delta)$ works out as being an
exponential tower of 2s with a height that is itself exponential in $\Delta$.
The situation was remedied somewhat by Eaton \cite{E98}, who proved, using a
variant of the regularity lemma, that the function $c(\Delta)$ can be taken to
be of the form $2^{2^{c \Delta}}$. Soon after, Graham, R\"{o}dl and Ruci\'nski
proved \cite{GRR00}, by a beautiful method which avoids any use of the
regularity lemma, that there exists a constant $c$ for which
\[c(\Delta) \leq 2^{c \Delta \log^2 \Delta}.\]
For bipartite graphs, they were able to do even better \cite{GRR01}, showing
that if $H$ is a bipartite graph with $n$ vertices and maximum degree $\Delta$
then $r(H) \leq 2^{c \Delta \log \Delta} n$. They also proved that there are
bipartite graphs with $n$ vertices and maximum degree $\Delta$ for which the
Ramsey number is at least $2^{c' \Delta} n$. Recently, Conlon \cite{C07} and,
independently, Fox and Sudakov \cite{FS07} have shown how to remove the $\log
\Delta$ factor in the exponent, achieving an essentially best possible bound of
$r(H) \leq 2^{c \Delta} n$ in the bipartite case. These results were jointly
extended to hypergraphs in \cite{CFS09}, after several proofs
\cite{CFKO07,CFKO072,NORS07} using the hypergraph regularity lemma.
Unfortunately, if one tries to use these recent techniques to treat general
graphs, the best one seems to be able to achieve is $c(\Delta) \leq 2^{c
\Delta^2}$. In this paper we take a different approach, more closely related to
that of Graham, R\"{o}dl and Ruci\'nski \cite{GRR00}. Improving on their bound,
we show that $c(\Delta) \leq 2^{c \Delta \log \Delta}$, which brings us a step
closer to matching the lower bound of $2^{c' \Delta}$.
\begin{theorem} \label{MaxDegreeIntro}
There exists a constant $c$ such that, for every graph $H$ with $n$ vertices
and maximum degree $\Delta$,
\[r(H) \leq 2^{c \Delta \log \Delta} n.\]
\end{theorem}
A graph $H$ is said to be an {\it induced subgraph} of $H$ if $V(H) \subset
V(G)$ and two vertices of $H$ are adjacent if and only if they are adjacent in
$G$. The {\it induced Ramsey number} $r_{ind} (H)$ is the smallest natural
number $N$ for which there is a graph $G$ on $N$ vertices such that in every
two-coloring of the edges of $G$ there is an induced monochromatic copy of $H$.
The existence of these numbers was independently proven by Deuber \cite{De},
Erd\H{o}s, Hajnal and P\'osa \cite{ErHaPo} and R\"{o}dl \cite{Ro1}. The bounds
that these original proofs give on $r_{ind} (H)$ are enormous, but it was
conjectured by Erd\H{o}s \cite{Er2} that the actual values should be more in
line with ordinary Ramsey numbers. More specifically, he conjectured the
existence of a constant $c$ such that every graph $H$ with $n$ vertices
satisfies $r_{ind} (H) \leq 2^{c n}$. If true, the complete graph shows that it
would be best possible.
In a problem paper, Erd\H{o}s \cite{Er1} stated that he and Hajnal had proved a
bound of the form $r_{ind} (H) \leq 2^{2^{n^{1 + o(1)}}}$. This remained the
state of the art for some years until Kohayakawa, Pr\"{o}mel and R\"{o}dl
\cite{KoPrRo} proved that there was a constant $c$ such that every graph $H$ on
$n$ vertices satisfies $r_{ind} (H) \leq 2^{c n \log^2 n}$. As in the
bounded-degree problem, we remove one of the logarithms in the exponent.
\begin{theorem} \label{InducedIntro}
There exists a constant $c$ such that every graph $H$ with $n$ vertices
satisfies
\[r_{ind} (H) \leq 2^{c n \log n}.\]
\end{theorem}
It is worth noting that the graph $G$ that Kohayakawa, Pr\"{o}mel and R\"{o}dl
use in their proofs is a random graph constructed with projective planes. This
graph is specifically designed so as to contain many copies of our target graph
$H$. Recently, Fox and Sudakov \cite{FS08} showed how to prove the same bounds
as Kohayakawa, Pr\"{o}mel and R\"{o}dl using explicit pseudo-random graphs. We
will follow a similar path.
A graph is said to be pseudo-random if it imitates some of the properties of a
random graph. One such random-like property, introduced by Thomason \cite{T87,
T872}, is that of having approximately the same density between any pair of
large disjoint vertex sets. More formally, we say that a graph $G = (V, E)$ is
{\it $(p, \lambda)$-pseudo-random} if, for all subsets $A, B$ of $V$, the
density of edges $d(A,B)$ between $A$ and $B$ satisfies
\[|d(A, B) - p| \leq \frac{\lambda}{\sqrt{|A||B|}}.\]
The usual random graph $G(N,p)$, where each edge is chosen independently with
probability $p$, is itself a $(p, \lambda)$-pseudo-random graph where $\lambda$
is on the order of $\sqrt{N}$. A well-known explicit example, known to be
$(\frac{1}{2}, \sqrt{N})$-pseudo-random, is the Paley graph $P_N$. This graph
is defined by setting $V$ to be the set $\mathbb{Z}_N$, where $N$ is a prime
which is congruent to 1 modulo 4, and taking two vertices $x, y \in V$ to be
adjacent if and only if $x - y$ is a quadratic residue. For further information
on this and other pseudo-random graphs we refer the reader to \cite{KS06}. Our
next theorem states that, for $\lambda$ sufficiently small, a $(\frac{1}{2},
\lambda)$-pseudo-random graph has very strong Ramsey properties. Theorem
\ref{InducedIntro} follows by applying this theorem to the particular examples
of pseudo-random graphs given above.
\begin{theorem} \label{maininduced}
There exists a constant $c$ such that, for any $n \in \mathbb{N}$ and any
$(\frac{1}{2}, \lambda)$-pseudo-random graph $G$ on $N$ vertices with $\lambda
\leq 2^{-c n \log n} N$, every graph on $n$ vertices occurs as an induced
monochromatic copy in all 2-edge-colorings of $G$. Moreover, all of these
induced monochromatic copies can be found in the same color.
\end{theorem}
The theme that unites these two, apparently disparate, questions is the method
we employ in our proofs. A simplified version of this method is the following.
In the first color we attempt to find a large subset in which this color is
very dense. If such a set can be found, we can easily embed the required graph.
If, on the other hand, this is not the case, then there is a large subset in
which the edges of the second color are
well-distributed. Again, this allows us to prove an embedding lemma. Such ideas
are already explicit in the work of Graham,
R\"{o}dl and Ruci\'nski and, arguably, implicit in that of Kohayakawa,
Pr\"{o}mel and R\"{o}dl. The advantage of our method, which extends upon these
ideas, is that it is much more symmetrical between the colors. It is this
symmetry which allows us
to drop a log factor in each case.
In the next section, we will prove Theorem \ref{MaxDegreeIntro}. Section
\ref{InducedSection} contains the proof of Theorem \ref{maininduced}. The last
section contains some concluding remarks together with a discussion of a few
conjectures and open problems. Throughout the paper, we systematically omit
floor and ceiling signs whenever they are not crucial for the sake of clarity
of presentation. All logarithms, unless otherwise stated, are to the base $2$.
We also do not make any serious attempt to optimize absolute constants in our
statements and proofs.
\section{Ramsey number of bounded-degree graphs} \label{BoundedSection}
The {\it edge density} $d(X,Y)$ between two disjoint vertex subsets $X,Y$ of a
graph $G$ is the fraction of pairs $(x,y)
\in X \times Y$ that are edges of $G$. That is,
$d(X,Y)=\frac{e(X,Y)}{|X||Y|}$, where $e(X,Y)$ is the number of edges with one
endpoint in $X$ and
the other in $Y$. In a graph $G$, a vertex subset $U$ is called {\it
bi-$(\epsilon,\rho)$-dense} if, for all disjoint pairs $A,B \subset U$ with
$|A|,|B| \geq \epsilon |U|$, we have $d(A,B) \geq \rho$. We call a graph $G$
{\it bi-$(\epsilon,\rho)$-dense} if its vertex set $V(G)$ is
bi-$(\epsilon,\rho)$-dense. Trivially, if $\epsilon' \leq \epsilon$ and $\rho'
\geq \rho$, then a bi-$(\epsilon',\rho')$-dense graph is also
bi-$(\epsilon,\rho)$-dense. Moreover, if $\epsilon>1/2$, then every graph is
vacuously bi-$(\epsilon,\rho)$-dense as there is no pair of disjoint subsets
each with more than half of the vertices.
Before going into the proof of Theorem \ref{MaxDegreeIntro}, we first sketch
for
comparison the original idea of Graham, R\"odl, and Rucinski \cite{GRR00} which
gives a weaker bound. We then discuss our proof technique. They noticed that if
a graph $G$ on $N$ vertices is bi-$(\epsilon,\rho)$-dense with
$\epsilon=\rho^{\Delta}/(\Delta+1)$ and $N \geq 2\rho^{-\Delta}(\Delta+1)n$,
then $G$ contains every $n$-vertex graph $H$ of maximum degree $\Delta$. This
can be shown by embedding $H$ one vertex at a time. In particular, if a
red-blue edge-coloring of $K_N$ does not contain a monochromatic copy of $H$,
then the red graph is not bi-$(\epsilon,\rho)$-dense, and there are disjoint
vertex subsets $A$ and $B$ with $|A|,|B| \geq \epsilon N$ such that the red
density between them at most $\rho$. It is then possible to iterate, at the
expense of another factor in the exponent of roughly $\log (1/\rho)$, to get a
subset $S$ of size roughly $\epsilon^{\log (1/\rho)}N$ with red edge density at
most $2\rho$ inside. Picking $\rho = \frac{1}{16\Delta}$, a simple greedy
embedding then shows that inside $S$ we can find a blue copy of any graph with
at
most $|S|/4$ vertices and maximum degree $\Delta$.
To summarize, the proof finds a vertex subset $S$ which is either
bi-$(\epsilon,\rho)$-dense in the red graph or is very dense in the blue graph.
In either case, it is easy to find a monochromatic copy of any $n$-vertex graph
$H$ with maximum degree $\Delta$.
We will instead find a sequence of large vertex subsets $S_1,\ldots,S_t$ such
that, in one of the two colors, each of the subsets satisfies some bi-density
condition and the graph between these subsets is very dense. The bi-density
condition inside each $S_i$ is roughly the condition which ensures that we can
embed any graph on $n$ vertices with maximum degree $d_i$, where
$d_1+\ldots+d_t=\Delta-t+1$. A simple lemma of Lov\'asz guarantees that we can
partition $V(H)=V_1 \cup \ldots \cup V_t$ such that the induced subgraph of $H$
with vertex set $V_i$ has maximum degree at most $d_i$. Our embedding lemma
shows that we can embed a monochromatic copy of $H$ with the image of $V_i$
being in
$S_i$. We now proceed to the details of the proof.
\begin{definition} A graph on $N$ vertices is {\it
$(\alpha,\beta,\rho,\Delta)$-dense} if there is a sequence $S_1,\ldots,S_t$ of
disjoint vertex subsets each of cardinality at least $\alpha N$ and nonnegative
integers $d_1,\ldots,d_t$ such that $d_1+\cdots+d_t=\Delta-t+1$, and the
following holds:
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(\rho^{2d_i},\rho)$-dense, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ has at least
$(1-\beta)|S_j|$ neighbors in $S_j$.
\end{itemize}
\end{definition}
Note that since $d_1+\cdots+d_t=\Delta-t+1$ and each $d_i$ is
nonnegative, we must have $t \leq \Delta+1$.
Trivially, if a graph is $(\alpha',\beta',\rho,\Delta')$-dense and $\alpha'
\geq \alpha$, $\beta' \leq \beta$, and $\Delta' \geq \Delta$,
then it is also $(\alpha,\beta,\rho,\Delta)$-dense.
We say a red-blue edge-coloring of the complete graph $K_N$ is {\it
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$-dense} if the red graph is
$(\alpha,\beta,\rho,\Delta_1)$-dense or the blue graph is
$(\alpha,\beta,\rho,\Delta_2)$-dense. We say that
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$ is {\it universal} if, for every $N$,
every red-blue edge-coloring of $K_N$ is
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$-dense.
\begin{lemma}\label{firstlemma}
If $\beta \geq 4(\Delta_2+1)\rho$ and $(\alpha,\beta,\rho,\Delta_1,\Delta_2)$
is universal, then
$(\frac{1}{2}\rho^{2\Delta_1}\alpha,\beta,\rho,\Delta_1,2\Delta_2+1)$ is also
universal.
\end{lemma}
\begin{proof}
Consider a red-blue edge-coloring of a complete graph $K_N$. If the red graph
is bi-$(\rho^{2\Delta_1},\rho)$-dense, then, taking $t=1$, $S_1=V(K_N)$ and
$d_1=\Delta_1$, we see that the red graph is
$(\alpha,\beta,\rho,\Delta_1)$-dense and we are done. So we may suppose that
there are disjoint vertex subsets $V_0,V_1$ with $|V_0|,|V_1| \geq
\rho^{2\Delta_1}N$ such that the red density between them is less than $\rho$.
Delete from $V_0$ all vertices in at least $2\rho|V_1|$ red edges with vertices
in $V_1$; the remaining subset $V_0'$ has cardinality at least
$\frac{1}{2}|V_0| \geq \frac{1}{2}\rho^{2\Delta_1}N$. Since $(\alpha, \beta,
\rho, \Delta_1, \Delta_2)$ is universal, the coloring restricted to $V_0'$ is
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$-dense. Thus, the
red graph is $(\alpha,\beta,\rho,\Delta_1)$-dense (in which case we are again
done)
or the blue graph is $(\alpha,\beta,\rho,\Delta_2)$-dense. We may suppose the
latter holds, and there are subsets $S_1,\ldots,S_t$ each of cardinality at
least $\alpha |V_0'| \geq \frac{1}{2}\rho^{2\Delta_1}\alpha N$ and nonnegative
integers $d_1,\ldots,d_t$ such that $d_1+\cdots+d_t=\Delta_2-t+1$, and the
following holds:
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(\rho^{2d_i},\rho)$-dense, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ has at least
$(1-\beta)|S_j|$ neighbors in $S_j$.
\end{itemize}
Since each vertex in $V_0'$ (and hence in each $S_i$) is in at most
$2\rho|V_1|$ red edges with vertices in $V_1$, there are
at most $2\rho|S_i||V_1|$ red edges between $S_i$ and $V_1$. For $1\leq i \leq
t$, delete from $V_1$ all vertices
in at least $4(\Delta_2+1)\rho|S_i|$ red edges with vertices in $S_i$. For any
given $i$, there can be at most
$\frac{1}{2(\Delta_2 + 1)} |V_1|$ such vertices. Therefore, since $t \leq
\Delta_2 + 1$, the set
$V_1'$ of remaining vertices has cardinality at least $|V_1|-t \cdot
\frac{1}{2(\Delta_2+1)}|V_1| \geq |V_1|/2$.
Since $(\alpha, \beta, \rho, \Delta_1, \Delta_2)$ is universal, the coloring
restricted to $V_1'$ is
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$-dense. Thus, the red graph is
$(\alpha,\beta,\rho,\Delta_1)$-dense (in which case we are done) or the blue
graph is $(\alpha,\beta,\rho,\Delta_2)$-dense. We may suppose the latter holds,
and there are subsets $T_1,\ldots,T_u$ each of cardinality at least $\alpha
| V_1'| \geq \frac{1}{2}\rho^{2\Delta_1}\alpha N$ and nonnegative integers
$e_1,\ldots,e_u$ such that $e_1+\cdots+e_u=\Delta_2-u+1$, and the following
holds:
\begin{itemize}
\item for $1 \leq i \leq u$, $T_i$ is bi-$(\rho^{2e_i},\rho)$-dense, and
\item for $1 \leq i < j \leq u$, each vertex in $T_i$ has at least
$(1-\beta)|T_j|$ neighbors in $T_j$.
\end{itemize}
Note that
$e_1+\cdots+e_u+d_1+\cdots+d_t=\Delta_2-u+1+\Delta_2-t+1=(2\Delta_2+1)-(u+t)+1$.
Moreover,
$\beta \geq 4(\Delta_2+1)\rho$, implying that for all $1 \leq i \leq u$ and all
$1 \leq j \leq t$ every vertex in $T_i$ has at least $(1 - \beta) |S_j|$
neighbors in $S_j$.
Therefore, the sequence $T_1,\ldots,T_u,S_1,\ldots,S_t$ implies that the blue
graph is
$(\frac{1}{2}\rho^{2\Delta_1}\alpha,\beta,\rho,2\Delta_2+1)$-dense, completing
the proof.
\end{proof}
By symmetry, the above lemma implies that if $\beta \geq 4(\Delta_1+1)\rho$ and
$(\alpha,\beta,\rho,\Delta_1,\Delta_2)$
is universal, then
$(\frac{1}{2}\rho^{2\Delta_2}\alpha,\beta,\rho,2\Delta_1+1,\Delta_2)$ is also
universal.
As already mentioned, if $\epsilon>1/2$, every graph $G$ is vacuously
bi-$(\epsilon,\rho)$-dense. As $\rho^{2 \cdot 0}=1>1/2$, setting $t=1$ and
$S_1=V(G)$, we have that every graph $G$ is $(\alpha,\beta,\rho,0)$-dense. This
shows that $(1,2\rho,\rho,0,0)$ is universal, which is the base case $h=0$ in
the induction proof of the next lemma.
\begin{lemma}\label{secondlemma}
Let $h$ be a nonnegative integer and $D:=2^h-1$. Then
$(2^{-2h}\rho^{6D-4h},2(D+1)\rho,\rho,D,D)$ is universal.
\end{lemma}
\begin{proof}
As mentioned above, the proof is by induction on $h$, and the base case $h=0$
is satisfied. Suppose it is satisfied for $h$, and we wish
to show it for $h+1$. Let $D=2^h-1$, $D'=2D+1=2^{h+1}-1$, and
$\beta=4(D+1)\rho=2(D'+1)\rho \geq 2(D+1)\rho$. Recall that, for $\beta \geq
\beta'$, if $(\alpha, \beta', \rho, \Delta_1, \Delta_2)$ is universal then so
is $(\alpha, \beta, \rho, \Delta_1, \Delta_2)$. Therefore, since $(2^{-2h}
\rho^{6D-4h}, 2(D+1)\rho, \rho, D, D)$ is universal, $(2^{-2h} \rho^{6D-4h},
\beta, \rho, D, D)$ is also. Applying Lemma \ref{firstlemma}, we have that
$(\frac{1}{2}\rho^{2D}2^{-2h}\rho^{6D-4h},\beta,\rho,D,2D+1)$ is universal.
Applying the symmetric version of Lemma \ref{firstlemma} mentioned above,
we have that
$$\Big(\frac{1}{2}\rho^{2(2D+1)}\frac{1}{2}\rho^{2D}2^{-2h}\rho^{6D-4h},\beta,\rho,2D+1,2D+1\Big)=(2^{-2(h+1)}\rho^{6D'-4(h+1)},\beta,\rho,D',D'),$$
is universal, which completes the proof by induction.
\end{proof}
We will use the following lemma of Lov\'asz \cite{L66}.
\begin{lemma}\label{basiclemma}
If $H$ has maximum degree $\Delta$ and $d_1,\ldots,d_t$ are nonnegative
integers satisfying
$d_1+\cdots+d_t=\Delta-t+1$, then there is a partition $V(H)=V_1 \cup \ldots
\cup V_t$ such that for $1 \leq i \leq t$, the induced subgraph
of $H$ with vertex set $V_i$ has maximum degree at most $d_i$.
\end{lemma}
The next simple lemma shows that in a bi-$(\epsilon,\rho)$-dense graph, for any
large vertex subset $B$, there are few vertices with few neighbors in $B$.
\begin{lemma}\label{previous}
If $G$ is a bi-$(\epsilon,\rho)$-dense graph on $n$ vertices with $\epsilon
\geq 1/n$ and $B \subset V(G)$ with $|B| \geq 2\epsilon n$, then there are less
than $3\epsilon n$ vertices in $G$ with fewer than $\frac{\rho}{2} |B|$
neighbors
in $B$.
\end{lemma}
\begin{proof}
Suppose for contradiction that the set $A$ of vertices in $G$ with fewer than
$\frac{\rho}{2}|B|$ neighbors in $B$ satisfies
$|A| \geq 3\epsilon n$. Partition $A \cap B=C_1 \cup C_2$ with $|C_1| \leq
| C_2|$ into two sets of size as equal as possible. Then the sets $A'=A
\setminus C_2$ and $B'= B \setminus C_1$ are disjoint, $|A'| \geq \lfloor |A|/2
\rfloor \geq \epsilon n$, $|B'| \geq |B|/2 \geq \epsilon n$, the number of
edges between $A'$ and $B'$ is less than $|A'|\frac{\rho}{2}|B|$, and the edge
density between $A'$ and $B'$ is less than
$\frac{|A'|\frac{\rho}{2}|B|}{|A'||B'|}=\frac{\rho}{2}\frac{|B|}{|B'|} \leq
\rho$, contradicting $G$ is bi-$(\epsilon,\rho)$-dense.
\end{proof}
The following embedding lemma is the last ingredient for the proof of Theorem
\ref{MaxDegreeIntro}.
\begin{lemma}\label{embedlemma}
If $\rho \leq 1/30$ and $G$ is a graph on $N \geq 4(2/\rho)^{2
\Delta}\alpha^{-1}
n$ vertices which is $(\alpha,\frac{1}{2\Delta},\rho,\Delta)$-dense, then $G$
contains every graph $H$ on $n$ vertices with maximum degree at most $\Delta$.
\end{lemma}
\begin{proof}
Since $G$ is $(\alpha,\frac{1}{2\Delta},\rho,\Delta)$-dense, there is a
sequence $S_1,\ldots,S_t$ of disjoint vertex subsets each of cardinality at
least $\alpha N$ and nonnegative integers $d_1,\ldots,d_t$ such that
$d_1+\cdots+d_t=\Delta-t+1$, and the following holds:
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(\rho^{2d_i},\rho)$-dense, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ has at least
$(1-\frac{1}{2\Delta})|S_j|$ neighbors in $S_j$.
\end{itemize}
By Lemma \ref{basiclemma}, there is a vertex partition $V(H)=V_1 \cup \ldots
\cup V_t$ such that the maximum degree of the induced subgraph of $H$ with
vertex set $V_i$ is at most $d_i$ for $1 \leq i \leq t$. Let $v_1,\ldots,v_n$
be an ordering of the vertices in $V(H)$ such that the vertices in $V_i$ come
before the vertices in $V_j$ for $i<j$. Let $N(h,k)$ denote the set of
neighbors $v_i$ of $v_k$ with $i \leq h$. For $v_k \in V_j$, let $M(h,k)$
denote the set of neighbors $v_i \in V_j$ of $v_k$ with $i \leq h$, that is,
$M(h,k) = N(h,k) \cap V_j$. Notice that
$|M(h,k)| \leq d_j$ for $v_k \in V_j$ since the induced subgraph of $H$ with
vertex set $V_j$ has maximum degree at most $d_j$.
We will find an embedding $f:V(H) \rightarrow V(G)$ of $H$ in $G$ such that
$f(V_i) \subset S_i$ for each $i$. We will embed the vertices in increasing
order of their indices.
The embedding will have the property that after embedding the first $h$
vertices, if $k>h$ and $v_k \in V_j$, then the set $S(h,k)$ of vertices in
$S_j$ adjacent to all vertices in $f(N(h,k))$ has cardinality at least
$\frac{1}{2}(\rho/2)^{|M(h,k)|}|S_j|$. Notice that this condition is trivially
satisfied when $h=0$. Suppose that this condition is satisfied after embedding
the first $h$ vertices. The set $S(h,k)$ are the potential
vertices in which to embed $v_k$ after the first $h$ vertices have been
embedded, though
this set may already contain embedded vertices.
Let $j$ be such that $v_{h+1} \in V_{j}$. We need to find a vertex in
$S(h,h+1)$ to embed the copy of $v_{h+1}$. We have $$|S(h,h+1)| \geq
\frac{1}{2}(\rho/2)^{|M(h,h+1)|}|S_j| \geq \frac{1}{2}(\rho/2)^{d_j}|S_j|$$
since $|M(h,h+1)| \leq d_j$. If $d_j = 0$, we may pick $f(v_{h+1})$ to be any
element of the set $S(h,h+1) \char92 \{f(v_1), \dots, f(v_h)\}$. We may assume,
therefore, that $1 \leq d_j \leq \Delta$. In this case we know, for each of the
at most $d_j$ neighbors $v_k$ of $v_{h+1}$ with $k > h+1$ that are in $V_j$,
that the set $S(h,k)$ has cardinality at least
$\frac{1}{2}(\rho/2)^{d_j}|S_j|$. Let $\epsilon = \rho^{2 d_j}$.
Since, for $1 \leq d_j \leq \Delta$ and $\rho \leq 1/30$, $S_j$ is
bi-$(\rho^{2d_j},\rho)$-dense,
$|S(h,k)| \geq \frac{1}{2} (\rho/2)^{d_j} |S_j| \geq 2 \rho^{2 d_j} |S_j| = 2
\epsilon |S_j|$
and $\epsilon |S_j| = \rho^{2 d_j} |S_j| \geq \rho^{2 \Delta} \alpha N \geq
1$, we may apply Lemma \ref{previous} in $S_j$ with $B = S(h,k)$.
Therefore, for each vertex $v_k \in V_j, k>h+1$ adjacent to $v_{h+1}$, at most
$3 \rho^{2d_j}|S_j|$ vertices in $S_j$ have fewer than
$\frac{\rho}{2}|S(h,k)|$ neighbors in $S(h,k)$. Thus, all but at most $d_j
\cdot 3 \rho^{2d_j}|S_j|$ vertices
in $S_j$ have at least $\frac{\rho}{2}|S(h,k)|$ neighbors in $S(h,k)$ for all
$v_k \in V_j, k>h+1$ that are neighbors of $v_{h+1}$.
Since, for $\rho \leq 1/30$, we have $d_j \cdot 3 \rho^{2 d_j} \leq \frac{1}{4}
(\rho/2)^{d_j}$, there are at least
\begin{eqnarray*}
| S(h,h+1)|-d_j \cdot 3\rho^{2d_j}|S_j|-h &\geq&
\frac{1}{2}(\rho/2)^{d_j}|S_j|-d_j \cdot 3\rho^{2d_j}|S_j|-h \geq
\frac{1}{4}(\rho/2)^{d_j}|S_j|-h \\
&\geq& \frac{1}{4}(\rho/2)^{\Delta}\alpha N-h \geq (2/\rho)^{\Delta}n-h > 0
\end{eqnarray*}
such vertices that are not
already embedded. We can pick any of these vertices to be $f(v_{h+1})$. To
continue, it remains to check that any such choice preserves the properties of
our embedding. Indeed,
\begin{itemize}
\item for any $k<h+1$ for which $v_{h+1}$ is adjacent to $v_k$, $f(v_{h+1})$ is
adjacent to $f(v_k)$;
\item if $k > h+1$ and $v_k$ and $v_{h+1}$ are not adjacent, then
$S(h+1,k)=S(h,k)$ and $M(h+1,k)=M(h,k)$;
\item if, for some $k > h+1$, $v_k$ and $v_{h+1}$ are adjacent and $v_k \in
V_{\ell}$ with $\ell \neq j$, then
$M(h+1,k)=0$ since vertices of $V_j$ are embedded before vertices of
$V_{\ell}$, $\ell>j$, so no vertex of $V_{\ell}$ was embedded yet.
Also, $|S(h+1,k)|\geq \frac{1}{2}|S_{\ell}|$ since
$|N(h+1,k)| \leq \Delta$, the vertices in $f(N(h+1,k))$ each have at least
$(1-\frac{1}{2\Delta})|S_{\ell}|$ neighbors in $S_{\ell}$, and hence
$|S(h+1,k)| \geq |S_{\ell}|-\Delta \cdot
\frac{1}{2\Delta}|S_{\ell}|=\frac{1}{2}|S_{\ell}|$;
\item if $k > h+1$, $v_k$ and $v_{h+1}$ are adjacent and $v_k
\in V_j$, then $|M(h+1,k)| = |M(h,k)| + 1$. Moreover, by our choice of the
vertex $f(v_{h+1})$, it has at least $\frac{\rho}{2}|S(h,k)|$ neighbors in
$S(h,k)$.
Therefore
$|S(h+1,k)|\geq \frac{\rho}{2}|S(h,k)| \geq \frac{1}{2} (\rho/2)^{|M(h,k)| + 1}
| S_j| = \frac{1}{2} (\rho/2)^{|M(h+1, k)|} |S_j|$, as required.
\end{itemize}
As we supposed there is an embedding of the first $h$ vertices with the desired
property, the above four facts imply
that there is an embedding of the first $h+1$ vertices with the desired
property. By induction on $h$, we find an embedding of $H$ in $G$.
\end{proof}
We can now prove the following theorem, which implies Theorem
\ref{MaxDegreeIntro}.
\begin{theorem}
For every $2$-edge-coloring of $K_N$ with $N=2^{84\Delta+2}\Delta^{32\Delta}n$,
at least one of the color classes contains a copy of every graph on $n$
vertices with
maximum degree $\Delta \geq 2$.
\end{theorem}
\begin{proof}
Let $h$ be the smallest positive integer such that $D:=2^h-1 \geq \Delta$. By
the definition of $D$, $\Delta \leq D<2\Delta$. Let $\rho=\frac{1}{8D^2}$,
$\alpha=2^{-2h}\rho^{6D-4h} \geq \rho^{6D}$, and $\beta=2(D+1)\rho \leq
\frac{1}{2D}$. Lemma \ref{secondlemma} implies that every red-blue coloring of
the edges of the complete graph $K_N$ is $(\alpha,\beta,\rho,D,D)$-dense. By
Lemma
\ref{embedlemma},
since
\begin{eqnarray*}
4(2/\rho)^{2D}\alpha^{-1} n & \leq & 4(16D^2)^{2D} \cdot (8D^2)^{6D} n \leq
4 (16 (2 \Delta)^2)^{4 \Delta} (8 (2\Delta)^2)^{12 \Delta} n\\
& = & 2^2 (2^6 \Delta^2)^{4 \Delta} (2^5 \Delta^2)^{12 \Delta} n = 2^{84 \Delta
+ 2} \Delta^{32 \Delta} n = N,
\end{eqnarray*}
at least one of the color classes contains a
copy of every graph on $n$ vertices with maximum degree $\Delta$.
\end{proof}
\section{Induced Ramsey numbers} \label{InducedSection}
The goal of this section is to prove Theorem \ref{maininduced}. We will do this
by
finding, in any $2$-edge-coloring of the pseudo-random graph $G$, a collection
of vertex subsets
$S_1, \ldots, S_t$ satisfying certain conditions. The conditions in question
are closely
related to the notion of density that we applied in the last section. Now, as
then, we demand
that the graph of one particular color satisfies a certain bi-density condition
within each $S_i$. In addition, we
demand that between the different $S_i$ the other color be sparse. This may
look like a
simple rearrangement of the condition from the previous section, but, given
that we are now
looking at colorings of a pseudo-random graph $G$ rather than the complete
graph $K_N$, the condition is
more general. Moreover, it is exactly what we need to make our embedding
lemma work.
\begin{definition}
An edge-coloring of a graph $G$ on $N$ vertices with colors
$1$ and $2$ is $(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$-dense if there is a
color $q \in \{1,2\}$, disjoint vertex subsets $S_1,\ldots,S_t$ each of
cardinality at least $\alpha N$ and nonnegative integers $d_1,\ldots,d_t$ with
$d_1+\cdots+d_t=\Delta_q-t+1$ such that the following holds:
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(f(\rho,d_i),\rho)$-dense in the
graph
of color $q$, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ is in at most
$\beta|S_j|$ edges of color $3-q$ with vertices in $S_j$.
\end{itemize}
\end{definition}
We say that $(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$ is {\it universal} if,
for every graph $G$,
every edge-coloring of $G$ with colors $1$ and $2$ is
$(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$-dense.
Note that the density condition used in the last section corresponds to the
case when $G=K_N$ and
$f(\rho,d_i)=\rho^{2d_i}$. Essentially the same proofs as Lemmas
\ref{firstlemma} and \ref{secondlemma} give the
following two more general lemmas. We include the proofs for completeness.
\begin{lemma}\label{firstlemmaind}
If $\beta \geq 4(\Delta_2+1)\rho$ and $(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$
is universal, then
$(\frac{1}{2}f(\rho,\Delta_1)\alpha,\beta,\rho,f,\Delta_1,2\Delta_2+1)$ is also
universal.
\end{lemma}
\begin{proof}
Consider an edge-coloring of a graph $G$ with colors $1$ and $2$. If the graph
of color $1$
is bi-$(f(\rho,\Delta_1),\rho)$-dense, then, taking, $q=1$, $t=1$, $S_1=V(G)$
and
$d_1=\Delta_1$, we are done. So we may suppose that there are disjoint vertex
subsets $V_0,V_1$ with $|V_0|,|V_1| \geq
f(\rho,\Delta_1)N$ such that the density of color $1$ between them is less than
$\rho$.
Delete from $V_0$ all vertices in at least $2\rho|V_1|$ edges of color $1$ with
vertices
in $V_1$; the remaining subset $V_0'$ has cardinality at least
$\frac{1}{2}|V_0| \geq \frac{1}{2}f(\rho,\Delta_1)N$. Since $(\alpha, \beta,
\rho, f, \Delta_1, \Delta_2)$ is universal, the coloring restricted to the
induced subgraph of $G$ with vertex set $V_0'$ is
$(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$-dense. Thus, there is $q \in
\{1,2\}$, disjoint vertex subsets $S_1,\ldots,S_t \subset V_0'$ each of
cardinality at least $\alpha |V_0'|$ and nonnegative integers $d_1,\ldots,d_t$
with
$d_1+\cdots+d_t=\Delta_q-t+1$ such that the following holds:
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(f(\rho,d_i),\rho)$-dense in the
graph
of color $q$, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ is in at most
$\beta|S_j|$ edges of color $3-q$ with vertices in $S_j$.
\end{itemize}
If $q=1$, we are done. Therefore, we may suppose $q=2$.
Since each vertex in $V_0'$ (and hence in each $S_i$) is in at most
$2\rho|V_1|$ edges of color $1$ with vertices in $V_1$, then there are
at most $2\rho|S_i||V_1|$ edges of color $1$ between $S_i$ and $V_1$. For
$1\leq i \leq
t$, delete from $V_1$ all vertices in at least $4(\Delta_2+1)\rho|S_i|$ edges
of color $1$ with vertices in $S_i$. For any
given $i$, there can be at most $\frac{1}{2(\Delta_2 + 1)} |V_1|$ such
vertices. Therefore, since $t \leq
\Delta_2 + 1$, the set $V_1'$ of remaining vertices has cardinality at least
$|V_1|-t \cdot
\frac{1}{2(\Delta_2+1)}|V_1| \geq |V_1|/2$.
Since $(\alpha, \beta, \rho, f, \Delta_1, \Delta_2)$ is universal, the coloring
restricted to the induced subgraph of $G$ with vertex set $V_1'$ is
$(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$-dense. Thus, there is $q' \in
\{1,2\}$, disjoint vertex subsets $T_1,\ldots,T_u \subset V_1'$ each of
cardinality at least $\alpha |V_1'|$ and nonnegative integers $e_1,\ldots,e_u$
with
$e_1+\cdots+e_u=\Delta_{q'}-u+1$ such that the following holds:
\begin{itemize}
\item for $1 \leq i \leq u$, $T_i$ is bi-$(f(\rho,e_i),\rho)$-dense in the
graph
of color $q'$, and
\item for $1 \leq i < j \leq u$, each vertex in $T_i$ is in at most
$\beta|T_j|$ edges of color $3-q'$ with vertices in $T_j$.
\end{itemize}
If $q'=1$, we are done. Therefore, we may suppose $q'=2$.
Note that
$e_1+\cdots+e_u+d_1+\cdots+d_t=\Delta_2-u+1+\Delta_2-t+1=(2\Delta_2+1)-(u+t)+1$.
Moreover,
$\beta \geq 4(\Delta_2+1)\rho$, implying that for all $1 \leq i \leq u$ and all
$1 \leq j \leq t$ every vertex in $T_i$ is in at most $\beta|S_j|$
edges of color $1$ with vertices in $S_j$. Therefore, the sequence
$T_1,\ldots,T_u,S_1,\ldots,S_t$ implies that the edge-coloring of $G$
is
$(\frac{1}{2}f(\rho,\Delta_1)\alpha,\beta,\rho,f,\Delta_1,2\Delta_2+1)$-dense,
completing
the proof.
\end{proof}
By symmetry, the above lemma implies that if $\beta \geq 4(\Delta_1+1)\rho$ and
$(\alpha,\beta,\rho,f,\Delta_1,\Delta_2)$
is universal, then
$(\frac{1}{2}f(\rho,\Delta_2)\alpha,\beta,\rho,f,2\Delta_1+1,\Delta_2)$ is
also
universal.
\begin{lemma}\label{generallemma} Let $h$ be a nonnegative integer and $f$ be
such that $f(\rho,0)=1$.
Define $$\alpha_h=2^{-2h}f(\rho,0)^{-1}f(\rho,2^{h}-1)^{-1}\prod_{i=0}^h
f(\rho,2^i-1)^2.$$ Then $(\alpha_h,2^{h+1}\rho,\rho,f,2^{h}-1,2^{h}-1)$ is
universal.
\end{lemma}
\begin{proof}
The proof is by induction on $h$. As already mentioned, if $\epsilon>1/2$,
every graph $G$ is vacuously bi-$(\epsilon,\rho)$-dense. Since
$\alpha_0=1>1/2$, setting $t=1$ and
$S_1=V(G)$, we have $(1,2\rho,\rho,f,0,0)$ is universal, which is the base case
$h=0$.
Suppose the lemma is satisfied for $h$, and we wish to show it for $h+1$. Let
$D=2^h-1$, $D'=2D+1=2^{h+1}-1$, and
$\beta=4(D+1)\rho=2(D'+1)\rho=2^{h+2}\rho$. Note that, for $\beta \geq
\beta'$, if $(\alpha, \beta', \rho, f, \Delta_1, \Delta_2)$ is universal then
so
is $(\alpha, \beta, \rho, f, \Delta_1, \Delta_2)$. Therefore, since $(\alpha_h,
2(D+1)\rho, \rho,f, D, D)$ is universal, $(\alpha_h,
\beta, \rho, f, D, D)$ is also. Applying Lemma \ref{firstlemmaind}, we have
that
$(\frac{1}{2}f(\rho,D)\alpha_h,\beta,\rho,f,D,2D+1)$ is universal.
Applying the symmetric version of Lemma \ref{firstlemmaind} mentioned above,
we have that
$$\Big(\frac{1}{2}f(\rho,2D+1)\frac{1}{2}f(\rho,D)\alpha_h,\beta,\rho,f,2D+1,2D+1\Big)=(\alpha_{h+1},\beta,\rho,f,D',D'),$$
is universal, which completes the proof by induction.
\end{proof}
A graph $G$ is {\it $n$-Ramsey-universal} if, in any $2$-edge-coloring of $G$,
there are monochromatic induced copies of every graph on $n$ vertices all of
the same color. The following lemma implies Theorem \ref{maininduced}.
\begin{lemma}\label{inducedembeddinglemma} If $G$ is
$(1/2,\lambda)$-pseudo-random on $N$ vertices with $\lambda \leq
2^{-140 n} n^{-40n}N$, then $G$ is $n$-Ramsey-universal.
\end{lemma}
The set-up for the proof of this lemma is roughly similar to the one presented
in the previous section. We start with a collection of bi-dense sets, in say
blue, such that the density of red edges between each pair of sets is
small. The goal is to embed a blue induced copy of a given graph $H$ on
vertices $1,\ldots,n$. We embed vertices one at a time, always maintaining
large sets in which we may embed later vertices. Suppose that at
step $i$ of our embedding, after $v_1, v_2, \dots, v_i$ are chosen, we have
sets $V_{j,i}$ for $j>i$ corresponding to the possible choices for future
$v_j$. If the vertices $j,\ell>i$ are not adjacent, then, by the
pseudo-randomness of $G$, the density of nonedges between any two large sets is
roughly $1/2$, and it is therefore easy to guarantee that we can pick $v_j$ and
$v_{\ell}$ so that they are nonadjacent. On the other hand, if the vertices
$j,\ell>i$ are adjacent, then we need to guarantee that $v_j$ and $v_{\ell}$
will be joined by a blue edge. Thus, it would be helpful to ensure that the
density of blue edges between $V_{j,i}$ and $V_{\ell,i}$ is not too small. In
the bounded-degree case we maintain such a property by
exploiting the fact that the blue density between any two large sets is large.
Here, we do not have this luxury in the case that $V_{j,i}$ and $V_{\ell,i}$
are subsets of different bi-dense sets in the collection. It is instead
necessary to use the fact that the underlying graph $G$ is pseudo-random.
To see how this helps, suppose that we now wish to embed $v_{i+1}$. This will
affect the sets $V_{j,i}$ and $V_{\ell,i}$, resulting in
subsets $V_{j,i+1}$ and $V_{\ell,i+1}$. We would like these subsets to mirror
the density
properties between $V_{j,i}$ and $V_{\ell,i}$. The way we proceed is to show
that using pseudo-randomness we can choose $v_{i+1}$ such that the density of
red edges between the sets $V_{j,i+1}$ and $V_{\ell,i+1}$ remains small. Since
$G$ is pseudo-random, the total density between large sets is roughly
$1/2$ and therefore there will still be many blue edges between these two sets.
{\bf Proof of Lemma \ref{inducedembeddinglemma}:}
We split the proof into four steps.
{\bf Step 1:} We will first choose appropriate constants and prepare $G$ for
embedding monochromatic induced subgraphs.
Any $(1/2,\lambda)$-pseudo-random graph on at least two vertices must satisfy
$\lambda \geq 1/2$. Indeed, letting $A$ and $B$ be distinct vertex subsets each
of cardinality $1$, we have $$1/2=|d(A,B)-1/2| \leq
\frac{\lambda}{\sqrt{|A||B|}} =\lambda.$$ It follows that $N \geq
2^{140n}n^{40n}\lambda \geq 2^{138n}n^{40n}$.
We will start by picking some constants. Pick $\rho = 2^{-13}n^{-3}$, $h=\lceil
\log n \rceil \leq \log 2n$,
$\beta=2^{h+1}\rho \leq 8n\rho=2^{-10}n^{-2}$, $f(\rho,0)=1$ and
$f(\rho,d)=2^{-5n}\rho^d$ if $d>0$, so
\begin{eqnarray*}
\alpha & = & 2^{-2h} f(\rho,0)^{-1}f(\rho,2^{h}-1)^{-1}\prod_{i=0}^{h}
f(\rho,2^i-1)^2 = 2^{-2h} f(\rho,2^{h}-1)\prod_{i=1}^{h-1} f(\rho,2^i-1)^2 \\ &
= &
2^{-2h -5n} \rho^{2^h - 1} \prod_{i=1}^{h-1} 2^{-10n} \rho^{2(2^i - 1)}
= 2^{-2h-(2h-1)5n}\rho^{3 \cdot 2^h-2h-3} \\ & \geq & (2n)^{-12n}\rho^{6n} =
2^{-90n}n^{-30n}.
\end{eqnarray*}
Lemma \ref{generallemma} implies that $(\alpha,\beta,\rho,f,2^{h}-1,2^h-1)$ is
universal. As $2^h \geq n$, it follows that $(\alpha,\beta,\rho,f,n-1,n-1)$ is
also universal. Let $\epsilon_1=\frac{1}{2n}$,
$\epsilon_2=\frac{\epsilon_1\rho}{32n} = 2^{-19}n^{-5}$,
$\epsilon_3=\epsilon_4=\frac{1}{8n}$, $\epsilon_5=\frac{1}{8n^2}$,
$\epsilon_6=\epsilon_2\epsilon_5 = 2^{-22}n^{-7}$ and
$\beta'=2n\beta \leq 2^{-9}n^{-1}$.
Since every red-blue edge of $G$ is $(\alpha,\beta,\rho, f,n-1,n-1)$-dense, we
may assume that there are disjoint vertex subsets $S_1,\ldots,S_t$ each of
cardinality at least $\alpha N$ and nonnegative integers $d_1,\ldots,d_t$ with
$d_1+\cdots+d_t=n-t$ such that
\begin{itemize}
\item for $1 \leq i \leq t$, $S_i$ is bi-$(f(\rho,d_i),\rho)$-dense in the blue
graph, and
\item for $1 \leq i < j \leq t$, each vertex in $S_i$ is in at most
$\beta|S_j|$ red edges with vertices in $S_j$.
\end{itemize}
We will show that we can find a monochromatic blue induced copy of each graph
$H$ on $n$ vertices. We may suppose the vertex set of $H$ is
$V(H)=[n]:=\{1,\ldots,n\}$. Partition $[n]=U_1 \cup \ldots \cup U_t$, with the
vertices in $U_i$ coming before the vertices in $U_j$ for $i<j$ and
$|U_i|=d_i+1$ for $1 \leq i \leq t$. For $j \in U_l$, let $D(i,j)$ denote the
number of neighbors $h$ of $j$ with
$h \leq i$ and $h \in U_l$. Arbitrarily partition $S_i$ into $d_i+1$
sets $\bigcup_{k \in U_i} V_k$ each of cardinality at least $\lfloor
\frac{|S_i|}{d_i+1}\rfloor \geq \lfloor
\frac{|S_i|}{n}\rfloor \geq \frac{|S_i|}{2n} \geq \frac{\alpha}{2n} N$, where
we use $d_i+1 \leq n$, $|S_i| \geq \alpha N$, and the lower bounds on $\alpha$
and $N$.
{\bf Step 2:} We now describe our strategy for constructing induced blue copies
of $H$. In broad outline, we
proceed by induction, embedding each successive vertex $i$ in the set $V_i$.
To achieve this, we have to maintain several conditions which allow us to embed
future vertices.
At the end of step $i$, we will have vertices $v_1,\ldots,v_i$ and, for $j >
i$, subsets
$V_{j,i} \subset V_j$ such that the following four conditions hold.
\begin{enumerate}
\item for $j, \ell \leq i$, if $(j,{\ell})$ is an
edge of $H$, then $(v_j,v_{\ell})$ is a blue edge of $G$, otherwise
$v_j$ and $v_{\ell}$ are not adjacent in $G$;
\item for $ j \leq i < \ell$, if $(j,{\ell})$ is an
edge of $H$, then $v_j$ is adjacent to all vertices in $V_{\ell,i}$
by blue edges, otherwise there are no edges of $G$ from $v_j$ to
$V_{\ell,i}$;
\item for $j > i$, we have $|V_{j,i}| \geq 4^{-i}\rho^{D(i,j)}|V_j|$;
\item for $\ell>j>i$, if $j \in U_{q_1}$ and $\ell \in U_{q_2}$ with
$q_1<q_2$, then each vertex in $V_{j,i}$ is in at most $(1+\epsilon_1)^i
\beta'|V_{\ell,i}|$ red edges with vertices in $V_{\ell,i}$.
\end{enumerate}
Note that $V_{j,i}$ is a subset of vertices of $G$ in which we can still embed
vertex $j$ from $H$ after $i$ steps of our embedding procedure.
Clearly, at the end of the first $n$ steps of this process we obtain the
required copy of $H$. For $i=0$ and $j \in [n]$, define
$V_{j,0}=V_j$. Notice that the above four properties are satisfied
for $i=0$. Indeed, the first two properties are vacuously satisfied, the third
property follows from $V_{j,0}=V_j$, and the last property follows from the
simple inequality $\beta'|V_{\ell,0}|
= 2n\beta|V_{\ell}|\geq \beta|S_{q_2}|$. We now assume that the above four
properties are satisfied at the end
of step $i$ and show how to complete step $i+1$ by finding a vertex
$v_{i+1} \in V_{i+1,i}$ and, for $j>i+1$, subsets $V_{j,i+1} \subset V_{j,i}$
such that conditions 1-4 still hold.
Before we begin the next step of the proof, we need to introduce some notation.
For a vertex $w \in V_j$ and a
subset $S \subset V_{\ell}$ with $j \not =\ell$, let
\begin{itemize}
\item $N(w,S)$ denote the set of vertices $s \in S$ such that $(s,w)$ is an
edge of
$G$,
\item $R(w,S)$ denote the set of vertices $s \in S$ such that $(s,w)$ is a
red edge of $G$,
\item $B(w,S)$ denote the set of vertices $s \in S$ such that $(s,w)$ is a
blue edge of $G$,
\item $\tilde{N}(w,S)=N(w,S)$ if $(j, \ell)$ is an edge of $H$ and
$\tilde{N}(w,S)=S
\setminus N(w,S)$ otherwise,
\item $\tilde{B}(w,S)=B(w,S)$ if $(j, \ell)$ is an edge of $H$ and
$\tilde{B}(w,S):=S
\setminus N(w,S)$ otherwise.
\end{itemize}
Note, for all $S \subset V_{\ell}$ and $w \in V_j$, that
$\tilde{B}(w,S)=\tilde{N}(w,S) \setminus R(w,S)$.
Moreover, since the graph $G$ is pseudo-random with
edge density $1/2$, we expect that for every large subset $S \subset V_{\ell}$
and for most vertices $w \in V_j$ the size of $\tilde{N}(w,S)$ will be roughly
$|S|/2$.
{\bf Step 3:} We next show that if there is a vertex satisfying certain
conditions, then we can continue our embedding. In the last step we show that
there is such a ``good'' vertex.
Let $q$ be the index such that $i+1 \in U_q$. Call a vertex $w \in V_{i+1,i}$
{\it good}
if
\begin{enumerate}
\item for all $j >i+1$ such that $(j,i+1)$ is an edge of $H$ and $j \in U_q$,
$|B(w,V_{j,i})|\geq \rho|V_{j,i}|$,
\item for all $j >i+1$, $|\tilde{N}(w,V_{j,i})|\geq
\left(\frac{1}{2}-\frac{\epsilon_1}{20}\right)|V_{j,i}|$,
\item for all $\ell>j>i+1$ with $j \in U_{q_1}$, $\ell \in U_{q_2}$, and $q_1 <
q_2$,
there are at most $\epsilon_2|V_{j,i}|$ vertices $y \in V_{j,i}$ such that $y$
is in at least
$\beta'\left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)|V_{\ell,i}|$ red edges
with vertices in $V_{\ell,i}$ and $y$ is in at least
$\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,V_{\ell,i})|$ red edges
with vertices of
$\tilde{N}(w,V_{\ell,i})$.
\end{enumerate}
Note that, because the graph $G$ is pseudo-random with edge density $1/2$, we
expect a typical vertex in $V_{i+1,i}$ to be adjacent (and also nonadjacent) to
roughly $1/2$ of the vertices in $V_{j,i}$ and $V_{\ell,i}$. Moreover,
condition $3$ roughly says that for a typical vertex, the density of red edges
between its neighborhoods in $V_{j,i}$ and $V_{\ell,i}$ is not much larger
than the overall density of red edges between these two sets.
We will now show that if there is a good vertex $w \in V_{i+1,i}$, then we may
continue the
embedding by taking $v_{i+1}=w$ and, for $j > i + 1$ with $j \in U_{q_1}$,
letting
$V_{j,i+1}$ be the subset of $\tilde B(w,V_{j,i})$ formed by deleting all
vertices $y$ for which there is $\ell>j$ with $\ell \not \in U_{q_1}$ such that
$y$ is in at least
$\beta'\left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)|V_{\ell,i}|$ red edges
with vertices in $V_{\ell,i}$ and $y$ is in at least
$\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,V_{\ell,i})|$ red edges
with vertices of $\tilde{N}(w,V_{\ell,i})$. Note that, by the third property of
good vertices,
\begin{equation}\label{eq1234}|V_{j,i+1}| \geq |\tilde
B(w,V_{j,i})|-n\epsilon_2|V_{j,i}|.\end{equation}
Let us verify each of the required properties of our embedding in turn.
To verify the first property, we need to show that if $j \leq i$ and $(j, i+1)$
is
an edge of $H$ then $(v_j, v_{i+1})$ is a blue edge and, if $(j,i+1)$ is not an
edge
of $H$, then $(v_j, v_{i+1})$ is not in $G$. But this follows by induction
since, when the first
$i$ vertices were embedded, we had that for all $j \leq i < l$, if $(j,l)$ was
an edge of $H$, then $v_j$
was adjacent to all edges of $V_{l,i}$ by blue edges. Otherwise, there were no
edges between $v_j$ and $V_{l,i}$.
Taking $l = i+1$, the necessary property follows.
For the second property, we would like to show that for $j \leq i+1 < l$, if
$(j,l)$ is an edge of $H$, then $v_j$ is
adjacent to all vertices in $V_{l,i+1}$ by blue edges and, otherwise, there are
no edges between $v_j$ and $V_{l,i+1}$.
Observe that, for all $l > i+1$, the set $V_{l,i+1}$ is a subset of the set
$V_{l,i}$. Therefore, by induction,
we only need to check the condition for $j = i+1$. But $V_{l,i+1}$ is a subset
of $\tilde{B}(v_{i+1}, V_{l,i})$, so this
follows by definition.
We now wish to prove that, for all $j > i+1$, $|V_{j,i+1}| \geq 4^{-(i+1)}
\rho^{D(i+1,j)} |V_j|$. Inequality (\ref{eq1234}) together with the first
property of good vertices
implies that if $j>i+1$, $(j,i+1)$ is an edge of $H$ and $j \in U_q$ (recall
that also $i+1 \in U_q$), then,
since $\epsilon_2 \leq \rho/(2n)$ and $D(i+1,j) = D(i,j) + 1$,
$$|V_{j,i+1}| \geq (\rho-n\epsilon_2)|V_{j,i}| \geq \frac{\rho}{2}|V_{j,i}|
\geq
\frac{\rho}{2}4^{-i}\rho^{D(i,j)}|V_j| \geq 4^{-(i+1)}\rho^{D(i+1,j)}|V_j|.$$
Inequality (\ref{eq1234}), the second property of good vertices and the
inductive assumption that $w$ has at most $(1+\epsilon_1)^i \beta' |V_{j,i}|$
red neighbors in $V_{j,i}$ if $j \not\in U_q$
together imply that for all other $j>i+1$, we have
\begin{eqnarray*} |V_{j,i+1}| & \geq &
| \tilde{B}(w,V_{j,i})|-n\epsilon_2|V_{j,i}| = |\tilde{N}(w,V_{j,i}) \setminus
R(w,V_{j,i})| -n\epsilon_2|V_{j,i}| \\ & \geq &
\left(\frac{1}{2}-\frac{\epsilon_1}{20}\right)|V_{j,i}|-(1+\epsilon_1)^i\beta'|V_{j,i}|
-n\epsilon_2|V_{j,i}|
\geq
\left(\frac{1}{2}-\frac{\epsilon_1}{20}-3\beta'-n\epsilon_2\right)|V_{j,i}|
\\ & \geq & \left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)|V_{j,i}|
\geq \frac{1}{4}4^{-i}\rho^{D(i,j)}|V_j| = 4^{-(i+1)}\rho^{D(i+1,j)}|V_j|.
\end{eqnarray*}
Here we use that $\epsilon_1 = 1/2n$,
$\beta' \leq 2^{-9} n^{-1}$, $\epsilon_2 \leq \epsilon_1/32 n$, and $D(i+1,j)=
D(i,j)$ (since $i+1$ and $j$ are either nonadjacent or belong to different
$U$s).
In either case, the required lower bound on the cardinality of $V_{j,i+1}$
holds. Note the intermediate inequality that $|V_{l,i+1}| \geq
\left(\frac{1}{2} -
\frac{\epsilon_1}{10}\right) |V_{l,i}|$ whenever $l \not\in U_q$.
If $i+1<j<\ell$ is such that $j \in U_{q_1}$ and $\ell \in U_{q_2}$ with
$q \leq q_1<q_2$, our deletion of vertices from $\tilde{B}(w,V_{j,i})$ implies
that each vertex in $V_{j,i+1}$ is in less than
$$\beta'\left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)|V_{\ell,i}| \leq \beta'
| V_{\ell,i+1}|$$ red edges with vertices in $V_{\ell,i}$
or each vertex in $V_{j,i+1}$ is in less than
\begin{eqnarray*}
\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,V_{\ell,i})| & \leq &
\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)(1+\epsilon_1)^i\beta'|V_{\ell,i}|
\leq
\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)(1+\epsilon_1)^i\beta'|V_{\ell,i+1}|/\left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)
\\ & \leq &
(1+\epsilon_1)^{i+1}\beta'|V_{\ell,i+1}|
\end{eqnarray*}
red edges with vertices of $\tilde{N}(w,V_{\ell,i})$. In either case, we see
that the last desired condition of the embedding is satisfied.
{\bf Step 4:} We have shown that if there is a good vertex, then we can
continue the embedding. In this step we show that there is a good vertex in
$V_{i+1,i}$, which completes the proof.
The next three claims imply that the fraction of vertices in $V_{i+1,i}$ that
are
good is at least $1-n\epsilon_3-n\epsilon_4-n^2\epsilon_5>1/2$, i.e., more than
half of the vertices of $V_{i+1,i}$ are good. Indeed, Claim \ref{claim1} shows
that the first property of good vertices is satisfied for all but at most
$n\epsilon_3|V_{i+1,i}|$
vertices in $V_{i+1,i}$. Claim \ref{claim2} shows that the second property of
good
vertices is satisfied for all but at most $n\epsilon_4|V_{i+1,i}|$ vertices in
$V_{i+1,i}$. Claim 3 shows that the third property of good vertices is
satisfied for all but at most $n^2\epsilon_5|V_{i+1,i}|$ of the vertices in
$V_{i+1,i}$. These three claims therefore complete the proof. \ifvmode\mbox{ }\else\unskip\fi\hskip 1em plus 10fill$\Box$
\begin{claim}\label{claim1}
For $j >i+1$ such that $(j,i+1)$ is an edge of $H$ and $j \in U_q$, let $Q_j$
denote the set of vertices $w \in V_{i+1,i}$ such that
$|B(w,V_{j,i})|< \rho|V_{j,i}|$. Then $|Q_j| < \epsilon_3|V_{i+1,i}|$.
\end{claim}
\begin{proof}
Suppose, for contradiction, that $|Q_j| \geq \epsilon_3 |V_{i+1,i}|$. As $j,i+1
\in U_q$ and $|U_q|=d_q+1$, we have $d_q \geq 1$ and
$f(\rho,d_q)=2^{-5n}\rho^{d_q}$. Since $2^{3n} \geq 8n^2$, $|V_{i+1, i}|
\geq 4^{-i} \rho^{D(i, i+1)} |V_{i+1}|$ and $|V_{i+1}| \geq |S_q|/2n$, we have
$$|Q_j| \geq
\epsilon_3 4^{-i}\rho^{D(i,i+1)}|V_{i+1}| \geq \epsilon_3
4^{1-n}\rho^{d_q}|V_{i+1}| \geq
\frac{\epsilon_3}{n}4^{-n}\rho^{d_q}|S_q| \geq f(\rho,d_q)|S_q|.$$
We also have $$|V_{j,i}| \geq 4^{-i}\rho^{D(i,j)}|V_j| \geq
4^{1-n}\rho^{d_q}|V_j| \geq n^{-1}4^{-n}\rho^{d_q}|S_q| \geq
f(\rho,d_q)|S_q|.$$
Since $S_q$ is bi-$(f(\rho,d_q),\rho)$-dense in blue, the blue edge
density between $Q_j$ and $V_{j,i}$ is at least $\rho$, contradicting the
definition of $Q_j$.
\end{proof}
\begin{claim}\label{claim2}
For $j >i+1$, let $P_j$ denote the set of vertices $w \in V_{i+1,i}$ such that
$$|\tilde{N}(w,V_{j,i})|<
\left(\frac{1}{2}-\frac{\epsilon_1}{20}\right)|V_{j,i}|.$$ Then $|P_j| <
\epsilon_4|V_{i+1,i}|$.
\end{claim}
\begin{proof}
The definition of $P_j$ implies that the density of edges between $P_j$ and
$V_{j,i}$ is either less than $\frac{1}{2} - \frac{\epsilon_1}{20}$ or more
than
$\frac{1}{2}+\frac{\epsilon_1}{20}$ (depending on whether or not $(i+1,j)$ is
an edge of $H$). Therefore,
since $G$ is $(1/2,\lambda)$-pseudo-random, we have
$\frac{\epsilon_1}{20}<\frac{\lambda}{\sqrt{|P_j||V_{j,i}|}}$. Note that, for
$j > i$, since $\rho = 2^{-13} n^{-3}$ and $\alpha \geq 2^{-90n} n^{-30 n}$,
\begin{equation}\label{eqn2} |V_{j,i}| \geq 4^{-i} \rho^{D(i,j)} |V_j| \geq
4^{-n} \rho^{n} |V_j| \geq 2^{-15n} n^{-3n} \frac{\alpha}{2n}
N \geq 2^{-106n} n^{-34n} N.\end{equation}
Hence, since we also have $\epsilon_1 = 1/2n$, $\epsilon_4 = 1/8n$ and $\lambda
\leq 2^{-140n} n^{-40n} N$,
\begin{eqnarray}\label{eqn3}
| P_j| & < & \frac{400\lambda^2}{\epsilon_1^2 |V_{j,i}|} < \frac{2^{9}
| 2^{-280n}
n^{-80n} N^2}{(2n)^{-2} 2^{-106n} n^{-34n} N} = 2^{11} n^2 2^{-174 n}
n^{-46n} N\\
& \leq & 2^{-163n} n^{-44n} N \leq (8n)^{-1} 2^{-106n} n^{-34n}N \leq
\epsilon_4|V_{i+1,i}|.\nonumber
\end{eqnarray}
\end{proof}
\begin{claim}\label{claim3}
Fix a pair $j$ and $\ell$ with $i+1<j<\ell$, $j \in U_{q_1}$, $\ell \in
U_{q_2}$, and $q_1 < q_2$. Let $X=V_{i+1,i}$, $Y=V_{j,i}$, and $Z=V_{\ell,i}$.
Define the bipartite graph $F=F_{j,\ell}$ with parts $X$ and $Y$ where $(x,y)
\in X \times Y$ is an edge if
$$|R(y,Z)| \geq \beta' \left(\frac{1}{2}-\frac{\epsilon_1}{10}\right)|Z|$$ and
$$|R(y,Z) \cap \tilde{N}(x,Z)| >
\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,Z)|.$$ Let $T_{j,\ell}$
denote the set of vertices in
$X$ with degree at least $\epsilon_2|Y|$ in $F$. Then $|T_{j,\ell}| \leq
\epsilon_5|X|$.
\end{claim}
\begin{proof}
For $y \in Y$, let $X_y \subset X$ denote the neighbors of $y$ in graph $F$.
Note that, for
every $x \in X_y$, the fact that $|R(y,Z) \cap \tilde{N}(x,Z)| >
\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,Z)|$ implies that, in either
the graph $G$ or its complement,
$x$ has at least $\left(\frac{1}{2}+\frac{\epsilon_1}{10}\right)|R(y,Z)|$
neighbors in $R(y,Z)$ (this is again because $\tilde{N}(x,Z)$ is either the
neighborhood of $x$ or its complement depending on whether or not $(i+1,\ell)$
is an edge of $H$).
Therefore, since $G$ is $(1/2,\lambda)$-pseudo-random, $$\frac{\epsilon_1}{10}
\leq
\frac{\lambda}{\sqrt{|X_y||R(y,Z)|}},$$
Note that, by the first condition on $F$, if $y$ has any
neighbors in $X$, $|R(y,Z)| \geq \beta' |Z|/4$. Therefore,
$$|X_y| \leq \frac{100\lambda^2}{\epsilon_1^2|R(y,Z)|} \leq
\frac{400\lambda^2}{\epsilon_1^2\beta'|Z|} \leq \epsilon_6|X|.$$
This last inequality follows as in the previous claim. Indeed, since
$\beta' = 2^{-9} n^{-1}$, $\epsilon_6 = 2^{-22} n^{-7}$, $Z=V_{\ell,i}$, and
$X=V_{i+1,i}$, using inequalities (\ref{eqn2},\ref{eqn3}), we have
$$\frac{400 \lambda^2}{\epsilon_1^2 \beta' |Z|} \leq \beta'^{-1} 2^{-163n}
n^{-44n} N \leq 2^{-154 n} n^{-43 n} N \leq 2^{-22} n^{-7} 2^{-106n} n^{-34n} N
\leq \epsilon_6 |X|.$$
Therefore, the edge density of $F$ between $X$ and $Y$ is at most $\epsilon_6$
and there are at most
$\frac{\epsilon_6|X||Y|}{\epsilon_2|Y|}=\epsilon_5|X|$ vertices in $X$ with
degree at least $\epsilon_2|Y|$ in $F$.
\end{proof}
\section{Concluding remarks}
Another interesting concept of sparseness, introduced by Chen and Schelp
\cite{CS93}, is that of arrangeability. A graph $H$ is said to be {\it
$p$-arrangeable} if there is an ordering of the vertices of $H$ such that, for
any vertex $v_i$, the set of neighbors to the right of $v_i$ in the ordering
have at most $p$ neighbors to the left of $v_i$ (including $v_i$ itself).
Extending the result of Chv\'atal, R\"{o}dl, Szemer\'edi and Trotter
\cite{CRST83}, Chen and Schelp showed that for every $p$ there is a constant
$c(p)$ such that, for any $p$-arrangeable graph $H$ with $n$ vertices, $r(H)
\leq c(p) n$. This result has several consequences. Planar graphs, for example,
may be shown to be $10$-arrangeable \cite{KT94}, so their Ramsey numbers grow
linearly. The best bound that is known for $c(p)$, again due to Graham,
R\"{o}dl and Ruci\'nski \cite{GRR00}, is $c(p) \leq 2^{c p (\log p)^2}$.
Unfortunately, it
is unclear whether the bounds that we have given for bounded-degree graphs can
be extended to the class of arrangeable graphs. It would be interesting to
prove such a bound.
An even more problematic notion is that of degeneracy. A graph $H$ is said to
be {\it $d$-degenerate} if there is an ordering of the vertices of $H$ such
that any vertex $v_i$ has at most $d$ neighbors that precede it in the
ordering. Equivalently, every subgraph of $H$ has a vertex of degree at most
$d$. A conjecture of Burr and Erd\H{o}s \cite{BE75} states
that for every $d$ there should be a constant $c(d)$ such that, for any
$d$-degenerate graph $H$ with $n$ vertices, $r(H) \leq c(d) n$. This
conjecture,
which is still open, is a substantial generalization of the results on Ramsey
numbers of bounded-degree graphs. The best result that is known, due to Fox and
Sudakov \cite{FS209}, is $r(H) \leq 2^{c(d) \sqrt{\log n}} n$.
An old related problem is to bound the Ramsey number of graphs with $m$ edges.
Erd\H{o}s and Graham \cite{ErGr} conjectured that among all graphs with
$m= \binom{n}{2}$ edges and no isolated vertices, the complete graph on
$n$ vertices has the largest Ramsey number. Motivated by the lack of
progress on this conjecture, Erd\H{o}s \cite{Er1} asked whether one could
at least show that the Ramsey number of any graph with $m$ edges is not
much larger than that of the complete graph with the same size. Since the
number of vertices in a complete graph with $m$ edges is on the order of
$\sqrt{m}$, Erd\H{o}s conjectured that $r(H) \leq 2^{c\sqrt{m}}$ holds for
every graph $H$ with $m$ edges and no isolated vertices. Until recently the
best known bound
for this problem was $2^{c\sqrt{m}\log m}$ (see \cite{AlKrSu}).
To attack Erd\H{o}s' conjecture one can try to use the result on Ramsey numbers
of bounded-degree graphs.
Indeed, given a graph $H$ with $m$ edges, one can first embed the $2\sqrt{m}$
vertices of largest
degree in $H$ using the standard pigeonhole argument of Erd\H{o}s and
Szekeres \cite{ErSz}. The remaining vertices of $H$ span a graph with maximum
degree $\sqrt{m}$. Hence, one may apply the arguments used to prove the upper
bound for Ramsey numbers of bounded-degree graphs to embed the rest of $H$.
However, this approach will likely require an upper bound of
$2^{c\Delta}n$ on the Ramsey number for graphs on $n$
vertices of maximum degree $\Delta$, which we do not have yet.
Recently, the third author \cite{Su} was able to circumvent this difficulty and
prove Erd\H{o}s' conjecture.
Finally, we would like to stress that the proofs given in this paper are highly
specific to the 2-color case. The best results that are known in the $q$-color
case are obtained by an entirely different method \cite{FS07} and are
considerably worse. For example, the $q$-color Ramsey number $r_q(H)$ of a
graph on $n$ vertices with maximum degree $\Delta$ is only known to satisfy the
inequality $r_q(H) \leq 2^{c_q \Delta^2} n$. It would be of considerable
interest to improve this latter bound to $r_q(H) \leq 2^{c_q \Delta^{1+o(1)}}
n$.
|
\section*{}
Einstein's dream of a unification of all forces of nature including gravity is still very far away, however much progress has been done in understanding how to formulate quantum field theories in curved space-time and in treating general relativity as an effective field theory. We are used to thinking of the Planck scale $M$ as a fundamental scale of Nature, indeed as the scale at which quantum gravitational effects become important. However, this parameter does get renormalized when quantum fluctuations are taken into account. In other words, Newton's constant and hence the Planck mass are scale dependent like any other coupling constant of a quantum field theory. The true scale $\mu_*$ at which quantum gravity effects are large is one at which
\begin{equation}
\label{strong}
M (\mu_*) \sim \mu_*.
\end{equation}
This condition implies that quantum fluctuations in spacetime geometry at length scales $\mu_*^{-1}$ will be unsuppressed.
A Wilsonian Planck mass $M( \mu )=/\sqrt{G_N(\mu)}$ can be introduced. The contributions of spin 0, spin 1/2 and spin 1 particles to the running of $M( \mu )$ can easily be calculated using the heat kernel method. This regularization procedure insures that the symmetries of the theory are preserved by the regulator. One finds \cite{Vassilevich:1994cz,Larsen:1995ax,Kabat:1995eq,Calmet:2008tn}
\begin{eqnarray}
\label{Nrunning}
M ( \mu )^2=M(0)^2 - \frac{\mu^2}{12 \pi} N
\end{eqnarray}
where $M(0)$ is the Planck mass measured in long distance (astrophysical) experiments and where $N=N_0+N_{1/2}-4 N_1$. The parameters $N_0$, $N_{1/2}$ and $N_1$ are respectively the number of scalar fields, Weyl fermions and gauge bosons in the theory. Note that this calculation relies on quantum field theory in curved spacetime and does not require any assumption about quantum gravity. Furthermore, as noted in ref \cite{Vassilevich:1994cz}, the contribution of the photon is gauge independent. It is possible to calculate the contribution of the graviton to the running of the Planck mass treating general relativity as an effective theory. One finds \cite{BjerrumBohr:2002ks} that the graviton contribution to the running of the Planck mass has the same sign as that of the vector field. In other words, quantum gravitational interactions make the Planck mass bigger at high energy. It is worth noticing that if the number of matter fields $N$ is large, quantum gravitational effects are a $1/N$ correction and thus under control since they represent a small correction to the one loop result. This limit has been studied already by Tomboulis \cite{Tomboulis:1977jk} and later by Smolin \cite{Smolin:1981rm}.
One may wonder about higher order loop corrections such as diagrams depicted in Figures (\ref{FG1}) and (\ref{FG2}). A back of the envelope calculation shows that diagrams of the type depicted in Figure (\ref{FG1}) lead to a contribution of the type:
\begin{eqnarray}
\sim \frac{1}{(4 \pi^2)^{l+2}} N^l \left (\frac{\Lambda}{M(0)} \right)^{2l} \frac{\Lambda^4 N}{M(0)^2}
\end{eqnarray}
where $l$ is the number of matter field loops on the graviton line and $\Lambda$ is a dimensionful cutoff. These contributions are small compared to the first loop result ($M(0) \sim 10^{18}$ GeV and $\Lambda \sim 10^3$ GeV). In other words if we were able to resum the diagrams on the graviton line, we would obtain a graviton line with a coupling to matter only suppressed $M(\mu_*)$ but not enhanced by $N$.
Diagrams with more graviton propagators (see e.g. Figure (\ref{FG2})), for a given number of matter field loops, are suppressed compared to those shown in Figure (\ref{FG1}):
\begin{eqnarray}
\sim \frac{1}{(4 \pi^2)^{l+3}} N^l \left (\frac{\Lambda}{M(0)} \right)^{2l} \frac{\Lambda^6 N}{M(0)^4}.
\end{eqnarray}
\begin{figure}
\begin{minipage}[t]{0.49\linewidth}
\includegraphics[width=5cm]{diagGrav1}
\caption{\label{FG1}\em
Higher loop contributions to the renormalization of the Planck mass. The wavy lines represent gravitons, whereas continuous lines are matter field loops. For a given number of matter loops, the most important contribution comes from the diagram where a single graviton propagator contains all the matter field loops}
\end{minipage}
\begin{minipage}[t]{0.49\linewidth}
\includegraphics[width=5cm]{diagGrav2}
\caption{\label{FG2}\em
Higher loop contributions to the renormalization of the Planck mass. For a fixed number of loops this topology is suppressed compared to the diagram shown in Figure (\ref{FG1}).}
\end{minipage}
\end{figure}
These results can be applied to design models able to solve the hierarchy problem. A large hidden sector of some $10^{33}$ particles of spin 0 and/or 1/2 reduces the scale of strong gravity $\mu_\star$ to the TeV region \cite{Calmet:2008tn}. This could lead to the production of small four dimensional non thermal black holes \cite{Calmet:2008dg} at the LHC or in cosmic ray experiments \cite{Calmet:2008rv} and also to the emission of massless gravitons at the LHC \cite{Calmet:2009gn}.
This running of the Planck mass also has implications for grand unified theories. It has been shown \cite{Calmet:2008df} that in typical supersymmetric grand unified theories based on for example SO(10), the large number of particles with masses close to the unification scale ($N\sim 1000$) leads to a shift of the strength at which gravity becomes strong. One finds $\mu_\star \sim 10^{17}$ GeV rather than $\mu_\star \sim 10^{18}$ GeV which implies that operators induced by strong gravitational effects can dramatically impact the unification conditions of the gauge couplings of the Standard Model. For example SUSY SU(5) may not unify the couplings of the standard model properly. On the other hand, these quantum gravitational effects could also lead to grand unification by the same mechanism in models which naively would not lead to unification of the coupling constants of the standard model such as non supersymmetric SO(10) models\cite{Calmet:2009hp}.
Finally the renormalization of Newton's constant has deep consequences for the validity of linearized general relativity. As shown in ref. \cite{AC}, a comparison of the scale at which unitarity is violated in gravitational scattering to the scale at which quantum gravitational effects become large leads to a bound on the particle content of a model coupled to linearized general relativity.
\bibliographystyle{ws-procs975x65}
|
\section{Introduction}\label{intro}
\IEEEPARstart{C}{onsider} an estimation problem in which
we want to estimate $\theta
\in \Theta$ based on an observation $X$ from $\left\{ P_{\theta} ,
\theta \in \Theta \right\}$ where each $P_{\theta}$ is a probability measure
on a sample space ${\mathcal{X}}$. Suppose that estimators are allowed to
take values in ${\mathcal{A}} \supseteq \Theta$ and
that the loss function is of the form $\ell(\rho)$ where $\rho$ is a metric on
${\mathcal{A}}$ and $\ell: [0, \infty) \rightarrow [0, \infty)$ is a nondecreasing
function. The minimax risk for this problem is defined by
\begin{equation*}
R := \inf_{\hat{\theta}} \sup_{\theta \in \Theta}
{\mathbb E}_{\theta} \ell( \rho(\theta, \hat{\theta}(X)) ),
\end{equation*}
where the infimum is over all measurable functions $\hat{\theta} :
{\mathcal{X}} \rightarrow {\mathcal{A}}$ and the expectation is taken under the
assumption that $X$ is distributed according to $P_{\theta}$.
In this article, we are concerned with the problem of obtaining lower
bounds for the minimax risk $R$. Such bounds are useful in assessing the quality
of estimators for $\theta$. The standard approach to these bounds is to obtain a
reduction to the more tractable problem of bounding from below the
minimax risk of a multiple hypothesis testing problem. More specifically, one
considers a finite subset $F$ of the parameter space $\Theta$ and a
real number $\eta$ such that $\rho(\theta, \theta') \geq \eta$ for
$\theta, \theta ' \in F, \theta \neq \theta'$ and employs the
inequality $R \geq \ell(\eta/2)r$,
where
\begin{equation}\label{rdef}
r := \inf_T \sup_{\theta
\in F} P_{\theta} \left\{T \neq \theta \right\},
\end{equation}
the infimum being over all estimators $T$ taking values in $F$. The
proof of this inequality relies on the triangle inequality satisfied
by the metric $\rho$ and can be found, for example, in~\cite[Page
1570, Proof of Theorem 1]{YangBarron} (Let us note, for the
convenience of the reader, that the notation employed by Yang and
Barron~\cite{YangBarron} differs from ours in that they use $d$ for
the metric $\rho$, $\epsilon_{n, d}$ for our $\eta$ and
$N_{\epsilon_n, d}$ for the finite set $F$. Also the proof
in~\cite{YangBarron} involves a positive constant $A$ which
can be taken to be 1 for our purposes. The constant $A$ arises because
Yang and Barron~\cite{YangBarron} do not require that $d$ is a metric
but rather require it to satisfy a weaker local triangle inequality
which involves the constant $A$.)
The next step is to note that $r$ is bounded from below by Bayes
risks. Let $w$ be a probability measure on $F$. The Bayes risk $\bar{r}_w$
corresponding to the prior $w$ is defined by
\begin{equation}\label{minbayes}
\bar{r}_w := \inf_T \sum_{\theta \in F}
w_{\theta}P_{\theta} \left\{ T \neq \theta \right\},
\end{equation}
where $w_{\theta} := w \left\{\theta \right\}$ and the infimum is over
all estimators $T$ taking values in $F$. When $w$ is the discrete uniform
probability measure on $F$, we simply write $\bar{r}$ for $\bar{r}_w$. The
trivial inequality $r \geq \bar{r}_w$ implies that lower bounds for
$\bar{r}_w$ are automatically lower bounds for $r$.
The starting point for the results described in this paper is
Theorem~\ref{maha}, which provides a lower
bound for $\bar{r}_w$ involving $f$-divergences of the probability
measures $P_{\theta}, \theta \in
F$. The $f$-divergences (\cite{AliSilvey, Csiszar66, Csiszar67,
Csiszar67fdiv}) are a general
class of divergences between probability measures which include many
common divergences\slash distances like the Kullback Leibler
divergence, chi-squared divergence, total variation distance, Hellinger
distance etc. For a \textit{convex} function $f:[0, \infty) \rightarrow {\mathbb R}$
satisfying $f(1) = 0$, the $f$-divergence between two probabilities $P$ and
$Q$ is given by
\begin{equation*}
D_f(P||Q) := \int f\left( \frac{dP}{dQ} \right) dQ
\end{equation*}
if $P$ is absolutely continuous with respect to $Q$ and $\infty$
otherwise.
Our proof of Theorem~\ref{maha} presented in section~\ref{test} is
extremely simple. It just relies on the convexity of the function
$f$ and the standard result that $\bar{r}_w$ has the following exact
expression:
\begin{equation}\label{betexp}
\bar{r}_w = 1 - \int_{{\mathcal{X}}} \max_{\theta \in F}\left\{
w_{\theta}p_{\theta}(x)\right\} d\mu(x),
\end{equation}
where $p_{\theta}$ denotes the density of $P_{\theta}$ with respect to
a common dominating measure $\mu$ (for example, one can take $\mu :=
\sum_{\theta \in F}P_{\theta}$).
We show that Fano's inequality is a special case (see Example~\ref{kld}) of
Theorem~\ref{maha}, obtained by taking $f(x) = x \log x$. Fano's
inequality is used extensively in the
nonparametric statistics literature for obtaining minimax lower
bounds, important works being~\cite{IbragimovHasminskii77, IbragimovHasminskii80,
Ibragimov:Hasminskii:81book, Hasminskii78, Birge83, Birge:86newZW,
YangBarron}. In the special case when $F$ has only
two points, Theorem~\ref{maha} gives a sharp inequality relating the
total variation distance between two probability measures to
$f$-divergences (see Corollary~\ref{Neq2}). When $f(x)= x\log x$,
Corollary~\ref{Neq2} implies an inequality due to
Tops{\o}e~\cite{TopsoeCapDiv} from which Pinsker's inequality can be
derived. Thus Theorem~\ref{maha} can be viewed as a generalization of
both Fano's inequality and Pinsker's inequality.
The bound given by Theorem~\ref{maha} involves the quantity $J_f :=
\inf_Q \sum_{\theta \in F} D_f(P_{\theta}||Q)/|F|$, where the infimum
is over all probability measures $Q$ and $|F|$ denotes the cardinality
of the finite set $F$. It is usually not possible to calculate $J_f$
exactly and in section~\ref{jsbounds}, we provide upper bounds for
$J_f$. The main result of this section, Theorem~\ref{ybgen}, provides
an upper bound for $J_f$ based on approximating the set of $|F|$
probability measures $\left\{P_{\theta}, \theta \in F\right\}$ by a
smaller set of probability measures. This result is motivated by and a
generalization to $f$-divergences of a result of Yang and
Barron~\cite{YangBarron} for the Kullback-Leibler divergence.
In section~\ref{gent}, we use the inequalities proved in
sections~\ref{test} and~\ref{jsbounds} to
obtain minimax lower bounds involving only global metric entropy
attributes. Of all the lower bounds presented in this paper,
Theorem~\ref{myb.thm}, the main result of section~\ref{gent}, is the most
application-ready method. In order to apply this in a particular
situation, one only needs to determine suitable bounds on global
covering and packing numbers of the parameter space $\Theta$ and the
space of probability measures $\left\{ P_{\theta}, \theta \in \Theta
\right\}$ (see section~\ref{suppest} for an application).
Although the main results of sections~\ref{test}
and~\ref{jsbounds} hold true for all $f$-divergences,
Theorem~\ref{myb.thm} is stated only for the Kullback-Leibler divergence,
chi-squared divergence and the divergences based on $f(x) = x^l -
1$ for $l > 1$. The reason behind this is that Theorem~\ref{myb.thm}
is intended
for applications where it is usually the case that the underlying
probability measures $P_{\theta}$ are product measures and divergences
such as the Kullback-Leibler divergence and chi-squared divergence can
be computed for product probability measures.
The inequalities given by Theorem~\ref{myb.thm} for the chi-squared
divergence and divergences based on $f(x) = x^l - 1$ for $l > 1$ are
new while the inequality for the Kullback-Leibler divergence is due to
Yang and Barron~\cite{YangBarron}. There turn out to be qualitative
differences between these inequalities in the case of estimation
problems involving finite dimensional parameters where the inequality
based on chi-squared divergence gives minimax lower bounds having the
optimal rate while the one based on the Kullback-Leibler divergence
only results in sub-optimal lower bounds. We shall explain this
happening in section~\ref{gent} by means of elementary examples.
We shall present two applications of our bounds. In
section~\ref{suppest}, we shall prove a new lower bound for the minimax
risk in the problem of estimation/reconstruction of a $d$-dimensional
convex body from noisy measurements of its support function in $n$
directions. In section~\ref{covmat}, we shall provide a different
proof of a recent result by Cai, Zhang and
Zhou~\cite{CaiZhangZhou2009} on covariance matrix estimation.
\section{Lower bounds for the testing risk $\bar{r}_w$}\label{test}
We shall prove a lower bound for $\bar{r}_w$ defined
in~\eqref{minbayes} in terms of $f$-divergences. We shall assume that
the $N := |F|$ probability measures $P_{\theta}, \theta \in F$ are all
dominated by a sigma finite measure $\mu$ with densities $p_{\theta},
\theta \in F$. In terms of these densities, $\bar{r}_w$ has the exact
expression given in~\eqref{betexp}. A trivial consequence
of~\eqref{betexp} that we shall often use in the sequel is that
$\bar{r} \leq 1 - 1/N$ (recall that $\bar{r}$ is $\bar{r}_w$ in the
case when $w$ is the uniform probability measure on $F$).
\begin{theorem}\label{maha}
Let $w$ be a probability measure on $F$. Define $T: {\mathcal{X}} \rightarrow
F$ by $T(x) := \arg
\max_{\theta \in F} \left\{w_{\theta}
p_{\theta}(x)\right\}$, where $w_{\theta} := w\left\{ \theta\right\}$.
For every convex function $f: [0, \infty) \rightarrow {\mathbb R}$ and every
probability measure $Q$ on ${\mathcal{X}}$, we have
\begin{equation}\label{maha.eq}
\sum_{\theta \in F}w_{\theta}D_f(P_{\theta}||Q) \geq
Wf\left( \frac{1-\bar{r}_w}{W}\right) + (1-W)f\left(
\frac{\bar{r}_w}{1-W}\right),
\end{equation}
where $W := \int_{{\mathcal{X}}} w_{T(x)}dQ(x)$. In particular,
taking $w$ to be the uniform probability measure, we get that
\begin{equation}\label{maha.eq1}
\sum_{\theta \in F} D_f(P_{\theta}||Q) \geq
f\left( N(1 - \bar{r})\right) + (N - 1)f\left(
\frac{N \bar{r}}{N - 1}\right).
\end{equation}
\end{theorem}
The proof of this theorem relies on a simple application of the
convexity of $f$ and it is presented below.
\begin{IEEEproof}
We may assume that all the weights $w_{\theta}$ are strictly
positive and that the probability measure $Q$ has a density $q$ with
respect to $\mu$. We start with a simple inequality for nonnegative
numbers $a_{\theta}, \theta \in F$ with $\tau := \arg \max_{\theta \in
F} \left\{w_{\theta} a_{\theta} \right\}$. We first write
\begin{equation*}
\sum_{\theta \in F} w_{\theta}f(a_{\theta}) = w_{\tau}f(a_{\tau}) + (1-w_{\tau})\sum_{\theta \neq \tau}
\frac{w_{\theta}}{1-w_{\tau}}f(a_{\theta})
\end{equation*}
and then use the convexity of $f$ to obtain that the quantity
$\sum_{\theta}w_{\theta} f(a_{\theta})$ is bounded from below by
\begin{equation*}
w_{\tau}f(a_{\tau}) + (1-w_{\tau}) f \left(
\frac{\sum_{\theta \in F} w_{\theta} a_{\theta} - w_{\tau}a_{\tau}}{1-w_{\tau}}\right).
\end{equation*}
We now fix $x \in {\mathcal{X}}$ such that $q(x) > 0$ and apply the inequality
just derived to $a_{\theta} := p_{\theta}(x)/q(x)$. Note that in this
case $\tau = T(x)$. We get that
\begin{equation}\label{aux1}
\sum_{\theta \in F} w_{\theta}f \left( \frac{p_{\theta}(x)}{q(x)} \right)
\geq A(x) + B(x),
\end{equation}
where
\begin{equation*}
A(x) := w_{T(x)} f\left( \frac{p_{T(x)}(x)}{q(x)} \right)
\end{equation*}
and
\begin{equation*}
B(x) := (1-w_{T(x)}) f \left( \frac{\sum_{\theta \in
F}w_{\theta}p_{\theta}(x)
- w_{T(x)}p_{T(x)}(x)}{(1-w_{T(x)})q(x)} \right).
\end{equation*}
Integrating inequality~\eqref{aux1} with respect to the probability
measure $Q$, we get that the term $\sum_{\theta \in F}
w_{\theta}D_f(P_{\theta}||Q)$ is bounded from below by
\begin{equation*}
\int_{{\mathcal{X}}}A(x)q(x)d\mu(x) + \int_{{\mathcal{X}}}B(x)q(x)d\mu(x).
\end{equation*}
Let $Q'$ be the probability measure on ${\mathcal{X}}$ having the density
$q'(x) := w_{T(x)}q(x)/W$ with respect to $\mu$. Clearly
\begin{equation*}
\int_{{\mathcal{X}}}A(x)q(x)d\mu(x) = W \int_{{\mathcal{X}}} f \left(\frac{p_{T(x)}(x)}{q(x)} \right) q'(x)
d\mu(x),
\end{equation*}
which, by Jensen's inequality, is larger than or equal to
$Wf((1-\bar{r}_w)/W)$. It follows similarly that
\begin{equation*}
\int_{{\mathcal{X}}}B(x) q(x) d\mu(x) \geq (1-W) f\left(
\frac{\bar{r}_w}{1-W} \right).
\end{equation*}
This completes the proof of inequality~\eqref{maha.eq}. When $w$ is
the uniform probability measure on the finite set $F$, it is obvious
that $W$ equals $1/N$ and this leads to inequality~\eqref{maha.eq1}.
\end{IEEEproof}
Let us denote the function of $\bar{r}$ on the right hand side
of~\eqref{maha.eq1} by $g$:
\begin{equation}\label{gintang}
g(a) := f\left(N(1-a)\right) + (N-1) f \left(\frac{Na}{N-1} \right).
\end{equation}
Inequality~\eqref{maha.eq1} provides an implicit lower bound for
$\bar{r}$. This is because $\bar{r} \in [0, 1-1/N]$ and $g$ is
non-increasing on $[0, 1-1/N]$ (as can be seen in the proof of the
next corollary in the case when $f$ is differentiable; if $f$ is not
differentiable, one needs to work with right and left derivatives
which exist for convex functions).
The convexity of $f$ also implies trivially that $g$ is
convex, which can be used to convert the implicit
bound~\eqref{maha.eq1} into an explicit lower bound. This is the
content of the following corollary. We assume differentiability for
convenience; to avoid working with one-sided derivatives.
\begin{corollary}\label{explow}
Suppose that $f: [0, \infty)$ is a differentiable convex function and
that $g$ is defined as in~\eqref{gintang}. Then, for every $a \in
[0,1-1/N]$, we have
\begin{equation}\label{explow.eq}
r \geq \bar{r} \geq a + \frac{\inf_{Q}\sum_{\theta \in F}
D_f(P_{\theta}||Q) - g(a)}{g'(a)},
\end{equation}
where the infimum is over all probability measures $Q$.
\end{corollary}
\begin{IEEEproof}
Fix a probability measure $Q$. Inequality~\eqref{maha.eq1} says that
$\sum_{\theta \in F} D_f(P_{\theta} || Q) \geq g(\bar{r})$. The
convexity of $f$ implies that $g$ is also convex and hence, for every
$a \in [0, 1-1/N]$, we can write
\begin{equation}\label{explow.temp}
\sum_{\theta \in F} D_f(P_{\theta}||Q) \geq g(\bar{r}) \geq g(a) +
g'(a)(\bar{r} - a).
\end{equation}
Also,
\begin{equation*}
\frac{g'(a)}{N} = f'\left(\frac{Na}{N-1}\right) - f'\left(N(1-a)\right).
\end{equation*}
Because $g$ is convex, we have $g'(a) \leq g'(1-1/N) = 0$ for $a \leq
1-1/N$ (this proves that $g$ is non-increasing on $[0,
1-1/N]$). Therefore, by rearranging~\eqref{explow.temp}, we
obtain~\eqref{explow.eq}.
\end{IEEEproof}
Let us now provide an intuitive understanding of
inequality~\eqref{maha.eq1}. When the probability measures
$P_{\theta}, \theta \in F$ are tightly packed i.e., when they are
close to one another, it is hard to distinguish between them
(based on the observation $X$) and hence, the testing Bayes risk $\bar{r}$
will be large. On the other hand, when the probability measures are
well spread out, it is easy to distiguish between them and
therefore, $\bar{r}$ will be small. Indeed, $\bar{r}$ takes on its
maximum value of $1-1/N$ when the probability measures $P_{\theta},
\theta \in F$ are all equal to one another and it takes on its
smallest value of 0 when $\max p_{\theta} = \sum p_{\theta}$ i.e.,
when $P_{\theta}, \theta \in F$ are all mutually singular.
Now, one way of measuring how packed/spread out the probability
measures $P_{\theta}, \theta \in F$ are is to consider the quantity
$\inf_Q \sum_{\theta \in F} D_f(P_{\theta}||Q)$, which is small when
the probabilities are tightly packed and large when they are spread
out. It is therefore reasonable to expect a connection between this
quantity and $\bar{r}$. Inequality~\eqref{maha.eq1} makes
this connection explicit and precise. The fact that the function $g$
in~\eqref{gintang} is non-increasing means that when $\inf_Q
\sum_{\theta \in F} D_f(P_{\theta}||Q)$ is small, the lower bound on
$\bar{r}$ implied by~\eqref{maha.eq1} is large and when $\inf_Q
\sum_{\theta \in F} D_f(P_{\theta}||Q)$ is large, the lower bound on
$\bar{r}$ is small.
Theorem~\ref{maha} implies the following corollary which provides
sharp inequalities between total variation distance and
$f$-divergences. The total variation distance between two probability
measures is defined as \emph{half} the $L^1$ distance between their
densities.
\begin{corollary}\label{Neq2}
Let $P_1$ and $P_2$ be two probability measures on a space
${\mathcal{X}}$ with total variation distance $V$. For every convex function $f:
[0, \infty) \rightarrow {\mathbb R}$, we have
\begin{equation}\label{tvf.eq}
\inf_Q \left( D_f(P_1||Q) + D_f(P_2||Q) \right) \geq
f\left( 1 + V \right) + f\left( 1 - V \right),
\end{equation}
where the infimum is over all probability measures $Q$. Moreover this
inequality is sharp in the sense that for every $V \in [0, 1]$, the
infimum of the left hand side of~\eqref{tvf.eq} over all probability
measures $P_1$ and $P_2$ with total variation distance $V$ equals the
right hand side of~\eqref{tvf.eq}.
\end{corollary}
\begin{IEEEproof}
In the setting of Theorem~\ref{maha}, suppose that $F= \left\{1,
2\right\}$ and that the two probability measures are $P_1$ and
$P_2$ with densities $p_1$ and $p_2$ respectively. Since $2\max (p_1,
p_2)$ equals $p_1+p_2+|p_1-p_2|$, it follows that $2\bar{r}$ equal
$1-V$. Inequality~\eqref{tvf.eq} is then a direct consequence
of inequality~\eqref{maha.eq1}.
The following example shows that~\eqref{tvf.eq} is sharp. Fix $V \in
[0, 1]$. Consider the space ${\mathcal{X}} = \left\{1, 2 \right\}$ and define
the probabilities $P_1$ and $P_2$ by $P_1\left\{ 1 \right\} =
P_2\left\{2 \right\} = (1+V)/2$ and of course $P_1\left\{ 2
\right\} = P_2\left\{1 \right\} = (1-V)/2$. Then the total
variation distance between $P_1$ and $P_2$ equals $V$. Also if we take
$Q$ to be the uniform probability measure $Q \left\{1 \right\} =
Q\left\{2 \right\} = 1/2$, then one sees that $D_f(P_1||Q) +
D_f(P_2||Q)$ equals $f(1+V) + f(1-V)$ which is same as the right hand
side in~\eqref{tvf.eq}.
\end{IEEEproof}
What we have actually shown in the above proof is that
inequality~\eqref{tvf.eq} is sharp for the space ${\mathcal{X}} = \left\{1, 2
\right\}$. However, the result holds in more general spaces as
well. For example, if the space is such that there exist two disjoint
nonempty subsets $A_1$ and $A_2$ and two probability measures $\nu_1$
and $\nu_2$ concentrated on $A_1$ and $A_2$ respectively, then we can
define $P_1 := \nu_1(1+V)/2 + \nu_2 (1-V)/2$ and $P_2 := \nu_1(1-V)/2
+ \nu_2 (1+V)/2$ so that $V(P_1, P_2) = V$ and~\eqref{tvf.eq} becomes
an equality (with $Q = \nu_1/2 + \nu_2/2$).
There exist many inequalities in the literature relating the
$f$-divergence of two probability measures to their total variation
distance. We refer the reader to~\cite{genpin} for the sharpest
results in this direction and for earlier
references. Inequality~\eqref{tvf.eq}, which is new, can be trivially
converted into an inequality between $D_f(P_1||P_2)$ and $V$ by taking
$Q = P_2$. The resulting inequality will not be sharp however and
hence will be inferior to the inequalities in~\cite{genpin}. As
stated, inequality~\eqref{tvf.eq} is a sharp inequality relating not
$D_f(P_1||P_2)$ but a symmetrized form of $f$-divergence between $P_1$
and $P_2$ to their total variation distance.
In the remainder of this section, we shall apply Theorem~\ref{maha} and
Corollary~\ref{Neq2} to specific $f$-divergences.
\begin{example}[Kullback-Leibler Divergence]\label{kld}\normalfont
Let $f(x) := x \log x$. Then $D_f(P||Q)$ becomes the Kullback-Leibler
divergence $D(P||Q)$ between $P$ and $Q$. The quantity $\sum_{\theta \in
F}D(P_{\theta}||Q)$ is minimized when $Q =
\bar{P} := (\sum_{\theta \in F}P_{\theta})/N$. This is a consequence
of the following identity which is sometimes referred to as the
\textit{compensation identity}, see for example~\cite[Page 1603]{TopsoeCapDiv}:
\begin{equation*}
\sum_{\theta \in F}D(P_{\theta}|| Q) = \sum_{\theta \in F}
D(P_{\theta}|| \bar{P}) + N D(\bar{P} || Q).
\end{equation*}
Using inequality~\eqref{maha.eq1} with $Q = \bar{P} = (\sum_{\theta \in
F}P_{\theta})/N$, we obtain
\begin{equation*}
\frac{1}{N} \sum_{\theta \in F} D(P_{\theta}||\bar{P}) \geq
(1-\bar{r})\log(N(1-\bar{r})) + \bar{r} \log \left( \frac{N
\bar{r}}{N-1} \right) .
\end{equation*}
The quantity on the left hand side is known as the
Jensen-Shannon divergence. It is also Shannon's mutual
information~\cite[Page 19]{CoverThomas} between the random parameter
$\theta$ distributed according to the uniform distribution on $F$ and
the observation $X$ whose conditional distribution given $\theta$
equals $P_{\theta}$. The above inequality is stronger
than the version of Fano's inequality commonly used in nonparametric
statistics. It is
implicit in~\cite[Proof of Theorem 1]{HanVerdu} and is explicitly stated in
a slightly different form in~\cite[Theorem
3]{BirgeFano}. The proof in~\cite{HanVerdu} is based on the Fano's inequality from
information theory~\cite[Theorem 2.10.1]{CoverThomas}. To obtain the
usual form of Fano's inequality as used in statistics, we turn to
inequality~\eqref{explow.eq}. For $a_0 := (N-1)/(2N-1) \leq 1-1/N$ and the
function $g$ in~\eqref{gintang}, it can be checked that
\begin{equation*}
g(a_0) = \frac{N^2}{2N-1} \log N + N \log \left( \frac{N}{2N-1}
\right)
\end{equation*}
and $g'(a_0) = -N \log N$. Using inequality~\eqref{explow.eq} with
$a = a_0$, we get that
\begin{equation*}
\bar{r} \geq 1 - \frac{\log ((2N-1)/N) + \frac{1}{N}\sum_{\theta \in
F}D(P_{\theta}||\bar{P}) }{\log N}.
\end{equation*}
Since $\log ((2N-1)/N) \leq \log 2$, we have obtained
\begin{equation}\label{normfano}
r \geq \bar{r} \geq 1 - \frac{\log 2 + \frac{1}{N}\sum_{\theta \in
F}D(P_{\theta}||\bar{P}) }{\log N},
\end{equation}
which is the commonly used version of Fano's inequality.
By taking $f(x) = x\log x$ in Corollary~\ref{Neq2}, we get that
\begin{equation*}
D(P_1||\bar{P}) + D(P_2||\bar{P}) \geq (1+V)\log (1 + V) + (1-V)\log (1 - V).
\end{equation*}
This inequality relating the Jensen-Shannon divergence between two
probability measures (also known as
capacitory discrimination) to their total variation distance is due to
Tops{\o}e~\cite[Equation (24)]{TopsoeCapDiv}. Our proof is slightly simpler
than Tops{\o}e's. Tops{\o}e~\cite{TopsoeCapDiv} also explains how to
use this inequality to deduce Pinsker's inequality with sharp
constant: $D(P_1|| P_2) \geq 2 V^2$. Thus, Theorem~\ref{maha} can be
considered as a generalization of both Fano's inequality and Pinsker's
inequality to $f$-divergences.
\end{example}
\begin{example}[Chi-Squared Divergence]\label{csd}\normalfont
Let $f(x) = x^2-1$. Then $D_f(P||Q)$ becomes the chi-squared
divergence $\chi^2(P|| Q) := \int p^2/q -1$. The function $g$ can be
easily seen to satisfy
\begin{equation*}
g(a) = \frac{N^3}{N-1} \left(1-\frac{1}{N} - a \right)^2 \geq N^2
\left(1-\frac{1}{N} - a \right)^2.
\end{equation*}
Because $\bar{r} \leq 1-1/N$, we can invert the inequality
$\inf_Q\sum_{\theta \in F} \chi^2(P_{\theta}||Q) \geq g(\bar{r})$ to
obtain
\begin{equation}\label{chifano}
r \geq \bar{r} \geq 1 - \frac{1}{N} - \frac{1}{\sqrt{N}}
\sqrt{\frac{\inf_Q \sum_{\theta \in F} \chi^2(P_{\theta}|| Q)}{N}}.
\end{equation}
Also it follows from Corollary~\ref{Neq2} that for every two probability
measures $P_1$ and $P_2$,
\begin{equation}\label{tridis}
\inf_Q \left( \chi^2(P_1||Q) + \chi^2(P_2||Q) \right) \geq 2 V^2.
\end{equation}
The weaker inequality $\chi^2(P_1||\bar{P}) + \chi^2(P_2||\bar{P})
\geq 2V^2$ can be found in~\cite[Equation (11)]{TopsoeCapDiv}.
\end{example}
\begin{example}[Hellinger Distance]\label{helld}\normalfont
Let $f(x) = 1-\sqrt{x}$. Then $D_f(P||Q) = 1 - \int \sqrt{pq} d\mu =
H^2(P, Q)/2$, where $H^2(P,Q) = \int (\sqrt{p} - \sqrt{q})^2 d
\mu$ is the square of the Hellinger distance between $P$ and $Q$. It
can be shown, using the Cauchy-Schwarz inequality, that
$\sum_{\theta \in F}D_f(P_{\theta}||Q)$ is minimized when $Q$ has a
density with respect to $\mu$ that is proportional to $(\sum_{\theta
\in F}\sqrt{p_{\theta}})^2$. Indeed if $u := \sum_{\theta \in F}
\sqrt{p_{\theta}}$, then
\begin{align*}
\sum_{\theta \in F} D_f(P_{\theta}||Q) = N - \int \sqrt{qu^2}d\mu
\geq N - \sqrt{\int u^2 d\mu},
\end{align*}
by the Cauchy-Schwarz inequality with equality when $q$ is
proportional to $u^2$. The inequality~\eqref{maha.eq1} can then be
simplified to
\begin{equation}\label{helldtemp}
\sqrt{1-\bar{r}} + \sqrt{(N-1)\bar{r}} \geq \sqrt{\frac{\int u^2 d\mu}{N}}.
\end{equation}
We now observe that
\begin{equation*}
\int u^2 d\mu = N + \sum_{\theta \neq \theta '} \int
\sqrt{p_{\theta} p_{\theta '}} d\mu = N^2 - \frac{1}{2} \sum_{\theta
\neq \theta '} H^2(P_{\theta}, P_{\theta '}).
\end{equation*}
We let $h^2 := \sum_{\theta, \theta'}H^2(P_{\theta},P_{\theta'})/N^2$ so
that $\int u^2 d\mu = N^2(1-h^2/2)$. As a consequence, we have $\int
u^2 d\mu \leq N^2$. Also note that $\int u^2 d\mu \geq \int (\sum_{\theta}p_{\theta})
d\mu = N$. Therefore, the right hand side of the
inequality~\eqref{helldtemp} lies between 1 and $\sqrt{N}$. On the
other hand, it can be checked that, as a function of $\bar{r}$, the
left hand side of~\eqref{helldtemp} is strictly increasing from $1$ at
$\bar{r}=0$ to $\sqrt{N}$ at $\bar{r} = 1-1/N$. It therefore follows that
inequality~\eqref{helldtemp} is equivalent to $\bar{r} \geq \breve{r}$ where
$\breve{r} \in [0, 1-1/N]$ is the solution to the equation obtained by
replacing the inequality in~\eqref{helldtemp} with an equality.
This equation can be solved in the usual way by squaring etc., until we
get a quadratic equation in $\bar{r}$ which can be solved resulting in
two solutions. One of the two solutions can be discarded by continuity
considerations (the solution has to be continuous in $\int u^2
d\mu/N$) and the fact that $\bar{r} \leq 1-1/N$. The other solution
equals $\breve{r}$ and is given by
\begin{equation*}
\breve{r} = 1-\frac{1}{N} - \frac{N-2}{N}\frac{h^2}{2} -
\frac{\sqrt{N-1}}{N} \sqrt{h^2(2-h^2)}.
\end{equation*}
We have thus shown that
\begin{equation*}
r \geq \bar{r} \geq 1-\frac{1}{N} - \frac{N-2}{N}\frac{h^2}{2} -
\frac{\sqrt{N-1}}{N} \sqrt{h^2(2-h^2)}.
\end{equation*}
In the case when $N=2$ and $F = \left\{1, 2 \right\}$, it is clear
that $h^2 =(H^2(P_1, P_2) + H^2(P_2, P_1))/4= H^2(P_1, P_2)/2$. Also
since $2 \bar{r}$ equals $1-V$,
where $V$ denotes the total variation distance between $P_1$ and
$P_2$, the above inequality implies that for every pair of probability
measures $P_1$ and $P_2$, we have
\begin{equation*}
V \leq H(P_1, P_2) \sqrt{1- \frac{H^2(P_1, P_2)}{4}}.
\end{equation*}
This inequality is usually attributed to Le Cam~\cite{LeCam:86book}.
\end{example}
\begin{example}[Total Variation Distance]\normalfont
Let $f(x) = |x-1|/2$. Then $D_f(P||Q)$ becomes the total variation
distance between $P$ and $Q$. The function $g$ satisfies
\begin{equation*}
g(\bar{r}) = \frac{1}{2} |N(1-\bar{r}) - 1| + \frac{N-1}{2}
\left|\frac{N\bar{r}}{N-1} -1 \right|.
\end{equation*}
Since $\bar{r} \leq 1-1/N$, we have $N(1-\bar{r}) \geq 1$ and
$N\bar{r}/(N-1) \leq 1$ so that the above expression for $g(\bar{r})$
simplifies to $N-1-N\bar{r}$. Inequality~\eqref{maha.eq1}, therefore,
results in
\begin{equation*}
r \geq \bar{r} \geq 1 - \frac{1}{N} - \frac{\inf_Q \sum_{\theta \in
F}V_{\theta}}{N}.
\end{equation*}
where $V_{\theta}$ denotes the total variation distance between
$P_{\theta}$ and $Q$.
\end{example}
\begin{example}\label{lchi}\normalfont
Let $f(x) = x^l - 1$ where $l>1$. The case $l = 2$ has already
been considered in Example~\ref{csd}. The function $g$ has the
expression
\begin{equation*}
g(\bar{r}) = N^l(1-\bar{r})^l - N + (N-1)\left(\frac{N\bar{r}}{N-1}\right)^l.
\end{equation*}
It therefore follows that $\inf_Q \sum_{\theta \in F}
D_f(P_{\theta}||Q) \geq g(\bar{r}) \geq N^l(1-\bar{r})^l - N$ which
results in the inequality
\begin{equation}\label{lchi.eq}
r \geq \bar{r} \geq 1 - \left(\frac{1}{N^{l-1}}+ \frac{\inf_Q \sum_{\theta \in F}
D_f(P_{\theta}||Q)}{N^l} \right)^{1/l} .
\end{equation}
When $l = 2$, inequality~\eqref{lchi.eq} results in a bound that is weaker than
inequality~\eqref{chifano} although for large $N$, the two bounds are
almost the same.
\end{example}
\begin{example}[\textit{Reverse} Kullback-Leibler divergence]\normalfont
Let $f(x) = -\log x$ so that $D_f(P||Q) = D(Q|| P)$. Then from
Corollary~\ref{Neq2}, we get that for every two probability measures
$P_1$ and $P_2$,
\begin{equation*}
\inf_Q \left\{ D(Q|| P_1) + D(Q|| P_2) \right\} \geq \log \left(
\frac{1}{1 - V^2}\right).
\end{equation*}
This can be rewritten to get
\begin{equation}\label{myrevpins}
V \leq \sqrt{1 - \exp \left(- \inf_Q \left\{ D(Q|| P_1) + D(Q|| P_2)
\right\}\right)}.
\end{equation}
Unlike Example~\ref{kld}, it is not true that $D(Q|| P_1) + D(Q||
P_2)$ is minimized when $Q = \bar{P}$. This is easy to see because
$D(\bar{P}, P_1) + D(\bar{P}, P_2)$ is finite only when $P_1 << P_2$
and $P_2 << P_1$. By taking $Q = P_1$ and $Q = P_2$, we get
that
\begin{equation*}
V \leq \sqrt{1 - \exp \left( - \min \left( D(P_1|| P_2), D(P_2||
P_1)\right) \right)}.
\end{equation*}
The above inequality, which is clearly weaker than
inequality~\eqref{myrevpins}, can also be found in~\cite[Proof of
Lemma 2.6]{Tsybakovbook}.
\end{example}
\section{Bounds for $J_f$}\label{jsbounds}
In order to apply the minimax lower bounds of the previous section in
practical situations, we
must be able to bound the quantity $J_f := \inf_Q \sum_{\theta
\in F} D_f(P_{\theta}||Q)/N$ from above. We shall provide such bounds in
this section. It should be noted that for some functions $f$,
it may be possible to calculate $J_f$ directly. For example, the
quantity $\inf_Q \sum_{\theta \in F}H^2(P_{\theta},Q)$ can be written
in terms of pairwise Hellinger distances (Example~\ref{helld}) and may
be calculated exactly for certain probability measures
$P_{\theta}$. This is not the case for most functions $f$ however.
The following is a simple upper bound for $J_f$ which, in the case
when $f(x) = x \log x$ or Kullback-Leibler divergence, has been
frequently used in the literature (see for example~\cite{Birge83}
and~\cite{nemirovski2000}).
\begin{align*}
J_f &\leq \frac{1}{N} \sum_{\theta \in F}D_f(P_{\theta}||\bar{P})
\\
&\leq \frac{1}{N^2} \sum_{\theta, \theta ' \in
F}D_f(P_{\theta}||P_{\theta '}) \leq \max_{\theta, \theta ' \in F}
D_f(P_{\theta}||P_{\theta '}).
\end{align*}
We observed in section~\ref{test} that $J_f$ measures the
\textit{spread} of the probability measures $P_{\theta}, \theta \in F$
i.e., how tightly packed/spread out they are. It should be clear that
the simple bound $\max_{\theta, \theta '}D_f(P_{\theta}||P_{\theta
'})$ does not adequately describe this aspect of $P_{\theta}, \theta
\in F$ and it is therefore desirable to look for alternative upper
bounds for $J_f$ that capture the notion of spread in a better way.
In the case of the Kullback-Leibler divergence, Yang and
Barron~\cite[Page 1571]{YangBarron} provided such an upper bound for
$J_f$. They showed that for any finite set $\left\{Q_{\alpha}: \alpha
\in G \right\}$ of probability measures,
\begin{equation}\label{jsyb}
\frac{1}{N} \sum_{\theta \in F} D(P_{\theta}||\bar{P}) \leq
\log |G| + \max_{\theta \in F} \min_{\alpha \in
G}D(P_{\theta}||Q_{\alpha}).
\end{equation}
Let us now take a closer look at this beautiful inequality of Yang and
Barron~\cite{YangBarron}. The $|G|$ probability measures $Q_{\alpha},
\alpha \in G$ can be viewed as an approximation of the $N$ probability
measures $P_{\theta}, \theta \in F$. The term $\max_{\theta}
\min_{\alpha}D(P_{\theta}||Q_{\alpha})$ then denotes the approximation
error, measured via the Kullback-Leibler divergence. The right hand
side of inequality~\eqref{jsyb} can therefore be made small if it is
possible to choose not too many probability measures $Q_{\alpha}$ which
well approximate the given set of probability measures $P_{\theta}$.
It should be clear how the upper bound~\eqref{jsyb} measures the spread of
the probability measures $P_{\theta}, \theta \in F$. If the
probabilities are tightly packed, it is possible to approximate them
well with a smaller set of probabilities and then the bound will be
small. On the other hand, if $P_{\theta}, \theta \in F$ are well
spread out, we need more probability measures for approximation and
consequently the bound will be large.
Another important aspect of inequality~\eqref{jsyb} is that it can be
used to obtain lower bounds for $R$ depending only on global metric
entropy properties of the parameter space $\Theta$ and the space of
probability measures $\left\{P_{\theta}, \theta \in \Theta
\right\}$ (see section~\ref{gent}). On the other
hand, the evaluation of inequalities resulting
from the use of $J_f \leq \max_{\theta,
\theta '} D(P_{\theta}||P_{\theta '})$ requires knowledge of both metric
entropy and the existence of certain special localized subsets. We
refer the reader to~\cite{YangBarron} for a detailed
discussion of these issues.
The goal of this section is to generalize inequality~\eqref{jsyb}
to $f$-divergences. The main result is given below. In
section~\ref{gent}, we shall use this theorem along with the results
of the previous section to come up with minimax lower bounds involving
global entropy properties.
\begin{theorem}\label{ybgen}
Let $Q_{\alpha}, \alpha \in G$ be $M := |G|$ probability measures
having densities $q_{\alpha}, \alpha \in G$ with respect to $\mu$ and
let $j:F \rightarrow G$ be a mapping from $F$ to $G$. For every convex
function $f: [0, \infty) \rightarrow {\mathbb R}$, we have
\begin{equation}\label{ybgen.eq}
J_f \leq
\frac{1}{N}\sum_{\theta \in F} \int_{{\mathcal{X}}} \frac{q_{j(\theta)}}{M} f \left(
\frac{Mp_{\theta}}{q_{j(\theta)}} \right) d\mu + \left( 1 - \frac{1}{M}
\right) f(0).
\end{equation}
\end{theorem}
\begin{IEEEproof}
Let $\bar{Q} := \sum_{\alpha \in G} Q_{\alpha}/M$ and $\bar{q} :=
\sum_{\alpha \in G} q_{\alpha}/M$. Clearly for each $\theta \in F$, we
have
\begin{equation*}
D_f(P_{\theta}||\bar{Q}) = \int_{{\mathcal{X}}} \bar{q} \left[f \left( \frac{p_{\theta}}{\bar{q}}\right)
- f(0)\right] d\mu + f(0).
\end{equation*}
The convexity of $f$ implies that the map $y
\mapsto y[f(a/y) - f(0)]$ is non-increasing for every nonnegative
$a$. Using this and the fact that $\bar{q} \geq q_{j(\theta)}/M$, we
get that for every $\theta \in F$,
\begin{equation*}
D_f(P_{\theta}||\bar{Q}) \leq
\int_{{\mathcal{X}}} \frac{q_{j(\theta)}}{M} \left[ f \left(
\frac{Mp_{\theta}}{q_{j(\theta)}} \right) - f(0) \right]d\mu + f(0).
\end{equation*}
Inequality~\eqref{ybgen.eq} now follows as a consequence of the inequality $J_f \leq \sum_{\theta \in
F} D_f(P_{\theta}||\bar{Q})/N$.
\end{IEEEproof}
In the following examples, we shall demonstrate that Theorem~\ref{ybgen}
is indeed a generalization of the bound~\eqref{jsyb} to
$f$-divergences. We shall also see that Theorem~\ref{ybgen}
results in inequalities that have the same qualitative structure
as~\eqref{jsyb}, at least for the convex functions $f$ of interest such as
$x^l-1, l > 1$ and $(\sqrt{x} - 1)^2$.
\begin{example}[Kullback-Leibler divergence]\normalfont
Let $f(x) = x\log x$. In this case, $J_f$ equals $\sum_{\theta \in F}
D(P_{\theta}||\bar{P})/N$ and invoking inequality~\eqref{ybgen.eq}, we get that
\begin{equation*}
\frac{1}{N} \sum_{\theta \in F} D(P_{\theta}||\bar{P}) \leq \log M +
\frac{1}{N} \sum_{\theta \in F} D(P_{\theta}||Q_{j(\theta)}).
\end{equation*}
Inequality~\eqref{jsyb} now follows if we choose $j(\theta) := \arg \min_{\alpha \in
G} D(P_{\theta}||Q_{\alpha})$. Hence Theorem~\ref{ybgen} is indeed a
generalization of~\eqref{jsyb}.
\end{example}
\begin{example}\label{chijsd}\normalfont
Let $f(x) = x^l -1$ for $l > 1$. Applying
inequality~\eqref{ybgen.eq}, we get that
\begin{equation*}
J_f \leq M^{l-1}\left( \frac{1}{N} \sum_{\theta \in F}
D_f(P_{\theta}||Q_{j(\theta)}) + 1 \right) - 1.
\end{equation*}
By choosing $j(\theta) = \arg \min_{\alpha \in G}
D_f(P_{\theta}||Q_{\alpha})$, we get that
\begin{equation}\label{ljsd.eq}
J_f \leq M^{l-1} \left( \max_{\theta \in F} \min_{\alpha \in G}
D_f(P_{\theta}||Q_{\alpha}) + 1 \right) - 1.
\end{equation}
In particular, in the case of the chi-squared divergence i.e., when
$l=2$, the quantity $J_f = \inf_Q \sum_{\theta \in F}
\chi^2(P_{\theta}||Q)/N$ is bounded from above by
\begin{equation}\label{chijsd.eq}
M \left( \max_{\theta \in F} \min_{\alpha \in
G}\chi^2(P_{\theta}||Q_{\alpha}) + 1 \right) - 1.
\end{equation}
Just like~\eqref{jsyb}, each of the above two inequalities is also a
function of the number of probability measures $Q_{\alpha}$ and the
approximation error which is now measured in terms of the chi-squared
divergence.
\end{example}
\begin{example}[Hellinger distance]\normalfont
Let $f(x) = (\sqrt{x} - 1)^2$ so that $D_f(P|| Q) = H^2(P, Q)$, the
square of the Hellinger distance between $P$ and $Q$. Using
inequality~\eqref{ybgen.eq}, we get that
\begin{equation*}
J_f\leq 2 - \frac{1}{\sqrt{M}} \left(2 - \frac{1}{N} \sum_{\theta \in
F} H^2(P_{\theta}, Q_{j(\theta)}) \right).
\end{equation*}
If we now choose $j(\theta) := \arg \min_{\alpha \in G}
H^2(P_{\theta}, Q_{\alpha})$, then we get
\begin{equation*}
J_f \leq 2 - \frac{1}{\sqrt{M}} \left(2 - \max_{\theta \in F}
\min_{\alpha \in G} H^2(P_{\theta}, Q_{\alpha}) \right).
\end{equation*}
Notice, once again, the trade-off between $M$ and the
approximation error which is measured in terms of the Hellinger
distance.
\end{example}
\section{Bounds involving global entropy}\label{gent}
In this section, we shall apply the results of the previous two
sections to obtain lower bounds for the minimax risk $R$ depending
only on global metric entropy properties of the parameter space. The
theorem is stated below, but we shall need to establish some notation
first.
\begin{enumerate}
\item For $\eta > 0$, let $N(\eta) \geq 1$ be a real number for
which there exists a finite subset $F \subseteq \Theta$ with
cardinality $\geq N(\eta)$ satisfying $\rho(\theta, \theta ') \geq \eta$ whenever
$\theta, \theta ' \in F$ and $\theta \neq \theta '$. In other words,
$N(\eta)$ is a lower bound on the $\eta$-packing number of the metric
space $(\Theta, \rho)$.
\item For a convex function $f: [0, \infty) \rightarrow {\mathbb R}$ satisfying $f(1)=0$, a subset $S
\subseteq \Theta$ and a positive real number $\epsilon$, let
$M_f(\epsilon, S)$ be a positive real number for which there
exists a finite set $G$ with cardinality $\leq M_f(\epsilon, S)$
and probability measures $Q_{\alpha}, \alpha \in G$ such that
$\sup_{\theta \in S} \min_{\alpha \in G} D_f(P_{\theta}||Q_{\alpha})
\leq \epsilon^2$. In other words, $M_{f}(\epsilon, S)$ is an upper bound on the
$\epsilon$-covering number of the space $\left\{P_{\theta}: \theta \in
S \right\}$ when distances are measured by the square root of the
$f$-divergence. For purposes of clarity, we write $M_{KL}(\epsilon,
S), M_C(\epsilon, S)$ and $M_l(\epsilon, S)$ for $M_f(\epsilon, S)$
when the function $f$ equals $x \log x$, $x^2 - 1$ and $x^l - 1$ and respectively.
\end{enumerate}
We note here that the probability measures $Q_{\alpha}, \alpha \in G$
in the definition of $M_f(\epsilon, S)$ do not need to be included in
the set $\left\{P_{\theta}, \theta \in \Theta \right\}$ and the set
$G$ just denotes the index set and need not have any relation to $S$
or $\Theta$.
\begin{theorem}\label{myb.thm}
The minimax risk $R$ satisfies the inequality $R \geq \sup_{\eta>0,
\epsilon >0} \ell(\eta/2) (1-\star)$ where $\star$ stands for
any of the following quantities
\begin{equation}\label{myb}
\frac{\log 2 + \log M_{KL}(\epsilon, \Theta) + \epsilon^2}{\log
N(\eta)}
\end{equation}
\begin{equation}\label{myb.chi}
\frac{1}{N(\eta)} +
\sqrt{\frac{(1+\epsilon^2)M_C(\epsilon, \Theta)}{N(\eta)}}
\end{equation}
and for $l > 1, l \neq 2$,
\begin{equation}\label{myb.l}
\left( \frac{1}{N(\eta)^{l-1}} +
\frac{(1+\epsilon^2)M_l(\epsilon, \Theta)^{l-1}}{N(\eta)^{l-1}} \right)^{1/l}.
\end{equation}
\end{theorem}
In the sequel, by inequality~\eqref{myb.chi}, we mean the inequality
$R \geq \sup_{\eta >0, \epsilon >0} \ell(\eta/2)(1 - \star)$ with $\star$
representing~\eqref{myb.chi} and similarly for inequalities~\eqref{myb}
and~\eqref{myb.l}.
\begin{IEEEproof}
We shall give the proof of inequality~\eqref{myb.chi}. The remaining
two inequalities are proved in a similar manner.
Fix $\eta > 0$. By the definition of $N(\eta)$, one can find a
finite subset $F \subset \Theta$ with cardinality $|F| \geq N(\eta)$
such that $\rho(\theta, \theta ') \geq \eta$ for $\theta, \theta
' \in F$ and $\theta \neq \theta '$. We then employ the inequality $R
\geq \ell(\eta/2) r$, where $r$ is defined as
in~\eqref{rdef}. Inequality~\eqref{chifano} can now be used to obtain
\begin{equation*}
r \geq 1 - \frac{1}{\sqrt{|F|}}
\sqrt{\frac{\inf_Q \sum_{\theta \in F} \chi^2(P_{\theta}|| Q)}{|F|}} - \frac{1}{|F|}.
\end{equation*}
We now fix $\epsilon > 0$ and use the definition of $M_C(\epsilon,
F)$ to get a finite set $G$ with cardinality $\leq M_C(\epsilon, F)$
and probability measures $Q_{\alpha}, \alpha \in G$ such that
$\sup_{\theta \in S} \min_{\alpha \in G} \chi^2(P_{\theta}||Q_{\alpha})
\leq \epsilon^2$.
We then use inequality~\eqref{chijsd.eq} to get that
\begin{equation*}
\inf_Q \frac{1}{|F|} \sum_{\theta \in F} \chi^2(P_{\theta}||Q) \leq M_C(\epsilon, F)
\left( 1 + \epsilon^2 \right) - 1.
\end{equation*}
The proof is complete by the trivial observation $M_C(\epsilon, F)
\leq M_C(\epsilon, \Theta)$.
\end{IEEEproof}
The inequality~\eqref{myb} is due to Yang and Barron~\cite[Proof of
Theorem 1]{YangBarron}. In their paper, Yang and Barron mainly
considered the problem of estimation from $n$ independent and
identically distributed observations. However their method results in
inequality~\eqref{myb} which applies to every estimation
problem. Inequalities~\eqref{myb.chi} and~\eqref{myb.l} are new.
Note that the lower bounds for $R$ given in
Theorem~\ref{myb.thm} all depend only on the
quantities $N(\eta)$ and $M_f(\epsilon, \Theta)$, which describe
packing/covering properties of the entire parameter space
$\Theta$. Consequently, these inequalities only involve global metric entropy
properties. This is made possible by the use of
inequalities in Theorem~\ref{ybgen}. In applications of Fano's
inequality~\eqref{normfano} with the standard bound $J_f \leq
\max_{\theta, \theta ' \in F} D(P_{\theta}||P_{\theta '})$ as
well as in the application of other popular methods for obtaining
minimax lower bounds like Le Cam's method or Assouad's lemma, one
needs to construct the finite subset $F$ of the parameter space in a
very special way: the parameter values in $F$ should be reasonably
separated in the metric $\rho$ and also, the probability measures
$P_{\theta}, \theta \in F$ should be close in some probability
metric. In contrast, the application of Theorem~\ref{myb.thm}
does not require the construction of such a special subset $F$.
Yang and Barron~\cite{YangBarron} have successfully applied
inequality~\eqref{myb} to achieve minimax lower bounds of the optimal
rate for
many nonparametric density estimation and regression problems where
$N(\eta)$ and $M_{KL}(\epsilon, \Theta)$ can be deduced from standard
results in approximation theory for function classes. We
refer the reader to~\cite{YangBarron} for
examples. In some of these examples, inequality~\eqref{myb.chi} can
also be applied to get optimal lower bounds. In section~\ref{suppest},
we shall employ inequality~\eqref{myb.chi} to obtain a new minimax
lower bound in the problem of reconstructing convex bodies from noisy
support function measurements.
But prior to that, let us assess the performance of
inequality~\eqref{myb.chi} in certain standard parametric estimation
problems. In these problems, an interesting contrast arises between
the two minimax lower bounds~\eqref{myb} and~\eqref{myb.chi}: the
inequality~\eqref{myb} only results in a sub-optimal lower bound on
the minimax risk (this observation, due to Yang and Barron~\cite[Page
1574]{YangBarron}, is also explained in Example~\ref{parnorm} below)
while~\eqref{myb.chi} produces rate-optimal lower bounds.
Our intention here is to demonstrate, with the help of the subsequent three
examples, that inequality~\eqref{myb.chi} works even
for finite dimensional parametric estimation problems, a scenario in which
it is already known~\cite[Page 1574]{YangBarron} that inequality~\eqref{myb}
fails. Of course, obtaining optimal minimax rates in
such problems is facile in most situations. For
example, a two-points argument based on Hellinger distance gives the
optimal rate, as is widely recognized since Le
Cam~\cite{LeCam:73AnnStat}. But the point here is that even in finite
dimensional situations, global metric entropy features are adequate
for obtaining rate-optimal minimax lower bounds. This is contrary to
the usual claim that in order to establish rate-optimal lower bounds
in parametric settings, one needs more information than global entropy
characteristics~\cite[Page 1574]{YangBarron}.
In each of the ensuing three examples, we take the parameter space
$\Theta$ to be a bounded interval of the real line and we consider the problem of
estimating a parameter $\theta \in \Theta$ from $n$ independent
observations distributed according to $m_{\theta}$, where $m_{\theta}$
is a probability measure on the real line. The probability measure
$P_{\theta}$ accordingly equals the $n$-fold product of
$m_{\theta}$. We shall work with the squared error loss so that $\ell(x)
= x^2$, $\rho$ is the Euclidean distance on the real line and
$N(\eta)$ can be taken to $c_1/\eta$ for $\eta \leq \eta_0$
where $c_1$ and $\eta_0$ are positive constants depending on the
bounded parameter space alone. We shall encounter more positive
constants $c, c_2, c_3, c_4, c_5, \epsilon_0$ and $\epsilon_1$ in the
examples all of which depend possibly on the parameter space alone and
thus, independent of $n$.
\begin{example}\normalfont
\label{parnorm}
Suppose that $m_{\theta}$ equals the normal distribution with mean $\theta$
and variance 1. It can be readily verified that, for $\theta, \theta '
\in \Theta$, one has
\begin{equation*}
D(P_{\theta}||P_{\theta '}) = \frac{n}{2}|\theta - \theta '|^2
\end{equation*}
and
\begin{equation*}
\chi^2(P_{\theta}||P_{\theta '}) = \exp \left(n|\theta
- \theta '|^2\right)-1.
\end{equation*}
It follows directly that $D(P_{\theta}||P_{\theta '}) \leq
\epsilon^2$ if and only if
$|\theta - \theta '| \leq \sqrt{2} \epsilon/\sqrt{n}$ and
$\chi^2(P_{\theta}||P_{\theta '}) \leq \epsilon^2$ if and only if
$|\theta - \theta '| \leq \sqrt{\log(1+\epsilon^2)}/\sqrt{n}$. As a
result, we can take
\begin{equation}\label{ucp}
M_{KL}(\epsilon, \Theta) = \frac{c_2
\sqrt{n}}{\epsilon} \text{ and } M_C(\epsilon, \Theta) = \frac{c_2
\sqrt{n}}{\sqrt{\log (1+\epsilon^2)}}
\end{equation}
for $\epsilon \leq \epsilon_0$.
Now, inequality~\eqref{myb} says that the minimax risk $R_n$ satisfies
\begin{equation*}
R_n \geq \sup_{\eta \leq \eta_0, \epsilon \leq \epsilon_0} \frac{\eta^2}{4}
\left( 1 - \frac{\log 2 + \log (c_2 \sqrt{n}/\epsilon) +
\epsilon^2}{\log (c_1/\eta)}\right).
\end{equation*}
The function $\epsilon \mapsto \epsilon^2 - \log \epsilon$ is
minimized on $[0, \epsilon_0]$ at, say, $\epsilon = \epsilon_1$ and we
then get
\begin{equation}\label{noryb}
R_n \geq \sup_{\eta \leq \eta_0} ~\frac{\eta^2}{4} \left(1 -
\frac{\log n + c_3}{2\log c_1 + 2 \log(1/\eta)} \right),
\end{equation}
where $c_3$ is a function of $c_2$ and $\epsilon_1$. We now note that
when $\eta = c/\sqrt{n}$ for a constant $c$, the
quantity inside the parantheses on the right hand side
of~\eqref{noryb} converges to 0 as $n$ goes to $\infty$. This means
that inequality~\eqref{myb} only gives lower bounds of inferior order
for $R_n$, the optimal order being, of course, $1/n$.
On the other hand, we shall show below that
inequality~\eqref{myb.chi} gives $R_n \geq c/n$ for a positive
constant $c$. Indeed, inequality~\eqref{myb.chi} says that
\begin{equation*}
R_n \geq \sup_{\eta \leq \eta_0, \epsilon \leq \epsilon_0}~\frac{\eta^2}{4}
\left(1 - \frac{\eta}{c_1} - \sqrt{\eta
\sqrt{n}}\sqrt{\frac{c_2(1+\epsilon^2)}{c_1\sqrt{\log(1+\epsilon^2)}}}\right).
\end{equation*}
Taking $\epsilon = \epsilon_0$ and $\eta = c_3/\sqrt{n}$, we get
\begin{equation}
R_n \geq \frac{c_3^2}{4n} \left(1 -
\frac{c_3}{c_1\sqrt{n}} - c_4 \sqrt{c_3} \right),
\end{equation}
where $c_4$ depends only on $c_1, c_2$ and $\epsilon_0$. Hence by
choosing $c_3$ small, we get that $R_n \geq c/n$ for all large $n$.
\end{example}
\begin{example}\normalfont
Suppose that $\Theta$ is a compact interval of the positive real line
that is bounded away from
zero and suppose that $m_{\theta}$ denotes the uniform distribution on
$[0, \theta]$. It is then elementary to check that the chi-squared
divergence between $P_{\theta}$ and $P_{\theta '}$ equals $(\theta
'/\theta)^n-1$ if $\theta \leq \theta '$ and $\infty$ otherwise. It
follows accordingly that $\chi^2(P_{\theta}||P_{\theta '}) \leq
\epsilon^2$ provided
\begin{equation}\label{mj}
0 \leq \theta ' - \theta \leq \frac{\theta \log (1+\epsilon^2)}{n}.
\end{equation}
Because the parameter space is a compact interval bounded away from
zero, in order to ensure~\eqref{mj}, it is enough to require that $0
\leq \theta ' - \theta \leq c_2 \log(1+\epsilon^2)/n$. Therefore, we
can take
\begin{equation*}
M_C(\epsilon, \Theta) = \frac{c_3n}{\log(1+\epsilon^2)}
\end{equation*}
for $\epsilon \leq \epsilon_0$. Inequality~\eqref{myb.chi} now implies
that
\begin{equation*}
R_n \geq \sup_{\eta \leq \eta_0, \epsilon \leq \epsilon_0} \frac{\eta^2}{4} \left(1 -
\frac{\eta}{c_1} - \sqrt{\eta n}
\sqrt{\frac{c_3(1+\epsilon^2)}{c_1\log(1+\epsilon^2)}} \right).
\end{equation*}
Taking $\epsilon = \epsilon_0$ and $\eta = c_4/n$, we get that
\begin{equation*}
R_n \geq \frac{c_4^2}{4n^2} \left(1 - \frac{c_4}{nc_1} - \sqrt{c_4}
c_5 \right),
\end{equation*}
where $c_5$ depends only on $c_1, c_3$ and $\epsilon_0$.
Hence by choosing $c_4$ sufficiently small, we get that $R_n \geq
c/n^2$ for all large $n$. This is the optimal minimax rate for this
problem as can be seen by estimating $\theta$ by the maximum of the
observations.
\end{example}
\begin{example}\normalfont
Suppose that $m_{\theta}$ denotes the uniform distribution on the
interval $[\theta, \theta + 1]$. We shall argue that $M_C(\epsilon,
\Theta)$ can be chosen to be
\begin{equation}\label{lara}
M_C(\epsilon, \Theta) = \frac{c_2}{(1+\epsilon^2)^{1/n}-1}
\end{equation}
for a positive constant $c_2$ at least for large $n$. To see this, let
us define $\epsilon '$ so that $2\epsilon
':=(1+\epsilon^2)^{1/n}-1$ and let $G$ denote an $\epsilon '$-grid of
points in the interval $\Theta$; $G$ would contain at most
$c_2/\epsilon '$ points when $\epsilon \leq \epsilon_0$. For a point
$\alpha$ in the grid, let $Q_{\alpha}$ denote the $n$-fold product of
the uniform distribution on the interval $[\alpha, \alpha + 1 +
2\epsilon ']$. Now, for a fixed $\theta \in \Theta$, let $\alpha$
denote the point in the grid such that $\alpha \leq \theta \leq \alpha
+ \epsilon '$. It can then be checked that the chi-squared divergence
between $P_{\theta}$ and $Q_{\alpha}$ is equal to $(1+2 \epsilon ')^n
- 1 = \epsilon^2$. Hence $M_C(\epsilon, \Theta)$ can be taken to be
the number of probability measures $Q_{\alpha}$, which is the same as the
number of points in $G$. We thus have~\eqref{lara}. It can be checked
by elementary calculus (Taylor expansion, for example) that the
inequality
\begin{equation*}
(1+\epsilon^2)^{1/n} - 1 \geq \frac{\epsilon^2}{n} -
\frac{1}{2n}\left( 1 - \frac{1}{n}\right)\epsilon^4
\end{equation*}
holds for $\epsilon \leq \sqrt{2}$ (in fact for all $\epsilon$, but
for $\epsilon > \sqrt{2}$, the right hand side above may be
negative). Therefore for $\epsilon \leq \min(\epsilon_0, \sqrt{2})$,
we get that
\begin{equation*}
M_C(\epsilon, \Theta) \leq \frac{2nc_2}{2\epsilon^2 - (1-1/n)\epsilon^4}.
\end{equation*}
From inequality~\eqref{myb.chi}, we get that for every $\eta \leq
\eta_0$ and $\epsilon \leq \min(\epsilon_0, \sqrt{2})$,
\begin{equation*}
R_n \geq \frac{\eta^2}{4} \left(1 - \frac{\eta}{c_1} - \sqrt{n \eta}
\sqrt{\frac{2(1+\epsilon^2)c_2}{c_1 \left(2\epsilon^2 -
(1-1/n)\epsilon^4 \right)}}\right).
\end{equation*}
If we now take $\epsilon = \min(\epsilon_0, 1)$ and $\eta = c_3/n$, we
see that the quantity inside the parantheses converges to
$1-\sqrt{c_3}c_4$ where $c_4$ depends only on $c_1, c_2$ and
$\epsilon_0$. Therefore by choosing $c_3$ sufficiently small, we get
that $R_n \geq c/n^2$. This is the optimal minimax rate for this
problem as can be seen by estimating $\theta$ by the minimum of the
observations.
\end{example}
The fact that inequality~\eqref{myb.chi} produced optimal lower bounds for
the minimax risk in each of the above three examples is reassuring but
not really exciting because, as we mentioned before, there are other simpler
methods of obtaining such bounds in these examples. We presented them
as simple toy examples to evaluate the performance of~\eqref{myb.chi},
to present a difference between~\eqref{myb} and~\eqref{myb.chi} (which
provides a justification for using divergences other than the
Kullback-Leibler divergence for lower bounds) and also to stress the
fact that global packing and covering characteristics
are enough to obtain optimal minimax lower bounds. In order to
convince the reader of the effectiveness of~\eqref{myb.chi} in more
involved situations, we now apply it to obtain the optimal minimax
rate in a $d$-dimensional normal mean estimation problem. We
are grateful to an anonymous referee for communicating this example to
us. Another non-trivial application of~\eqref{myb.chi} is presented in
the next section.
\begin{example}\normalfont
\label{ddimnorm}
Let $\Theta$ denote the ball in ${\mathbb R}^d$ of radius
$\Gamma$ centered at the origin. Let us consider the problem of estimating
$\theta \in \Theta$ from an observation $X$ distributed according to
the normal distribution with mean $\theta$ and variance covariance
matrix $\sigma^2 I_d$, where $I_d$ denotes the
identity matrix of order $d$. Thus $P_{\theta}$ denotes the $N(\theta,
\sigma^2 I_d)$ distribution. We assume squared error loss so that
$\ell(x) = x^2$ and $\rho$ is the Euclidean distance on ${\mathbb R}^d$.
We shall use inequality~\eqref{myb.chi} to show that the minimax risk
$R$ for this problem is larger than or equal to a constant multiple of
$d \sigma^2$ when $\Gamma \geq \sigma \sqrt{d}$. We begin by arguing that
we can take
\begin{equation}\label{pete}
N(\eta) = \left( \frac{\Gamma}{\eta}\right)^d,
M_C(\epsilon, \Theta) = \left(\frac{3\Gamma}{\sigma
\sqrt{\log(1+\epsilon^2)}} \right)^d
\end{equation}
whenever $\sigma \sqrt{\log(1+\epsilon^2)} \leq \Gamma$.
For $N(\eta)$, we first note that the $\eta$-packing number
of the metric space $(\Theta, \rho)$ is bounded from below by its
$\eta$-covering number. Now, for any $\eta$-covering set, the space
$\Theta$ is contained in the union of the balls of radius $\eta$ with
centers in the covering set and hence the volume of $\Theta$ must be
smaller than the sum of the volumes of these balls. Therefore, the
number of points in the $\eta$-covering set must be at least
$(\Gamma/\eta)^d$. Since this is true for every $\eta$-covering set,
it follows that the $\eta$-covering number and hence the
$\eta$-packing number is not smaller than $(\Gamma/\eta)^d$.
For $M_C(\epsilon, \Theta)$, we first observe that for $\theta, \theta
' \in \Theta$, the chi-squared divergence between
$P_{\theta}$ and $P_{\theta '}$ can be easily computed (because they
are normal distributions with the same covariance matrix) to be
$\chi^2(P_{\theta}||P_{\theta '}) = \exp \left(\rho^2(\theta, \theta
')/\sigma^2 \right) - 1$. Therefore $\chi^2(P_{\theta}||P_{\theta
'}) \leq \epsilon^2$ if and only if $\rho(\theta, \theta ') \leq
\epsilon ' := \sigma \sqrt{\log(1+\epsilon^2)}$. As a result,
$M_C(\epsilon, \Theta)$ can be taken to be any upper bound on the
$\epsilon '$-covering number of $(\Theta, \rho)$. The $\epsilon
'$-covering number, as noted previously, is bounded from above by the
$\epsilon '$-packing number. Now, for any $\epsilon '$-packing set,
the balls of radius $\epsilon '/2$ with centers in the packing set are
all disjoint and their union is contained in the ball of radius
$\Gamma + (\epsilon '/2)$ centered at the origin. Consequently, the
sum of the volumes of these balls is smaller than the volume of the
ball of radius $\Gamma + (\epsilon '/2)$ centered at the
origin. Therefore, the number of points in the $\epsilon '$-packing
set is at most $(1+(2\Gamma/\epsilon '))^d \leq (3\Gamma/\epsilon
')^d$ provided $\epsilon ' \leq \Gamma$. Since this is true for every
$\epsilon '$-packing set, it follows that the $\epsilon '$-packing
number and hence the $\epsilon '$-covering number is not larger than
$(3\Gamma/\epsilon ')^d$.
We can thus apply inequality~\eqref{myb.chi} with~\eqref{pete} to get
that, for every $\eta > 0$ and $\epsilon > 0$ such that $\sigma
\sqrt{\log (1+\epsilon^2)} \leq \Gamma$, we have
\begin{equation*}
R \geq \frac{\eta^2}{4} \left(1 - \left(\frac{\eta}{\Gamma}\right)^d
- \left(\frac{3\eta}{\sigma}\right)^{d/2}
\frac{\sqrt{1+\epsilon^2}}{(\log (1+\epsilon^2))^{d/4}} \right).
\end{equation*}
Now by elementary calculus, it can be checked that the function
$\epsilon \mapsto \sqrt{1+\epsilon^2}/(\log (1+\epsilon^2))^{d/4}$ is
minimized (subject to $\sigma \sqrt{\log(1+\epsilon^2)} \leq \Gamma$)
when $1+\epsilon^2 = e^{d/2}$. We then get that
\begin{equation*}
R \geq \sup_{\eta > 0} \frac{\eta^2}{4} \left(1 -
\left(\frac{\eta}{\Gamma}\right)^d -
\left(\frac{18e\eta^2}{\sigma^2d} \right)^{d/4}\right).
\end{equation*}
We now take $\eta = c_1 \sigma \sqrt{d}$ and since $\Gamma \geq \sigma
\sqrt{d}$, we obtain
\begin{equation*}
R \geq \frac{c_1^2\sigma^2d}{4} \left(1 - c_1^d - (18ec_1^2)^{d/4} \right).
\end{equation*}
We can therefore choose $c_1$ small enough (independent of $d$) to
obtain that $R \geq cd\sigma^2$. Note that, up to constants, this
lower bound is optimal for $R$ because ${\mathbb E} \rho^2(X, \theta) =
d\sigma^2$.
\end{example}
\section{Reconstruction of convex bodies from noisy support function
measurements}\label{suppest}
In this section, we shall present a novel application of the global minimax
lower bound~\eqref{myb.chi}. Let $d \geq 2$ and let $K$ be a convex
body in ${\mathbb R}^d$, i.e., $K$ is compact, convex and has a non-empty
interior. The support function of $K$, $h_K : S^{d-1} \rightarrow {\mathbb R}$,
is defined by
\begin{equation*}
h_K(u) := \sup \left\{\left<x, u \right> : x \in K
\right\} \text{ for } u \in S^{d-1},
\end{equation*}
where $S^{d-1} := \left\{ x \in {\mathbb R}^d : \sum_i x_i^2 = 1
\right\}$ is the unit sphere. We direct the reader to~\cite[Section
1.7]{Schneider} or~\cite[Section
13]{Rockafellar70book} for basic properties
of support functions. An important property is that the support
function uniquely determines the convex body, i.e., $h_K = h_L$
if and only if $K = L$.
Let $\left\{u_i, i \geq 1 \right\}$ be a sequence of $d$-dimensional
unit vectors. Gardner, Kiderlen and
Milanfar~\cite{GardnerKiderlenMilanfar} (see their paper for earlier
references) considered the problem of reconstructing an unknown convex
body $K$ from noisy measurements of $h_K$ in the directions $u_1,
\dots, u_n$. More precisely, their problem was to estimate $K$ from
observations $Y_1, \dots, Y_n$ drawn according to the model $Y_i=
h_K(u_i) + \xi_i, i = 1, \dots, n$ where $\xi_1, \dots, \xi_n$ are
independent and identically distributed mean zero gaussian random
variables. They constructed a convex body (estimator) $\hat{K}_n =
\hat{K}_n(Y_1, \dots, Y_n)$ having the property that, for
\textit{nice} sequences $\left\{ u_i, i \geq 1 \right\}$, the $L^2$
norm $||h_K - h_{\hat{K}_n}||_2$ (see~\eqref{lpnorm} below) converges
to zero at the rate $n^{-2/(d+3)}$ for dimensions $d = 2, 3, 4$ and at
a slower rate for dimensions $d \geq 5$ (see~\cite[Theorem
6.2]{GardnerKiderlenMilanfar}).
We shall show here that in the same setting, it is impossible in the
minimax sense to construct estimators
for $K$ converging at a rate faster than $n^{-2/(d+3)}$. This
implies that the least squares estimator
in~\cite{GardnerKiderlenMilanfar} is rate optimal for
dimensions $d = 2, 3, 4$. We shall need some
notation to describe our result.
Let ${\mathcal{K}}^d$ denote the set of all convex bodies in ${\mathbb R}^d$ and
for $\Gamma > 0$, let ${\mathcal{K}}^d(\Gamma)$ denote the set of all convex bodies
in ${\mathbb R}^d$ that are contained in the closed ball of radius $\Gamma$ centered
at the origin so that ${\mathcal{K}}^d(1)$ denotes the set of all convex bodies
contained in the unit ball, which we shall denote by $B$. Note that
estimating $K$ is equivalent to estimating
the function $h_K$ because the support function uniquely determines
the convex body. Thus we shall focus on the problem of estimating
$h_K$.
An estimator for $h_K$ is allowed to be a bounded
function on $S^{d-1}$ that depends on the data $Y_1, \dots, Y_n$. The
loss functions that we shall use are the $L^p$ norms for $p \in [1,
\infty]$ defined by
\begin{equation}\label{lpnorm}
||h_K - \hat{h}||_p := \left(\int_{S^{d-1}} |h_K(u) -
\hat{h}(u)|^p du
\right)^{1/p}
\end{equation}
for $p \in [1, \infty)$ and $||h_K - \hat{h}||_{\infty} := \sup_{u \in
S^{d-1}} |h_K(u) - \hat{h}(u)|$. For convex bodies $K$ and $L$ and
$p \in [1, \infty]$, we shall also write $\delta_p(K, L)$ for $||h_K -
h_L||_p$ and refer to $\delta_p$ as the $L^p$ distance between $K$ and
$L$.
We shall consider the minimax risk of the problem of estimating $h_K$ from
$Y_1, \dots, Y_n$ when $K$ is assumed to belong to ${\mathcal{K}}^d(\Gamma)$ i.e., we
are interested in the quantity
\begin{equation*}
r_n(p, \Gamma) := \inf_{\hat{h}} \sup_{K \in {\mathcal{K}}^d(\Gamma)} {\mathbb E}_K||h_K -
\hat{h}(Y_1, \dots, Y_n)||_p.
\end{equation*}
The following is the main theorem of this section.
\begin{theorem}\label{suppthm}
Fix $p \in [1, \infty)$ and $\Gamma > 0$. Suppose the errors $\xi_1, \dots,
\xi_n$ are independent normal random variables with mean zero and
variance $\sigma^2$. Then the minimax risk $r_n(p, \Gamma)$ satisfies
\begin{equation}\label{suppthm.eq}
r_n(p, \Gamma) \geq c \sigma^{4/(d+3)} \Gamma^{(d-1)/(d+3)} n^{-2/(d+3)},
\end{equation}
for a constant $c$ that is independent of $n$.
\end{theorem}
\begin{remark}\normalfont
In the case when $p = 2$, Gardner, Kiderlen and
Milanfar~\cite{GardnerKiderlenMilanfar} showed
that the least squares estimator converges at the rate given by the
right hand side of~\eqref{suppthm.eq} for dimensions $d = 2, 3,
4$. Thus, at least for $p=2$, the lower bound given
by~\eqref{suppthm.eq} is optimal for dimensions $d = 2, 3, 4$.
\end{remark}
We shall use inequality~\eqref{myb.chi} to
prove~\eqref{suppthm.eq}. First, let us put the support function
estimation problem in the general estimation setting of the last
section. Let $\Theta := \left\{h_K : K \in {\mathcal{K}}^d(\Gamma) \right\}$ and let
${\mathcal{A}}$ be the collection of all bounded functions on the unit sphere
$S^{d-1}$. The metric $\rho$ on ${\mathcal{A}}$ is just the $L^p$ norm and
$\ell(x) = x$.
Finally, let ${\mathcal{X}} = {\mathbb R}^n$ and for $f \in \Theta$, let $P_f$ be the
$n$-variate normal distribution with mean vector $(f(u_1), \dots,
f(u_n))$ and variance-covariance matrix $\sigma^2 I_n$, where $I_n$ is
the identity matrix of order $n$.
In order to apply inequality~\eqref{myb.chi}, we need to determine
$N(\eta)$ and $M_C(\epsilon, \Theta)$. The quantity $N(\eta)$ is a
lower bound on the $\eta$-packing number of the set ${\mathcal{K}}^d(\Gamma)$ under
the $L^p$ norm. When $p = \infty$, Bronshtein~\cite[Theorem 4 and
Remark 1]{Bronshtein76} proved that there exist positive constants
$c'$ and $\eta_0$ depending only on $d$ such that $\exp\left( c'
(\eta/\Gamma)^{(1-d)/2}\right)$ is a lower bound for the $\eta$-packing
number of $\Theta$ for $\eta \leq \eta_0$. It is a standard fact that
$p = \infty$ corresponds to the Hausdorff metric on ${\mathcal{K}}^d(\Gamma)$.
It turns out that Bronshtein's result is actually true for every $p
\in [1, \infty]$ and not just for $p = \infty$. However, to the best
of our knowledge, this has not been proved anywhere in the literature. By
modifying Bronshtein's proof appropriately and using the
Varshamov-Gilbert lemma (see for example~\cite[Lemma
4.7]{Massart03Flour}), we provide, in Theorem~\ref{Rpackresult}, a
proof of this fact. Therefore from Theorem~\ref{Rpackresult}, we can
take
\begin{equation}\label{Neta}
\log N(\eta) = c' \left(\frac{\Gamma}{\eta} \right)^{(d-1)/2}
\text{ for } \eta \leq \eta_0,
\end{equation}
where $c'$ and $\eta_0$ are constants depending only on $d$ and $p$.
Now let us turn to $M_C(\epsilon, \Theta)$. For $f, g \in \Theta$,
$P_f$ and $P_g$ are normal distributions with the same covariance
matrix and hence the chi-squared divergence between $P_f$ and $P_g$
can be seen to be
\begin{align*}
\chi^2(P_f||P_g) & = \exp \left[ \frac{1}{\sigma^2} \sum_{i=1}^n \left(f(u_i) - g(u_i)
\right)^2 \right] - 1 \\
&\leq \exp \left[ \frac{n||f-g||^2_{\infty}}{\sigma^2} \right] - 1.
\end{align*}
It follows that
\begin{equation}\label{klhaus}
||f - g||_{\infty} \leq \epsilon '\Longrightarrow \chi^2(P_f||P_g) \leq \epsilon^2.
\end{equation}
where $\epsilon ':= \sigma \sqrt{\log (1 +
\epsilon^2)}/\sqrt{n}$. Let $W_{\epsilon '}$ be the smallest $W$ for
which there exist sets $K_1, \dots, K_W$ in ${\mathcal{K}}^d(\Gamma)$ having the
property that for every set $K
\in {\mathcal{K}}^d(\Gamma)$, there exists a $K_j$ such that the Hausdorff distance
between $K$ and $K_j$ is less than or equal to $\epsilon '$. It must
be clear from~\eqref{klhaus} that $M_C(\epsilon, \Theta)$ can be taken to be
a number larger than $W_{\epsilon '}$. Bronshtein~\cite[Theorem
3 and Remark 1]{Bronshtein76} showed that there exist positive
constants $c''$ and $\epsilon_0$ depending only on $d$ such that
\begin{equation*}
\log W_{\epsilon '} \leq c'' \left(\frac{\Gamma}{\epsilon '}
\right)^{(d-1)/2} \text{ for } \epsilon ' \leq \epsilon_0.
\end{equation*}
Hence for all $\epsilon$ such that $\log(1 + \epsilon^2) \leq n \epsilon_0^2/\sigma^2$, we
can take
\begin{equation}\label{Meps}
\log M_C(\epsilon, \Theta) = c'' \left(\frac{\Gamma \sqrt{n}}{\sigma
\sqrt{\log(1 + \epsilon^2)}} \right)^{(d-1)/2}.
\end{equation}
We are now ready to prove inequality~\eqref{suppthm.eq}. We shall
define two quantities
\begin{equation*}
\eta(n):= c \sigma^{4/(d+3)} \Gamma^{(d-1)/(d+3)} n^{-2/(d+3)}
\end{equation*}
and
\begin{equation*}
u(n) := \left(\frac{\Gamma\sqrt{n}}{\sigma}
\right)^{(d-1)/(d+3)}.
\end{equation*}
where $c = c(d, p)$ will be specified shortly. Also let
$\epsilon(n)$ be such that $\log (1 + \epsilon^2(n)) = u^2(n)$.
Clearly as $n \rightarrow \infty$, we have $\eta(n) \rightarrow 0$,
$u(n) \rightarrow \infty$ and $u(n)/\sqrt{n} \rightarrow
0$. As a result $\eta(n) \leq \eta_0$ and
$u^2(n) \leq n\epsilon_0^2/\sigma^2$ for large $n$ and therefore
from~\eqref{Neta} and~\eqref{Meps}, we get that
\begin{equation*}
\log N(\eta(n)) = c' \left(\frac{\Gamma}{\eta(n)} \right)^{(d-1)/2}
= \frac{c'}{c^{(d-1)/2}} u^2(n).
\end{equation*}
and
\begin{equation*}
\log M_C(\epsilon(n), \Theta) = c'' \left(\frac{\Gamma
\sqrt{n}}{\sigma u(n)} \right)^{(d-1)/2} = c'' u^2(n).
\end{equation*}
We now apply inequality~\eqref{myb.chi} (recall that $\ell(x) = x$) to
obtain that $r_n(p, \Gamma)$ is bounded from below by
\begin{equation*}
\frac{\eta(n)}{2} \left[1 - \frac{1}{N(\eta(n))} - \exp
\left(\frac{u^2(n)}{2} \left(1 + c'' - \frac{c'}{c^{(d-1)/2}}
\right) \right) \right]
\end{equation*}
for all large $n$. If we now choose $c$ so that $c^{(d-1)/2} =
c'/(2+2c'')$, we get that
\begin{equation*}
r_n(p, \Gamma) \geq \frac{\eta(n)}{2} \left[1 - \frac{1}{N(\eta(n))}
- \exp \left(\frac{-u^2(n)}{2}(1+c'') \right)\right].
\end{equation*}
Now observe that as $n \rightarrow \infty$, the quantity $\eta(n)$
goes to 0 and hence $N(\eta(n))$ goes to $\infty$. Further, as we have
already noted, $u(n)$ goes to $\infty$. It follows hence that $r_n(p,
\Gamma) \geq \eta(n)/4$ for all large $n$. By choosing $c$ even smaller, we
can make inequality~\eqref{suppthm.eq} true for all $n$.
\section{A covariance matrix estimation example}\label{covmat}
In the previous section, we have used the global minimax lower
bound~\eqref{myb.chi}. However, in some situations, the global entropy
numbers might be difficult to bound. In such cases,
inequalities~\eqref{myb} and~\eqref{myb.chi} are, of course, not
applicable and we are unaware of the use of inequality~\eqref{jsyb} in
conjuction with Fano's inequality~\eqref{normfano} in the
literature. The standard examples use~\eqref{normfano} with the bound
$J_f \leq \min_{\theta, \theta ' \in F}D(P_{\theta}||P_{\theta '})$
while the examples in~\cite{YangBarron} all deal with the
case when global entropies are available. In this section, we shall
demonstrate how a recent minimax lower bound due to Cai, Zhang and
Zhou~\cite{CaiZhangZhou2009} can also be proved using
inequalities~\eqref{normfano} and~\eqref{jsyb}.
Cai, Zhang and Zhou~\cite{CaiZhangZhou2009} considered $n$ independent $p
\times 1$ random vectors $X_1, \dots, X_n$ distributed according to
$N_p(0, \Sigma)$. Suppose that the entries of the $p \times p$
covariance matrix $\Sigma = (\sigma_{ij})$ decay at a certain rate as
we move away from the diagonal. Specifically, let us suppose that for
a fixed positive constant $\alpha > 0$, the entries $\sigma_{ij}$ of
$\Sigma$ satisfy the inequality $\sigma_{ij} \leq |i-j|^{-\alpha-1}$ for $i
\neq j$. Cai, Zhang and Zhou~\cite{CaiZhangZhou2009} showed that when $p$ is large compared
to $n$, it is impossible to estimate $\Sigma$ from $X_1, \dots, X_n$
in the spectral norm at a rate faster than $n^{-\alpha/(2\alpha +
1)}$. More precisely, they showed that when $p \geq Cn^{1/(2\alpha +
1)}$,
\begin{equation}
\label{covres}
R_n(\alpha) := \inf_{\hat{\Sigma}} \sup_{\Sigma \in \Theta}
{\mathbb E}_{\Sigma}||\hat{\Sigma} -
\Sigma || \geq c~n^{-\alpha/(2\alpha + 1)},
\end{equation}
where $c$ and $C$ denote positive constants depending only on $\alpha$. Here
$\Theta$ denotes the collection of all covariance matrices
$\Sigma = (\sigma_{ij})$ satisfying $\sigma_{ij} \leq |i-j|^{-\alpha-1}$ for $i
\neq j$ and the norm $||.||$ is the spectral norm (largest
eigenvalue).
Cai, Zhang and Zhou~\cite{CaiZhangZhou2009} used Assouad's lemma
for the proof of the inequality~\eqref{covres}. We shall use
inequalities~\eqref{normfano} and~\eqref{jsyb}. Moreover, the choice of
the finite subset $F$ that we use is different from the one used
in~\cite[Equation (17)]{CaiZhangZhou2009}. This
makes our approach different from the general method, due to
Yu~\cite{Yu97lecam}, of replacing Assouad's lemma by Fano's inequality.
Throughout, $\Delta$ denotes a constant that depends on $\alpha$
alone. The value of the constant might vary from place to place.
Consider the matrix $A = (a_{ij})$ with $a_{ij} = 1$ for $i=j$
and $a_{ij} = 1/(\Delta|i-j|^{\alpha+1})$ for $i \neq j$. For $\Delta$
sufficiently large (depending on $\alpha$ alone), $A$ is positive
definite and belongs to $\Theta$. Let
us fix a positive integer $k \leq p/2$ and partition $A$ as
\begin{equation*}
A = \left[\begin{array}{c|c} A_{11} & A_{12}
\\ \hline \rule[13pt]{0pt}{0pt} A_{12}^T &
A_{22} \end{array} \right],
\end{equation*}
where $A_{11}$ is $k \times k$ and $A_{22}$ is $(p-k) \times (p-k)$.
For each $\tau \in {\mathbb R}^k$, we define the matrix
\begin{equation*}
A(\tau) := \left[\begin{array}{c|c} A_{11} & A_{12}(\tau)
\\ \hline \rule[13pt]{0pt}{0pt} \left(A_{12}(\tau)\right)^T &
A_{22} \end{array} \right],
\end{equation*}
where $A_{12}(\tau)$ is the $k \times (p-k)$ matrix obtained by
premultiplying $A_{12}$ with the $k \times k$ diagonal matrix with
diagonal entries $\tau_1, \dots, \tau_k$. Clearly, $A(\tau) \in
\Theta$ for all $\tau \in \left\{0, 1 \right\}^k$. We shall need the
following two lemmas in order to prove inequality~\eqref{covres}.
\begin{lemma}\label{hamlb}
For $\tau, \tau '\in \left\{0, 1 \right\}^k, \tau \neq \tau '$, we have
\begin{equation}
\label{hamlb.eq}
||A(\tau) - A(\tau ')|| \geq
\frac{1}{\Delta k^{\alpha}} \sqrt{\frac{\Upsilon(\tau, \tau ')}{k}},
\end{equation}
where $\Upsilon(\tau, \tau ') := \sum_{r=1}^k \left\{\tau_r \neq
\tau'_r \right\}$ denotes the Hamming distance between $\tau$ and
$\tau'$.
\end{lemma}
\begin{IEEEproof}
Fix $\tau, \tau' \in \left\{0, 1 \right\}^k$ with $\tau \neq \tau'$.
Let $v$ denote the $p \times 1$ vector
$\left(0_k, 1_k, 0_{p-2k} \right)^T$, where $0_k$ denotes the
$k\times 1$ vector of zeros etc. Clearly $||v||^2 = k$ and $(A(\tau) -
A(\tau ')) v$ will be a vector of the form $(u, 0)^T$ for
some $k \times 1$ vector $u = (u_1, \dots, u_k)^T$. Moreover $u_r =
\sum_{s=1}^k (\tau_r - \tau_r') a_{r, k+s}$ and hence
\begin{align*}
|u_r| &= \frac{ \left\{ \tau_r \neq \tau_r' \right\}}{\Delta}
\sum_{s=1}^k \frac{1}{|r-k-s|^{\alpha+1}} \\
& \geq \frac{\left\{ \tau_r
\neq \tau_r' \right\}}{\Delta} \sum_{i=k}^{2k -1} \frac{1}{i^{\alpha+1}}
\geq \frac{\left\{ \tau_r \neq \tau_r'
\right\}}{\Delta}\frac{1}{k^{\alpha}}.
\end{align*}
Therefore,
\begin{equation*}
||\left(A(\tau) - A(\tau ')\right)v||^2 \geq \sum_{r=1}^k u_r^2 \geq
\frac{1}{\Delta^2 k^{2 \alpha}} \Upsilon(\tau, \tau ').
\end{equation*}
The proof is complete because $||v||^2 = k$.
\end{IEEEproof}
\begin{lemma}\label{covublem}
Let $1 \leq m < k, \tau \in \left\{0, 1 \right\}^k$ and $\tau'
:= (0, \dots, 0, \tau_m, \dots, \tau_k)$. Then
\begin{equation*}
D \left(N(0, A(\tau)) || N(0, A(\tau')) \right) \leq
\frac{\Delta}{(k-m)^{2 \alpha}}.
\end{equation*}
\end{lemma}
\begin{IEEEproof}
The key is to note that one has the inequality $D \left(N(0, A(\tau))
|| N(0, A(\tau')) \right)
\leq \Delta ||A(\tau) - A(\tau')||^2_F$, where $||A||_F := \left(\sum_{i,
j} a_{ij}^2 \right)^{1/2}$ denotes the Frobenius norm. The proof
of this assertion can be found in~\cite[Proof of
Lemma 6]{CaiZhangZhou2009}. We can now bound
\begin{align*}
||A(\tau) - A(\tau')||^2_F &\leq 2\sum_{r=1}^{m-1} \tau_r^2
\sum_{j=1}^{p-k} a_{r, k+j}^2 \\
&\leq \Delta \sum_{r=1}^{m-1}\sum_{j=1}^{p-k} \frac{1}{|r-k-j|^{2\alpha+2}}
\\
&\leq \Delta \sum_{r=1}^{m-1} \sum_{j=1}^{\infty}
\frac{1}{|k-r+j|^{2\alpha+2}}\\
& \leq \Delta \sum_{r=1}^{m-1}
\frac{1}{(k-r)^{2\alpha + 1}} \leq \frac{\Delta}{(k-m)^{2\alpha}}.
\end{align*}
The proof is complete.
\end{IEEEproof}
The Varshamov-Gilbert lemma (see for example~\cite[Lemma
4.7]{Massart03Flour}) asserts the existence of a subset
$W$ of $\left\{0, 1\right\}^k$ with $|W| \geq \exp(k/8)$ such that
$\Upsilon(\tau, \tau') \geq k/4$ for all $\tau, \tau' \in W$ with $\tau
\neq \tau'$. Let $F := \left\{A(\tau) : \tau \in W \right\}$. From
inequality~\eqref{normfano} and Lemma~\ref{hamlb}, we get that
\begin{equation}\label{emp2}
R_n(\alpha) \geq \frac{1}{\Delta} \frac{1}{k^{\alpha}}
\left(1 - \frac{\log 2 + \frac{1}{|W|}\sum_{A \in
F}D(P_A||\bar{P})}{k/8} \right),
\end{equation}
where $P_A$ denotes the $n$-fold product of the $N(0, A)$ probability
measure and $\bar{P} :=
\sum_{A\in F} P_A/|W|$. Now for $1 \leq m < k$ and for $t \in
\left\{0, 1 \right\}^{k-m+1}$, let $Q_t$ denote the $n$-fold product
of the $N(0, A(0, \dots, 0, t_1, \dots, t_{k-m+1}))$ probability
measure. Applying inequality~\eqref{jsyb}, we get the quantity
$\sum_{A \in F} D(P_A||\bar{P})/|W|$ is bounded from above by
\begin{equation*}
(k-m+1) \log 2 +
\max_{A \in F} \min_{t \in \left\{ 0, 1\right\}^{k-m+1}} D(P_A||Q_t).
\end{equation*}
Now we use Lemma~\ref{covublem} to obtain
\begin{equation*}
\frac{1}{|W|} \sum_{A \in F}D(P_A||\bar{P})
\leq \Delta \left[(k-m) + \frac{n}{(k-m)^{2\alpha}} \right].
\end{equation*}
Using the above in~\eqref{emp2}, we get
\begin{equation*}
R_n(\alpha) \geq \frac{1}{\Delta} \frac{1}{k^{\alpha}}
\left[1 - \frac{\Delta}{k}\left((k-m) + \frac{n}{(k-m)^{\alpha}}
\right) \right].
\end{equation*}
Note that the above lower bound for $R_n(\alpha)$ depends on $k$ and
$m$, which are constrained to satisfy $2k \leq p$ and $1 \leq m <
k$. To get the best lower bound, we need to optimize the right hand
side of the above inequality over $k$ and $m$. It should be obvious
that in order to prove~\eqref{covres}, it is enough to take $k-m =
n^{1/(2 \alpha + 1)}$ and $k = 4\Delta n^{1/(2\alpha + 1)}$. The condition
$2k \leq p$ will be satisfied if
$p \geq Cn^{1/(2\alpha + 1)}$ for a large enough $C$. It is elementary
to check that with these choices of $k$ and $m$,
inequality~\eqref{covres} is established.
\section{A Packing Number Lower Bound}
In this section, we shall prove that for every $p \in [1, \infty]$ the
$\eta$-packing number $N(\eta; p, \Gamma)$ of ${\mathcal{K}}^d(\Gamma)$ under the $L^p$
metric is at least $\exp \left(c (\eta/\Gamma)^{(1-d)/2} \right)$ for a positive
$c$ and sufficiently small $\eta$.
This means that there exist at least $\exp \left(c (\eta/\Gamma)^{(1-d)/2}
\right)$ sets in ${\mathcal{K}}^d(\Gamma)$ separated by at least $\eta$ in the $L^p$
metric. This result was needed in the proof of
Theorem~\ref{suppthm}. Bronshtein~\cite[Theorem 4 and Remark
1]{Bronshtein76} proved this for $p = \infty$ (the case of the
Hausdorff metric).
\begin{theorem}\label{Rpackresult}
Fix $p \in [1, \infty]$. There exist positive constants $\eta_0$ and
$C$ depending only on $d$ and $p$ such that for every $\eta \leq
\eta_0$, we have
\begin{equation}\label{packresulteq}
N(\eta; p, \Gamma) \geq \exp \left(C \left(
\frac{\Gamma}{\eta}\right)^{(d-1)/2}\right).
\end{equation}
\end{theorem}
\begin{IEEEproof}
Observe that by scaling, it is enough to prove for the case $\Gamma =
1$. We loosely follow Bronshtein~\cite[Proof of Theorem
4]{Bronshtein76}. Fix $\epsilon \in (0,
1)$. For each point $x \in S^{d-1}$, let $S_x$ denote the supporting
hyperplane to the unit ball $B$ at $x$ and let $H_x$ be the hyperplane
intersecting the sphere that is parallel to $S_x$ and at a distance of
$\epsilon$ from $S_x$. Let $H_x^+$ and $H_x^-$ denote the two
halfspaces bounded by $H_x$ where we assume that $H_x^+$ contains the
origin. Let $T_x := S^{d-1} \cap H_x^-$ and $A_x := B \cap H_x$, where
$B$ stands for the unit ball. It can be checked that the
(Euclidean) distance between $x$ and every point in $T_x$ (and $A_x$) is
less than or equal to $\sqrt{2}\sqrt{\epsilon}$. It follows that if
the distance between two points $x$ and $y$ in $S^{d-1}$ is strictly
larger than $2\sqrt{2}\sqrt{\epsilon}$, then the sets $T_x$ and $T_y$ are
disjoint.
By standard results (see for example~\cite[Proof of Theorem
4]{Bronshtein76} where it is referred to as Mikhlin's result), there
exist positive constants $C_1$, depending
only on $d$, and $\epsilon_0$ such that for every $\epsilon \leq
\epsilon_0$, there exist $N \geq C_1(\sqrt{\epsilon})^{1-d}$ points
$x_1, \dots, x_N$ in $S^{d-1}$ such that the Euclidean distance
between $x_i$ and $x_j$ is strictly larger than
$2\sqrt{2}\sqrt{\epsilon}$ whenever $i \neq j$. From now on, we assume that
$\epsilon \leq \epsilon_0$. We then consider a mapping $\Phi : \left\{
0, 1\right\}^N \rightarrow {\mathcal{K}}^d(1)$, which is defined, for $\tau = (\tau_1,
\dots, \tau_N) \in \left\{0, 1 \right\}^N$, by
\begin{equation*}
\Phi(\tau) := B \cap D_1(\tau_1) \cap D_2(\tau_2) \cap \dots \cap D_N(\tau_N),
\end{equation*}
where for $i = 1, \dots, N$,
\begin{equation*}
D_i(0) := H_{x_i}^+ \text{ and } D_i(1) := B.
\end{equation*}
It must be clear that the Hausdorff distance between $\Phi(\tau)$ and
$\Phi(\tau ')$ is not less than $\epsilon$ (in fact, it is exactly equal
to $\epsilon$) if $\tau \neq \tau '$. Thus, $\left\{\Phi(\tau):
\tau \in \left\{0, 1\right\}^N \right\}$ is an $\epsilon$-packing
set for ${\mathcal{K}}^d(1)$ under the Hausdorff metric. However, it is \textit{not} an
$\epsilon$-packing set under the $L^p$ metric. Indeed, the $L^p$
distance between $\Phi(\tau)$ and $\Phi(\tau ')$ is not necessarily larger than
$\epsilon$ for all pairs $(\tau, \tau '), \tau \neq \tau '$. The
$L^p$ distance between $\Phi(\tau)$ and $\Phi(\tau ')$ depends on
the Hamming distance $\Upsilon(\tau, \tau ') = \sum_i \left\{ \tau_i
\neq \tau_i'\right\}$ between $\tau$ and $\tau '$. We make the claim that
\begin{equation}\label{claim}
\delta_p\left(\Phi(\tau), \Phi(\tau ') \right) \geq C_2 \epsilon
\left( \sqrt{\epsilon}\right)^{(d-1)/p} \left( \Upsilon(\tau, \tau ')
\right)^{1/p},
\end{equation}
where $C_2$ depends only on $d$ and $p$.
The claim will be proved later. Assuming it is true, we recall the
Varshamov-Gilbert lemma from the previous section to assert the
existence of a subset $W$ of $\left\{0, 1\right\}^N$ with $|W| \geq
\exp(N/8)$ such that $\Upsilon(\tau, \tau') \geq N/4$ for all $\tau,
\tau' \in W$ with $\tau \neq \tau'$. Because
$N \geq C_1(\sqrt{\epsilon})^{1-d}$, we get from~\eqref{claim} that
for all $\tau, \tau' \in W$ with $\tau \neq \tau'$, we have
\begin{equation*}
\delta_p\left(\Phi(\tau), \Phi(\tau') \right) \geq C_3 \epsilon \text{
where } C_3 := C_2 \left(
\frac{C_1}{4}\right)^{1/p}.
\end{equation*}
Taking $\eta := C_3 \epsilon$, we have obtained, for each $\eta \leq
\eta_0 := C_3 \epsilon_0$, an $\eta$-packing subset of ${\mathcal{K}}^d(1)$ with
size $M$, where
\begin{equation*}
\log M \geq N/8 \geq \frac{C_1}{8} \left(
\frac{1}{\sqrt{\epsilon}}\right)^{d-1} = C_4 \left(
\frac{1}{\sqrt{\eta}}\right)^{d-1}.
\end{equation*}
The constant $C_4$ only depends on $d$ and $p$ thereby
proving~\eqref{packresulteq}.
It remains to prove the claim~\eqref{claim}. Fix a point $x \in
S^{d-1}$ and $\epsilon \in (0,1)$. We first observe that it is enough
to prove that
\begin{equation}\label{tworeg}
\delta_p(A_x, T_x)^p \geq C_5 \epsilon^p
\left( \sqrt{\epsilon}\right)^{d-1},
\end{equation}
for a constant $C_5$ depending on just $d$ and $p$, where $A_x$ and
$T_x$ are as defined in the beginning of the proof. This is because of
the fact that for every $\tau, \tau' \in W$ with $\tau \neq \tau'$, we
can write
\begin{equation}\label{addit}
\delta_p\left(\Phi(\tau),\Phi(\tau')\right)^p =
\sum_{i \in I} \delta_p(A_{x_i}, T_{x_i})^p,
\end{equation}
where $I := \left\{1\leq i \leq N : \tau_i \neq \tau_i'\right\}$. The
equality~\eqref{addit} is a consequence of the fact that the points
$x_1, \dots, x_N$ are chosen so that $T_{x_1}, \dots, T_{x_N}$ are
disjoint.
We shall now prove the inequality~\eqref{tworeg} which will complete
the proof. Let $u_0$ denote the point in $A_x$ that is closest to the
origin. Also let $u_1$ be a point in $A_x \cap
S^{d-1}$. Let $\alpha$ denote the angle between $u_0$ and
$u_1$. Clearly, $\alpha$ does not depend on the choice of $u_1$ and
$\cos \alpha = 1 - \epsilon$. Now let $u$ be a fixed unit vector and
let $\theta$ be the angle between the vectors $u$ and $u_0$. By
elementary geometry, we deduce that
\begin{equation*}
h_{T_x}(u) - h_{A_x}(u) =
\begin{cases} 1 - \cos \left(\alpha - \theta \right) & \text{if $0
\leq \theta \leq \alpha$,}
\\
0 &\text{otherwise.}
\end{cases}
\end{equation*}
Because the difference of support functions only depends on the angle
$\theta$, we can write, for a constant $C_6$ depending only on $d$,
that
\begin{equation*}
\delta_p(A_x, T_x)^p = C_6 \int_0^{\alpha}
\left( 1 - \cos (\alpha - \theta)\right)^p \sin^{d-2} \theta d\theta.
\end{equation*}
Now suppose $\beta$ is such that $\cos (\alpha - \beta) = 1 -
\epsilon/2$. Then from above, we get that
\begin{align*}
\delta_p(A_x, T_x)^p &\geq C_6
\int_0^{\beta} \left( 1 - \cos (\alpha - \theta)\right)^p \sin^{d-2}\theta d\theta \\
&\geq C_6 \left(\frac{\epsilon}{2}\right)^p \int_0^{\beta}
\sin^{d-2}\theta d\theta \\
&\geq C_6 \left(\frac{\epsilon}{2}\right)^p \int_0^{\beta}
\sin^{d-2} \theta \cos \theta d\theta \\
&= \frac{C_6}{d-1} \left(\frac{\epsilon}{2}\right)^p \sin^{d-1}
\beta.
\end{align*}
We shall show that $\sin \beta \geq
\left(\sqrt{\epsilon}\right)/(2\sqrt{2})$ which will
prove~\eqref{tworeg}. Recall that $\cos \alpha = 1 - \epsilon$. Thus
\begin{align*}
1 - \frac{\epsilon}{2} &= \cos (\alpha - \beta) \\
&\leq \cos \alpha + \sin \alpha \sin \beta \\
&= 1 - \epsilon + \sqrt{1 - (1-\epsilon)^2} \sin \beta \\
&\leq 1 - \epsilon + \sqrt{2} \sqrt{\epsilon} \sin \beta,
\end{align*}
which when rearranged gives $\sin \beta \geq
\left(\sqrt{\epsilon}\right)/(2\sqrt{2})$. The proof is complete.
\end{IEEEproof}
\section{Conclusion}
By a simple application of convexity, we proved an inequality relating
the minimax risk in multiple hypothesis testing problems to
$f$-divergences of the probability measures involved. This inequality
is an extension of Fano's inequality. As another corollary, we obtained a sharp
inequality between total variation distance and $f$-divergences. We
also indicated how to control the quantity $J_f$ which appears in our
lower bounds. This leads to important global lower bounds for the minimax
risk. Two applications of our bounds are presented. In the first
application, we used the bound~\eqref{myb.chi} to prove a new lower
bound (which turns to be rate-optimal) for the minimax risk of
estimating a convex body from noisy measurements of the support
function in $n$ directions. In the second application, we employed
inequalities~\eqref{normfano} and~\eqref{jsyb} to give a
different proof of a recent lower bound for covariance matrix
estimation due to Cai, Zhang and Zhou~\cite{CaiZhangZhou2009}.
\section*{Acknowledgment}
The author is indebted to David Pollard for his insight and also for
numerous stimulating discussions which led to many of the ideas in
this paper; to Andrew Barron for his constant encouragement and
willingness to discuss his own work on minimax bounds. Thanks are also
due to Aditya Mahajan for pointing out to the author
that inequality~\eqref{maha.eq1} has the extension~\eqref{maha.eq} for
the case of non-uniform priors $w$; to an anonymous referee for
helpful comments, for pointing out an error and for
Example~\ref{ddimnorm}; to Richard Gardner for comments that
greatly improved the quality of the paper and to Alexander Gushchin
for informing the author about his paper~\cite{Gushchin} and for
sending him a scanned copy of it.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\def\noopsort#1{}
|
\section{Introduction}
Metastability has given rise to a profuse literature in the Statistical Physics and Probability Theory areas during the last decades. It concerns a wide range of models, from the Curie-Weiss model to lattice gas and spin models and to diffusion processes. See ~\cite{Olivieri:05} and ~\cite{Bovier:06} for an overview of the subject. Metastability can be roughly described as the phenomenon occuring when a physical system stays a very long time in some abnormal state before reaching its normal - under the prevailing conditions - equilibrium state. The normal situation is only restored after, under some random perturbation or some other external provision of energy, the system can get over some energy barrier. Mathematically, it is usually formalized through exponential growth of some exit times under some asymptotic, supposed to correspond to the physics of the system.
While the potential theoretic approach, more recently developed (see~\cite{Bovier:06}), provides sharp estimations on crossover times for reversible dynamics, a lot is still due to the large deviation approach developed by Freidlin and Wentzell in ~\cite {Freidlin:84} (see ~\cite{Olivieri:05} for an outline of the application range). It has lead to some quite complete descriptions of the \emph{exit path} from a metastable state, e.g.~for the Ising model under Glauber dynamics at low temperature. Yet again, reversibility is often crucial, even in this context.
Metastability is expected to occur for some specific models in communication networks, but a formal proof is still lacking. Namely, ~\cite{Gibbens:90} and ~\cite{Antunes:08} analyze two loss systems with local interactions, that admit a mean field limit as the number of nodes goes to infinity. The model in ~\cite{Gibbens:90} is ruled by an alternative routing procedure for blocked calls, adapted from ~\cite{Marbukh:85}, while ~\cite{Antunes:08} analyzes a simple model for a loss network with mobile customers. In both papers, the limiting dynamics are shown to exhibit a phase with several stable equilibrium points. ~\cite{Gibbens:90} provides estimations of certain exit times for a one dimensional diffusion approximation of the model, suggesting that metastability occurs. This should also be the case for the model in ~\cite{Antunes:08}, as suggested by simulations.
Yet, much less is known for these network models than for the classical examples cited above. In particular, their Markovian dynamics are not reversible and computing the invariant distributions is out of reach. Such useful quantities as Hamiltonian, energy barrier etc... are thus not available. However, it is expected that as the number $N$ of nodes grows, the invariant distribution should be approximately given by $ Z_N^{-1} \exp (- N h)$ for some energy function $h$ that would play the role of the Hamiltonian in the limit. In ~\cite{Antunes:08} a Lyapunov function for the limiting dynamical system is exhibited (and used for proving multistability), but this function has no reason to describe the correct energy landscape. In particular, it is not known which equilibrium points are asymptotically relevant in the invariant distribution (i.e., correspond to global minima of the possible Hamiltonian).
\\
The main purpose of the present paper is to show that the models in ~\cite{Gibbens:90} and ~\cite{Antunes:08} essentially fit the scheme of Freidlin and Wentzell (\cite {Freidlin:84}), as - in these authors' terminology - \emph{locally infinitely divisible processes}. This is the object of Section ~\ref{sec:quasipot}. As a result, exponential growth (as $N$ gets large) of exit times from neighborhoods of stable equilibrium points, that is, metastability is obtained. The location of the exit points can also be described.
It must be pointed out that the large deviations results stated in ~\cite {Freidlin:84} are not rigorously applicable to the processes of interest in this paper, since these evolve in some compact subset of $\mathds{R}^d$, on the frontier of which some of the technical hypothesis required in ~\cite {Freidlin:84} are no longer valid. Yet, it seems that there is no fundamental reason for these restrictions. Extension of the results in ~\cite {Freidlin:84} to this slightly more general context will thus be used without proof. Such a proof
is beyond the scope of this paper, that aims at opening a way for understanding the stochastic behavior of systems for which only the deterministic limiting evolution has been described so far.
A second issue of this paper is to decrypt the Lyapunov function exhibited in ~\cite{Antunes:08}. Two answers are given in this direction. One, presented in Section ~\ref{sec:quasipot}, is related to the \emph{quasipotential} introduced in ~\cite {Freidlin:84}, which is the crucial quantity involved in estimations of exit times and exit points. Moreover, as suggested in ~\cite {Freidlin:84} from the thoroughly analyzed case of diffusion processes, when there is just one equilibrium point, the quasipotential should represent the underlying energy function associated with the stationary state. It is here proved that the Lyapunov function exhibited in ~\cite{Antunes:08} coincides with the quasipotential of a slightly modified version of the process of interest. In addition, a heuristic argument suggests that the modified process is \emph{asymptotically reversible} as the number of nodes gets large.
Section ~\ref{sec:lyapunov} is devoted to another interpretation of this Lyapunov function, related to the well known decrease of relative entropy along a semi-group. This relies on a very particular feature of the model in ~\cite{Antunes:08}. Yet, two other models (particular in some other sense) can then be introduced, to which this principle for obtaining a Lyapunov function can be exported. This helps proving convergence of their invariant distribution to a Dirac mass.
Section ~\ref{sec:mean_field} recalls the two models of interest and main results from ~\cite{Gibbens:90} and ~\cite{Antunes:08}.
\section{Mean field limits, multistability} \label{sec:mean_field}
This section gives a short review of the models in ~\cite{Gibbens:90} and ~\cite{Antunes:08} and of the main results therein. These models exhibit a mean field limiting dynamics of some generic form, of which two other examples will be given in Section~\ref{sec:lyapunov}. But only the two models from ~\cite{Gibbens:90} and ~\cite{Antunes:08} exhibit a phase with several stable points.
\subsection{Multistability due to rerouting}\label{subsec:GHK}
A first example of bistability in queueing networks is given in ~\cite{Gibbens:90}. The analysis is formalized through convergence of some family of empirical processes to some dynamical system. All the subsequent examples of this section and Section~\ref{sec:lyapunov} will fit this frame.
This first model is a simplified version of a network with alternative routing proposed in ~\cite{Marbukh:85}. In ~\cite{Anantharam:91}, a lattice version, with long range rerouting, is proposed.
The model in ~\cite{Gibbens:90} is the following: The network considered consists of $N$ nodes offering the same finite (integer) capacity $C \ge 1$. Customers enter the network at the different nodes according to $N$ independent Poisson processes with intensity $\lambda >0$. When some customer arrives at some node where no more than $C-1$ customers are present, he occupies one unit of capacity for an exponentially distributed service time with mean one. When some customer arrives at some saturated node, he is rerouted to two other nodes, chosen uniformly among the $N-1$ possible nodes. If both chosen nodes have one unit of capacity available, the customer then behaves like two independent customers, leaving the nodes after two independent exponential times with mean one. In the contrary (i.e., if at least one of the two nodes is saturated), the customer is definitively rejected from the system.
Due to symmetry with respect to the $N$ nodes, the quantity of interest is the empirical distribution of the nodes as function of time, that is
\begin{align} \label{eq:Y} Y^N(t)=\frac{1}{N} \sum _{i=1} ^N \delta _{X^N_i(t) },\end{align}
where $X_i^N(t)$ denotes the number of customers present at node $i$ at time $t$ and $\delta _x$ is the Dirac mass at $x$.
The $X^N_i(t)$ ($N \ge 1$, $t\ge 0$ and $i=1, \dots ,N$) evolve in the set $\mathcal{X} = \{ 0, \dots ,C\}$. Identifying the set $\mathcal P(\mathcal{X})$ of probability measures on $\mathcal{X}$ with the set
\[ \mathcal{Y}= \left \{ y=(y_0, \dots , y_C) \in [0,+ \infty[^{C+1}: \sum _{n=0} ^C y_n=1\right \},\]
where $y_n$ represents the mass at $n$ ($n=0, \dots ,C$) of the probability measure $y$,
one can write $Y^N(t)= (Y^N_n(t), 0 \le n \le C)$, with
\begin{align} \label{eq:Y_n} Y^N_n(t)=\frac{1}{N} \sum _{i=1} ^N \indicator {X^N_i(t) =n }.\end{align}
In other words, $Y^N_n(t)$ is the proportion of nodes that are in state $n$ at time $t$.\\
For any fixed $N\ge 3$, $(Y^N(t), t\ge 0)$ is a Markov jump process with the following transitions and rates, where $e_n$ denotes the $n^{th}$ unit vector in $\mathds{R}^{C+1}$:
\[ y \longrightarrow \left \{ \begin{array}{l} y+\frac{1}{N} (e_{n+1}-e_n) \\ \\
y+\frac{1}{N} (e_{n-1}-e_n)\\
\\ y+\frac{1}{N} \big( e_{m+1}-e_m +e_{n+1}-e_n\big) \\
\\y+\frac{2}{N} \big(e_{n+1}-e_n\big)
\end{array} \right.
\text{at rates} \begin{array}{ll}
\lambda N y_n & (0 \le n \le C-1) \\
\\ Nn y_n & (0 < n \le C)\\
\\ 2 \lambda \frac{ N^3 y_C y_ny_m}{(N-1)(N-2)} & (0 \le m \neq n \le C-1)\\
\\ \lambda \frac{N^2 y_C y_n (Ny_n-1)}{(N-1)(N-2)} &(0 \le n \le C-1)
\end{array} \]
The first jump corresponds to some arrival at some node with $n$ customers, the second one to some departure from some node with $n$ customers and the two last jumps correspond to rerouting of some customer to two nodes with, respectively, different or equal numbers of customers. \\
~\cite{Gibbens:90} proves that for any $T>0$, the process $(Y^N(t), 0 \le t \le T)$ converges in distribution, as $N$ goes to infinity, to the solution with initial value $y(0)$ of the following differential system of equations: for $n=0, \dots ,C$
\begin{multline*} y_n'(t)=\lambda \Big( 1+2y_C(t)(1-y_C(t)\Big) y_{n-1}(t) \mathds 1_ {n\ge 1} + (n+1) y_{n+1} (t) \mathds 1_ {n\le C-1}
\\ - \Big(\lambda (1+2y_C(t)(1-y_C(t)) +n \Big) y_n(t).
\end{multline*}
provided that $Y^N(0)$ converges in distribution to $y(0)$.
The vector field characterizing the limiting dynamical system, given by the right hand sides of this differential system, can be heuristically obtained by computing the infinitesimal mean jump from position $y$ (suming up the jump amplitudes multiplied by the corresponding rates) and letting $N$ grow to infinity.
Introducing the family of infinitesimal generators $(L_y, y\in \mathcal{Y})$ defined by
\[ L_yf(n)= \lambda (1+2 y_C(1-y_C))(f(n+1) -f(n)) \mathds 1_ {n\le C-1} + n(f(n-1)-f(n)) \quad (f \in \mathds{R} ^{ \mathcal{X}}, n \in \mathcal{X}),\]
the above differential
system rewrites as one unique differential equation on $\mathcal{Y}$:
\begin{align} \label{eq:S} \dot y = y L_y,
\end{align}
where the second member is the product of probability measure (or row vector) $y$ by the infinitesimal generator (or rate matrix) $L_y$. For $y\in \mathcal{Y}$, $L_y$ is the generator of an $M/M/C/C$ queue with arrival rate $ \lambda (1+2 y_C(1-y_C))$ and service rate $1$.
Equation ~\eqref{eq:S} precisely conveys the mean field property of the model: it tells that, in the limit $N \to \infty$, the empirical distribution $y(t)$ of the nodes evolves in time as the distribution of some non-homogeneous Markov process on $\mathcal{X}=\{0, \dots ,C\}$, whose jump rates at time $t$ are given by $L_{y(t)}$, being hence constantly updated according to the current distribution, or ``mean field'' $y(t)$. These jump rates are those of an $M/M/C/C$ queue. Only the arrival rate is time dependent, and more precisely depends on $y_C(t)$, that is, on the proportion of saturated nodes. This $M/M/C/C$ queue can be viewed as representing the instantaneous evolution of a ``typical node'' under the global influence of the other nodes. Due to symmetry, for large $N$, this virtual node summarizes the whole network.
All the forecoming examples (Sections~\ref{sec:mean_field} and~\ref{sec:lyapunov}) will follow this scheme, with different types of dependency on $y$ of the mean field arrival rate.
\\
The equilibrium points of the limiting dynamical system ~\eqref{eq:S} are the solutions of $y L_y=0$. Since for all $y$, $L_y$ is the generator of some ergodic Markov process on $\mathcal{X}$, this means that $y$ is equal to the unique invariant distribution associated to $L_y$.
The $M/M/C/C$ queue with arrival rate $\rho$ and service rate $1$ is known to be reversible, having invariant distribution given by the well known Erlang distribution:
\begin{align} \label{eq:erlang} \nu _{\rho}(n)=\frac{1}{Z(\rho)}\frac { \rho ^n}{n!} \quad (n=0, \dots , C),
\end{align}
where $Z(\rho)=\sum _{n=0}^C \rho ^n/n!$ is a normalizing constant. Equilibrium points are thus given by the solutions of the fixed point equation
\begin{align}\label{eq:fix} y= \nu _{\rho(y)},
\end{align}
where $\rho (y)= \lambda (1+2 y_C(1-y_C))$.
~\cite{Gibbens:90} shows numerically that for certain values of $\lambda$, this equation exhibits several solutions, namely three, among which two are stable and one is unstable.
This suggests that for $N$ large and for suitable values of $\lambda$, the system should be attracted to one of two possible states, both of the Erlang form but with two different values of $\rho$. Intuitively, the system can either fall into a heavy loaded regime or into a light loaded one. In the first one, the heavy load maintains itself by inducing many reroutings, while in the opposite way, a small proportion of saturated nodes maintains a low rate of rerouting.
\begin{remark}
A slight variant of this model consists in rerouting customers to only one other node, but instead, changing their service rate from $1$ to some value $\mu <1$. One can prove that multistability still occurs for well chosen values of $\mu$ and $\lambda$. Notice that, in order to preserve the Markov property, it is here necessary to introduce two types of customers: those with service rate $1$ and the rerouted ones, with service rate $\mu$.
\end{remark}
\subsection{Multistability due to coexistence} \label{subsec:AFRT}
A second example of multistability in the networks context is given by ~\cite{Antunes:08}. One major difference with the previous model is that here multistability can occur only when different classes of customers coexist having different capacity requirements. Besides that, the model deals with mobile customers travelling from one node to another (and being possibly rejected during their service).
Here again, the network consists of $N$ nodes with capacity $C$, now offered to $K$ different classes of customers. For $k=1, \dots ,K$, each customer of class $k$ occupies $A_k$ units of capacity at each node he visits ($1 \le A_k \le C$). Class $k$ customers arrive at each node according to some Poisson process with intensity $\lambda_k$, and have service times exponentially distributed with parameter $\mu _k$. When some class $k$ customer arrives at some node where his capacity requirement is not available, he is definitively rejected from the network. Otherwise, he begins to be served at this node, and then moves at rate $\gamma _k$ during service. At each move, a new node is chosen uniformly among the $N-1$ possible nodes. Customers either leave the system through rejection at some node along their route, or through end of service.
All arrival processes, service durations and sojourn times of customers at the different nodes are assumed independent.
The empirical distribution $Y^N(t)$ of the nodes at time $t$ is still given by equation ~\eqref{eq:Y}, but here the state $X_i^N(t)$ of node $i$ at time $t$ is $K$-dimensional, consisting of the different numbers of customers of each class present at node $i$ at time $t$. The state spaces $\mathcal{X}$ of variables $X^N_i(t), 1 \le i \le N, t\ge 0,$ and $\mathcal{Y}$ of $Y^N(t), t\ge 0,$ are now
\[ {\mathcal X} = \{ n\in \mathds{N}^K : \sum _{k=1}^K n_k A_k \leq C \} \]
\[ \text{ and } \quad \mathcal{Y}= \mathcal P (\mathcal{X})= \left \{ y=(y_n,n \in \mathcal{X}) \in [0,+ \infty[^{\mathcal{X}}: \sum _{n\in \mathcal{X}} y_n=1\right \}.\]
Here again $Y^N(t)=(Y^N_n(t), n\in \mathcal{X})$ for $t \ge 0$, where $Y^N_n(t)$ is given by equation ~\eqref{eq:Y_n}.
Note that the process $Y^N$ actually evolves in some finite subset $\mathcal{Y} ^N$ of $ \mathcal{Y}$:
\[ \mathcal{Y} ^N = \{y=(y_n,n \in \mathcal{X}) \in \mathcal{Y}: Ny_n \in \mathds{N} \text{ for } n\in \mathcal{X} \} .\]
$Y^N$ is a Markov jump process with the following transitions, where $f_k$ denotes the $k^{th}$ unit vector in $\mathds{R} ^K$ ($k=1, \dots ,K$) and $e_n$ is the $n^{th}$ unit vector in $\mathds{R}^{\mathcal{X}}$ ($n \in \mathcal{X}$):
for $1 \leq k\leq K$, $n,m \in \mathcal{X}$,
\[ y \longrightarrow \left \{ \begin{array}{l} y+\frac{1}{N} (e_{n+f_k}-e_n) \\ \\
y+\frac{1}{N} (e_{n-f_k}-e_n)\\
\\ y+\frac{1}{N} \big( (e_{m+f_k}-e_m) {\mathds 1}_{m+f_k \in {\mathcal X}} +e_{n-f_k}-e_n\big)
\end{array} \right.
\text{at rates} \begin{array}{l}
\lambda _k N y_n {\mathds 1}_{n+f_k \in {\mathcal X}} \\
\\ \mu _k Nn_k y_n \\
\\ \frac{\gamma _k N n_k y_n}{N-1}(Ny_m-{\mathds 1} _{m=n})
\end{array} \]
The first jump corresponds to the arrival of some class $k$ customer at some node in state $n$, the second one to the
end of service of some class $k$ customer at some node in state $n$, and the last one to a move of some class $k$ customer from some node in state $n$ to some node in state $m$ (possibly saturated, implying rejection). \\
It is proved in ~\cite{Antunes:08} that for $T>0$, the process $ (Y^N(t), 0 \le t \le T)$ converges in distribution to $ (y(t), 0 \le t \le T)$
solving the differential system of equations in $\mathcal{Y}$:
\[ y'_n(t)=\sum _{k=1}^K \Big[ \Big( \lambda _k+\gamma _k [ I_k,y(t)] \Big) y_{n-f_k}(t) {\bf 1}_{n_k\geq 1} +(\mu _k +\gamma _k)(n_k+1) y_{n+f_k}(t) {\bf 1}_{n+f_k\in {\mathcal X}}\]
\[-\Big( (\lambda _k +\gamma _k [ I_k,y(t) ]) {\bf 1}_{n+f_k\in {\mathcal X}}+(\mu_k +\gamma _k)n_k\Big) y_n(t) \Big] \quad (n\in \mathcal{X}),\]
if $Y^N(0)$ converges in distribution to $y(0)$.
Here $\displaystyle{ [ I_k,y ]= \sum _{n\in {\mathcal X}} n_k y_n}$ is the mean of the $k^{th}$ marginal of $y$ (in particular $ [ I_k,Y^N(t) ]$ is the number of class $k$ customers per node at time $t$ in the network, that is, the density of class $k$ customers present). Convergence in distribution refers to the Skorohod topology on $\mathcal D([0,T])$.\\
This system can here again be written as ~\eqref{eq:S} where $L_y$ is now the infinitesimal generator of an $M/M/C/C$ queue with $K$ classes of customers having different arrival rates $ \lambda _k+\gamma _k [ I_k,y]$, service rates $\mu _k +\gamma _k$ and capacity requirements $A_k$ ($k=1, \dots ,K$).
Note that in ~\cite{Dawson:05} a similar mean field limiting dynamics, described by a non linear equation in the form of \eqref{eq:S}, is obtained for a system of $N$ interacting queues. The $L_y$ involved are birth and death generators with parameters depending on the mean of $y$, as in our present case. However, a major difference with the present model is that the stochastic dynamics itself involves a mean field interaction: The arrival rates of customers in the network depend on the global density of occupation. On the contrary, in the model from ~\cite{Antunes:08}, interaction between customers is \emph{local} (only due to saturation at some node) and the mean field evolution only appears in the limit.
\\
Erlang formula ~\eqref{eq:erlang} for the invariant distribution of the $M/M/C/C$ queue extends to the case of $K$ classes of customers with capacity requirements $A_k$ and arrival rate-to-service rate ratios $\rho _k$ ($k=1, \dots ,K$), where $n!$ and $\rho ^n$ ($n \in \mathcal{X}$) now hold for
\[n!=\prod _{k=1}^K n_k! \quad \text{ and } \rho ^n= \prod _{k=1}^K \rho _k ^{n_k} \quad (n=(n_1, \dots ,n_K) \in \mathcal{X}).\]
$Z(\rho)$ is here given by $\displaystyle{ Z(\rho)= \sum _{n \in \mathcal{X}} \frac{\rho ^n}{n!}}$. Equilibrium points are then again characterized by ~\eqref{eq:fix}, where $\rho(y)$ is now multidimensional: $\displaystyle{\rho(y)= \left (\frac{ \lambda _k+\gamma _k [ I_k,y] }{\mu _k + \gamma _k}\right )_{ 1 \le k \le K }}$.\\
In ~\cite{Antunes:08}, coexistence of several equilibrium points is proved to occur when $K=2$, $A_1=1$ and $A_2=C$, for $C$ sufficiently large and for certain values of parameters $\lambda _k, \mu _k$ and $\gamma _k$ ($k=1,2$). A key argument is the determination of a Lyapunov function for dynamical system ~\eqref{eq:S}, that is, some continuously differentiable, bounded from below, function $g$ defined on $[0, + \infty[^{\mathcal{X}}$ such that
\[ y L_y \nabla g(y) \le 0 \quad (y \in \mathcal{Y} \subset [0, + \infty[^{\mathcal{X}}),\]
where equality holds only if $yL_y=0$, i.e., if $y$ is an equilibrium point of the dynamics.
$g$ is explicitely given by
\begin{align} \label{eq:g} g(y)= \sum _{n \in \mathcal{X}} y_n \log(n!y_n) - \sum _{k=1} ^ K \int _0^{ [ I_k,y ]} \log \frac{ \lambda _k+\gamma _k x}{\mu _k +\gamma _k} dx.\end{align}
Moreover, $g$ satisfies: for $y \in \mathcal{Y}$,
\[ yL_y=0 \Longleftrightarrow \nabla g(y)\perp \mathcal{Y} ,\]
so that equilibrium points are characterized as the critical points of $g_{|\mathcal{Y}}$. An analytic function argument shows that these critical points are isolated.
These properties of $g$ allow one to discriminate \emph{stable} (local minima of $g$) from \emph{unstable} (local maxima and saddle points of $g$) equilibrium points.
Multistability is then proved (for certain values of the parameters) by proving existence of a saddle point for $g$, and then showing that two local minima are necessarily present, one on each side of some line crossing the saddle point.\\
Besides the multistability issue, a Lyapunov function is a tool for showing that equilibrium points of the limiting dynamical system are the concentration points, as $N \to \infty$, of the invariant measures $\pi^N$ of processes $Y^N$ (note that $Y^N$ is an irreducible Markov jump process on the finite state space $\mathcal{Y}^N$, so that it admits a unique invariant measure denoted $\pi ^N$). This means in some sense commutation of limits as $N \to \infty$ and $t\to + \infty$.
More precisely, it is proved in ~\cite{Antunes:08} that the infinitesimal generator $\Omega ^N$ of $Y^N$ converges to the generator $\Omega $ of the limiting (degenerated) Markov process given by ~\eqref{eq:S}. $\Omega$ is defined, at any $C^1$ function $f$ on $\mathds{R}^{\mathcal{X}}$, by
\[\Omega f(y)= yL_y \nabla f(y).\]
It is then standard that any weak limit of the sequence $(\pi ^N)$ is an invariant measure for $\Omega$.
The Lyapunov function then makes it possible to show that the invariant measures of $\Omega$, hence the weak limits of $(\pi ^N)$, are precisely the convex combinations of Dirac masses at equilibrium points of ~\eqref{eq:S}. In particular, when the equilibrium point $\bar y$ is unique, $\pi^N$ converges to the Dirac mass at $\bar y$.
We refer to ~\cite{Antunes:08} for details, or to the proof of Proposition ~\ref{prop:closed} hereafter.
\section{Large deviations, quasipotential} \label{sec:quasipot}
As just recalled, the Lyapunov function $g$ associated with the model in ~\cite{Antunes:08} helps describing the weak limits of the stationary distributions $\pi ^N$ of the corresponding processes $Y^N$. Such a function is not available in the case of ~\cite{Gibbens:90}. Indeed, Sections ~\ref{sec:quasipot} and ~\ref{sec:lyapunov} will emphasize the specificity of the model of ~\cite{Antunes:08} that makes the explicit formulation of $g$ possible.
Note however that the previous analysis of the model in ~\cite{Antunes:08} does not tell which stable equilibria remain significant in the limit, in the multistable phase. It is not even proved that the weak limits of $\pi ^N$ have positive mass only at \emph{stable} equilibrium points.\\
In this respect, a more precise issue would be to find, for both models of interest, an {\em energy function} $h$ describing their invariant distribution $\pi ^N$ as $N \to \infty$ in the sense that
\[ \pi ^N(y) \approx \frac {e^{-N h(y)}}{Z_N} \quad \mathrm { for } \ y \in \mathcal Y ^N.\]
Such an $h$ is expected to be a Lyapunov function for the limiting dynamical system, or at least to satisfy $ yL_y \nabla h(y) \le 0$ (see Remark ~\ref{remark:balance} below).
This amounts to stating a Large Deviation Principe for measures $\pi ^N$ (with action functional $I(y)= h(y)-\min_{_{\mathcal Y}} h$). Global minima of $h$ would then provide the concentration points of $\pi ^N$ in the limit (the global minimum should be unique, for most values of the parameters).\\
The forecoming analysis will not directly address this question.
We will focus on another issue, which is not elucidated in ~\cite{Antunes:08}: Metastability. In ~\cite{Gibbens:90} metastability is proved to hold for a rough one-dimensional diffusion approximation of the model considered. It tells that as $N$ gets large, the process stays trapped for some long time, of exponential order in $N$, in the neighborhood of any stable equilibrium point (regardless of its asymptotic significance in the invariant distribution).
The present section will show that both models in ~\cite{Antunes:08} and ~\cite{Gibbens:90} exhibit a metastable behavior, as a consequence of the Large Deviations results of Freidlin and Wentzell (ref ~\cite{Freidlin:84}), here supposed to be still valid under slightly enlarged conditions.
A central notion in ~\cite{Freidlin:84} is the {\em quasipotential}. For diffusion-like perturbations of dynamical systems, this quantity appears in ~\cite{Freidlin:84} as the energy function mentioned above, under suitable hypothesis, among which uniqueness of the equilibrium point. It is not clear that this holds for Markov jump processes as those considered here. Nevertheless, the quasipotential is involved in estimations of exit times that give evidence of metastability.
Here, it will be moreover shown that the Lyapunov function exhibited in ~\cite{Antunes:08} is actually equal to the quasipotential of a slight variant of the model of ~\cite{Antunes:08} associated with the same dynamical system.\\
The following very simple observation opens the way to using the Large Deviation results of Freidlin and Wentzell. The two models in ~\cite{Antunes:08} and ~\cite{Gibbens:90}, recalled in ~\ref{subsec:GHK} and ~\ref{subsec:AFRT}, have a similar structure: For each $N$, they are described by some irreducible Markov processes $(Y^N(t), t \ge 0)$ on the finite subset $\mathcal Y ^N=\{ y \in\mathcal{Y}: Ny_n \in \mathds{N} \text{ for all } n\in \mathcal{X} \}$ of $\mathcal{Y} =\mathcal P(\mathcal{X})$, with transitions $\ y \longrightarrow y+\frac{z}{N} $, where $z$ lives in some finite set $Z \subset \{(z_n) _{n \in \mathcal{X}} \in \mathds{Z}^{\mathcal{X}}: \sum_n z_n=0\}$ which is independent of $y$ and $N$. Moreover the rate of jump from $y$ to $ y+\frac{z}{N} $ is given by
\begin{align} \label{eq:rates} Q^N(y, y+\frac{z}{N})= N \Big( \mu_y(z) +O(1/N) \Big),\end{align}
where $(\mu _y)_{y \in\mathcal Y}$ is a family of positive measures on $Z$ with index $y$ in $\mathcal{Y}= \mathcal P (\mathcal{X})$ such that
\begin{itemize}
\item $ \mu_y(z)$ is continuous with respect to $y\in \mathcal{Y}$ for fixed $z$,
\item $ \mu_y(z)=0$ whenever $z_n <0$ for some $n\in \mathcal{X}$ such that $y_n=0$, that is, if $y \in \partial \mathcal Y \equiv \left \{ y\in \mathcal Y: \exists n, y_n=0\right \}$ and if the jump $(y, y+z)$ is directed toward the outside of $\mathcal Y$.
\end{itemize}
In the above expression of jump rates, $O(1/N)$ denotes functions of $N$, $y$ and $z$, whose product by $N$ is uniformly bounded in $y\in \mathcal Y$.
Jump rates ~\eqref{eq:rates} tell that in the neighborhood of some $y$, the process is approximately described by the random walk with jump rates $\mu _y(z)$, $z \in Z$, rescaled by accelerating time and shrinking space by the same factor $N$.
\\
For $y \in \mathcal Y$, denote $m_y$ the mean of measure $\mu _y$ ($\mu _y$ is not a probability measure in general):
\[ m_y = \sum _{z\in Z} z \mu _y(z) \quad (m_y \in \{(z_n) _{n \in \mathcal{X}}\in \mathds{R}^{\mathcal{X}}: \sum _{n \in \mathcal{X}} z_n=0\}) .\]
Then for $T<+\infty$, $(Y^N(t))_{0 \le t \le T}$ converges in distribution as $N$ tends to infinity to the dynamical system:
\begin{align} \label{eq:systdyn} \dot y (t) = m_{y(t)}.
\end{align}
In both examples of interest, $m_y=yL_y$ for some family $(L_y)$ of reversible generators. This will be essential for both interpretations of the Lyapunov function of ~\cite{Antunes:08}, respectively given in Theorem ~\ref{thm:modif} and Proposition ~\ref{prop:lyap}.
As mentioned in Section ~\ref{sec:mean_field}, the models in ~\cite{Gibbens:90} and ~\cite{Antunes:08} exhibit three equilibrium points (among which two are stable) for suitable values of the parameters.\\
If terms $O(1/N)$ in the jump rates ~\eqref{eq:rates} are omitted and $\mathcal{Y}$ is replaced by $\mathds{R} ^d$, the resulting processes belong to a class introduced by Freidlin and Wentzell in~\cite{Freidlin:84}, Chapter 5, as \emph{locally infinitely divisible processes}.
Their context is more general, since the infinitesimal generators they consider include both jump and diffusion terms, writing for $C^2$ functions $f$,
\[\Omega^N(f)= \langle m_y,\nabla f(y)\rangle + N \int _{z \in \mathds{R}^d} \left [ f(y+ \frac{z}{N}) -f(y) -\frac{1}{N} \langle z,\nabla f(y)\rangle \right ] d\mu _y(z) +\frac{1}{2N} \sum _{i,j} a_{ij}\frac{\partial ^2 f(y)}{ \partial y_i \partial y_j},\]
where, here and after, $\langle \, , \, \rangle$ denotes the usual scalar product in $\mathds{R} ^d$.
The main result in Chapter 5 of ~\cite{Freidlin:84} is that these processes satisfy a Large Deviation Principle on any finite time interval $[0,T]$, with scaling coefficient $N$ and action functional given by
\[S_{0T}(\varphi)= \left \{ \begin{array}{ll} \int _0 ^T L(\varphi _t, \dot {\varphi} _t) \, dt & \text{ for absolutely continuous } \varphi \text{ s.t.\ the integral is well defined},\\ + \infty & \text{ otherwise} \end{array} \right. \]
where $L(y,\cdot)$ is the Legendre transform of some $H(y,\cdot)$ given in our discrete pure jump case by:
\begin{align} \label{eq:H} H(y,\alpha)= \sum _{z \in Z} \mu _y(z) \Big( e^{\langle \alpha, z \rangle} -1 \Big) \quad (\alpha \in \mathds{R}^d) .\end{align}
Recall that this Legendre transform is defined, for $y \in \mathcal{Y}$ and $\beta \in \mathds{R} ^d$, by
\[ L( y, \beta)= \sup _{\alpha \in \mathds{R}^d} [\langle \alpha, \beta \rangle - H(y,\alpha)]. \]
Large Deviation estimates are then deduced for derived quantities as the \emph {exit point}, \emph {exit path} and \emph {exit time} from some domain included in the attraction basin of a stable equilibrium point $y_0$. The estimations for the exit point and exit time both involve the \emph {quasipotential} $V$ relative to $y_0$, which is defined as:
\[ V(y_0,y) =\inf \{ S_{0T}(\varphi): T>0, \varphi \text{ absolutely continuous }, \varphi _0=y_0, \varphi _T=y \}.\]
The exit point from some attracted domain $D$ is shown to be concentrated around the point $y$, if unique, minimizing $ V(y_0,y) $ on the boundary $\partial D$ of $D$, while the logarithm of the exit time is close to $N$ times the minimum value of $ V(y_0,y) $ on $\partial D$.
A Large Deviation Principle is stated for the invariant distribution, with action functional precisely given by $ V(y_0,y) $, but only in the pure diffusion case and when the equilibrium point $y_0$ is unique. This cannot thus be applied to the multistable models of interest to us.\\
The main large deviation result for sample paths (hence all its subsequent results) relies on a set of technical hypothesis about the function $H$ and its Legendre transform $L$. They can be summarized as follows:
I. finiteness of $\sup _ y H(y, \alpha)$;
II. finiteness of $L(y,\beta)$; local in $\beta$, uniform in $y$ boundedness of $L(y,\beta)$ and $\nabla _{\beta}L(y,\beta)$; strict, uniform in $y$ convexity of functions $L(y,.)$;
III. some specific equicontinuity property of functions $L(\cdot,\beta)$.
The last condition is essentially used for replacing arbitrary sample paths by polygons. The mere proof of lower semicontinuity of $S_{0T}$ makes use of it.
\\
The models in ~\cite{Antunes:08} and ~\cite{Gibbens:90} differ from the pure-jump processes of ~\cite{Freidlin:84} both through the second order terms $O(1/N)$ appearing in the jump rates, and through the compact state space $\mathcal{Y}$ standing in place of $\mathds{R}^d$.
Dropping one component of $y=(y_n)_{ n \in \mathcal{X}}$ ($\sum _n y_n=1$), $\mathcal{Y}$ can be identified with a compact subset of $\mathds{R}^{d}$ with $d= |\mathcal{X}|-1$. Transitions can then be extended outside $\mathcal{Y}$ to the entire space $\mathds{R}^d$. But II and III then fail to be true, due to the fact that for $y$ on the frontier of $\mathcal{Y}$, $L(y, \beta)= + \infty$ for all $\beta$'s outside some cone, depending on $y$ (and not only consisting of those $\beta$'s pointing out of $\mathcal{Y}$). ~\cite{Wentzell:76} slightly relax conditions I and II, but this is still not sufficient for our purpose. New arguments need thus be found to free from these conditions, in order to deal with sample paths hitting or starting on the boundary of $\mathcal{Y}$. Note that condition I is satisfied, due to compactness of $\mathcal{Y}$ and continuity of each $\mu _y(z)$ with respect to $y$.
As for the second order terms $O(1/N)$, here again the proof of the Large Deviation principle for sample paths (with action functional $S_{0T}$) needs to be adapted. The same change of measure technique might be used, but with $H^N$ and its Legendre transform $L^N$ instead of $H$ and $L$, where $H^N$ is defined analogously to $H$ in ~\eqref{eq:H}, replacing $\mu _y(z)$ by $\mu _y(z) +O(1/N)=1/N Q_N(y,y+z/N)$. The terms $O(1/N)$ should finally not infer on the estimations.
All this constitutes a challenging problem which is not addressed in this paper. For the present discussion, it is admitted that all the above mentioned results from Chapter 5 of ~\cite{Freidlin:84} (that is, the Large Deviation Principle for sample paths, from which all other results derive) still hold for the two models of interest in this paper.
\\
In all what follows, the state space $\mathcal{Y}$ is identified with the compact subset of $\mathds{R} ^d$, where $d=| \mathcal{X} | -1$: $\{ y \in [0, + \infty[^d : \sum _{i=1}^d y_i \le 1 \}$, dropping for example coordinate $y_0$ of $y = (y_n)_{ n \in \mathcal{X}} \in \mathcal{Y}$. Similarly, the set $Z$ of possible jumps is identified with a finite subset of $\mathds{R}^d$. $H(y, \alpha)$ is then defined by ~\eqref{eq:H} for $y \in \mathcal{Y}$ and $\alpha \in \mathds{R}^d$. This embedding of $\mathcal{Y}$ in the correct dimension space $\mathds{R}^d$ in which the process lives ensures at least that the conditions I to III of ~\cite{Freidlin:84} are satisfied in the interior of $\mathcal{Y}$.
Let us just mention that the lower semicontinuity of $S_{0T}$ can be proved in our context by describing $S_{0T}$ as the supremum of functionals $S^{\varepsilon}_{0T}$ that satisfy the conditions in ~\cite{Freidlin:84}. Indeed, extend the $\mu_y(z)$ to all $y \in \mathds{R}^{d}$ so that, for all $z \in Z$, $y \mapsto \mu_y(z)$ is continuous with compact support. Then introduce the following perturbations of the resulting $H$: For $\varepsilon >0$,
\[ H^{\varepsilon}(y, \alpha)= H(y,\alpha) + \varepsilon \sum _{z \in Z'} \Big( e^{\langle \alpha, z \rangle} + e^{-\langle \alpha, z \rangle} -2 \Big), \quad (y, \alpha \in \mathds{R}^{d}),\]
where $Z'$ is some finite subset of $\mathds{R}^{d}$ whose generated convex cone is the whole space. It can then be proved that the $H ^{\varepsilon}$ and their Legendre transforms $L ^{\varepsilon}$ satisfy conditions I to III (following the lines in the last section of ~\cite{Wentzell:76} for III). According to ~\cite{Freidlin:84} (theorem 2.1 of Chapter 5) or ~\cite{Wentzell:76}, the associated $S ^{\varepsilon} _{0T}$ are then lower semicontinuous on the set of continuous paths in $\mathds{R} ^d$ endowed with the uniform norm topology.
Now $H ^{\varepsilon}$ clearly decreases to $H$ as $\varepsilon$ decreases to $0$, so that $L ^{\varepsilon}$ increases and
\[ L(y, \beta) = \sup _{\alpha} [\langle \alpha, \beta \rangle - H(y,\alpha)]= \sup _{\alpha} \sup _{\varepsilon} [\langle \alpha, \beta \rangle - H ^{\varepsilon} (y,\alpha)] = \sup _{\varepsilon} L ^{\varepsilon} (y,\beta). \]
It follows by monotone convergence that $ S _{0T} = \sup _{\varepsilon}S ^{\varepsilon} _{0T}$, from which the lower semicontinuity of $S_{0T}$ follows.
Notice that semicontinuity of $ S _{0T} $ together with condition I imply (see ~\cite{Wentzell:76}) that the level sets of $ S _{0T} $ (that is, the sets $\{ \varphi : S_{0T}(\varphi) \le s \}$ for $s \in [0, + \infty[$) are compact sets in the topology of uniform convergence. This is technically important for proving the Large Deviation Principle along the classical scheme.
\\
Metastability refers to the large deviation result for exit times from neighborhoods of stable equilibrium points.
Theorem 4.3 of Chapter 4 of ~\cite{Freidlin:84} gives the main estimate of such exit times for Gaussian perturbations of dynamical systems. It is then indicated in Chapter 5 how to adapt the proof to a jump-like perturbation of the above described form. In both situations, the result is derived from the Large Deviation Principle for sample paths.
Recall that an \emph {asymptotically stable} equilibrium point is an equilibrium point $y_0$ such that for any neighborhood $\mathcal N$ of $y_0$, there exists some smaller neighborhood $\mathcal N'$ such that any trajectory initiated in $\mathcal N'$ converges to $y_0$ without leaving $\mathcal N$.
Also, a domain $D$ is said to be \emph{attracted} to some equilibrium point $y_0$ if any trajectory initiated in $D$ converges to $y_0$ without leaving $D$.
For the model in ~\cite{Antunes:08}, the Lyapunov function $g$ ensures that all stable equilibrium points (that is, local minima of $g$) are asymptotically stable.
From the above discussion, one can deduce from Chapter 5 of ~\cite{Freidlin:84} the following:
\begin{corollary}
Assume that the Large Deviation Principle for sample paths, with action functional $S_{0T}$, is valid for the model in ~\cite{Antunes:08}. Let $y_0$ be any stable equilibrium point for the limiting dynamics, and let $g$ be the Lyapunov function given by ~\eqref{eq:g}. For any positive $\delta$, define $B_{\delta}$ as the connected component of $y_0$ in the set $\{y \in \mathcal{Y}: g(y) <g(y_0) + \delta \}$. Define
\[\tau ^N_{\delta} = \inf \{t>0: Y^N(t) \notin B_{\delta} \}.\]
If $\delta$ is small enough, then for any $\alpha >0$ and $y \in B_{\delta}$,
\[\lim _{N \to \infty} \mathds{P} _y(e^{N(V_0- \alpha)} <\tau^N_{\delta} < e^{N(V_0+ \alpha)}) =1\]
where $V_0= \min _{y' \in \partial B_{\delta}} V(y_0,y')$.
\end{corollary}
\begin{proof}
In order to apply Theorem 4.3 of Chapter 4 of ~\cite{Freidlin:84}, in its modified version discussed in Chapter 5, it must be proved that $y_0$ is an asymptotically stable equilibrium point for the dynamical system $\dot y =m_y$, that for small enough $\delta$, the domain $ B_{\delta}$ is attracted to $y_0$ and has smooth boundary $\partial B_{\delta}$, and that $\langle n(y),m_y \rangle <0$ for $y \in \partial B_{\delta}$, where $n(y)$ is the exterior normal at $y$.
As indicated in Section ~\ref{sec:mean_field}, the critical points of $g$ are isolated, so that the stable equilibria, that is, the local minima of $g$, are necessarily {\em strict} local minima. It is then classical (see for example ~\cite{Perko:91}) that, due to the Lyapunov property of $g$, $y_0$ is asymptotically stable.
Now choose some open ball $B(y_0,r)$ centered at $y_0$, with radius $r>0$ small enough so that $g(y)>g(y_0)$ for all $y \in ( B(y_0,r) \cup \partial B(y_0,r)) \setminus \{y_0 \}$ ($y_0$ is a strict minimum of $g$) and that any trajectory initiated in $B(y_0,r)$ converges to $y_0$ ($y_0$ is asymptotically stable). Denote $\delta _0= \inf \{g(y)-g(y_0): y \in \partial B(y_0,r) \}$. Compactness of $\partial B(y_0,r)$ implies that $\delta _0>0$.
For any $\delta >0$, by the decreasing property of $g$, the set $ \{y \in \mathcal{Y}: g(y) <g(y_0) + \delta \}$ is invariant under the flow. The same is then true for the connected component $B_{\delta}$, by continuity of trajectories. In addition, for $\delta \le \delta _0$, by connectedness of $B_{\delta}$, $B_{\delta } \subset B(y_0,r)$ (if this were not the case, $B_{\delta}$ would meet $\partial B(y_0,r)$, contradicting $g(y)-g(y_0) < \delta$ on $B_{\delta}$).
It then results that $B_{\delta}$ is attracted to $y_0$.
It now only remains to prove, still assuming $\delta \le \delta _0$, that $\partial B_{\delta}$ is smooth and that $\langle n(y),m_y \rangle <0$ for $y \in \partial B_{\delta}$, where $n(y)$ the exterior normal of $\partial B_{\delta}$ at $y$. Both properties derive from the fact that $\partial B_{\delta}$ is a level surface for $g$. In particular, $ \nabla g(y) = \alpha _y n(y)$ for some non-negative $\alpha _y$, which is not zero since $y$ is not an equibrium point (take $\delta < \delta _0$ here). Then $\langle n(y),m_y \rangle= \langle \alpha_y ^{-1} \nabla g(y),m_y \rangle <0$ from the Lyapunov property of $g$.
The proof is achieved by applying the above cited theorem of ~\cite{Freidlin:84} in its version corresponding to the jump-process case of Chapter 5.
\end{proof}
\begin{remark}
For the model in ~\cite{Gibbens:90}, the same kind of estimation holds for exit times from neighborhoods of asymptotically stable points. Note however that little is known for this model: Multistability is obtained numerically, and asymptotical stability of stable points remains to be proved.
\end{remark}
A companion corollary can be stated, telling that the exit point from a neighborhood $\mathcal N$of $y_0$ is, with high probability as $N \to \infty$, close to the point that minimizes $V(y_0,y)$ on $\partial \mathcal N$, when this one is unique. However, this last condition is not guaranteed for $B_{\delta}$. As an example, for the modified model introduced below, $V(y_0,.)$ and $g$ coincide up to a constant on some neighborhood of $y_0$ (Theorem ~\ref{thm:modif}), so that $B_{\delta}$ is not a good choice in this case, $V(y_0,.)$ being \mbox {constant on $\partial B_{\delta}$. }
\\
It will now be proved that $g$ is equal to the quasipotential of a process obtained from the model in ~\cite{Antunes:08} by modifying the jumps in Section ~\ref{subsec:AFRT} as follows: Moves of customers from one node to another are replaced by departures and arrivals now occuring independently. In other words, jumps of the form $ y \longrightarrow y+\frac{1}{N} (e_{m+f_k}-e_m +e_{n-f_k}-e_n)$ are now split into jumps $ y \longrightarrow y+\frac{1}{N} (e_{m+f_k}-e_m)$ and $ y \longrightarrow y+\frac{1}{N} (e_{n-f_k}-e_n)$, each keeping the original rate. This modifies the Markovian dynamics, but not the limiting dynamical system.
The new transitions and rates are easily checked: For $1 \leq k\leq K$, $n \in \mathcal{X}$,
\[ y \longrightarrow \left \{ \begin{array}{l} y+\frac{1}{N} (e_{n+f_k}-e_n) \\ \\
y+\frac{1}{N} (e_{n-f_k}-e_n)
\end{array} \right.
\text{at rates} \begin{array}{l}
N \Big( (\lambda _k + \gamma _k [I_k,y])y_n {\mathds 1}_{n+f_k \in {\mathcal X}} + O(1/N) \Big) \\
\\ N \Big( (\mu _k + \gamma _k) n_k y_n + O(1/N) \Big)
\end{array} \]
Formally, the initial Markov process on $\mathcal{Y}^N$ has jump rates as in ~\eqref{eq:rates} with
\newline $\displaystyle{m_y= \sum _{z \in Z} z \mu _y(z) } =yL_y$ ($y \in \mathcal{Y}$) where $L_y$ is some infinitesimal generator (namely, that of an $M/M/C/C$ queue with $K$ classes of customers and for $k=1, \dots ,K$, arrival rates $ \lambda _k+\gamma _k [ I_k,y]$, service rates $\mu _k +\gamma _k$ and capacity requirements $A_k$).
So
\[ m_y= \sum _{(m,n) \in \mathcal{X}^2} (y_m L_y(m,n) -y_n L_y(n,m))e_n = \sum _{(m,n) \in \mathcal{X}^2} y_m L_y(m,n) (e_n -e_m). \]
For the modified process just described, transitions are reduced to ``elementary'' jumps of the form $y \longrightarrow y+\frac{1}{N} (e_n -e_m)$, with the associated rates $N (y_m L_y(m,n) + O(1/N))$. In other words, for the new process, \eqref{eq:rates} is still satisfied with $\mu _y(z)$ now defined for $y \in \mathcal{Y}$ and $z=e_n-e_m$, $n,m \in \mathcal{X}$, by
\begin{align} \label{eq:modifrates} \mu _y(e_n-e_m)= y_m L_y(m,n). \end{align}
This gives the same limiting vector field $(m_y)_{y\in \mathcal{Y}}$ as for the original process.
\noindent (Note that $L_y(m,n)$ is non zero for $m\neq n$ only if $n=m \pm f_k$ for some $k=1, \dots ,K$.)
\begin{thm} \label{thm:modif} (i) Let $H$ be the functional defined by ~\eqref{eq:H} for the above modified process. The Lyapunov function $g$ given by ~\eqref{eq:g} satisfies
\begin{align} \label{eq:nabla} H(y, \nabla g(y))=0 \quad (y \in \stackrel{\circ}{\mathcal{Y})},\end{align}
and
\[\nabla g(y)=0 \Longleftrightarrow m_y=0.\]
(ii) Let $y_0$ be a stable equilibrium of the dynamical system $\dot y= yL_y$ associated with the model of ~\cite{Antunes:08} or equivalently with its modified version.
Denote $V(y_0,y)$ the quasipotential of the modified process, relative to $y_0$.
The following equality holds on some neighborhood of $y_0$:
\[V(y_0,y)=g(y) -g(y_0). \]
\end{thm}
\begin{remark} \label {remark:nabla}
Before proving the theorem, let us make a remark on equation ~\eqref{eq:nabla}.
Assume that $g$ solves ~\eqref{eq:nabla} and satisfies
\begin{align} \label{eq:strict} \nabla g(y)=0 \Longrightarrow m_y=0,\end{align}
then $g$ is a Lyapunov function for the dynamical system $\dot y=m_y$.
Indeed due to strict convexity of $H(y,\cdot )$, the solutions $\alpha$ of $H(y,\alpha)\le 0$ are a strictly convex subset $C$ of $\mathds{R}^{d}$. Since the boundary $\partial C$ contains $0$, and $m_y = \nabla _{\alpha} H(y, 0) $ is the exterior normal at $0$, then $\langle m_y, \alpha\rangle \le 0$ for any $\alpha $ in $C$, and equality holds only for $\alpha =0$. In particular, equation ~\eqref{eq:nabla} implies that $\langle m_y,\nabla g(y)\rangle \le 0$, and that $\langle m_y,\nabla g(y)\rangle = 0$ holds only if $\nabla g(y)=0$, hence only if $m_y=0$ under assumption ~\eqref{eq:strict}.
\end{remark}
\begin{proof}
(i) $H$ writes:
\[ H(y, \alpha)= \sum _{(m,n) \in \mathcal{X}^2} y_m L_y(m,n) \Big(e^{\langle \alpha, e_n -e_m \rangle} -1\Big). \]
(Note that in order to fit conditions I to III of ~\cite{Freidlin:84} on $\stackrel{\circ}{\mathcal{Y}}$, $\alpha$ should vary in $\mathds{R}^{|\mathcal{X}|-1}$, $e_n-e_m$ being replaced by its projection on $\mathds{R}^{|\mathcal{X}|-1}$, and scalar products being understood in $\mathds{R}^{|\mathcal{X}|-1}$. But for computing $ H(y, \nabla g(y))$, it is easily checked that the result in unchanged if one keeps the original definition of $g$ as a function of $y \in \mathds{R}^{|\mathcal{X}|}$ and uses the scalar product in $\mathds{R}^{|\mathcal{X}|}$).
Denote $\nu _{\rho (y)}$, as in Section ~\ref{sec:mean_field}, the reversible distribution associated with generator $L_y$, and set $q_y(m,n)= \nu _{\rho (y)}(m) L_y(m,n)$ so that $q_y$ is symmetric in $(m,n)\in \mathcal{X} ^2$. Then
\begin{multline*} H(y, \nabla g(y))= \sum _{(m,n) \in \mathcal{X}^2} y_m L_y(m,n) \, e^{-\langle \nabla g(y), e_m \rangle} \Big(e^{\langle \nabla g(y), e_n \rangle} - e^{\langle \nabla g(y), e_m \rangle}\Big) \\
= \frac{1}{2} \sum _{(m,n) \in \mathcal{X}^2} \Big(y_m L_y(m,n) \, e^{-\langle \nabla g(y), e_m \rangle} - y_n L_y(n,m) \, e^{-\langle \nabla g(y), e_n \rangle} \Big) \Big(e^{\langle \nabla g(y), e_n \rangle} - e^{\langle \nabla g(y), e_m \rangle}\Big) \\
= \frac{1}{2} \sum _{(m,n) \in \mathcal{X}^2} q_y(m,n) \Big(\frac{y_m}{\nu _{\rho (y)}(m)} \, e^{-\langle \nabla g(y), e_m \rangle} - \frac{ y_n}{\nu _{\rho (y)}(n)} \, e^{-\langle \nabla g(y), e_n \rangle} \Big) \Big(e^{\langle \nabla g(y), e_n \rangle} - e^{\langle \nabla g(y), e_m \rangle}\Big)
. \end{multline*}
The result follows from the fact that ${\langle \nabla g(y), e_n \rangle}= \log \frac{ y_n}{\nu _{\rho (y)}(n)} +1- \log Z( \rho (y))$, as can easily be computed from equation ~\eqref{eq:g}, so that
$ \frac{ y_n}{\nu _{\rho (y)}(n)} \, e^{-\langle \nabla g(y), e_n \rangle}$ is independent of $n\in \mathcal{X}$.
(ii) is then essentially a consequence of a result by Freidlin and Wentzell, stated in ~\cite{Freidlin:84} as Theorem 4.3 of Chapter 5. Since a ``local'' variant of this result is actually needed, an independent proof is given, for completeness.
Recall that
\[V(y_0,y)= \inf \left \{ S_{0T}(\varphi): T>0, \varphi \text{ absolutely continuous}, \varphi _0=y_0, \varphi _T=y \right \} ,\]
and consider any $T>0$ and any absolutely continuous $\varphi$ on $[0,T]$ such that $ \varphi _0=y_0, \varphi _T=y$. Reversing time variable $t$ in the integral
$S_{0T}(\varphi)= \int _0 ^T L(\varphi _t, \dot {\varphi} _t) \, dt $ gives $S_{0T}(\varphi)= \int _0 ^T L(\psi _t, -\dot {\psi} _t) \, dt $ where $\psi _t= \varphi _{T-t}$ for $0 \le t \le 0$.
Introduce the following family of measures $(\tilde \mu _y)_{y \in \mathcal{Y}}$ on $-Z$, defined by
\[ \tilde \mu _y (z) = e^{- \langle \nabla g(y),z \rangle} \mu _y(-z) \quad (z \in- Z). \]
(Considering $\mu _y$ as the jump length distribution for the local random walk approximating $Y^N$ in the neighborhood of $y$, $\tilde \mu _y$ corresponds to the reversed random walk with respect to the measure $ ( e^{- \langle \nabla g(y),z \rangle})_{z \in Z}$, which is stationary by ~\eqref{eq:nabla} and ~\eqref{eq:H}.)
Define $\tilde H$ and $\tilde L$ associated with $(\tilde \mu _y)$ in the same way as $H$ and $L$ with $(\mu _y )$. The following relations are easily checked: for any $y \in \mathcal{Y}$ and $\alpha \in \mathds{R} ^{\mathcal{X}}$,
\[ \tilde H(y, \alpha)= H(y, \nabla g(y) - \alpha) \quad \text{ and } \quad \tilde L(y, \beta)= L(y, -\beta) + \langle \nabla g(y), \beta \rangle. \]
One gets
\[S_{0T}(\varphi)= \int _0 ^T \tilde L(\psi _t, \dot {\psi} _t) \, dt -\int _0 ^T \langle \nabla g(\psi_t), \dot {\psi} _t \rangle \, dt =\int _0 ^T \tilde L(\psi _t, \dot {\psi} _t) \, dt + g(y) -g(y_0) \]
since $\psi _0= y$ and $\psi _T=y_0$.
All that is left to prove now is that for $y$ close enough to $y_0$,
\begin{align} \label{eq:inf} \inf \left \{ \int _0 ^T \tilde L(\psi _t, \dot {\psi} _t) \, dt: T>0, \psi \text{ abs.~continuous}, \psi _0=y, \psi _T=y_0 \right \} =0 .
\end{align}
Using Remark~\ref{remark:nabla}, it can be shown that $g$ is a Lyapunov function for the ``locally reversed" dynamical system
\[\dot y= \tilde m_y \quad \text{ where } \tilde m_y= \sum _{z\in -Z} z \tilde \mu _y(z). \]
Indeed $\tilde H(y ,\nabla g(y) )= H(y, 0)=0$ for all $y$
and ~\eqref{eq:strict} is satisfied with $\tilde m_y$ in place of $m_y$ since $\nabla g(y)=0 $ both implies that $m_y=0$ by (i) and that $\tilde m_y=- m_y$ from the definition of $\tilde \mu _y$.
One can deduce from this that $y_0$ is also a stable equilibrium point for the dynamical system $\dot y= \tilde m_y$. As a consequence, there exists a neighborhood $\mathcal N$ of $y_0$ such that all trajectories initiated in $\mathcal N$ converge to $y_0$ at infinity. Assume from now on that $y \in \mathcal N$ and consider the trajectory $\psi$ initiated at $y $, then $\lim _{t \to + \infty} \psi _t = y_0$ and $ \int _0 ^{+ \infty} \tilde L(\psi _t, \dot {\psi} _t) \, dt =0$.
It is then easy to derive that the infimum in ~\eqref{eq:inf} is non-positive, using boundedness of $\tilde L(y, \beta)$ on compact subsets of $\stackrel {\circ}{\mathcal{Y}} \times \mathds{R} ^{\mathcal{X}}$. The result follows since this infimum is clearly non-negative (as $\tilde L \ge 0$).
\end{proof}
\begin{remark} \label{remark:balance}
We conclude this section with two heuristic remarks.
First, ~\eqref{eq:nabla} can be seen as the limiting balance equation satisfied by $\pi ^N$ if the approximation
\begin{align} \label{eq:pi} \pi ^N(y) \approx \frac {e^{-N g(y)}}{Z_N} \quad \text{ as } N \to \infty \end{align}
is valid with enough accuracy for some differentiable $g$.
Indeed, \mbox{the balance equations}
\[ \forall y \in \mathcal Y ^N \quad \sum _{z \in Z} \Big( \pi ^N(y-\frac{z}{N}) Q^N(y-\frac{z}{N},y) - \pi ^N(y) Q^N(y,y+\frac{z}{N}) \Big)=0,\]
become
$ \quad \displaystyle{ \frac {e^{-N g(y)}}{Z_N} \sum _{z \in Z} \mu _y(z) \Big( e^{\langle \nabla g(y), z \rangle} -1 \Big)=0} \ $ under ~\eqref{eq:pi},
since
\[ \pi ^N(y-\frac{z}{N})/\pi ^N(y) \approx e^{\langle \nabla g(y), z \rangle} \quad \text{ and } \quad Q^N(y-\frac{z}{N},y)\approx Q^N(y,y+\frac{z}{N}) \approx \mu _y(z).\]
This intuitively confirms the relation between solutions of ~\eqref{eq:nabla} and the quasipotential stated in (ii) of Theorem ~\ref{thm:modif} (or in Theorem 4.3 of Chapter 5 of ~\cite{Freidlin:84}), since in the unique equilibrium case, the quasipotential is expected to be the action functional associated to the invariant distribution.
Secondly, the above proof of (i) suggests that for the modified process, the measure $(e^{-Ng(y)})_{y\in \mathcal{Y}^N}$ solves the {\em local} balance equations in the limit. Indeed, from ~\eqref{eq:modifrates} and the proof of (i), $g$ solves
\begin{align} \label{eq:rev} \mu _y(z) - \mu _y(-z) e^{-\langle \nabla g(y), z \rangle}=0 \quad (y \in \mathcal{Y}, z \in Z).\end{align}
This relation appears as the limiting identity obtained from
\[ e^{-Ng(y)} Q^N(y,y+\frac{z}{N}) = e^{-Ng(y+\frac{z}{N})} Q^N(y+\frac{z}{N},y)\]
via the same approximations as in the first part of the present remark.
The modified process can thus be considered as asymptotically reversible. This is a particularity of the dynamical system in ~\cite{Antunes:08}. For the analogous modified version of the process in ~\cite{Gibbens:90}, no $g$ satisfies equation ~\eqref{eq:rev}.
\end{remark}
\section{Lyapunov function and relative entropy} \label{sec:lyapunov}
This section is devoted to a rereading of the Lyapunov function $g$, given by ~\eqref{eq:g}, for the model in ~\cite{Antunes:08}. It is connected to a well known decreasing property of relative entropy. Two other models can then be introduced, for which a similar relative entropy argument for constructing a Lyapunov function applies. Convergence of their stationary measure to a Dirac mass can then be derived.
\subsection{A reformulation of the Lyapunov function} \label{subsec:g}
Section ~\ref{sec:quasipot} has provided an interpretation of $g$ in terms of the quasipotential associated with another Markov random perturbation of the same dynamical system.
The Lyapunov property of $g$ is now given an interpretation in terms of relative entropy.
Underlying both interpretations, it appears that the model in ~\cite{Antunes:08} is very particular, due to the quantities $[I_k,y]$ that drive the $M/M/C/C$ generators $L_y$ involved in equation ~\eqref{eq:S}.\\
For $y,y' \in \mathcal{Y}$, two probability distributions on $\mathcal{X}$, the relative entropy of $y$ with respect to $y'$ is defined as:
\[h(y|y') =\sum _{n \in \mathcal{X}} y_n \log \frac{y_n}{y'_n}. \]
It is non-negative, and finite if $y' \in \, \stackrel {\circ}{\mathcal{Y}} = \{ y \in \mathcal{Y}: y_n >0 \text{ for all } n \in \mathcal{X} \}$.
Relative entropy appears in the following, easily checked, expression of $g$:
\begin{align} \label{eq:entg} g(y) = h(y|\nu _{\rho (y)})-\log Z(\rho(y)) +\sum _{k=1}^K \psi _k( [ I_k,y ]) \quad \quad (y \in \mathcal{Y})
\end{align}
where $\nu _{\rho (y)}$ is the Erlang reversible distribution of $L_y$ (here $\displaystyle{\rho _k(y)= \frac{ \lambda _k+\gamma _k [ I_k,y] }{\mu _k + \gamma _k}}$ for $1 \le k \le K$) and:
\[\psi _k(x)= \int _0^x \frac{\gamma _k u}{\lambda _k+ \gamma _k u} \, du \quad \quad (k=1, \dots ,K).\]
Note that, though no mention of relative entropy is made in ~\cite{Antunes:08}, it is there noticed that critical points of $g$ on $\mathcal{Y}$ correspond, through $\rho \mapsto \nu _{\rho}$, to critical points of some function of $\rho$ (given by the two last terms in the right hand side of ~\eqref{eq:entg}). This was the clue for a dimension reduction argument.
The Lyapunov property of $g$ will be explained from the classical decreasing property of the relative entropy between the distribution at time $t$ of some Markov ergodic process and its invariant distribution. Entropy dissipation and its quantification are widely present in the literature. Classicaly, estimating the entropy dissipation is an alternative to logarithmic Sobolev inequalities for obtaining exponential rates of decay to equilibrium (see for example ~\cite{Caputo:07}). In a different context, ~\cite{Yau:91} introduces a method for deriving hydrodynamical limits (see also ~\cite{Olla:93}, ~\cite{Kipnis:99}) by controlling the variations of the relative entropy between the current distribution and the so-called local equilibrium, varying in time. Our situation differs from the usual ones, in the sense that the generator itself varies in time: $y$ solves $\dot y (t)=y(t)L_{y(t)}$. However, this can be compensated by adding appropriate terms. This is made possible because $\rho _k(y)$'s depend on $y$ only through quantities $[ I_k,y ]$, that naturally appear in the expression of $ h(y|\nu _{\rho (y)})$. This method for building a Lyapunov function is thus restricted to very special dynamics. A different example will be given in ~\ref{subsec:closed}, for which the method applies due to invariance of $[I,y]$ along the flow.
In the following proof, the reversibility of generators $L_y$ manifests itself, as it is classical, through the Dirichlet form.\\
\begin{prop} \label{prop:lyap} Assume that for $y \in \mathcal{Y}$, $L_y$ is the infinitesimal generator of an $M/M/C/C$ queue with $K$ classes of customers having capacity requirements $A_k$ and arrival-to-service rate ratios $ \rho _k(y)= \varphi _k ( [ I_k,y])$ for $k=1, \dots ,K$, where the $\varphi _k$ are positive $C^1$ functions on $\mathds{R}$. Set $\rho(y)=(\rho _k(y), 1\le k \le K)$ for $y \in \mathcal{Y}$.
Then the following function is a Lyapunov function for dynamical system ~\eqref{eq:S}:
\[g(y)= h(y| \nu _{\rho(y)}) -\log Z(\rho(y)) +\sum _{k=1}^K \psi _k( [ I_k,y ]), \]
where, for $k=1, \dots ,K$, $\psi _k$ is some primitive of $\displaystyle{x \mapsto x \frac{\varphi ' _k(x)}{ \varphi _k (x)}}$.
\end{prop}
\begin{proof}
Using the definition of relative entropy together with ~\eqref{eq:erlang} gives for $y \in \mathcal{Y}$,
\[g(y)= \sum _{n \in \mathcal{X}} y_n \log \frac{n! y_n}{ \rho (y) ^n } + \sum _k \psi _k ( [ I_k,y ])= \sum _{n \in \mathcal{X}} y_n \log (n! y_n) + \sum _k \Big( \psi _k ( [ I_k,y ]) - [ I_k,y ] \log \rho _k (y) \Big).\]
Due to the relation, for $k=1, \dots ,K$, between $\psi _k$ and the $\varphi _k$ satisfying $\rho _k(y)=\varphi _k ( [ I_k,y ])$, derivation with respect to $y_n$ $(n\in \mathcal{X})$ simply yields:
\begin{align} \label{eq:nablag} \frac{\partial g(y)}{\partial y_n}= \log\frac{n! y_n}{\rho (y)^n } +1 .\end{align}
It must be proved that $ yL_y \nabla g(y) \le 0$ for all $y \in \mathcal{Y}$, equality holding only when $yL_y=0$.
Now $ yL_y \nabla g(y) $ can be expressed using the following identity due to reversibility of generator $L_y$ with respect to distribution $\nu _{\rho (y)}$:
for $u \in \mathds{R} ^{\mathcal{X}}$,
\begin{align} \label{eq:u} yL_y u = - \frac{1}{2} \sum _{(m,n) \in \mathcal{X} ^2} q_y(m,n) \Big(\frac{y_m}{\nu _{\rho(y)}(m)} -\frac{y_n}{\nu _{\rho(y)}(n)} \Big)( u_n -u_m)
\end{align}
where $q_y(m,n)= \nu _{\rho(y)}(m) L_y(m,n)$ is non-negative and symmetric in $(m,n)$.\\
(The last term is the Dirichlet form associated to $L_y$ evaluated at vectors $\displaystyle{ \Big (\frac{y_n}{ \nu _{\rho (y)}(n)} \Big)}$ and $u$.)
Then using ~\eqref{eq:nablag},
\[ yL_y \nabla g(y) = - \frac{1}{2} \sum _{(m,n) \in \mathcal{X} ^2} q_y(m,n) \Big(\frac{y_m}{\nu _{\rho(y)}(m)} -\frac{y_n}{\nu _{\rho(y)}(n)} \Big)\Big(\log \frac{y_m}{\nu _{\rho(y)}(m)} - \log \frac{y_n}{\nu _{\rho(y)}(n)}\Big).
\]
This shows that $ yL_y \nabla g(y) \le 0$ for all $y$. It can be zero only if $y_m/\nu _{\rho(y)}(m) =y_n/\nu _{\rho(y)}(n)$ for all pair $m,n$ such that $q_y(m,n)>0$. By irreducibility of $L_y$, this is possible only if $y_n/\nu _{\rho(y)}(n)$ does not depend on $n$, which means that $y=\nu _{\rho(y)}$.
\end{proof}
\subsection{Some Statistical Mechanics formalism} \label{mecastat}
Measures $\nu _{\rho}$ with $\rho \in ]0,+\infty[^K$, that include the fixed points of all our models, can be written in the following Gibbs form:
\[ \nu _{\rho}(n)= \frac{1}{Z(\rho)}\exp \sum _k (n_k \log \rho _k - \log n_k!)= \frac{1}{Z(\theta)}\exp (\langle \theta, n \rangle - \log n!) \quad (n\in \mathcal{X}).\]
where $\theta = (\theta _1, \dots ,\theta_K)$ is defined by $\theta _k = \log \rho _k$ ($1 \le k \le K$) and appears as the natural Gibbs parameter. It is then convenient to re-parametrize the family $(\nu _{\rho})$ as $(\nu _{\theta})$ for $\theta \in \mathds{R} ^K$ (abusively writing $Z(\theta)$ for $Z(\rho)$). This labelling is \mbox{clearly one-to-one.}
We now address the problem of minimizing the relative entropy distance of a given probability measure $y$ on $\mathcal{X}$ to the set of $\nu_{\theta}$'s.
It is related to some classical results in Statistical Mechanics. Yet, our arguments may not be standard. A more complete overview can be found in ~\cite {Ellis:85}, where this problem underlies some contraction principles related to Large Deviations of i.i.d.~random vectors.
Only the case $K=1$ is relevant for the present paper, since the two next models are one class, but Proposition ~\ref{prop:prox}, is worth mentioning in the multidimensional case for its own sake, or for possible extension of section ~\ref{subsec:open} to several classes.
First, it is classical that derivating the ``free energy" $\log Z$ gives the ``magnetization", that is, the expectation of the Gibbs measure:
\[ \nabla \log Z(\theta)=[ I, \nu _{\theta} ]. \]
Note that for $Z$ regarded as a function of $\rho$, it gives
\[ \frac{\partial \log Z(\rho)}{\partial \rho _k}= \frac{1}{\rho _k}[ I_k ,\nu _{\rho} ] = 1- B_k(\rho) \quad (k=1, \dots ,K), \]
using the well known relation $[ I_k ,\nu _{\rho} ] = \rho _k (1- B_k(\rho))$. Here for $k=1, \dots ,K$,
\[B_k(\rho)= \sum _{n \in \mathcal{X}: n+f_k \notin \mathcal{X}} \nu _{\rho}(n)\]
is the so-called ``blocking probability" corresponding to class $k$ for parameter $\rho$, that is, the probability that a new class $k$ customer is rejected in an $M/M/C/C$ queue with load $\rho$ in its stationary regime.\\
Next, $\log Z(\theta)$ is strictly convex with respect to $\theta$: This can be shown using the relative entropy between two Gibbs measures $\nu _{\theta}$. Indeed for $\theta, \theta ' \in \mathds{R} ^K$
\begin{align} \label{eq:gibbs_entropy} h(\nu _{\theta} | \nu _{\theta '}) = \sum _{n \in \mathcal{X}} \nu _{\theta}(n) \log \frac{\nu _{\theta}(n)}{\nu _{\theta '}(n)} = \log \frac{ Z(\theta ')}{Z(\theta)} + \langle [I, \nu _{\theta}], \theta '-\theta \rangle.
\end{align}
which rewrites
\[ \log Z(\theta ') -\log Z(\theta) = \langle \nabla \log Z_{| \theta}, \theta '-\theta \rangle + h(\nu _{\theta } | \nu _{\theta '}) .\]
This shows that the non-negative quantity $h(\nu _{\theta } | \nu _{\theta '}) $ (positive if $\theta ' \neq \theta$) measures the difference between the graph of $\log Z$ and its tangent hyperplane at $ \theta$. It proves strict convexity of $\log Z( \theta)$.
Properties of the relative entropy of some $y \in \mathcal{Y}$ with respect to some $\nu _{\theta}$ can then be derived. For $y \in \mathcal{Y}$ and $\theta \in \mathds{R} ^K$:
\begin{align} \label{eq:entropy} h(y| \nu _{\theta }) = \sum _{n \in \mathcal{X}} y_n \log \frac{y_n}{\nu _{\theta }(n)} = \log Z(\theta) +\sum _n y_n \log (n!y_n) -\langle [I, y], \theta \rangle,
\end{align}
so that for fixed $y$, the relative entropy $h(y| \nu _{\theta }) $ is also strictly convex in $\theta$ (as the sum of $\log Z(\theta)$ and an affine function of $\theta$).
It is not difficult to show that if $y \in \stackrel{\circ}{\mathcal{Y}}$ (that is $y \in\mathcal{Y}$ and $y_n>0$ for all $n$), then $h(y| \nu _{\theta }) $ tends to infinity as $\| \theta \| \to + \infty$: Indeed, for any $\theta \in \mathds{R} ^K$, the following inequalities hold:
\begin{multline*} \log Z(\theta) -\langle [I, y], \theta \rangle \ge \max _{n \in \mathcal{X}} \, (\langle \theta, n \rangle - \log n!) -\langle [I, y], \theta \rangle \\
\ge \max _{n \in \mathcal{X}} \, \langle n, \theta \rangle - \langle [I, y], \theta \rangle - \max _{n \in \mathcal{X}} (\log n!).
\end{multline*}
Now setting $h_y(\theta)= \max _{n \in \mathcal{X}} \langle n, \theta \rangle - \langle [I, y], \theta \rangle = \sum _m y_m ( \max _n \langle n, \theta \rangle - \langle m, \theta \rangle)$, $h_y(\theta)$ is positive for any $ \theta \ne 0$ since all $y_n$ are positive and the terms $ \langle m, \theta \rangle, m \in \mathcal{X},$ cannot be all equal (recall that, since $A_k \le C$ for $k =1, \dots , K$, the set $\mathcal{X}$ contains $0$ together with the canonical vectors $f_1, \dots , f_K$). Since $h_y$ is continuous on $\mathds{R}^K$, it is bounded from below by some positive constant on the unit sphere of $\mathds{R}^K$. Then using the fact that $h_y(\theta) = \| \theta \| \, h_y( \theta / \| \theta \|)$ for all non-zero $\theta$, it results that $h_y$, and hence $h(y| \nu _{\theta }) $, tends to infinity as $\| \theta \| \to + \infty$.
As a consequence, $h(y| \nu _{\theta }) $ attains one unique minimum on $\mathds{R} ^K$ at some value denoted $\bar {\theta} (y)$. Derivating $h(y| \nu _{\theta }) $ with respect to $\theta$ gives that $\bar {\theta} (y)$ is the unique solution of
\[ [I, \nu _{\theta }] = [I, y]. \]
(Note that unicity also appears on the following relation, itself derived from ~\eqref{eq:gibbs_entropy}:
\[ h(\nu _{\theta} | \nu _{\theta '}) + h(\nu _{\theta '} | \nu _{\theta }) = \langle [I, \nu _{\theta '}]- [I, \nu _{\theta}], \theta '-\theta \rangle,\]
which proves that equality $[I, \nu _{\theta '}]= [I, \nu _{\theta}]$ is possible only if $\nu _{\theta '}=\nu _{\theta }$, i.e., $\theta '= \theta$.)
One gets the following result:
\begin{prop} \label{prop:prox}
For any $y \in \stackrel{\circ}{\mathcal{Y}}$, there exists one unique $\theta \in \mathds{R} ^K$, denoted $\bar {\theta} (y)$ such that $[I, \nu _{\bar {\theta} (y)}] = [I, y]$.
This $\bar {\theta }(y)$ minimizes $\theta \mapsto h(y| \nu _{\theta }) $, and moreover satisfies
\[h(y| \nu _{\theta }) = h(y| \nu _{\bar {\theta} (y)}) + h(\nu _{\bar {\theta} (y)}| \nu _{\theta }) \quad \text{ for any } \theta \in \mathds{R} ^K. \]
\end{prop}
\begin{proof}
Only the last relation is left to prove, but it is a direct consequence of ~\eqref{eq:entropy}, ~\eqref{eq:gibbs_entropy} and $[I, \nu _{\bar {\theta} (y)}] = [I, y]$.
\end{proof}
\begin{remark}
From the additive formula in the previous proposition, it also results that for any fixed $ \theta, {\bar {\theta} } \in \mathds{R}^K$, the measure $ \nu _{\bar {\theta} }$ minimizes $y \mapsto h(y| \nu _{\theta }) $ on the set of probability measures $y$ having mean $[ I, \nu _{\bar {\theta} } ]$. Theorem VIII.4.1 in ~\cite {Ellis:85} shows that this minimum value can be read as some Legendre-Fenchel transform.
\end{remark}
As a corollary, getting back to the $\rho$-parametrization, one gets that the mapping $\rho \mapsto [I, \nu _{\rho}]$ is one to one from the set $]0 ,+ \infty[^K$ onto the set $\{ \sum _{n \in \mathcal{X}} ny_n, y \in \stackrel{\circ}{\mathcal{Y} } \}$. When $K=1$, the last set is simply the interval $]0,C[$, but for larger $K$ it is not easy to characterize. Like the convex hull of $\mathcal{X}$: $\{ \sum _{n \in \mathcal{X}} ny_n, y \in \mathcal \mathcal{Y} \}$, it arithmetically depends on integers $C$ and $A_k$'s in some intricate manner. In particular it does not coincide in general with the set $\{ m\in ]0, +\infty[^K: \sum _k A_k m_k <C \}$.
\begin{remark}
The term $\langle [I, y], \theta \rangle$ in relation ~\eqref{eq:entropy} explains the particularity of the case where $\rho _k(y)$'s are functions of $[I _k, y]$'s, as considered in Proposition ~\ref{prop:lyap}. Another nice situation, which is the case for the next model, is when $[I, y]$ remains constant along the trajectories of the dynamical system.
\end{remark}
\subsection{A closed system}\label{subsec:closed}
The next model again describes a system of $N$ nodes with the same capacity $C$, but here no rejection can occur: Customers are directly routed towards non saturated nodes (one can imagine that they are randomly rerouted as many times as necessary, at a null time cost, so as to find some node having the required free capacity).
In this model, there are no external arrivals nor departures. $M$ customers are present for ever in the system, with $M<NC$. Each customer spends an exponential time with mean one at each visited node, after which he chooses uniformly one new node among those, different from the current node, that are not saturated (if the current position is the only non saturated one, the customer does not move). All exponential variables and choices of successive nodes are assumed independent. The model is analyzed through the following asymptotics: $N$ and $M=M(N)$ tend to infinity with $M(N)/N$ converging to some $\lambda \in ]0,C[$.
For fixed $N$ and $M$ with $M<NC$, the empirical measure process is defined as in ~\eqref{eq:Y} and here denoted $(Y_M^N(t), t\ge 0)$. It is a Markov jump process on the finite subset $\mathcal{Y}_M^N = \{y=(y_n, 0 \le n \le C) \in \mathcal{Y}: Ny_n \in \mathds{N} \text{ for } n= 0, \dots ,C \text{ and } [I,y ]=\sum _{n=1}^C ny_n=M/N \}$ of $\mathcal{Y}= \{y=(y_n, 0 \le n \le C) \in [0, + \infty[^{C+1}: \sum _{n=0}^C y_n=1\} $. The transitions are the following ($e_n$ still denotes the $n^{th}$ unit vector in $\mathds{R}^{C+1}$):
\begin{align} \label{eq:closed} y \longrightarrow y+\frac{1}{N} \big( e_{m+1}-e_m +e_{n-1}-e_n\big) \text{ at rate } N ny_n \frac{Ny_m-\indicator {m=n}}{N(1-y_c)},
\end{align}
for $0 \le n \le C$ and $0 \le m \neq n \le C-1$.
(This corresponds to some move from some node in state $n$ to some node in state $m$. Note that due to the condition $M<NC$, $y_C<1$ for $y \in \mathcal{Y}^N_M$.)
It is clear that $Y_M^N$ is irreducible and thus admits one unique invariant distribution.
\begin{remark}
Extension of this model to the case with $K \ge 2$ classes of customers (with capacity requirements $A_1, \dots ,A_K$) is problematic, since in this case, the Markov process analogous to $Y^N_M$ is possibly non irreducible. For example, consider the case of $N$ nodes, each of capacity $6$, and $2N$ customers, $N$ of each of two classes with $A_1=3$ and $A_2=2$. The distribution with one customer of each class at each node does not communicate with any other configuration (e.g., for $N$ even, with the configuration with $N/2$ nodes occupied by two class $1$ customers and the $N/2$ others by two class $2$ customers), because only one unit of free capacity is available at each node. And yet, the total free capacity goes to infinity proportionally to $N$.
This problem could be solved in some sense by replacing this closed system by the open (irreducible) one next introduced in ~\ref{subsec:open}, extended to the multiclass case (see Remark ~\ref{remark:multiclass} below), that should have the same limiting dynamics in restriction to $\mathcal{Y} _{\lambda}$.
\end{remark}
Analogously to the previous examples, it can be shown that as $N,M$ go to infinity with $M/N $ converging to $\lambda$, the process $(Y^N_M(t), 0 \le t \le T)$ converges in distribution, for any finite $T$, to the solution with initial value $y(0)$ of the following differential system of equations:
\begin{multline} \label{eq:closed_dyn} y_n'(t)= \frac{\lambda}{1-y_C} y_{n-1}(t) \mathds 1_ {n\ge 1} + (n+1) y_{n+1} (t) \mathds 1_ {n\le C-1}
\\ - \Big(\frac{\lambda}{1-y_C} \mathds 1_ {n\le C-1} +n \Big) y_n(t) \quad (n=0, \dots ,C).
\end{multline}
provided that $Y^N_M(0)$ converges in distribution to $y(0)$.
Since $M/N$ converges to $\lambda$, the initial point $y(0)$ necessarily belongs to the set $\mathcal{Y}_{\lambda}= \{y \in \mathcal{Y}: \sum _n ny_n=\lambda \}$. As the next proposition will show, unsurprisingly,
the set $\mathcal{Y} _{\lambda}$ is invariant under the above system. Now since $\lambda <C$, $y_c<1$ for any $y\in \mathcal{Y}_{\lambda}$, so that for $y(0) \in \mathcal{Y}_{\lambda}$, the solution of \eqref{eq:closed_dyn} is defined on the whole time axis.
This system rewrites as ~\eqref{eq:S} where now, for $y \in \mathcal{Y} _{\lambda}$, $L_y$ is the rate matrix of an $M/M/C/C$ queue with arrival rate $\rho(y)=\frac{\lambda}{1-y_C} $ and service rate $1$.
\vspace{1mm}
Equilibrium points are again characterized by ~\eqref{eq:fix}, where $\rho(y)= \frac{\lambda}{1-y_C}$ and $\nu _{\rho}$ is defined by ~\eqref{eq:erlang} for $\rho>0$. This fixed point equation has one unique solution. Indeed, since $0< \lambda < C$, Proposition ~\ref{prop:prox} above shows that there is a unique $\rho \in ]0,+ \infty[$ such that $[ I,\nu _{\rho}]= \lambda$.
Denote it $\rho _{\lambda}$. Equation ~\eqref{eq:fix} rewrites: $y=\nu _{\rho}$ with $\rho= \frac{\lambda}{1-B(\rho)} $ and $B(\rho)= \nu _{\rho}(C)$, or equivalently, using the relation $[ I,\nu _{\rho} ]=\rho (1-B(\rho)) $: $[ I,\nu _{\rho } ]= \lambda$, so that $\nu _{\rho _{\lambda}}$ is the unique equilibrium point for ~\eqref{eq:S} on $\mathcal{Y} _{\lambda}$.
\begin{prop} \label{prop:closed} Fix $\lambda \in ]0,C[$ and assume that for $y \in \mathcal{Y} _{\lambda}$, $L_y$ is the rate matrix of an $M/M/C/C$ queue with arrival rate $\rho(y)=\frac{\lambda}{1-y_C} $ and service rate $1$. Then
(i) The set $\mathcal{Y}_{\lambda}$ is invariant under the dynamical system ~\eqref{eq:S}.
(ii) $g(y)=h(y| \nu _{\rho _{\lambda}}) $ is a Lyapunov function for this dynamical system on $\mathcal{Y}_{\lambda}$.
(iii) The sequence of invariant probability distributions $\pi ^N$ of processes $Y^N_{M(N)}$ converges weakly to the Dirac mass at $\nu _{\rho _{\lambda}}$ as $N,M(N) \to \infty$ with $M(N)/N \to \lambda$.
\end{prop}
\begin{proof}
(i) Set $f(y)= \sum _n ny_n$ ($y \in \mathcal{Y}$). Then $\nabla f(y)=(n, 0 \le n \le C)$. The derivative of $f$ along the dynamical system ~\eqref{eq:closed_dyn} is given by $yL_y \nabla f(y)$. Using ~\eqref{eq:u} with the present reversible generators $L_y$ and $u_n=n$ for $0 \le n \le C$ gives
\[ yL_y \nabla f(y)= - \sum _{n=0}^{C-1} q_y(n,n+1) \Big(\frac{y_{n+1}}{\nu _{\rho(y)}(n+1)} -\frac{y_n}{\nu _{\rho(y)}(n)} \Big),\]
where $q_y(n,n+1)= \nu _{\rho(y)}(n)L_y(n,n+1)= \nu _{\rho(y)}(n+1)L_y(n+1,n)$ so that
\begin{multline} \label{eq:mean} yL_y \nabla f(y)= -\sum _{n=0}^{C-1} \Big( y_{n+1} L_y(n+1,n) -y_nL_y(n,n+1) \Big)\\
=- \sum _{n=0}^{C-1} \Big( (n+1) y_{n+1} - \rho(y) y_n \Big)= (1-y_C) \rho(y) - \sum _n ny_n =0 \end{multline}
for $y \in \mathcal{Y} _{\lambda}$, which proves invariance of $\mathcal{Y} _{\lambda}$ under ~\eqref{eq:closed_dyn}.
(ii) Using relation ~\eqref{eq:entropy} together with invariance of $[I,y ]$ under the present dynamical system gives for $y \in \mathcal{Y} _{\lambda}$
\[ yL_y \nabla g(y)= yL_y \Big(\log (n!y_n)+1\Big) _{0 \le n \le C}= yL_y \Big(\log \frac{n!y_n}{Z(\rho (y))}\Big)_{0 \le n \le C} , \]
(as shown by ~\eqref{eq:u}, $ yL_y u $ is inchanged through adding some constant to $u$).
Then, again using invariance of $[ I,y ]$ under the dynamics, which writes $ yL_y (n,0 \le n \le C) =0$, one also has
\[ yL_y \nabla g(y) = yL_y \Big(\log \frac{n!y_n}{ \rho (y) ^n Z(\rho (y))}\Big) _{0 \le n \le C} = yL_y \Big(\log \frac{y_n}{ \nu_{\rho (y)} (n)}\Big)_{0 \le n \le C} .\]
The argument is then the same as in the proof of Proposition ~\ref{prop:lyap}.
\vspace{1mm}
(iii) We only give a sketch of the proof, which is the same as in ~\cite{Antunes:08}. It is first proved that the infinitesimal generator $ \Omega ^N$ of the Markov jump process given by ~\eqref{eq:closed} converges, as $N, M\to \infty$ with $M=M(N)$ and $M/N \to \lambda$, to the degenerate generator given by: $\Omega f (y) = y L_y \nabla f(y) $. This convergence holds for $C^2$ functions $f$ on $\mathcal{Y}$, and is uniform in $f$. It is then standard that any weak limit $\pi$ of the invariant distributions $(\pi ^N)$ of generators $\Omega ^N$, solves $\pi \Omega =0$.
The Lyapunov function $g$ helps then proving that the Dirac mass at the unique equilibrium point $\nu _{\rho _{\lambda}}$ of the dynamical system ~\eqref{eq:closed_dyn} is the only invariant probability measure for generator $\Omega$. Indeed, $\pi \Omega =0$ implies in particular that $\pi \Omega g = \int _{\mathcal{Y}} yL_y \nabla g(y) \, \pi (dy)=0$. The integrand being non-positive, and zero only at $\nu _{\rho _{\lambda}}$, $\pi$ needs be supported by this single state.
One concludes that the Dirac mass at $\nu _{\rho _{\lambda}}$ is the only weak limit of the sequence $(\pi ^N)$, so by compactness of $\mathcal P(\mathcal{Y})$, it is the limit of $(\pi ^N)$.
\end{proof}
\begin{remark}
Note that the above $g$ could be replaced by $h(y)=h(y| \nu _{\rho}) $ for any fixed $\rho$ since, by Proposition ~\ref{prop:prox}, $g$ and $h$ only differ by the constant $h(\nu _{\rho _{\lambda}}| \nu _{\rho}) $.
\end{remark}
\subsection{An open version} \label{subsec:open}
The next model is analogous to the previous one, in the sense that customers are instantly directed to available nodes. But now, new customers enter the network, are served at some node and then leave the network. Only one class of customers, requiring one unit of capacity, is considered (see Remark~\ref{remark:multiclass} below for possible extension to several classes). Customers enter the system, still consisting of $N$ nodes, according to some Poisson process with intensity $\lambda N$. Each customer is instantaneously directed, if possible, toward some node chosen uniformly among those having one free unit of capacity, that he then occupies during an exponentially distributed time with mean $1$. If no such node exists, the customer is definitively rejected.
We denote by $Y^N(t)$ ($t\ge 0$) the empirical distribution of the nodes at time $t$, defined as in ~\eqref{eq:Y}. Its state space is the finite subset $\mathcal{Y} ^N = \{y=(y_n,0 \le n\le C) \in \mathcal{Y}: Ny_n \in \mathds{N} \text{ for } n= 0, \dots , C\}$ of $\mathcal{Y}= \{y=(y_n,0 \le n\le C) \in [0, + \infty [ ^{C+1}: \sum _ny_n=1 \}$.
Notice that the total number of customers present in the system: $N \sum _n nY^N_n= N[I, Y^N]$ is simply an $M/M/CN/CN$ queue with arrival rate $\lambda N$ and service rate $1$. Considering $[I, Y^N]$ as $N$ grows to infinity is equivalent to Kelly scaling for the $M/M/C/C$ queue. When $\lambda \ge C$ the situation is simple: In the limit, the renormalized queue gets close to some deterministic trajectory that is constant equal to $C$ after some time. This means that after some time, for large $N$, the system is saturated.
It will be assumed that $0 <\lambda <C$, so as to maintain the system in a no-rejection regime. Indeed this case corresponds to the subcritical regime of Kelly's asymptotic: $[I, Y^N]$ converges to some process with values in $]0,C[$ for $t>0$ (having limit $\lambda $ at infinity) so that in the limit, no rejection occurs.\\
$(Y^N(t), t \ge 0)$ is a family of irreducible Markov processes on $\mathcal{Y}$ with the following transitions, respectively corresponding to some arrival or departure at some node in state $n \in \mathcal{X}$:
\[ y \longrightarrow \left \{ \begin{array}{l} y+\frac{1}{N} (e_{n+1}-e_n) \\ \\
y+\frac{1}{N} (e_{n-1}-e_n)
\end{array} \right.
\text{at rate} \begin{array}{ll}
\lambda N \frac{y_n}{1-y_C} \mathds 1 _{ y_C<1} & (0 \le n \le C-1) \\
\\ N n y_n & (0 \le n \le C)
\end{array} \]
Since each $Y^N$ actually evolves in a finite subset of $\mathcal{Y}$, these processes are ergodic; their invariant distributions will be denoted $\pi ^N$.
Apart from the particular case $C=1$, in which coordinate $Y_1^N$ of $Y^N$ (that is the proportion of occupied nodes) is itself a renormalized $M/M/N/N$ queue, $Y^N$ is non-reversible and its invariant distribution is not explicitly known.
The process $ Y^N(t)$ converges in distribution, on any finite time interval $[0,T]$, to the solution of the same differential system ~\eqref{eq:closed_dyn} as in the closed case, here considered on the enlarged
space $\mathcal{Y} \setminus \{\delta _C\}$, instead of $\mathcal{Y}_{\lambda}$.
This holds provided that $Y^N(0)$ converges in distribution to some $y(0)$ such that $y_C(0) <1$. The assumption that $\lambda <C$ is crucial here, ensuring that the condition $y_C <1$ is preserved in time: Indeed the last equation in the above differential system writes $y'_C(t)= \lambda \frac{y_{C-1}(t)}{1-y_C(t)} -C y_C(t) \le \lambda -C y_C(t)<0$ for $ y_C(t)$ close to $1$, so that $y_C$ cannot reach the value $1$, guaranteeing existence and unicity of a solution to ~\eqref{eq:closed_dyn} defined for all positive times.
\\
This system can again be read as ~\eqref{eq:S} where $L_y$ is the infinitesimal generator of an $M/M/C/C$ queue with arrival rate $ \frac{\lambda }{1-y_C} $ and service rate $1$. Equilibrium points are thus characterized by ~\eqref{eq:fix}, where $\rho (y)= \frac{\lambda }{1-y_C}$, or by $y=\nu _{\rho}$ with $\rho= \frac{\lambda}{1-B(\rho)} $. Then as in the previous model, using the relation $[ I,\nu _{\rho}]=\rho (1-B(\rho)) $ together with Proposition ~\ref{prop:prox}, $\nu _{\rho _{\lambda}}$ is the unique equilibrium point for ~\eqref{eq:S} on $\mathcal{Y} _{\lambda}$, where $\rho _{\lambda}$ is characterized by $[I, \nu _{\rho _{\lambda}}]=\lambda$.
\begin{prop}
As $N$ goes to infinity, the invariant distribution $\pi ^N$ of process $Y^N$ converges to the Dirac mass at $\nu _{\rho _{\lambda}}$.
\end{prop}
\begin{proof}
The Lyapunov function $g(y)=h(y| \nu _{\rho _{\lambda}}) $ for the previous closed system is no longer a Lyapunov function for the present model, though the differential system is formally the same. Indeed the state space is here larger, consisting of $\mathcal{Y} \setminus \{\delta _C\}$ instead of $\mathcal{Y}_{\lambda}$, and the Lyapunov property of $g$ relied on invariance of $[I,y]$ along the dynamical system ~\eqref{eq:closed_dyn} restricted to $\mathcal{Y}_{\lambda}$: This is no longer valid on $\mathcal{Y} \setminus \{\delta _C\}$.
Nevertheless one can restrict to the set $\mathcal{Y}_{\lambda}$ once the following is noticed: The quantity $l(y)=([I,y]- \lambda)^2$ decreases along the flow. This results from relation ~\eqref{eq:mean} in the proof of Proposition ~\ref{prop:closed}, that remains valid except for the last equality. Here
\[ yL_y \nabla f(y)= \lambda - [I,y] \quad \text{ where } \quad f(y) = [I,y] , \]
so that $ yL_y \nabla l(y)= -2(\lambda - [I,y] )^2 \le 0.$
Arguing as in the proof of (iii) of Proposition ~\ref{prop:closed} gives that any weak limit of the sequence $(\pi ^N)$ is supported by the set $\{y \in \mathcal{Y}: l(y)=0 \}$, that is $\mathcal{Y}_{\lambda }$. Then using the Lyapunov function $g(y)=h(y| \nu _{\rho _{\lambda}}) $ for the system restricted to $\mathcal{Y}_{\lambda }$ shows as previously that the Dirac mass at $\nu _{\rho _{\lambda}}$ is the only weak limit for the sequence $(\pi ^N)$, which proves convergence by compactness of $\mathcal P (\mathcal{Y})$.
\end{proof}
\begin{remark} \label{remark:multiclass}
Extension to several classes of customers here preserves irreducibility but raises the question: Under what condition, analogous to $\lambda<C$, should the previous results generalize?
Consider $K$ classes of customers with arrival rates $\lambda _k N$, service rates $\mu _k$ and capacity requirements $A_k$ ($A_k \ge 1$) for $k=1, \dots ,K$ (customers of class $k$ being here again, if possible, directed toward some node chosen uniformly among those having $A_k$ free units of capacity, and being rejected otherwise).
Now the total number of customers is no longer an $M/M/CN/CN$ queue: Indeed when some $A_k$ is larger than $1$, some class $k$ customer can be rejected though there are $A_k$ free units of capacity in the system, because no such volume of capacity is available at a single node. The condition $\sum _k A_k\frac{\lambda _k}{\mu _k} <C$ (ensuring that the $M/M/CN/CN$ queue with parameters $\lambda _k$'s, $\mu _k$'s, $A_k$'s and $C$ is subcritical) is thus irrelevant. (As an example, consider two classes of customers respectively requiring $A_1=3$ and $A_2=2$ units of capacity, while $C=3$. The process is exactly the same as if parameters were $A_1=A_2=C=1$, since each node can be occupied by at most one customer).
The right condition should be $ (\lambda _k/\mu _k)_k \in \{ [I,y], y \in \stackrel{\circ} {\mathcal{Y}} \}$, which is equivalent to existence of a $ \rho $ such that $[I, \nu _{\rho }] = (\lambda _k/\mu _k)_k$, by Proposition ~\ref{prop:prox}. This condition says that $ \lceil N \lambda _k/\mu _k \rceil$ customers of each class $k$ can be simultaneously accomodated for $N$ sufficiently large.
But contrary to the one class case, this condition does not ensure that the system stays in a non blocking regime in the limit. For example in the very particular case when $A_1= \dots = A_k =C=1$, the empirical process (simply consisting of the densities of customers of the different classes) is itself a renormalized $M/M/N/N$ queue. The previous condition writes $\sum _k \lambda _k/\mu _k <1$, which is the subcritical regime of Kelly. The limiting dynamics is known (see for example ~\cite{Fricker:03}) and for some values of the parameters, the trajectories can spend some non negligible time on the blocking region $y_0=0$.
\end{remark}
|
\section{Introduction}\label{Introduction}
Let $(X, d_X)$ and $(Y, d_Y)$ be metric spaces. A map $\phi \colon X \rightarrow Y$ is called a {\em quasi-isometry} if it satisfies the following two conditions:
\begin{enumerate}
\item There exists constants $a \geq 1, b \geq 0$ such that for $x_1, x_2 \in X$
\[ \frac{1}{a} d_X(x_1, x_2) - b \leq d_Y( \phi(x_1), \phi(x_2)) \leq a d_X(x_1, x_2) + b. \]
\item There exists a positive constant $c$ such that for each $y \in Y$, there exists an $x \in X$ that satisfies $d_Y(\phi(x), y) < c.$
\end{enumerate}
Let $G$ be a graph and let $x$ be a vertex of $G$. The set of neighbors of $x$ will be denoted by $N_x$ and $deg(x)$ will denote the number of neighbors of $x$. We shall say that $G$ is of {\em bounded degree} if there exists a positive integer $k$ such that $deg(x) \leq k$ for every vertex $x$ of $G$. A path in $G$ is a sequence of vertices $x_1, x_2, \dots, x_n$ where $x_{i+1} \in N_{x_i}$ for $1 \leq i \leq n-1$. A graph $G$ is connected if any two given vertices of $G$ are joined by a path. All graphs considered in this paper will be countably infinite, connected, of bounded degree with no self-loops. Two vertices $x$ and $y$ in $G$ are connected by an edge if and only if $y \in N_x$. Assign length one to each edge of $G$, then $G$ is a metric space with respect to the shortest path metric. Let $d_G(\cdot, \cdot)$ denote this metric. So if $x$ and $y$ are vertices in $G$, then $d_G(x,y)$ is the length of the shortest path joining $x$ and $y$. We will drop the subscript $G$ from $d_G( \cdot, \cdot)$ when it is clear what graph $G$ we are working with. If $A$ is a set of vertices of $G$ then $\#A$ will denote the cardinality of $A$.
Let $M$ be a complete, connected, and non-compact, smooth Riemannian manifold of dimension $n \geq 2$. Using the Riemannian distance, which we will denote by $d_M(\cdot, \cdot)$, $M$ is also a metric space. We will use $dx$ for the Riemannian volume element. For $x \in M, B_r(x)$ will denote the metric ball centered at $x$ of radius $r$; and $Vol(S)$ will be the volume of a measurable set $S \subseteq M$. In addition to the conditions in the first sentence of this paragraph, all manifolds considered in this paper will also have the following properties:
\begin{description}
\item[(V)] There are positive increasing functions $V_0(r)$ and $V_1(r)$ on $(0, \infty)$ that satisfy
\[ V_0(r) \leq Vol(B_r(x)) \leq V_1(r) \]
for all $x \in M$.
\item[(P)] For $r > 0$, there exists a real number $C_r$ such that for any $y \in M$ and any smooth function $f$ on $B_r(y)$
\[ \int_{B_r(y)} \vert f(x) - \bar{f} \vert \, dx \leq C_r\int_{B_r(y)} \vert \nabla f(x) \vert \, dx, \]
where $\bar{f} = (Vol (B_r(y)))^{-1} \int_{B_r(y)} f(x) \, dx.$
\end{description}
These properties are satisfied by any complete manifold $M$ where the Ricci curvature of $M$ is uniformly bounded from below by $-(n-1)K^2$, where $K > 0$, and the injective radius of $M$ is positive.
Let $p$ be a real number greater than one. In section \ref{preliminaries} we will define the $p$-harmonic boundary for both graphs and manifolds. It was shown in \cite[Theorem 2.7]{PulsPJM} that if $G$ and $H$ are quasi-isometric graphs, then their $p$-harmonic boundaries are homeomorphic. On the other hand Theorem 2 of \cite{Lee} says that if $M$ and $N$ are quasi-isometric manifolds, then their $p$-harmonic boundaries are homeomorphic. A reasonable question to ask is the following: if a graph $G$ of bounded degree is quasi-isometric with a complete Riemannian manifold $M$, how are their $p$-harmonic boundaries related? Reinterpreting Theorem 2 of \cite{Kanai} into our setting it was shown that if $G$ is quasi-isometric with $M$, then the $p$-harmonic boundary of $G$ is empty if and only if the $p$-harmonic boundary of $M$ is empty. Theorem 1.1 of \cite{Holasor} says that if there is a quasi-isometry from $G$ to $M$ then the $p$-harmonic boundary of $G$ contains one element if and only if the $p$-harmonic boundary of $M$ contains one element. In this paper we extend these results by proving
\begin{Thm} \label{boundhomeo}
Let $G$ be a graph of bounded degree and let $M$ be a complete Riemannian manifold with dimension at least two and that has properties (V) and (P). If $G$ and $M$ are quasi-isometric, then their $p$-harmonic boundaries are homeomorphic.
\end{Thm}
In Section \ref{preliminaries} we will also define what it means for a function to be $p$-harmonic on $G$ and $M$. Our other main result for this paper is:
\begin{Thm} \label{bijectionpharm}
Let $G$ be a graph of bounded degree and let $M$ be a complete Riemannian manifold with dimension at least two and that has properties (V) and (P). If $G$ and $M$ are quasi-isometric, then there is a bijection between the bounded $p$-harmonic functions on $G$ and the bounded $p$-harmonic functions on $M$.
\end{Thm}
This paper is organized as follows: In Section \ref{preliminaries} we give some preliminaries concerning the $p$-harmonic boundary and $p$-harmonic functions. In section \ref{nets} we define $\kappa$-nets and give some results concerning $\kappa$-nets that we will need. We prove our main results in Section \ref{proofsofmain}.
I would like to thank the referee for some useful remarks concerning this paper.
This work was partially supported by PSC-CUNY grants 60123-38 39 and 62598-00 40.
\section{Preliminaries} \label{preliminaries}
Let $1 < p \in \mathbb{R}$. In this section we will define the $p$-harmonic boundary and $p$-harmonic functions. Furthermore, we will state some properties of these concepts that will be needed later in the paper. We will also set some notation to be used in this paper, and give some facts that will be needed concerning estimates on volumes of metric balls. We begin by defining certain function spaces that will be used in our definitions. For more detailed explanations about these function spaces and about the $p$-harmonic boundary see Section 1 of \cite{PulsPJM} for graphs and for Riemannian manifolds see Section 1 of \cite{Lee} or Chapter III.1 of \cite{SarNak}.
Let $M$ be a complete Riemannian manifold. Denote by $D_p(M)$ the set of continuous real-valued functions on $M$ for which $\nabla f \in L^p(M)$, where $\nabla f$ is the distributional gradient of $f$. Set $BD_p(M)$ equal to the set of bounded functions in $D_p(M)$. Under the usual operations of function addition, pointwise multiplication of functions and scalar multiplication $BD_p(M)$ is a commutative algebra. Furthermore, $BD_p(M)$ is a Banach algebra with respect to the following norm
\[ \parallel f \parallel_{BD_p} = \parallel \nabla f \parallel_p + \parallel f \parallel_{\infty} \]
where $\parallel \cdot \parallel_{\infty}$ denotes the sup-norm and $\parallel \cdot \parallel_p$ is the $L^p$-norm. Let $C_c(M)$ be the set of continuous functions on $M$ with compact support. Denote by $B(\overline{C_c(M)}_{D_p})$ the closure of $C_c(M)$ in $BD_p(M)$ with respect to the following topology: A sequence $(f_n)$ in $C_c(M)$ converges to $f \in BD_p(M)$ if $\sup_K \mid f_n - f \mid \rightarrow 0$ as $n \rightarrow \infty$ for each compact subset $K$ in $M$, $(f_n)$ is uniformly bounded on $M$ and
\[ \lim_{n \rightarrow \infty} \int_M \mid \nabla (f_n - f)(x) \mid^p \, dx \rightarrow 0. \]
We now proceed to define analogous function spaces for a graph $G$ of bounded degree. Let $V$ be the vertex set of $G$ and let $x \in V$. For a real-valued function $f$ on $V$ we define the $p$-th power of the gradient and the $p$-Dirichlet sum by
\begin{equation*}
\begin{split}
\vert Df(x) \vert^p & = \sum_{y \in N_x} \vert f(y) - f(x) \vert^p, \\
I_p(f, V) & = \sum_{x \in V} \vert Df(x) \vert^p.
\end{split}
\end{equation*}
In this setting $D_p(G)$ will be the set of functions $f$ for which $I_p(f, V) < \infty$. Under the following norm $D_p(G)$ is a reflexive Banach space
\[ \parallel f \parallel_{D_p} = \left( I_p(f,V) + \vert f(o) \vert^p \right)^{1/p}, \]
where $o$ is a fixed vertex of $G$. Let $BD_p(G)$ be the set of bounded functions in $D_p(G)$. It is also the case that $BD_p(G)$ is a commutative algebra under the operations of function addition, pointwise multiplication and scalar multiplication. With respect to the following norm $BD_p(G)$ is a Banach algebra
\[ \parallel f \parallel_{BD_p} = \left( I_p(f,V) \right)^{1/p} + \parallel f \parallel_{\infty}, \]
where $\parallel \cdot \parallel_{\infty}$ is the usual sup-norm. The set $C_c(G)$ will consist of all functions on $V$ with compact support. Denote by $B(\overline{C_c(G)}_{D_p})$ the closure of $C_c(G)$ in $BD_p(G)$ with respect to the $D_p$-norm.
In what follows $X$ will be either $M$ or $G$. A character on $BD_p(X)$ is a nonzero homomorphism from $BD_p(X)$ into the complex numbers. Denote by $Sp( BD_p(X))$ the set of characters on $BD_p(X)$. With respect to the weak $\ast$-topology, $Sp(BD_p(X))$ is a compact Hausdorff space. The space $Sp (BD_p(X))$ is known as the spectrum of $BD_p(X)$. Let $C(Sp(BD_p(X)))$ denote the set of continuous functions on $Sp(BD_p(X))$. For each $f \in BD_p(X)$ a continuous functions $\hat{f}$ can be defined on $Sp(BD_p(X))$ by $\hat{f}(\tau) = \tau(f)$. Each $x \in X $ defines an element in $Sp(BD_p(X))$ via evaluation by $x$; that is, if $f \in BD_p(X)$, then $x(f) = f(x)$. It turns out that under this identification $X$ is an open dense subset of $Sp(BD_p(X))$. The compact Hausdorff space $Sp(BD_p(X))\setminus X$ is known as the {\em $p$-Royden boundary} of $X$, which we will denote by $R_p(X)$. Now, $B(\overline{C_c(X)}_{D_p})$ is closed in $BD_p(X)$ with respect to the $BD_p$-norm. The {\em $p$-harmonic boundary} of $X$ is the following subset of the $p$-Royden boundary
\[ \partial_p (X) \colon = \{ \tau \in R_p(X) \mid \hat{f}(\tau) = 0 \mbox{ for all } f \in B(\overline{C_c(X)}_{D_p})\}. \]
It can be shown that $\partial_p(X) = \emptyset$ if and only if $1 \in \overline{C_c(X)}_{D_p}$, where 1 is the constant function 1 on $X$. In this case it follows from Theorem 2 of \cite{Kanai} that $\partial_p(G) = \partial_p(M)$ because $X$ is $p$-parabolic if and only if $1 \in \overline{C_c(X)}_{D_p}$. For the rest of this paper it will be assumed that $X$ is not $p$-parabolic. From now on we will implicitly assume the following
\begin{Lem} \label{gotobound}
Let $x \in \partial_p(X)$ and let $(x_n)$ be a sequence in $X$ that converges to $x$. Then $d_X(o, x_n) \rightarrow \infty$ as $n \rightarrow \infty$, where $o$ is a fixed point in $X$.
\end{Lem}
\begin{proof}
Suppose there exists a real number $M$ such that $d_X(o, x_n) \leq M$ for all $n \in \mathbb{N}$. Define a function $\chi_M \in C_c(X)$ by $\chi_M (y) = 1$ if $d_X(o,y) \leq M$ and $\chi_M(y) = 0$ if $d_X(o,y) > M$. Then $\hat{\chi}_M(x) = \lim_{n \rightarrow \infty} \chi_M (x_n) = 1$, a contradiction. Thus $d_X(o, x_n) \rightarrow \infty$ as $n \rightarrow \infty.$
\end{proof}
Now suppose $X=M$ and let $W^{1,p}(M)$ be the set of functions $f \in L^p(M)$ for which $\nabla f \in L^p(M)$. If $h$ is a continuous function in $W^{1,p}_{loc}(M)$ that is a weak solution of
\[ -\mbox{div}(\vert \nabla h \vert^{p-2} \nabla h ) = 0, \]
then we shall say that $h$ is $p$-harmonic.
On the other hand if it is the case $X=G$, then $h$ is defined to be $p$-harmonic if
\[ \sum_{y \in N_x} \vert h(y) - h(x) \vert^{p-2}(h(y) - h(x)) = 0 \mbox{ for all }x \in V. \]
In the case $1 < p < 2$ we make the convention that $\vert h(y) - h(x) \vert^{p-2} (h(y) - h(x)) = 0$ if $h(y) - h(x) =0$.
Let $BHD_p(X)$ be the set that consists of all bounded $p$-harmonic functions on $X$ that are contained in $D_p(X)$. We now state some properties of $p$-harmonic functions and the $p$-harmonic boundary that will be needed in the sequel. For proofs of these and other properties see Section 4 of \cite{PulsPJM} for graphs and Section 2 of \cite{Lee} for manifolds.
\begin{Thm} \label{decomp} ($p$-Royden decomposition) Let $f \in BD_p(X)$. Then there exists a unique $u \in B(\overline{C_c(X)}_{D_p})$ and a unique $h \in BHD_p(X)$ such that $f = u + h.$
\end{Thm}
Using the $p$-Royden decomposition the following characterization of functions in $BD_p(X)$ that vanish on $\partial_p(X)$ can be obtained.
\begin{Thm} \label{charvan}
Let $f \in BD_p(X)$. Then $f \in B(\overline{C_c(X)}_{D_p})$ if and only if $\hat{f}(\tau) = 0$ for all $\tau \in \partial_p(X).$
\end{Thm}
Observe that it follows immediately from the theorem that if $h \in BHD_p(X)$ is the $p$-harmonic function in the $p$-Royden decomposition of $f \in BD_p(X)$, then $f(\tau) = h(\tau)$ for all $\tau \in \partial_p(X)$. Furthermore, the following is also a consequence of the previous theorem
\begin{Cor} \label{boundval} A function in $BHD_p(X)$ is uniquely determined by its values on $\partial_p(X)$.
\end{Cor}
Note that if $\partial_p(X)$ contains only one element, then $BHD_p(X)$ consists precisely of the the constant functions on $X$.
\section{$\kappa$-nets}\label{nets}
For the proofs of our main results we will need to use $\kappa$-nets. In this section we will explain what a $\kappa$-net is, and give some of its properties that will be useful for our needs.
A {\em net} is a countable set $\Gamma$ with a family $\{N_g\}_{g \in \Gamma}$ of finite subsets $N_g$ of $\Gamma$ such that for all $g, h \in \Gamma, g \in N_h$ if and only if $h \in N_g$. For $g \in \Gamma$, each element of $N_g$ is called a neighbor of $g$. It is possible to think of a net as a countable graph with vertex set $\Gamma$ by connecting vertices $g$ and $h$ in $\Gamma$ by an edge if $h \in N_g$. Thus the definitions given in Section \ref{Introduction} for properties of a graph, such as a path, carry over to a net.
Let $M$ be a complete Riemannian manifold. We shall say that a subset $\Gamma$ of $M$ is $\kappa$-separated for $\kappa > 0$ if $d_M(g, h) \geq \kappa$ whenever $g$ and $h$ are distinct points of $\Gamma$. Now assume that $\Gamma$ is a maximal $\kappa$-separated subset of $M$. By setting $N_g = \{ h \in \Gamma \mid 0 < d_M(g,h) \leq 3\kappa \}$ for each $g \in \Gamma$ we have a net structure on $\Gamma$. We define $\Gamma$ to be a {\em $\kappa$-net} if $\Gamma$ is a maximal $\kappa$-separated subset of $M$ with the net structure given above. It is easy to see that a $\kappa$-net in $M$ is connected due to our standing assumption that $M$ is connected. If $\Gamma$ is a $\kappa$-net, then it is also a countable, connected graph. Thus, $\Gamma$ is a metric space with respect to the shortest path metric. A couple of facts about $\kappa$-nets that we will need later are:
\begin{enumerate}
\item Let $\Gamma$ be a $\kappa$-net in $M$. Then for a given $r>0$, there exists a constant $C_r$ such that $\# (\Gamma \cap B_r(x)) \leq C_r$ for all $x \in M$.
\item Let $\Gamma$ be a $\kappa$-net in $M$. Then the inclusion map $\iota \colon \Gamma \rightarrow M$ is a quasi-isometry.
\end{enumerate}
The first fact can be proved by using the argument from Lemma 2.3 of \cite{Kanaicapp} since $Vol(B_R(x)) \leq \frac{V_1(R)}{V_0(r)} Vol (B_r(x)),$ where $0 < r < R < \infty$. This fact shows that $\Gamma$ is a graph of bounded degree; and that for each $x \in M$ there exists at most $C_r$ elements $g$ in $\Gamma$ for which $B_r(g)$ contains $x$. The second fact is \cite[Lemma 2.6]{Kanaicapp}, where it was assumed that the Ricci curvature was bounded from below; it was also proven in Lemma 2.13 of \cite{HoloRMI} without any curvature assumptions on $M$.
Let $\kappa$ be a small positive number and let $\Gamma$ be a $\kappa$-net in a complete Riemannian manifold $M$. Let $1 < p \in \mathbb{R}$. The rest of this section is devoted to describing how to map functions from $BD_p(\Gamma)$ to $BD_p(M)$, and vice-versa. For each $g \in \Gamma$, pick a smooth function $\eta_g \in C_c(M)$ such that $0 \leq \eta_g \leq 1, \eta_g = 1$ on $B_{\kappa}(g), \eta_g = 0$ outside of $B_{\frac{3\kappa}{2}} (g)$, and that $\vert \nabla \eta_g \vert \leq c$, where $c$ is a constant that does not depend on $g$. For $x \in M$ define
\[ \xi_g (x) = \frac{\eta_g(x)}{\sum_{ h \in \Gamma} \eta_h(x)} . \]
Now $\vert \nabla \xi_g \vert$ is uniformly bounded. Indeed,
\begin{eqnarray*}
\vert \nabla \xi_g \vert & \leq & \vert \nabla \eta_g \vert \left(\sum_{h \in \Gamma} \eta_h\right)^{-1} + \eta_g \sum_{h \in \Gamma} \vert \nabla \eta_h \vert \left(\sum_{h \in \Gamma} \eta_h \right)^{-2} \\
& \leq & \vert \nabla \eta_g \vert + \sum_{h \in \Gamma} \vert \nabla \eta_h \vert \\
& \leq &(k + 2)c,
\end{eqnarray*}
where $k$ is a constant that satisfies $\#N_g \leq k$ for all $g \in \Gamma$. Let $\bar{f} \colon \Gamma \rightarrow \mathbb{R}$. Define a smooth function $f \colon M \rightarrow \mathbb{R}$ by
\begin{equation}
f(x) = \sum_{g \in \Gamma} \bar{f}(g) \xi_g (x), \label{graphtorm}
\end{equation}where $x \in M$. We are now ready to state and prove
\begin{Prop} \label{definefrm}
If $\bar{f} \in BD_p(\Gamma)$, then $ f\in BD_p(M).$
\end{Prop}
\begin{proof}
Let $g \in \Gamma$ and suppose $x \in B_{\kappa}(g)$. Now
\begin{eqnarray*}
\nabla f(x) & = &\sum_{h \in N_g \cup \{g\}} \bar{f}(h) \nabla \xi_h (x) \\
& = &\sum_{h \in N_g} (\bar{f}(h) - \bar{f}(g)) \nabla \xi_h (x).
\end{eqnarray*}
The last equality is due to $\sum_{h \in N_g \cup \{g\}} \bar{f}(g) \xi_h(x) = \bar{f}(g)$ and $\sum_{h \in N_g \cup \{g\}} \nabla \xi_h (x) =0$. We now obtain
\begin{eqnarray*}
\vert \nabla f(x) \vert^p & \leq & \left( \sum_{h \in N_g} \vert (\bar{f}(h) - \bar{f}(g)) \nabla \xi_h (x) \vert \right)^p \\
& \leq & \left( \left( \sum_{h \in N_g} \vert \bar{f}(h) - \bar{f}(g) \vert^p \right)^{1/p} \left( \sum_{h \in N_g} \vert \nabla \xi_h (x) \vert^q \right)^{1/q} \right)^p \\
& \leq & \vert D\bar{f}(g) \vert^p (ck^{1/q})^p,
\end{eqnarray*}
where $\frac{1}{q} + \frac{1}{p} = 1$ and $k$ is a constant with $\#N_g \leq k$ for all $g \in \Gamma$. It now follows that,
\[ \int_M \vert \nabla f(x) \vert^p dx \leq \sum_{g \in \Gamma} \int_{B_{\kappa}(g)} \vert \nabla f(x) \vert^p dx \leq (ck^{1/q})^p V_1(\kappa) \sum_{g \in \Gamma} \vert D\bar{f}(g)\vert^p \]
Hence, $f \in BD_p(M)$.
\end{proof}
\begin{Cor} \label{compactfrm}
If $\bar{f} \in B(\overline{C_c(\Gamma)}_{D_p})$, then $f \in B(\overline{C_c(M)}_{D_p}).$
\end{Cor}
\begin{proof}
By the proposition $f \in BD_p(M)$. Now let $(\bar{f}_n)$ be a sequence in $C_c(\Gamma)$ that converges to $\bar{f}$. For each $n, f_n \in C_c(M)$ because $\bar{f}_n \in C_c( \Gamma)$. We will now show that $f \in B(\overline{C_c(M)}_{D_p})$. By using the argument from the above proposition we see that
\[ \int_M \vert \nabla (f_n -f)(x) \vert^p dx \leq C I_p(\bar{f}_n - \bar{f}, \Gamma), \]
where $C$ is a constant. Consequently, $\int_M \mid \nabla (f_n - f)(x) \mid^p dx \rightarrow 0$ as $n \rightarrow \infty$. Let $K$ be a compact subset in $M$. Set $\Gamma_K = \{ g \in \Gamma \mid d_M(g, K) < \frac{3\kappa}{2} \}$. Now $\# \Gamma_K$ is finite because $K$ is compact. Let $x \in K$ and let $\epsilon > 0$. Since $\xi_g (x) = 0$ if $g \notin \Gamma_K$ we obtain
\[ \vert f_n (x) - f(x) \vert \leq \sum_{g \in \Gamma_K} \vert \bar{f}_n(g) - \bar{f}(g) \vert. \]
For $h \in D_p(\Gamma)$ and $g \in \Gamma$, there exists a constant $C_g$ depending on $g$ such that $\vert h(g) \vert \leq C_g \parallel h \parallel_{D_p}$. Thus $(\bar{f}_n) \rightarrow \bar{f}$ pointwise. Hence for each $g \in \Gamma_K$ there exists a number $N(g)$ such that for $n > N(g), \vert \bar{f}_n (g) - \bar{f}(g) \vert < \frac{\epsilon}{\# \Gamma_K}$. Set $N = \max_{g \in \Gamma_K} \{ N(g)\}$. So for $n > N, \vert f_n(x) - f(x) \vert < \epsilon$ for all $x \in K$. Thus $\sup_K \vert f_n - f \vert \rightarrow 0$ as $n \rightarrow \infty$ for each compact subset $K$ in $M$. By making slight modifications to the proof of Theorem 1G from page 153 of \cite{SarNak} it follows that $f \in B(\overline{C_c(M)}_{D_p})$.
\end{proof}
Let $f \in BD_p(M)$. Define a function $f^{\ast} \colon \Gamma \rightarrow \mathbb{R}$ by
\begin{equation}
f^{\ast} (g) = \frac{1}{Vol (B_{4\kappa}(g))} \int_{B_{4\kappa}(g)} f dx. \label{rmtograph}
\end{equation}
We will now show $f^{\ast} \in BD_p(\Gamma)$.
\begin{Prop} \label{definefgraph}
Let $f \in BD_p(M)$, then $f^{\ast} \in BD_p(\Gamma)$.
\end{Prop}
\begin{proof}
Let $g \in \Gamma$. By H\"{o}lder's inequality and property (P) we get the following
\begin{eqnarray*}
V_1(4\kappa)^{p-1} \int_{B_{4\kappa}(g)} \vert \nabla f(x) \vert^p dx & \geq & Vol(B_{4\kappa}(g))^{p-1} \int_{B_{4\kappa}(g)} \vert \nabla f(x)\vert^p dx \\
& \geq & \left( \int_{B_{4\kappa}(g)} \vert \nabla f(x) \vert dx \right)^p \\
& \geq & C^p \left( \int_{B_{4\kappa}(g)} \vert f(x) - f^{\ast}(g) \vert dx \right)^p,
\end{eqnarray*}
where $C$ is a constant. Let $h \in N_g$, then $d_M(g,h) \leq 3\kappa$. Consequently both $B_{4\kappa}(g)$ and $B_{4\kappa}(h)$ are contained in $B_{7\kappa}(g)$. Letting $\beta = C^{-p}V_1(4\kappa)^{p-1}$ for convience we now obtain
\begin{eqnarray*}
2\beta\int_{B_{7\kappa}(g)} \vert \nabla f(x) \vert^p dx & \geq & \beta\left( \int_{B_{4\kappa}(g)}\vert \nabla f(x) \vert^p dx + \int_{B_{4\kappa}(h)} \vert \nabla f(x) \vert^p dx \right) \\
& \geq & \left( \int_{B_{4\kappa}(g)} \vert f(x) - f^{\ast}(g) \vert dx \right)^p \\ & & \hspace{.4in} + \left( \int_{B_{4\kappa}(h)} \vert f(x) - f^{\ast}(h) \vert dx \right)^p \\
& \geq & \frac{1}{2^{p-1}} \left( \int_{B_{4\kappa}(g)} \vert f(x) - f^{\ast}(g) \vert dx \right. \\ & & \hspace{.4in} \left. + \int_{B_{4\kappa}(h)} \vert f(x) - f^{\ast}(h) \vert dx \right)^p \\
& \geq & \frac{1}{2^{p-1}} \left( \int_{B_{4\kappa}(g) \cap B_{4\kappa}(h)} \vert f^{\ast} (g) - f^{\ast} (h) \vert dx \right)^p \\
& \geq & \frac{1}{2^{p-1}} V_0(\kappa)^p \left( \vert f^{\ast} (g) - f^{\ast} (h) \vert^p \right).
\end{eqnarray*}
The last inequality follows from $B_{\kappa}(g) \subseteq B_{4\kappa}(g) \cap B_{4\kappa}(h)$, and three inequalities up is Jensen's inequality. Due to $\Gamma$ having bounded degree there exists a constant $C_1$ that does not depend on $f$ or $g$ for which
\[ C_1 \int_{B_{7\kappa}(g)} \vert \nabla f(x) \vert^p dx \geq \sum_{h \in N_g} \vert f^{\ast}(g) - f^{\ast}(h) \vert^p. \]
Furthermore, we saw earlier that if $x \in M$ then there exists at most $C_{7\kappa}$ balls $B_{7\kappa}(g)$ that contain $x$, where $C_{7\kappa}$ is a constant that does not depend on $f$ or $g$. Hence,
\[ C_{7\kappa} \int_M \vert \nabla f(x) \vert^p dx \geq \sum_{g \in \Gamma} \int_{B_{7\kappa}(g)} \vert \nabla f(x) \vert^p dx. \]
Summing up we obtain
\[ C_2 \int_M \vert \nabla f(x) \vert^p dx \geq I_p (f^{\ast}, \Gamma), \]
where $C_2$ is a suitable constant. Therefore, $f^{\ast} \in BD_p(\Gamma)$.
\end{proof}
\begin{Cor} \label{compactfgraph} If $f \in B( \overline{C_c(M)}_{D_p})$, then $f^{\ast} \in B(\overline{C_c(\Gamma)}_{D_p})$.
\end{Cor}
\begin{proof}
Let $f \in B(\overline{C_c(M)}_{D_p})$, then $f^{\ast} \in BD_p(\Gamma)$ by the proposition. We will now show that $f^{\ast}$ is also an element of $B(\overline{C_c(\Gamma)}_{D_p})$. Let $(f_n)$ be a sequence in $C_c(M)$ that converges to $f$. For each $n, f_n^{\ast} \in C_c(\Gamma)$ since $f_n$ has compact support and $\Gamma$ is $\kappa$-separated. Arguing as in the proposition it can be shown that there exists a constant $C$ such that
\[ I_p(f_n^{\ast} - f^{\ast}, \Gamma) \leq C\int_M \vert \nabla(f_n - f)(x) \vert^p dx. \]
Thus $I_p( f_n^{\ast} - f^{\ast}, \Gamma) \rightarrow 0$ as $n \rightarrow \infty$. Let $o$ be a fixed vertex of $\Gamma$ and let $\epsilon > 0$. A calculation shows that
\[ \vert f_n^{\ast}(o) - f^{\ast}(o) \vert \leq \frac{1}{V_0(4\kappa)} \int_{B_{4\kappa}(o)} \vert (f_n - f)(x) \vert dx. \]
Now $\vert f_n(x) - f(x) \vert < \epsilon$ for large $n$ and all $x \in B_{4\kappa}(o)$ due to the closure of $B_{4\kappa}(o)$ being compact in $M$. Hence $\vert f_n^{\ast} (o) - f^{\ast}(o) \vert^p \rightarrow 0$ as $n \rightarrow \infty$. Therefore, $f^{\ast} \in B(\overline{C_c(\Gamma)}_{D_p})$.
\end{proof}
\section{Proofs of Theorems \ref{boundhomeo} and \ref{bijectionpharm}}\label{proofsofresults}\label{proofsofmain}
Let $G$ be a graph of bounded degree and let $M$ be a Riemannian manifold. Let $\Gamma$ be a maximal $\kappa$-separated net in $M$, where $\kappa$ is a small positive number. Recall that $\Gamma$ can also be considered as graph of bounded degree with vertex set $\Gamma$. Assume for now that $\partial_p(\Gamma)$ is homeomorphic to $\partial_p(M)$. We saw in Section \ref{nets} that the embedding $\iota \colon \Gamma \rightarrow M$ is a quasi-isometry, so the graph $G$ is quasi-isometric with $\Gamma$ because the composition of quasi-isometries is a quasi-isometry. It follows from Theorem 2.7 of \cite{PulsPJM} that $\partial_p(G)$ is homeomorphic to $\partial_p(M)$, as desired. In order to complete the proof of Theorem \ref{boundhomeo} we need to show that $\partial_p(M)$ is homeomorphic to $\partial_p(\Gamma)$, which we now proceed to do.
We begin with two crucial lemmas.
\begin{Lem}\label{infinityconv}
Let $(y_n)$ be a sequence in $M$ with $d_M(o, y_n) \rightarrow \infty$ as $n \rightarrow \infty$, where $o$ is a fixed point in $M$. For each $n \in \mathbb{N}$ let $x_n \in \Gamma$ that satisfies $d_M(x_n, y_n) < \kappa$. Let $1 < p \in \mathbb{R}$. If $f \in BD_p(M)$, then $\vert f^{\ast}(x_n) - f(y_n) \vert \rightarrow 0$ as $n \rightarrow \infty$, where $f^{\ast}$ is defined by (\ref{rmtograph}).
\end{Lem}
\begin{proof}
For each $n \in \mathbb{N}$ we have that
\begin{eqnarray*}
\vert f^{\ast}(x_n) - f(y_n) \vert & = & \left\vert \frac{1}{Vol(B_{4\kappa}(x_n))}\int_{B_{4\kappa}(x_n)} f(x) dx - f(y_n) \right\vert \\
& = & \left\vert \frac{1}{Vol(B_{4\kappa}(x_n))} \int_{B_{4\kappa}(x_n)} \left(f(x) - f(y_n)\right) dx \right\vert \\
& \leq & \frac{1}{Vol(B_{4\kappa} (x_n))} \int_{B_{4\kappa}(x_n)} \vert f(x) - f(y_n) \vert dx \\
& \leq & \left( \frac{1}{Vol(B_{4\kappa}(x_n))} \int_{B_{4\kappa}(x_n)} \vert f(x) - f(y_n) \vert^p dx \right)^{1/p}.
\end{eqnarray*}
Let $r_x = d_M(y_n,x)$ and let $\gamma_x \colon [0, \infty) \rightarrow M$ be a geodesic parameterized by arclength that satisfies $\gamma_x(0) = y_n$ and $\gamma_x(r_x) = x$. By independence of path
\[ \int_0^{r_x} \nabla f(\gamma_x(t)) \cdot \gamma'_x(t) dt = f(x) - f(y_n). \]
For notational convenience set $\beta = \frac{1}{Vol(B_{4\kappa}(x_n))}$. Consequently,
\begin{eqnarray*}
\left( \beta \int_{B_{4\kappa}(x_n)} \vert f(x) - f(y_n) \vert^p dx \right)^{\frac{1}{p}} & \leq & \left( \beta \int_{B_{4\kappa}(x_n)} \left( \int_0^{5\kappa} \vert \nabla f(\gamma_x(t)) \vert dt \right)^p dx \right)^{1/p} \\
& \leq & \int_0^{5\kappa} \left( \beta \int_{B_{4\kappa}(x_n)} \vert \nabla f(\gamma_x(t)) \vert^p dx \right)^{1/p} dt \\
& \leq & 5\kappa\beta^{1/p} \left( \int_{B_{5\kappa}(y_n)} \vert \nabla f(x) \vert^p dx \right)^{1/p}.
\end{eqnarray*}
The second to last inequality is Minkowski's Integral Inequality. By condition (V) we have a constant $V_0(5\kappa)$ such that $V_0(5\kappa) \leq Vol(B_{5\kappa}(x))$ for all $x \in M$. Hence,
\[ \vert f^{\ast} (x_n) - f(y_n) \vert \leq 5 \kappa V_0(5\kappa)^{-1/p} \left( \int_{B_{5\kappa}(y_n)} \vert \nabla f(x) \vert^p dx \right)^{1/p}. \]
Now $\int_{B_{5\kappa}(y_n)} \vert \nabla f(x) \vert^p dx \rightarrow 0$ as $d_M(0, y_n) \rightarrow \infty$ because $f\in BD_p(M)$. Therefore, $\vert f^{\ast} (x_n) - f(y_n) \vert \rightarrow 0$ as $n \rightarrow \infty.$
\end{proof}
\begin{Lem} \label{convinftyrm} Let $\bar{f} \in BD_p(\Gamma)$ and let $f \in BD_p(M)$ be defined by (\ref{graphtorm}). Let $x \in \partial_p(\Gamma)$ and let $(x_n)$ be a sequence in $\Gamma$ that converges to $x$. Then $\vert f(x_n) - \bar{f}(x_n) \vert \rightarrow 0$ as $n \rightarrow \infty$.
\end{Lem}
\begin{proof}
Let $n \in \mathbb{N}$ and let $\Gamma_{x_n} = \{g \in \Gamma \mid g \in B_{\frac{3\kappa}{2}}(x_n) \}$. Now,
\begin{eqnarray*}
\vert f(x_n) - \bar{f}(x_n) \vert & = & \left\vert \sum_{g \in \Gamma} (\bar{f}(g) - \bar{f}(x_n)) \xi_g (x_n) \right\vert \\
& \leq & \left( \sum_{g \in \Gamma} \vert \bar{f}(g) - \bar{f}(x_n) \vert^p \xi_g(x_n) \right)^{1/p} \\
& \leq & \left( \sum_{g \in \Gamma_{x_n}} \vert \bar{f}(g) - \bar{f}(x_n) \vert^p \right)^{1/p}.
\end{eqnarray*}
The last inequality follows from $\xi_g(x_n) = 0$ if $g \notin \Gamma_{x_n}$. If $g \in \Gamma_{x_n} \setminus \{x_n\}$ then $g \in N_{x_n}$, so it follows that $\sum_{g \in \Gamma_{x_n}} \vert \bar{f}(g) - \bar{f}(x_n) \vert^p \rightarrow 0$ as $d_{\Gamma}(o, x_n) \rightarrow 0$ due to $\bar{f} \in BD_p(\Gamma)$. Hence, $\vert f(x_n) -\bar{f}(x_n) \vert \rightarrow 0$ as $n \rightarrow \infty$.
\end{proof}
We will now proceed to define a function $\Phi \colon \partial_p(\Gamma) \rightarrow \partial_p(M).$ Let $x \in \partial_p(\Gamma)$ and let $(x_n)$ be a sequence in $\Gamma$ that converges to $x$. Since $Sp(BD_p(M))$ is a compact Hausdorff space we may assume, by passing to a subsequence if necessary, that the sequence $(x_n)$ converges to a unique element $y \in Sp(BD_p(M))$. Define $\Phi(x) = y$. We now prove that $\Phi$ is well-defined and $y \in \partial_p(M).$
\begin{Prop} \label{welldefined}
The map $\Phi$ is well-defined from $\partial_p(\Gamma)$ to $\partial_p(M)$.
\end{Prop}
\begin{proof}
Let $x$ and $y$ be as above. We first show that $\Phi$ is well-defined. Let $(x_n)$ and $(x'_n)$ be sequences in $\Gamma$ such that both $(x_n)$ and $(x'_n)$ converge to $x$. Suppose that $(x_n) \rightarrow y_1$ and $(x'_n) \rightarrow y_2$ in $Sp(BD_p(M))$ and further assume that $y_1 \neq y_2$. Let $f \in BD_p(M)$ that satisfies $f(y_1) =0$ and $f(y_2) = 1$. Define $f^{\ast}$ as in (\ref{rmtograph}). Setting $y_n = x_n$ in Lemma \ref{infinityconv} we obtain $\lim_{n \rightarrow \infty} f^{\ast}(x_n) = 0$ and $\lim_{n \rightarrow \infty} f^{\ast} (x'_n) = 1$. But $f^{\ast} \in BD_p(\Gamma)$ which means
\[ \lim_{n \rightarrow \infty} f^{\ast} (x_n) = f^{\ast}(x)= \lim_{n \rightarrow \infty} f^{\ast}(x'_n), \]
a contradiction. Thus $\Phi$ is well-defined.
We will now show that $\Phi(x) = y \in \partial_p(M)$. Suppose $y \notin \partial_p(M)$, then there exists an $f \in B(\overline{C_c(M)}_{D_p})$ such that $f(y) \neq 0$. Assume $f(y) > 0$ and define $f^{\ast}$ as in (\ref{rmtograph}). Since $f^{\ast} \in B(\overline{C_c(\Gamma)}_{D_p})$ it must be the case $\lim_{n \rightarrow \infty} f^{\ast}(x_n) = f^{\ast}(x) = 0$. However, Lemma \ref{infinityconv} says that $\lim_{n \rightarrow \infty} f^{\ast}(x_n) = f(y) > 0$, a contradiction. Thus $y \in \partial_p(M).$
\end{proof}
We now show that $\Phi$ is a bijection.
\begin{Prop} \label{bijection}
The map $\Phi$ is a bijection.
\end{Prop}
\begin{proof}
We will begin by showing that $\Phi$ is one-to-one. Let $x_1, x_2 \in \partial_p(\Gamma)$ that satisfy $\Phi(x_1) = \Phi(x_2)$. Assume $x_1 \neq x_2$. Then there exits an $\bar{f} \in BD_p(\Gamma)$ with $\bar{f}(x_1) = 1$ and $\bar{f}(x_2) = 0$. Let $(x_n)$ and $(x'_n)$ be sequences in $\Gamma$ such that $(x_n) \rightarrow x_1$ and $(x'_n) \rightarrow x_2$. Using $\bar{f}$ define a function $f \in BD_p(M)$ by (\ref{graphtorm}). By Lemma \ref{convinftyrm} we see that $\lim_{n \rightarrow \infty} f(x_n) = 1$ and $\lim_{n \rightarrow \infty}f(x'_n) = 0$. This contradicts the assumption
\[ \lim_{n \rightarrow \infty} (x_n) = \Phi(x_1) = \Phi(x_2) = \lim_{n \rightarrow \infty} (x'_n). \]
Thus, $\Phi$ is one-to-one.
We will now show that $\Phi$ is onto. Let $y \in \partial_p(M)$ and let $(y_n)$ be a sequence in $M$ with $(y_n) \rightarrow y$. For each $n \in \mathbb{N}$, choose $x_n \in \Gamma$ that satisfies $d(x_n, y_n) < \kappa$. We claim that $(x_n) \rightarrow y$ in $Sp(BD_p(M))$. To see the claim let $f\in BD_p(M)$ and define $f^{\ast} \in BD_p(\Gamma)$ by (\ref{rmtograph}). By Lemma \ref{infinityconv} both $\vert f(x_n) - f^{\ast}(x_n) \vert \rightarrow 0$ and $\vert f^{\ast} (x_n) - f(y_n) \vert \rightarrow 0$ as $n \rightarrow \infty$. Hence
\[ \vert f(x_n) - f(y_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty \]
for all $f \in BD_p(M)$. Thus $\lim_{n \rightarrow \infty} (x_n) = \lim_{n \rightarrow \infty} (y_n) = y$ and the claim is proved.
By passing to a subsequence if need be, we assume that the sequence $(x_n)$ converges to an unique element $x$ in the compact Hausdorff space $Sp(BD_p(\Gamma))$. To finish the proof we need to show $x \in \partial_p(\Gamma)$. Suppose $x \notin \partial_p(\Gamma)$, then $\bar{f}(x) \neq 0$ for some $\bar{f} \in B(\overline{C_c(\Gamma)}_{D_p})$. Using $\bar{f}$ define a function $f\in B(\overline{C_c(M)}_{D_p})$ via (\ref{graphtorm}). By Lemma \ref{convinftyrm}, $\lim_{n \rightarrow \infty} f(x_n) = \lim_{n \rightarrow \infty} \bar{f}(x_n)$, which implies $f(y) \neq 0$, contradicting $y \in \partial_p(M)$ and $f \in B(\overline{C_c(M)}_{D_p})$. Thus $x \in \partial_p(\Gamma)$ and $\Phi(x) = y$, which shows that $\Phi$ is onto.
\end{proof}
To finish the proof that the bijection $\Phi$ is a homeomorphism we only need to show that $\Phi$ is continuous, since both $\partial_p(\Gamma)$ and $\partial_p(M)$ are compact Hausdorff spaces. Let $U$ be an open subset of $\partial_p(M)$ and let $x \in \Phi^{-1}(U)$. Fix $\epsilon$ that satisfies $0 < \epsilon < 1$. Let $(x_n)$ be a sequence in $\Gamma$ for which $(x_n) \rightarrow x$ and let $y = \Phi(x) \in U$. By Proposition 1 of \cite{Lee} there exist a subset $\Omega$ of $M$ such that $y \in \overline{\Omega}$, where the closure is in $Sp(BD_p(M))$, and $\overline{\Omega} \cap \partial_p(M) \subseteq U$. It was shown in the proof of this proposition that $\Omega = \{ m \mid h(m) > \epsilon \}$, where $h \in BHD_p(M), 0 \leq h \leq 1, h(y) = 1$ and $h = 0$ on $\partial_p(M) \setminus U$. Using this $h$ a function $h^{\ast} \in BD_p(\Gamma)$ can be defined by (\ref{rmtograph}). Let $V = \{ z \in \partial_p(\Gamma) \mid h^{\ast}(z) > \frac{\epsilon}{2}\}$. The set $V$ is open due to $h^{\ast}$ being continuous. Moreover, it follows from Lemma \ref{infinityconv} that $x \in V$. Now let $z \in V$ and suppose $\Phi(z) \notin U$. Consequently, $h(z_n) \rightarrow 0$ where $(z_n)$ is a sequence in $\Gamma$ that converges to $z$.
It follows from Lemma \ref{infinityconv} that $h^{\ast}(z_n) < \frac{\epsilon}{2}$ for large $n$, contradicting $z \in V$. Thus $V \subseteq \Phi^{-1}(U)$ and the continuity of $\Phi$ is established. Therefore, $\Phi \colon \partial_p (\Gamma) \rightarrow \partial_p(M)$ is a homeomorphism and Theorem \ref{boundhomeo} is proved.
The proof of Theorem \ref{bijectionpharm} pretty much follows the same path as the proof of Theorem \ref{boundhomeo}. Let $G$ be a graph of bounded degree and let $M$ be a complete Riemannian manifold. Let $\Gamma$ be a maximal $\kappa$-separated net in $M$, where $\kappa$ is a small positive constant. The graph with bounded degree $\Gamma$ is quasi-isometric with $G$ so by Theorem \cite[Theorem 2.8]{PulsPJM} there is a bijection between $BHD_p(G)$ and $BHD_p(\Gamma)$. To complete the proof of the theorem we must establish a bijection between $BHD_p(\Gamma)$ and $BHD_p(M)$, which we will now proceed to do.
Let $\bar{h} \in BHD_p(\Gamma)$. Define a function $h \in BD_p(M)$ by (\ref{graphtorm}). Denote by $\pi(h)$ the unique element of $BHD_p(M)$ given by the Royden's decomposition of $h$. Define $\Psi \colon BHD_p(\Gamma) \rightarrow BHD_p(M)$ by $\Psi(\bar{h}) = \pi(h)$.
We will now show that $\Psi$ is one-to-one. Let $x \in \partial_p(\Gamma)$ and let $(x_n)$ be a sequence in $\Gamma$ that converges to $x$. Let $\Phi$ be the homeomorphism from $\partial_p(\Gamma)$ to $\partial_p(M)$ given in earlier in this section. Since $\Psi(\bar{h}) (\Phi(x)) = h (\Phi(x))$ for all $x \in \partial_p(\Gamma)$, it follows that
\begin{equation}
\vert \Psi(\bar{h})(x_n) - h(x_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty. \label{hconvbd}
\end{equation}
Combining this with Lemma \ref{convinftyrm} we obtain $\vert \Psi(\bar{h})(x_n) - \bar{h}(x_n)\vert \rightarrow 0$ as $n \rightarrow \infty$. Thus $\Psi (\bar{h})(\Phi(x)) = \bar{h}(x)$. Let $\bar{h}_1, \bar{h}_2 \in BHD_p(\Gamma)$ and assume $\Psi(\bar{h}_1) = \Psi(\bar{h}_2)$. Then $\bar{h}_1(x) = \bar{h}_2(x)$ for all $x \in \partial_p(\Gamma)$. Hence $\Psi$ is one-to-one because a $p$-harmonic function on $\Gamma$ is determined by its values on $\partial_p(\Gamma)$.
All that is left to do is show that $\Psi$ is onto. Let $u \in BHD_p(M)$ and define $u^{\ast} \in BD_p(\Gamma)$ by (\ref{rmtograph}). Denote by $\bar{h}$ the element in $BHD_p(\Gamma)$ given by the Royden decomposition or $u^{\ast}$. Let $y \in \partial_p(M)$ and let $x \in \partial_p(\Gamma)$ such that $\Phi(x) = y$. Pick a sequence $(x_n)$ in $\Gamma$ for which $(x_n) \rightarrow x$. Now $\Psi(\bar{h})(y) = \pi(h)(y) = h(y)$ since $y \in \partial_p(M)$. By (\ref{hconvbd}) and Lemma \ref{convinftyrm} we see that
\begin{equation}
\vert \pi(h)(x_n) - \bar{h}(x_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty. \label{hnetconv}
\end{equation}
By Lemma \ref{infinityconv}
\begin{equation}
\vert u^{\ast} (x_n) - u(x_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty. \label{convu}
\end{equation}
It is also true that
\begin{equation}
\vert \bar{h}(x_n) - u^{\ast}(x_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty, \label{huconv}
\end{equation}
due to $u^{\ast} = \bar{h}$ on $\partial_p(\Gamma)$. Combining (\ref{convu}) and (\ref{huconv}) with (\ref{hnetconv}) we obtain
\begin{equation}
\vert u(x_n) - \pi(h)(x_n) \vert \rightarrow 0 \mbox{ as } n \rightarrow \infty.
\end{equation}
Thus $u(y) = \Psi(\bar{h})(y)$ for all $y \in \partial_p(M)$. Hence $\Psi(\bar{h}) = u$ by \cite[Lemma1]{Lee}. The proof of Theorem \ref{bijectionpharm} is finished.
\bibliographystyle{plain}
|
\section{Introduction}
MV-algebras can be seen, in one of their facets, as a non-idempotent generalization of Boolean algebras.
The lack of idempotency gives to MV-algebras a strong monoidal character.
Nevertheless the typical lattice structure of Boolean algebras can be recovered inside MV-algebras,
by a suitable combination of the primitive connectives.
So, MV-algebras have a double behaviour: on the one hand, they look similar to
monoidal objects such as Abelian groups, on the other, they maintain a strong lattice structure (distributive and bounded), which makes many techniques of lattice theory be appropriate in their study. Such a twin complexion of MV-algebras is endorsed by Mundici's celebrated natural equivalence between MV-algebras and lattice ordered Abelian groups with strong unit.
Boolean algebras work as the \emph{equivalent algebraic semantics} of classical logic.
Boolean algebras also have a neat and intuitive depiction: modulo isomorphisms, any of
them is an algebra of \emph{special} subsets of some set w.r.t. the elementary operations of union, intersection and complement.
This triple scenario is nowadays well understood, the logical conjunction, the set intersection and the Boolean meet are just different personifications of the same concept. The \emph{special} property mentioned above is formalized through the topology of Stone spaces and allows to \emph{select} the right objects in the full powerset of some set.
MV-algebras are the equivalent algebraic semantics of {\L}ukasiewicz logic, one of the longest-known many-valued logic. Nevertheless the monoidal character of such systems
still escapes a complete understanding. One of the major achievements in this direction, was made by Di Nola who proved that any MV-algebra can be seen, up to isomorphisms, as a subalgebra of the algebra of functions $^{*}[0,1]^{X}$, where $^{*}[0,1]$ is some ultrapower of the real interval $[0,1]$ and $X$ is some suitable set. The structure carried by $^{*}[0,1]^{X}$ is defined pointwise upon the \emph{standard} MV-algebraic structure of $[0,1]$.
Di Nola's result is in complete accordance with the Stone representation Theorem for
Boolean algebras cited above, for subsets of a set $X$ can be seen as discrete functions in $\{0,1\}^{X}$.
Two specialisations of the above result are worth to be mentioned. In \cite{belsemisimple} Belluce proved that every semisimple MV-algebra
can be embedded in an algebra of functions from a set $X$ in $[0,1]$.
MV-algebras that are semisimple and linearly ordered admit a representation as subalgebras of $[0,1]$.
Under this perspective the above representation for MV-algebras suggests two things: it makes sense to consider MV-algebras as appropriate algebraic structures for \emph{fuzzy sets
; the \emph{infinitesimal} elements arising from the ultrapower construction are some sort of novelty which is produced by the monoidal structure inside MV-algebras and cannot be easily removed.
Concrete representations, as the aforementioned ones, are extremely useful in the study of abstract algebraic structures. Yet, to take full advantage of such correspondences, a full-flagged duality is called for. Unfortunately, Di Nola representation remains so far an ``embedding theorem'', or, in categorical terms, a faithful and dense functor.
Finding the set of conditions which enable to invert such a functor is probably one of the most interesting open problems regarding MV-algebras.
In this paper we revise some known results on \emph{non-standard} representation for MV-algebras and some subclasses and show some new ones, together with their corresponding results holding for Abelian $\ell u$-groups. In particular we study the subclasses of \emph{local} and \emph{perfect} MV-algebras and prove a uniform non-standard representation theorem for both of them. Our aim is to present a wide perspective of representations for this classes.
More in details, we show that any perfect MV-algebra can be embedded in an algebra of functions from the \emph{spectrum} of the algebra to an ultrapower of $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb R}, (1,0))$ (\autoref{perflex}).
After recalling a known result on the representation of local MV-algebras, we use it to prove a representation theorem for local Abelian $\ell u$-groups as the groups of \emph{quasi-constant} functions over an ultrapower of $\mathbb{R}$ (\autoref{groupquasiconstant}).
For both theorems we also give a uniform version parametrised only by an upper-bound on the cardinality of the algebras to be embedded (\autoref{uniformlocal}).
\section{Basics on MV-algebras}
\begin{definition}
An MV-algebra is an algebra $\langle A,\oplus ,\lnot ,0\rangle $ with a binary operation $\oplus$, a unary operation $\lnot $ and a constant $0$ satisfying the following equations:
\begin{enumerate}
\item $x\oplus \left( y\oplus z\right) =(x\oplus y)\oplus z$
\item $x\oplus y=y\oplus x$
\item $x\oplus 0=x$
\item $\lnot \lnot x=x$
\item $x\oplus \lnot 0=\lnot 0$
\item $\lnot \left( \lnot x\oplus y\right)
\oplus y=\lnot \left( \lnot y\oplus x\right) \oplus x.$
\end{enumerate}
\end{definition}
The real unit interval $[0,1]$ equipped with operations $x\oplus
y=\min (1,x+y)$ and $\lnot x=1-x$ is an MV-algebra, denoted henceforth by $[0,1]$. The MV-algebra $[0,1]$ generates the whole variety of MV-algebras and hence it has a prominent role in this theory.
Given an MV-algebra $A$ and a set $X$, the set $A^{X}$ of all functions $%
f:X\rightarrow A$ becomes an MV-algebra if the operations $\oplus $ and $%
\lnot $ and the element $0$ are defined pointwise.
In particular the set of all functions from a set $X$ to $[0,1]$
can be equipped with a structure of MV-algebra. We shall see that
MV-algebras of this kind can be characterized in terms of their ideals.
\medskip
A partially ordered Abelian group \cite{cdm} is an Abelian group
$\langle G,+,-,0\rangle $ endowed with a partial order relation
$\leq $ that is compatible with addition. When the order of $G$
defines a lattice structure, $G$ is called a \emph{lattice ordered
Abelian group} or $\ell$-group. A {\em strong unit} of an
$\ell$-group $G$ is an element $u \in G$ such that $u \geq 0$ and
for each $x \in G$ there is an integer $n \geq 0$ with $|x| \leq
nu$.
In the following by $\ell$- groups we shall always mean Abelian $\ell$- groups
and by $\ell u$-groups we shall mean Abelian $\ell$- groups with strong unit.
Let $G$ be an $\ell$-group. For any element $u\in G$, $u>0$, (not
necessarily $u $ being a strong unit of $G$) we let $[0,u] = \{x \in G : 0 \leq x \leq u\}$, and for each $x,y \in [0,u]$, $x\oplus y=u\wedge ( x+y)$ and $\neg x=u-x$.
The structure $\langle [0,u],\oplus, \lnot, 0\rangle$ denoted by $\Gamma (G,u)$, is an MV-algebra.
\medskip
Let $\mathcal{MV}$ denote the category of MV-algebras.
\begin{theorem}\label{thcdm}{\rm \cite{cdm}}
Let $\ell\mathcal{G}_u$ denote the category whose objects are pairs
$(G,u)$ with $G$ an $\ell u$-group and $u$ a
distinguished strong unit of $G$, and whose morphisms are unital
$\ell$-homomorphisms. Then $\Gamma $
is a natural equivalence (i.e., a full, faithful, dense functor) between
$\ell\mathcal{G}_u$ and $\mathcal{MV}$.
\end{theorem}
The inverse equivalence $\mathcal{MV} \longrightarrow \ell\mathcal{G}_u$ is usually denoted by $\Xi$.
\autoref{thcdm} allows to describe more examples of MV-algebras. Let ${\mathbb Z}$ be the $\ell$-group of integer numbers with natural order and denote by ${\mathbb Z} \times_{\operatorname{lex}} {\mathbb Z}$ the $\ell$-group given by the lexicographic product (the order is hence total). The MV-algebra $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb Z}, (n-1,0))$ is called {\em Komori chain of rank $n$} and is denoted by $S_n^\omega$ (\cite{cdm},\cite{komori}).
The MV-algebra $S_2^\omega$ is also known as {\em Chang MV-algebra} and it is denoted by $C$ (\cite{chang}).
On each MV-algebra $A$ we define
\[1:=\lnot 0,\quad x\odot y:=\lnot \left( \lnot x\oplus \lnot y\right),\quad x\ominus y:=x\odot \lnot y.\]
Note that in the MV-algebra $[0,1]$ we have $x\odot y=\max \left( 0,x+y-1\right) $ and $x\ominus y=\max (0,x-y).$
Let $A$ be an MV-algebra. For any two elements $x$ and $y$ of $A$
we write $x\leq y $ iff $x\ominus y=0$. It follows that $\leq $
is a lattice order in which the operations are given by
\begin{eqnarray}
\label{eq:lattice}
x\vee y&=&\left( x\odot \lnot y\right) \oplus y=\left(
x\ominus y\right) \oplus y, \\
\label{eq:lattice2}x\wedge y&=&\lnot \left( \lnot x\vee \lnot y\right) =x\odot \left(
\lnot x\oplus y\right) .
\end{eqnarray}
In the algebra $[0,1]$, such an order coincides with the natural one. An MV-algebra whose order is total is called an MV-chain.
\begin{definition}
An ideal of an MV-algebra $A$ is a subset $I$ of $A$ such that
$0\in I$, if $x, y\in I$ then $x\oplus y\in I$ and
if $x\in I$, $y\in A$ and $y\leq x$ then $y\in I$.
\end{definition}
Any ideal $I$ of $A$ induces a congruence $\equiv _{I}$
on $A$ such that for each $x$ and $y\in A$ we have:
\begin{equation*}
x\equiv _{I}y\Longleftrightarrow d(x,y)=\left( x\ominus y\right)
\oplus \left( y\ominus x\right) \in I.
\end{equation*}
We indicate with $x/I$ or $[x]_I$ the equivalence class
of $x$ and with $A/I$ the set of all equivalence classes with
respect to $\equiv _{I}.$ The quotient MV-algebra is defined as
usual.
An ideal of an MV-algebra $A$ is \emph{proper} if
$I\neq A$. We say that $I$ is \emph{prime} iff it is proper and
for each $x$ and $y$ in $A$, either $\left( x\ominus y\right) \in
I$ or $\left( y\ominus x\right) \in I$, hence $A/I$ is an MV-chain.
An ideal of an MV-algebra $A$ is called \emph{maximal}
if it is proper and it is not strictly contained in any proper ideal.
Let us denote by $\operatorname{Spec} A$ the set of prime ideals and $\operatorname{Max} A$ the set of maximal ideals in $A$. $\operatorname{Spec} A$ is called the
prime spectrum of $A$ and $\operatorname{Max} A$ the maximal spectrum of $A$.
\begin{theorem}[Chang Subdirect Representation Theorem, \cite{chang59}]\label{theo:chan}
Every nontrivial MV-algebra is a subdirect product of MV-chains. In particular, any MV-algebra $A$ is embeddable in $\prod_{P \in \operatorname{Spec} A}A/P$.
\end{theorem}
By \autoref{theo:chan}, any MV-algebra $A$ is a subdirect product of $\{A/P \mid P \in \operatorname{Spec} A\}$. By
finding a common embedding for each $A/P$, Di Nola showed the
following representation theorem.
\begin{theorem}{\rm \cite{din1}}\label{theo:DNLT}
Up to isomorphisms, every MV-algebra $A$ is an algebra of $[ 0,1]^\ast$-valued functions over $\operatorname{Spec} A$, where $[0,1]^{\ast }$ is an ultrapower of $[0,1]$, only depending on the cardinality of $A$.
\end{theorem}
Recently, such a representation have been sharpened to a uniform version that asserts the existence, for any infinite cardinal $\alpha$, of a single MV-algebra of functions in $[0,1]^{\ast}$, which contains all MV-algebras of cardinality less than or equal to $\alpha$.
\begin{theorem}{\rm \cite{DNLS}}
For any infinite cardinal $\alpha$ there exists an MV-algebra of functions $A$ such that any MV-algebra of infinite cardinality at most $\alpha$ embeds in $A$.
\end{theorem}
Such a result is obtained by using standard techniques in model theory, involving $\alpha$-regular ultrafilter \cite[Section 4.3]{ChK}. A further application of known constructions in model theory (iterated ultrapowers \cite[Section 6.5]{ChK}) yields a sort of \emph{canonical} MV-algebra of values, parametrised only by some cardinal number $\alpha$.
\begin{theorem}{\rm \cite{DNLS}}
For every infinite cardinal $\alpha$ there is an iterated ultrapower $\Pi_\alpha$ of $[0,1]$, definable\footnote{By \emph{definable} we mean definable in ZFC} in $\alpha$, such that every MV-algebra of cardinality $\alpha$ embeds in an algebra of functions with values in $\Pi_\alpha$.
\end{theorem}
\section{Perfect MV-algebras}
An {\em infinitesimal} element in an MV-algebra is an element $x\neq 0$ such that $nx < \neg x$,
for every $n \in {\mathbb N}$. Clearly in the MV-algebra $[0,1]$ and in all MV-algebras of functions
taking values in $[0,1]$, there are no such elements. On the other hand,
there are MV-algebras that are generated by their infinitesimal elements, as for example
the Chang MV-algebra $S^2_\omega$. MV-algebras generated by their infinitesimals
are called {\em perfect} and they form a subcategory of $\mathcal{MV}$
that is equivalent to the category of $\ell$- groups.
The class of perfect MV-algebras is included in the variety generated by the Chang MV-algebra,
but it is not a variety: it is a universal class \cite{beldingerper}.
Infinitesimal elements and perfect MV-algebras are strongly related with the phenomenon of incompleteness
of first order {\L}ukasiewicz logic: indeed the subalgebra of the Lindenbaum algebra of first order {\L}ukasiewicz logic
generated by the classes of formulas which are valid but non-provable is a perfect MV-algebra (see \cite{beldin}, \cite{scarp}).
\begin{definition}
For any MV-algebra $A$, the \emph{radical} of $A$ (denoted by $\operatorname{Rad} A$)
is the intersection of all maximal ideals of $A$.
\end{definition}
All the non-zero elements of the radical of an MV-algebra are infinitesimal.
\begin{definition}
\label{def:simplesemisimple}
An MV-algebra is called {\em simple} if it has exactly two ideals, in other words if it is non trivial and $\{0\}$ is the only proper ideal.
An MV-algebra $A$ is said to be {\em semisimple} if $A$ is non trivial and $\operatorname{Rad} A = \{0\} .$
\end{definition}
Every simple MV-algebra is semisimple. The proof of the following proposition can be
found in \cite{belsemisimple,chang,chang59,cdm}.
\begin{proposition}
An MV-algebra is simple if and only if it is isomorphic to a subalgebra of $[0,1]$. An MV-algebra is semisimple if and only if it is isomorphic to a separating MV-algebra of $[0,1]$-valued continuous functions on some nonempty compact Hausdorff space, with pointwise operations.
\end{proposition}
\begin{definition}
An MV-algebra $A$ is called {\em perfect} if it is nontrivial and
\begin{equation*}
A=\operatorname{Rad} A\cup \lnot \operatorname{Rad} A,
\end{equation*}
where $\lnot \operatorname{Rad} A=\left\{ x\in A:\lnot x\in
\operatorname{Rad} A\right\} .$
\end{definition}
Let $\mathcal{MV}_P$ be the full subcategory of $\mathcal{MV}$ having as objects the perfect MV-algebras.
Following \cite{dinletPerf}, we can associate each perfect MV-algebra $A$ with an $\ell$-group
$G = \mathcal D(A)$ as follows. Let $\theta \subseteq \operatorname{Rad} A^2 \times \operatorname{Rad} A^2$ be defined by $(x,y) \theta (x',y')$ if and only if $x \oplus y' = x'\oplus y$; the relation $\theta$ is a congruence and
$$\mathcal{D}(A) = \langle \operatorname{Rad} A^2/\theta, +, \leq, [0,0]\rangle$$
is an $\ell$-group, with $-[x,y]=[y,x]$. It can be shown that this construction yields a functor between $\mathcal{MV}_P$ and the category $\lg$ of $\ell$-groups. Moreover, the functor $\mathcal D$ is a categorical equivalence whose inverse is given by
$$\mathcal G: \ G \in \lg \ \longmapsto \ \Gamma({\mathbb Z} \times_{\operatorname{lex}} G, (1,0)) \in \mathcal{MV}_P.$$
In particular, we have that for every perfect MV-algebra $A$ there exists an $\ell$- group $G$ such that
$A$ is isomorphic to $\Gamma({\mathbb Z} \times_{\operatorname{lex}} G,(1,0))$.
\begin{theorem}{\rm \cite{beldinger}}\label{th:MVultrapower}
Let $\prod_F(G_i)$ denote the ultraproduct of the $\ell$-groups $(G_i)_{i \in I}$ with respect to a non-principal ultrafilter $F$ of $2^I$ and $\prod_F(\mathcal G(G_i))$ denote the ultraproduct of the perfect MV-algebras $\mathcal G(G_i)$ with respect to $F$. Then $\mathcal G(\prod_F(G_i)) \cong\, \prod_F(\mathcal G(G_i))$.
\end{theorem}
\begin{theorem}\label{perflex}
Every perfect MV-algebra can be embedded into an algebra of functions taking values in
an ultrapower of $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb R}, (1,0))$.
\end{theorem}
\begin{proof}
If $A$ is a perfect MV-chain, then $\mathcal D(A)$ is a totally ordered group,
hence $\mathcal D(A)$ can be embedded in some ultrapower ${}^*{\mathbb R}$ of ${\mathbb R}$. Since $\mathcal G$ preserves embeddings, then $\mathcal G(\mathcal D(A))$ embeds in $\mathcal G({}^*{\mathbb R})$ and, since $\mathcal{G}$ and $\mathcal D$ form an equivalence between the categories of perfect MV-algebras and $\ell$-groups, we have an embedding of $A$ into $\mathcal{G}({}^*{\mathbb R}) = \Gamma({\mathbb Z} \times_{\operatorname{lex}} {}^*{\mathbb R},(1,0))$.
From \autoref{th:MVultrapower} it follows that every perfect MV-chain can be embedded into an ultrapower of the perfect MV-chain $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb R}, (1,0))$ (see also \cite[Theorem 23]{beldinger}).
Let $\mathbf{K}$ be the class of all ultrapowers of $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb R},(1,0))$ and $M$ be a perfect MV-algebra. By Chang representation theorem, $M$ is embeddable in the product
$\prod_{P \in \operatorname{Spec} M}M/P$ where each $M/P$ is a perfect MV-chain. Then each $M/P$ can be embedded into an element $N_P$ of $\mathbf{K}$. On the other hand, all algebras in $\mathbf{K}$ are elementarily equivalent hence, by the joint embedding property, there exists a perfect MV-algebra $N \in \mathbf{K}$ such that each $N_P$ embeds in $N$. Hence $A$ embeds in $\prod_{P \in \operatorname{Spec} M}N \cong N^{\operatorname{Spec} M}$ and $N$ is an ultrapower of $\Gamma({\mathbb Z} \times_{\operatorname{lex}} {\mathbb R},(1,0))$.
\end{proof}
We aim now at giving a uniform version of the previous theorem. The idea is basically the same of \cite{DNLS}, we only have to check that the construction goes through also in the restricted case of perfect MV-algebras.
\begin{definition}
Let $\alpha$ be a cardinal. A proper filter $D$ over $I$ is said to be \emph{$\alpha$-regular} if there exists a set $E\subseteq D$ such that $|E|=\alpha$ and each $i\in I$ belongs to only finitely many $e\in E$.
\end{definition}
For any set $I$ of infinite cardinality $\alpha$ there exists an $\alpha$-regular ultrafilter over $I$ \cite{ChK}. Given a cardinal $\alpha$, let $\alpha^{+}$ be the smallest cardinal greater than $\alpha$. We let $\equiv$ stand for elementary equivalence, $\hookrightarrow$ for embeddability and $\hookrightarrow_{\mathit{el}}$ for elementary embeddability.
\begin{definition}
Given a cardinal $\alpha$, we say that a model $\mathfrak{A}$ is \emph{$\alpha$-universal} if and only if for every model $\mathfrak{B}$ we have:
\[ \mathfrak{B}\equiv \mathfrak{A}\quad\text{ and }\quad |\mathfrak{B}| <\alpha\quad \text{ implies }\quad\mathfrak{B}\hookrightarrow_{\mathit{el}}\mathfrak{A}.\]
\end{definition}
\begin{theorem}{\rm \cite{ChK}}\label{thm:chang}
Let $\mathcal{L}$ be a first order language such that $|\mathcal{L}|\leq \alpha$ and $D$ be an ultrafilter which is $\alpha$-regular. Then, for every model $\mathfrak{A}$, the ultrapower $\prod_D \mathfrak{A}$ is $\alpha^+$-universal.
\end{theorem}
Let $[0,1]_{\rm lex}$ be the perfect MV-algebra $\Gamma({\mathbb Z}\times_{\operatorname{lex}} {\mathbb R},(1,0))$.
\begin{theorem}\label{theorem:9}
For every cardinal $\alpha$, there exists an ultrapower $[0,1]_{\rm lex}^{*}$ of the MV-algebra $[0,1]_{\rm lex}$
such that any perfect MV-algebra $A$ of cardinality smaller than or equal to $\alpha$, can be embedded
into an MV-algebra of functions from some set $X$ to $[0,1]_{\rm lex}^{*}$.
\end{theorem}
\begin{proof}
Let $A$ be a perfect MV-algebra such that $|A|=\alpha$, with $\alpha$ an infinite cardinal. Then, By Chang representation theorem, $A$ is embeddable in the product $\prod_{P \in \operatorname{Spec} A}A/P$ where each $A/P$ is a perfect MV-chain.
The $\ell$- group $\mathcal{D}(A/P)$ is a totally ordered group hence it is embeddable in
a divisible ordered group that in turns is elementary equivalent to ${\mathbb R}$.
Let $F$ be a $\alpha$-regular ultrafilter over $\alpha$; then, by \autoref{thm:chang}, $\prod_{F}{\mathbb R}$ is $\alpha^{+}$-universal, hence for every $P\in \operatorname{Spec} A$, $\mathcal{D(A/P)} \hookrightarrow \prod_{F}{\mathbb R}$, and so
$A/P \hookrightarrow \prod_F [0,1]_{\rm lex}$ (see \autoref{th:MVultrapower}). Combining the embeddings we have that
\[A \hookrightarrow \left(\prod_{F}[0,1]_{\rm lex}\right)^{\operatorname{Spec} A}.\]
\end{proof}
Such a result can be further improved by showing that, for any infinite cardinal $\alpha$, there exists an ultrapower of $[0,1]_{\rm lex}$ (more precisely, an iterated ultrapower), which is definable in $\alpha$ and such that any perfect MV-algebra of cardinality at most $\alpha$ embeds in the algebra of functions $\left(\prod[0,1]_{\rm lex}\right)^{X}$, for some set $X$.
A detailed proof of the result claimed above would require a full explanation on how iterated ultrapowers are constructed. Roughly, an iterated ultrapower is obtained by iterating the ultrapower construction using a set of linearly ordered ultrafilters $S$. The structure resulting from such a construction has the remarkable property of containing an isomorphic copy of every ultrapower defined on an ultrafilter in $S$. The theorem below formalizes the result claimed above and gives a sketch of proof; we suggest the interested reader to consult \cite{ChK,DNLS}.
\begin{theorem}\label{theorem:10}
For any infinite cardinal $\alpha$, there exists an ultrapower of $[0,1]_{\rm lex}$ definable in $\alpha$, denoted by $\prod_{\alpha}[0,1]_{\rm lex}$,
such that any perfect MV-algebra $A$, with $|A|\leq \alpha$, can be embedded into an MV-algebra of functions from some set $X$ to $\prod_{\alpha}[0,1]_{\rm lex}$.
\end{theorem}
\begin{proof}
The idea is to construct an iterated ultrapower of $[0,1]_{\rm lex}$ using the set of \emph{all} ultrafilters on $\alpha$, obviously such a structure will have the desired properties. The only obstruction is that there is no definable linear order on the set of ultrafilter on $\alpha$, if we do not admit repetitions. Since repetitions do not affect the contruction of the iterated ultrapower, we linearly order the ultrafilters as follows.
Let $Y$ be the set of all maps $y:|{\P}(\alpha)|\rightarrow {\P}(\alpha)$ such that the image
of $y$ is an ultrafilter on $\alpha$. Note that every ultrafilter on $\alpha$ appears as image of some (actually infinitely many) elements of $Y$.
The set $Y$ is totally ordered by setting $y<y'$ if there is an ordinal $\xi<| {\P}(\alpha)|$ such that
$y|_\xi=y'|_\xi$ (that is, $y$ and $y'$ coincide on all the ordinals less than $\xi$) and $y(\xi)<y'(\xi)$ in the lexicographic order of $P(\alpha)$.
For every $y\in Y$, let $D_y$ be the ultrafilter on $\alpha$ associated to $y$, i.e$.$ the image of $y$.
Then $D_\alpha$ is the resulting indexed family of ultrafilters on $\alpha$.
\end{proof}
\section{Local MV-algebras and $\ell u$-groups}
Local MV-algebras are MV-algebras with only one maximal ideal that, hence, contains all
infinitesimal elements.
This class of algebras includes MV-chains and perfect MV-algebras and it has an important
role in the representation theory of MV-algebras.
Indeed all MV-algebras can be represented as sections over sheaves whose stalks are local MV-algebras: see for example \cite{filgeo}, where the base topological space is the space of maximal ideals, and \cite{ferlet}, where the base space is $\operatorname{Spec}$. We recall some basic properties of local MV-algebras.
\begin{definition}\label{order}
For any $x\in A$, the order of $x$, in symbols $\operatorname{ord}(x)$, is the smallest
natural number $n$ such that $nx=1$.
If no such $n$ exists, then $\operatorname{ord}(x)=\infty .$
\end{definition}
Let $A$ be an MV-algebra. We recall that an element $x$ of $A$ is
called \emph{finite} iff $\operatorname{ord}(x)<\infty $ and $\operatorname{ord}(\lnot x)<\infty $. Of
course $x$ is finite iff $x\wedge \lnot x$ is finite.
\begin{definition}\label{localmv}
An MV-algebra $A$ is called \emph{local} if it has only one maximal ideal, coinciding with
$\operatorname{Rad}(A)=\{a \in A \mid \operatorname{ord}(a)=\infty\}$.
\end{definition}
If $A$ is a local MV-algebra then $A/\operatorname{Rad} A$ is a simple MV-algebra, since it does not have non-trivial ideals.
Recall that, for any $\ell$-group $G$, a subgroup $J$ of $G$ is called an \emph{$\ell$- ideal} if it satisfies the following additional condition:
$$\textrm{if $x \in J$ and $\left|y\right| \leq \left|x\right|$ then $y \in J$.}$$
It is well-known that $\ell$- ideals are in one-one correspondence with $\ell$- group congruences and that, if $G$ is an $\ell u$-group,
then $G/J$ is an $\ell u$-group with strong unit $u/J$.
\begin{definition}\label{locallg}
An $\ell$- group $G$ is called \emph{local} if it has only one maximal $\ell$- ideal.
\end{definition}
Perfect MV-algebras and MV-chains are local MV-algebras, while the converse is false. In fact,
if $G$ is an $\ell$-group, then
\begin{equation*}
\Gamma ({\mathbb Z}\times_{\operatorname{lex}} G,(2,0))
\end{equation*}
is a local MV-algebra, but if $G$ is not totally ordered, it is neither an
MV-chain nor a perfect MV-algebra.
\begin{proposition}
An MV-algebra $A$ is local if and only if for every $x \in A$, either
$\operatorname{ord}(x)<\infty$ or $\operatorname{ord}(\neg x)<\infty$.
\end{proposition}
We describe an example of local MV-algebra that will result as a sort of prototypical one. Let $X$ be an arbitrary non empty set, $U$ an MV-algebra, and $\mathbf{K}(U^{X})$ the subset of the MV-algebra $U^{X}$ defined as follows:
\begin{equation*}
\mathbf{K}(U^{X})=\{f\in U^{X}\mid f(X)\subseteq a/\operatorname{Rad}(U)\quad
\text{for some} \ a\in U\}.
\end{equation*}
$\mathbf{K}(U^{X})$
will be called the \emph{the full MV-algebra of quasi-constant
functions} from $X$ to $U$ and any element $f$ of
$\mathbf{K}(U^{X})$ will be called a \emph{quasi-constant}
function from $X$ to $U$.
With the above notations we have:
\begin{proposition}{\rm \cite{dinespger}}
\label{prop:esloc} $\mathbf{K}(U^{X})$ is a local MV-algebra.
\end{proposition}
\begin{proof}
Let us show that $\mathbf{K}(U^{X})$ is a subalgebra of $U^{X}$. The zero-constant function $f_0$ belongs to $\mathbf{K}(U^{X})$ because $f_0(X) = \{0\} \subseteq \operatorname{Rad}(U) = 0/\operatorname{Rad}(U)$. If $f$ satisfies $f(X) \subseteq a/\operatorname{Rad}(U)$, for some $a\in U$, then we have $\lnot f(X) \subseteq \lnot a/\operatorname{Rad}(U)$. Finally, let $f, g \in \mathbf{K}(U^{X})$ be such that $f(X) \subseteq a/\operatorname{Rad}(U)$ and $g(X) \subseteq b/\operatorname{Rad}(U)$, with $a, b\in U$. Then $(f \oplus g)(X) \subseteq (a \oplus b)/\operatorname{Rad}(U)$. Hence $\mathbf{K}(U^{X})$ is a subalgebra of $U^{X}$.
In order to show that $\mathbf{K}(U^{X})$ is local, let us consider $f \in \mathbf{K}(U^{X})$. If $f(X) \subseteq 0/\operatorname{Rad}(U)$, then $\lnot f(X) \subseteq 1/\operatorname{Rad}(U)$ and $\operatorname{ord}(\lnot f) < \infty$. If $f(X) \subseteq 1/\operatorname{Rad}(U)$, we have $\operatorname{ord}(f) < \infty$. Assume now that $f(X) \subseteq a/\operatorname{Rad}(U) \notin \{0/\operatorname{Rad}(U),1/\operatorname{Rad}(U)\}$. Then, for every $x \in X$, $f(x) \equiv_{\operatorname{Rad}(U)} a$ and $a \notin \operatorname{Rad}(U)$. Since $A/\operatorname{Rad}(U)$ is an MV-chain, then $\operatorname{ord}(f) < \infty$ and $\operatorname{ord}(\lnot f) < \infty$.
\end{proof}
\begin{lemma}{\rm \cite{dinlet}}\label{lem:simpleform}
Let $x,y\in [0,1]$, and $x<y$. Then there is a term $\varphi $ such that $\varphi (x)=0$ and $\varphi (y)=1$.
\end{lemma}
The following result is a representation based on the spectral properties of local MV-algebras.
\begin{proposition}\label{prop:spec}
Let $A$ be a local MV-algebra. Then for every $x \in A$ and for every $P, Q \in \operatorname{Spec} A$, we have
\begin{equation*}
\frac{(x/P)}{\operatorname{Rad}(A/P)}=\frac{(x/Q)}{\operatorname{Rad}(A/Q)}.
\end{equation*}
\end{proposition}
\begin{proof}
We observe that for every $P \in \operatorname{Spec} A$, $(A/P)/\operatorname{Rad}(A/P)$ is simple and then, up to an isomorphism, is a subalgebra of $[0,1]$.
Assume, by contradiction, that there are $r,s\in [0,1]$, with $r<s$, such that $r=(x/P)/\operatorname{Rad}(A/P)$ and $s=(x/P)/\operatorname{Rad}(A/P)$.
By Lemma \ref{lem:simpleform}, there is a simple term $\varphi$ such that $\varphi(r)=0$ and $\varphi (s)=1$.
Then $\varphi((x/P)/\operatorname{Rad}(A/P))=(\varphi(x)/P)/\operatorname{Rad}(A/P)=0$ hence
$\varphi(x)/P \in \operatorname{Rad}(A/P)$ while $\varphi((x/Q)/\operatorname{Rad}(A/Q))=(\varphi(x)/Q)/\operatorname{Rad}(A/Q)=1$ hence $\neg \varphi(x)/Q \in \operatorname{Rad}(A/Q)$.
Thus $\operatorname{ord}(\varphi (x))=\infty $ and $\operatorname{ord}(\lnot \varphi (x))=\infty $, in contrast with the assumption that $A$ is local.
\end{proof}
\begin{theorem}{\rm \cite{dinespger}}
\label{th:radconst} Every local MV-algebra can be embedded into an MV-algebra of quasi-constant functions on an ultrapower of $[0,1]$.
\end{theorem}
\begin{proof}
Let $A$ be a local MV-algebra.
By \autoref{theo:DNLT}, any element $x$ of $A$ is a function from $\operatorname{Spec} A$ into $[0,1]^*$,
hence $x/P \in [0,1]^*$ for every $P \in \operatorname{Spec} A$. Moreover, by Proposition \ref{prop:spec},
for any $x \in A$ there exists $r_x \in [0,1]$ such that $(x/P)/\operatorname{Rad}(A/P) = r_x$.
Now, since every $A/P$ is embeddable in $[0,1]^*$, $\operatorname{Rad}(A/P)$ is embeddable in $\operatorname{Rad}([0,1]^*)$.
Therefore, for every $P \in \operatorname{Spec} A$, we have $x/P \subseteq r_x/\operatorname{Rad}[0,1]^*$, whence $A$ is an algebra of quasi-constant functions.
\end{proof}
Next, we shall state and prove an analogous representation for local $\ell u$-groups.
\begin{theorem}{\rm \cite{cdm}}\label{ideals}
Let $(G, u)$ be an $\ell u$-group and $A = \Gamma(G,u)$. Then the posets $\mathcal I(G)$ of all $\ell$-ideals of $G$ and $\mathcal I(A)$ of all ideals of $A$, ordered by set inclusion, are isomorphic under the following maps:
\begin{enumerate}
\item[$\phi:$] $J \in \mathcal I(A) \longmapsto \{x \in G \mid \left|x\right| \wedge u \in J\} \in \mathcal I(G)$,
\item[$\psi:$] $H \in \mathcal I(G) \longmapsto H \cap [0,u] \in \mathcal I(A)$.
\end{enumerate}
\end{theorem}
\begin{corollary}\label{local}
An $\ell u$-group $(G, u)$ is local if and only if $\Gamma(G,u)$ is a local MV-algebra.
\end{corollary}
\begin{theorem}\label{groupquasiconstant}
Every local $\ell u$-group is embeddable in an $\ell$- group
of quasi-constant functions over an ultrapower of ${\mathbb R}$.
\end{theorem}
\begin{proof}
Let ${\mathbb R}$ be the totally ordered group of real numbers, $I$ an ordered set and $U$ an ultrafilter of $I$. Since the property of being a totally ordered group is a first order property, then the ultrapower $R^I/U$ is again a totally ordered group.
Let $X$ be any set and consider the set $K(X,{\mathbb R}^I,U)=$
$$
\left\{f: X \to {\mathbb R}^I \mid \textrm{ there exists } r \in {\mathbb R}^I \textrm{ such that } f(x) \in [r]_U \ \forall x \in X\right\}.
$$
$K(X,{\mathbb R}^I,U)$ is an $\ell$- group with pointwise defined operations and order. Denoting by $\underline{0}$ the identity of ${\mathbb R}^I$, the element $\mathbf{0}: x \in X \longmapsto \underline{0} \in {\mathbb R}^I$, is the identity of $K(X,{\mathbb R}^I,U)$. Analogously, $\mathbf{1}: x \in X \longmapsto \underline{1} \in {\mathbb R}^I$
is a strong unit of $K(X,{\mathbb R}^I,U)$.\footnote{Note that from the ultrapower construction we are not guaranteed that there exists a strong unit.}
Then $\Gamma(K(X,{\mathbb R}^I,U),{\bf 1})=$
$$
\left\{f: X \to [0,1]^I \mid \textrm{there exists } r \in {\mathbb R}^I \textrm{ such that }
f(x) \in [r]_U \ \forall x \in X\right\},
$$
is precisely the MV-algebra $\mathbf{K}(([0,1]^*)^X)$.
Now, if $(G,u)$ is a local $\ell u$-group, by Corollary \ref{local}, $A = \Gamma(G,u)$ is a local MV-algebra. Hence, by \autoref{th:radconst}, $A$ embeds in some $\mathbf{K}(([0,1]^*)^X)$. So, since $\Gamma$ and $\Xi$ are equivalences, $(G,u)$ embeds in $\Xi(\mathbf{K}(([0,1]^*)^X)) = K(X,{\mathbb R}^I,U)$.
\end{proof}
The machinery used at the end of the previous section for perfect MV-algebras can also be used here to find uniform and canonical representation. We state the theorem without proof as it is an easy adaptation of the proofs of \autoref{theorem:9} and \autoref{theorem:10}.
\begin{theorem}
\label{uniformlocal}
For any infinite cardinal $\alpha$, there exists an ultrapower of ${\mathbb R}$, definable in $\alpha$, which we call $\prod{\mathbb R}_{\alpha}$,
such that any $\ell u$-group of cardinality at most $\alpha$ is embeddable in the
$\ell u$-group of quasi-constant functions over $\prod{\mathbb R}_{\alpha}$.
\end{theorem}
|
\section{Introduction}
Given three points $x = (x_0{:}x_1{:}x_2)$,
$y = (y_0{:}y_1{:}y_2)$ and $z = (z_0{:}z_1{:}z_2)$
in the projective plane $\mathbb{P}^2$ over a field $K$, we are interested in
the space $L_{x,y,z}$ of all homogeneous quadrics
that vanish at $x$, $y$ and $z$. By definition, the vector space
$L_{x,y,z}$ is the kernel of the $3 \times 6$ matrix
\begin{equation}
\label{3by6matrix}
\begin{pmatrix}
x_0^2 & x_1^2 & x_2^2 & x_0 x_1 & x_0 x_2 & x_1 x_2 \\
y_0^2 & y_1^2 & y_2^2 & y_0 y_1 & y_0 y_2 & y_1 y_2 \\
z_0^2 & z_1^2 & z_2^2 & z_0 z_1 & z_0 z_2 & z_1 z_2
\end{pmatrix}.
\end{equation}
This defines a rational map which is a morphism on the open set of
non-collinear triples:
$$\Phi\,:\, \mathbb{P}^2 \times \mathbb{P}^2 \times \mathbb{P}^2 \,\longrightarrow \,{\rm Gr}(3,6) \subset \mathbb{P}^{19}$$
Algebraically, the map $\Phi$ is given by evaluating the twenty
$3 \times 3$ minors of the matrix (\ref{3by6matrix}).
This note concerns the {\em tropicalization} of the map $\Phi$.
We study the following inclusions
\begin{equation}
\label{inclusions}
{\rm image} (\rm trop(\Phi))
\subset {\rm trop}({\rm image}(\Phi))
\subset
{\rm trop}({\rm Gr}(3,6)) \subset {\mathbb T}{\mathbb{P}}^{19}.
\end{equation}
In general, naive tropicalization does not commute with morphisms; accordingly, we will see that the inclusions in (\ref{inclusions}) are both strict. We know from \cite[\S 5]{SS} and \cite[Table 2]{HJJS}
that
the {\em tropical Grassmannian} ${\rm trop}({\rm Gr}(3,6))$
has a coarsest fan structure, which is represented as a $3$-dimensional
polyhedral complex with $1005$ maximal polytopes,
namely $990$ tetrahedra and $15$ bipyramids,
which is homologically a bouquet of $126$ 3-spheres.
With the help of {\tt GFan} \cite{gfan}, we computed
the two nested subcomplexes on the left in (\ref{inclusions})
and we found that they are also pure of dimension $3$.
The main point of this note is to furnish the
combinatorial descriptions of these polyhedral complexes which are summarized in Proposition \ref{prop:I_3} and Theorem \ref{thm:IV_3}.
Many studies in tropical geometry \cite{Mik}
concern curves passing
through given points in the plane ${\mathbb T}{\mathbb{P}}^2$.
Our results complement these by
offering a precise analysis of the plane of conics
passing through three points, as in Figure~\ref{fig:plane1},
and how that plane depends on the points.
Our results on (\ref{inclusions}) will be stated and derived in Section 3.
In Section 2, we warm up by solving the same problem for
two points in $\mathbb{P}^2$, where we obtain the {\em tropical Hilbert scheme}
discussed in \cite[\S 6.2]{AN}. Note that
the case of four points in $\mathbb{P}^2$
was already treated in \cite[\S 6]{RST}.\\
\begin{figure}[h!]
\begin{center}
\vskip -0.6cm
\includegraphics[height=6.2in]{exampleplane.eps}
\vskip -1cm
\caption{The plane $L_{X,Y,Z}$ of conics passing through three points
$X, Y,Z \in {\mathbb T}{\mathbb{P}}^2$.}
\label{fig:plane1}
\end{center}
\end{figure}
\section{Two Points}
Given two distinct points $x = (x_0{:}x_1{:}x_2)$ and
$y = (y_0{:}y_1{:}y_2)$ in the projective plane $\mathbb{P}^2$ over a field $K$,
there is a four-dimensional space $L_{x,y}$
of quadrics that vanish at $x$ and $y$, namely
\begin{equation}
\label{2by6matrix}
L_{x,y} \quad = \quad {\rm kernel}
\begin{pmatrix}
x_0^2 & x_1^2 & x_2^2 & x_0 x_1 & x_0 x_2 & x_1 x_2 \\
y_0^2 & y_1^2 & y_2^2 & y_0 y_1 & y_0 y_2 & y_1 y_2
\end{pmatrix}.
\end{equation}
This defines a rational map which takes pairs of points into the Grassmannian:
$$
\Psi \,:\, \mathbb{P}^2 \times \mathbb{P}^2 \,\longrightarrow \,{\rm Gr}(4,6) \subset \mathbb{P}^{14}
$$
The map $\Psi$ blows up the diagonal in
$\mathbb{P}^2 {\times} \mathbb{P}^2$. The closure of its image
is isomorphic to the Hilbert scheme
of two points in $\mathbb{P}^2$. That is,
$\,{\rm image}(\Psi) = {\rm Hilb}_2(\mathbb{P}^2) $
is a smooth $4$-dimensional subvariety of the
$8$-dimensional Grassmannian ${\rm Gr}(4,6)$.
Representing points in ${\rm Gr}(4,6)$ by their dual
Pl\"ucker coordinates, the map
$\Psi$ is given algebraically by evaluating the fifteen
$2 {\times} 2$-minors of the matrix in (\ref{2by6matrix}).
Gr\"obner-based implicitization of $\Psi$ yields the prime ideal
$$
\begin{matrix}
I_2 = \! & \!\!\! \!
\langle
p_{03} p_{15} {+} p_{13} p_{34} ,
p_{03} p_{24} {+} p_{04} p_{45} ,
p_{03} p_{25} {+} p_{34} p_{45} ,
p_{04} p_{13} {+} p_{03} p_{35} ,
p_{04} p_{15} {-} p_{34} p_{35} ,
p_{04} p_{25} {-} p_{24} p_{34} , \\ &
p_{13} p_{24} - p_{35} p_{45} ,
p_{13} p_{25} - p_{15} p_{45} ,
p_{15} p_{24} - p_{25} p_{35} ,\,
p_{01} p_{02} - p_{05}^2 + p_{34}^2 ,
p_{01} p_{12} + p_{14}^2 - p_{35}^2 ,\\ &
p_{02} p_{12} - p_{23}^2 + p_{45}^2 , \,
p_{03} p_{23} - p_{34}^2 + p_{04} p_{35} ,
p_{04} p_{14} - p_{34}^2 + p_{03} p_{45} ,
p_{05} p_{15} - p_{35}^2 - p_{13} p_{45} , \\ &
p_{05} p_{25} - p_{24} p_{35} - p_{45}^2 ,\,
p_{13} p_{23} + p_{14} p_{35} - p_{35}^2 - p_{13} p_{45} ,\,
p_{14} p_{24} - p_{24} p_{35} + p_{23} p_{45} - p_{45}^2 , \\ & \!\!\!\!\!
p_{01} p_{04} {-} p_{03} p_{05} {-} p_{03} p_{34} ,
p_{01} p_{23} {-} p_{14} p_{34} {+} p_{05} p_{35} ,
p_{01} p_{24} {+} p_{05} p_{45} {+} p_{34} p_{45} ,
p_{01} p_{25} {+} p_{14} p_{45} {+} p_{35} p_{45} , \\ & \!\!\!\!\!
p_{02} p_{03} {-} p_{04} p_{05} {+} p_{04} p_{34} ,
p_{02} p_{13} {+} p_{05} p_{35} {-} p_{34} p_{35} ,
p_{02} p_{14} {+} p_{23} p_{34} {+} p_{05} p_{45} ,
p_{02} p_{15} {+} p_{23} p_{35} {+} p_{35} p_{45} , \\ & \!\!\!\!\!
p_{03} p_{12} {+} p_{14} p_{34} {-} p_{34} p_{35} ,
p_{03} p_{14} {-} p_{01} p_{34} {+} p_{03} p_{35} ,
p_{04} p_{12} {+} p_{23} p_{34} {-} p_{34} p_{45} ,
p_{04} p_{23} {+} p_{02} p_{34} {+} p_{04} p_{45} ,\\ & \!\!\!\!\!
p_{05} p_{12} {+} p_{23} p_{35} {-} p_{14} p_{45} ,
p_{05} p_{13} {+} p_{13} p_{34} {+} p_{01} p_{35} ,
p_{05} p_{14} {-} p_{34} p_{35} {+} p_{01} p_{45} ,
p_{05} p_{23} {+} p_{02} p_{35} {+} p_{34} p_{45} ,\\ & \!\!\!\!\!
p_{05} p_{24} {-} p_{24} p_{34} {+} p_{02} p_{45} ,
p_{05} p_{34} {-} p_{04} p_{35} {+} p_{03} p_{45} ,
p_{12} p_{13} {-} p_{14} p_{15} {+} p_{15} p_{35} ,
p_{12} p_{24} {+} p_{23} p_{25} {-} p_{25} p_{45} ,\\ & \!\!\!\!\!
p_{13} p_{14} {+} p_{01} p_{15} {+} p_{13} p_{35} ,
p_{14} p_{23} {+} p_{12} p_{34} {-} p_{35} p_{45} ,
p_{14} p_{25} {-} p_{25} p_{35} {-} p_{12} p_{45} ,
p_{15} p_{23} {+} p_{12} p_{35} {-} p_{15} p_{45} , \\ & \!\!\!\!\!
p_{15} p_{34} {-} p_{14} p_{35} {+} p_{13} p_{45} \,,\,\,
p_{23} p_{24} {+} p_{02} p_{25} {+} p_{24} p_{45} \,,\,\,
p_{25} p_{34} {-} p_{24} p_{35} {+} p_{23} p_{45}
\,\rangle.
\end{matrix}
$$
This is the ideal of the embedding of ${\rm Hilb}_2(\mathbb{P}^2) $ via $\Psi$
as a subscheme of degree $21$ in $\mathbb{P}^{14}$.
We identify the tropicalization of ${\rm Gr}(4,6)$ with the
space of tree metrics on six taxa \cite[\S 2.4]{ascb}.
The taxa are the quadratic monomials.
Combinatorially, this is a simplicial complex
with $25$ vertices, $105$ edges and $105$ triangles.
The $25$ vertices are the splits: trees with one internal edge.
The tropicalization of $\Psi$ is a piecewise-linear map
into that tree space:
$$
{\rm trop}(\Psi) \,:\, {\mathbb T}{\mathbb{P}}^2 \times {\mathbb T}{\mathbb{P}}^2 \,\longrightarrow \,{\rm trop}({\rm Gr}(4,6))
\,\subset \, {\mathbb T}{\mathbb{P}}^{14} .
$$
The coordinates of this map are the $15$ tropical $2 {\times} 2$-minors
of the $2 {\times} 6$-matrix in (\ref{2by6matrix}):
$$ {\rm trop}(\Psi)(X,Y) \,=\,
\bigl( {\rm max}(2 X_0 {+} 2Y_1, 2 X_1 {+} 2X_0), \ldots,
{\rm max}(X_0{+}X_2{+}Y_1{+}Y_2, X_1{+}X_2{+}Y_0{+}Y_2) \bigr). $$
We regard its image as a ``combinatorial Hilbert scheme''
that parametrizes pairs of points in ${\mathbb T}{\mathbb{P}}^2$
by the quadrics that pass through them.
We have the following strict inclusions:
\begin{equation}
\label{hilb2inclusions}
\begin{matrix} &
{\rm Hilb}_2({\mathbb T}{\mathbb{P}}^2) & := & {\rm image}({\rm trop}(\Psi)) & & & \\
\,\subset \, &
{\rm trop}({\rm Hilb}_2(\mathbb{P}^2)) & = & {\rm trop}({\rm image}(\Psi)) &
\,\subset \,& {\rm trop} ({\rm Gr}(4,6)) &\,\subset \,& {\mathbb T}{\mathbb{P}}^{14}.
\end{matrix}
\end{equation}
Working modulo the common lineality spaces,
the Hilbert schemes in (\ref{hilb2inclusions}) are one-dimensional complexes.
These graphs are geometrically embedded, but not as subcomplexes,
inside the two-dimensional simplicial complex
$\,{\rm trop} ({\rm Gr}(4,6))$.
Alessandrini and Nesci argued
in \cite[\S 6.2]{AN} that
${\rm Hilb}_2({\mathbb T}{\mathbb{P}}^2) $ is a cycle of length six.
The following proposition extends their findings.
\begin{proposition}
\label{prop:two}
The tropicalized Hilbert scheme ${\rm trop}({\rm Hilb}_2(\mathbb{P}^2))$
is the graph with $16$ nodes and $30$ edges depicted in Figure
\ref{fig:hilb22}. The outer $6$-cycle is the subgraph ${\rm Hilb}_2({\mathbb T}{\mathbb{P}}^2)$.
The labeling of the graph describes its embedding
in the space of trees on six taxa and is explained below.
\end{proposition}
\begin{center}
\begin{figure}
\vskip -0.4cm
\includegraphics[width=13.5cm]{hilb22.eps}
\vskip -0.2cm
\caption{Tropicalization of the Hilbert scheme of two points in $\mathbb{P}^2$}
\label{fig:hilb22}
\end{figure}
\end{center}
We proved Proposition \ref{prop:two} by applying the software {\tt GFan} \cite{gfan}
to the ideal $I_2$ and by carefully analyzing the
output of that computation. We now discuss the
outcome of that analysis.
The graph ${\rm trop}({\rm Hilb}_2(\mathbb{P}^2))$ has $16$ nodes.
Twelve of the nodes are also nodes in the space of trees,
${\rm trop}({\rm Gr}(2,6))$, so
they correspond to splits of the set
of taxa $\{x_0^2, x_1^2,x_2^2, x_0 x_1, x_0 x_2, x_1 x_2\}$.
Up to the action of the symmetric group $\mathfrak{S}_3$ by permuting the coordinates
of $\mathbb{P}^2$, there are
\begin{enumerate}
\item three splits like $\,\{\, \{x_0^2, x_1^2\}\, , \, \{x_2^2, x_0 x_1, x_0 x_2, x_1 x_2\} \,\}$,
\item three splits like $\,\{\, \{x_0^2, x_1 x_2\}\, , \, \{x_1^2 ,x_2^2, x_0 x_1, x_0 x_2 \} \,\}$,
\item three splits like $\,\{\, \{x_0 x_1, x_0 x_2\}\, , \, \{x_0^2, x_1^2,x_2^2, x_1 x_2\} \,\}$,
\item three splits like $\,\{\, \{x_0^2 ,x_0 x_1, x_1^2\}\, , \, \{x_2^2, x_0 x_2, x_1 x_2\} \,\}$.
\end{enumerate}
The nine trees with one split in (1)-(3) have one cherry, or set of edges paired together. We represent
that tree by drawing the cherry pair as a thick black segment in the corresponding
vertex label of Figure~\ref{fig:hilb22}.
The three trees in (4) are 3-3 splits so they have no cherry.
They appear alternatingly on the outer $6$-cycle in Figure~\ref{fig:hilb22},
where they are drawn by a long segment across a triangle.
The other three nodes on the outer $6$-cycle are trees with two interior nodes:
\begin{enumerate}
\item[(5)] three split pairs like
$\bigl\{\! \{\!\{x_0^2,x_0 x_2\}\!, \!\{ x_1^2,x_2^2, x_0 x_1, x_1 x_2\}\!\},
\{\!\{ x_1^2, x_1 x_2 \} \!, \!\{ x_0^2,x_2^2, x_0 x_1, x_0 x_2\}\!\} \! \bigr\}$.
\end{enumerate}
Finally, there is one special node that lies in the relative interior
of a triangle in ${\rm trop}({\rm Gr}(2,6))$:
\begin{enumerate}
\item[(6)] the unique trivalent {\em snowflake tree} with
cherries
$\{ x_0^2, x_1 x_2\}$,
$\{x_1^2, x_0 x_2\}$ and
$\{x_2^2, x_0 x_1\}$.
\end{enumerate}
The graph ${\rm trop}({\rm Hilb}_2(\mathbb{P}^2))$ has $30$ edges.
Twelve are interior to
triangles of ${\rm trop}({\rm Gr}(2,6))$,
so they correspond to trivalent trees.
Six of those are the edges of type (4-5) that form the
outer $6$-cycle. The others are the three (2-6) edges adjacent to the
snowflake tree, and the three (2-5) edges that appear as the longest edges in Figure \ref{fig:hilb22}.
The remaining $18$ edges of ${\rm trop}({\rm Hilb}_2(\mathbb{P}^2))$ are also edges
of ${\rm trop}({\rm Gr}(2,6))$, so they correspond to trees with two interior edges.
Those edges are three (1-2)s, six (1-3)s,
three (1-4)s, three (2-3)s, and three (3-4)s.
The tropical map ${\rm trop}(\Psi)$ amounts to a
double cover of the $6$-cycle ${\rm Hilb}_2({\mathbb T}{\mathbb{P}}^2)$.
To see this, we note that the Newton polytope
${\rm NP}(\Psi)$ of the rational map $\Psi$,
as defined in \cite[(3.38)]{ascb} is a centrally-symmetric $12$-gon.
Namely, ${\rm NP}(\Psi)$ is the Minkowski sum of the $15$ Newton polytopes of all
$2 {\times} 2$-minors of the $2 {\times} 6$-matrix (\ref{2by6matrix}), and these are
line segments
that lie in a common plane and involve six distinct directions.
According to \cite[Theorem 3.42]{ascb}, ${\rm trop}(\Psi)$
is linear on each of the twelve cones in the normal fan of ${\rm NP}(\Psi)$.
The $12$-gon formed by these cones is mapped
onto ${\rm Hilb}_2({\mathbb T}{\mathbb{P}}^2)$ by looping twice around the outer $6$-gon in Figure~\ref{fig:hilb22}.
\section{Three Points}
Guided by the above results for the
two-point map $\Psi$, we now investigate the
three-point map $\Phi$. We write $I_3$ for the
homogeneous prime ideal that represents the image of
$\Phi$. The following proposition summarizes
basic facts about the variety
$V(I_3) = {\rm image}(\Phi) \subset {\rm Gr}(3,6)$.
\begin{proposition} \label{prop:I_3}
The projective variety $V(I_3)$ has dimension $6$ and degree $57$.
Its ideal $I_3$ is minimally generated by $62$
homogeneous quadrics in the $20$ Pl\"ucker coordinates $p_{ijk}$.
Among these are $35$ quadrics that vanish on the Grassmannian
${\rm Gr}(3,6) \subset \mathbb{P}^{19}$.
The corresponding tropical variety ${\rm trop}(V(I_3))$ in ${\mathbb T}{\mathbb{P}}^{19}$,
with its Gr\"obner fan structure
and taken modulo the lineality space, is a $3$-dimensional polyhedral
complex with f-vector $(1095, 6621, 12830, 7649)$.
\end{proposition}
\begin{ex} The quadrics
$\,p_{045} p_{145}-p_{013} p_{245}-p_{345}^2\,$ and
$\, p_{025} p_{145}-p_{134}p_{245}-p_{015} p_{245} + p_{235} p_{345} \,$
are among the $27$ generators of the image of $I_3$
in the coordinate ring of ${\rm Gr}(3,6)$.
\qed
\end{ex}
There is a natural morphism from the Hilbert scheme
${\rm Hilb}_3(\mathbb{P}^2)$ onto our variety $V(I_3)$. However, unlike in Section 2,
this is not an isomorphism. Geometrically, the morphism from the Hilbert scheme
contracts all triples of points that lie on the same line.
The singular locus of $V(I_3)$ is a projective plane $\mathbb{P}^2$,
namely, it is the image of all collinear triples in ${\rm Hilb}_3(\mathbb{P}^2)$.
The Gr\"obner fan structure on the tropical variety in Propsition \ref{prop:I_3}
is not simplicial: among the $7649$ three-dimensional
polytopes, $876$ have five vertices
($105$ bipyramids and $773$ Egyptian pyraminds),
$27$ have six vertices (all triangular prisms), and $12$ have seven vertices
(two types). The number of
$\mathfrak{S}_3$-orbits of facets of
${\rm trop}({\rm image}(\Phi)) = {\rm trop}(V(I_3))$ is $1318$.
Our results are summarized in Theorem \ref{thm:IV_3}.
The role of the graph in Figure \ref{fig:hilb22}
is now played by the $3$-dimensional complex with $7649$ facets.
It is obviously too big to be fully displayed here. Instead, we shall now focus
on the leftmost complex in (\ref{inclusions}). The tropical morphism ${\rm trop}(\Phi)$ is a piecewise linear map.
As shown in \cite[\S 3.4]{ascb}, its domains of linearity are the normal cones
of the Newton polytope.
We begin by computing this Newton polytope.
\begin{lemma}
The Newton polytope of $\Phi$ is $4$-dimensional and
has f-vector $(504,1056,684,132)$.
\end{lemma}
The polytope ${\rm NP}(\Phi)$ plays the same role as
the $12$-gon ${\rm NP}(\Psi)$ in Section~2. The $504$ vertices of ${\rm NP}(\Phi)$ correspond
to distinct types of three labeled points in ${\mathbb T}{\mathbb{P}}^2$, where the type is the cell
of ${\rm Trop}({\rm Gr}(3,6))$ that contains the plane of quadrics through these points.
The map $\Phi$ is invariant under permuting the three points $x$, $y$ and $z$, and this reduces
the number of image cones to $504/6 = 84$. These $84$ cones in ${\mathbb T}{\mathbb{P}}^{19}$
are grouped into $17$ orbits of size six with respect to the common symmetry
group $\mathfrak{S}_3$ of $I_3$, $V(I_3)$ and ${\rm Trop}(V(I_3))$.
Those symmetries correspond to permuting the indices $0$, $1$ and $2$ of the coordinates on
$\mathbb{P}^2$ or ${\mathbb T}{\mathbb{P}}^2$.
\begin{table}[h!]
\caption{The $17$ generic types of triples in ${\mathbb T}{\mathbb{P}}^2$ and their planes of conics}.
\centering
\begin{tabular}[h!]{l c c c c r}
\hline
Type & Orbit Size & Valency & $(X_1,X_2) $ & $ (Y_1,Y_2)$ & Plane $L_{X,Y,Z}$ \\
\hline
1 & 12 & 6 & $(4,3)$ & $(3,-1)$ & FFFGG\\
2 & 36 & 5 & $(6,5)$ & $(3,-1)$ & EFFG\\
3 & 36 & 5 & $(4,5)$ & $(2,-1)$ & FFFGG\\
4 & 36 & 4 & $(6,7)$ & $(2,-1)$ & EFFG\\
5 & 36 & 4 & $(8,6)$ & $(5,1)$ & EFFG\\
6 & 36 & 4 & $(5,8)$ & $(-1,5)$ & EEFG\\
7 & 36 & 4 & $ (5,6)$ & $(2,1)$ & EEFF(a)\\
8 & 36 & 4 & $(3,7)$ & $(-1,2)$ & EEFG\\
9 & 36 & 4 & $(6,5)$ & $(4,2)$ & EEFF(a)\\
10 & 36 & 4 & $(5,6)$ & $(3,1)$ & EFFG\\
11 & 36 & 4 & $ (6,5)$ & $(5,2)$ & EFFG\\
12 & 36 & 4 & $(4,6 )$ & $(2,3)$ & EEFF(a)\\
13 & 12 & 4 & $(5,3)$ & $(3,-2)$ & EEEG\\
14 & 18 & 4 & $(3,6)$ & $(1,3)$ & EEFF(b)\\
15 & 18 & 4 & $(3,6)$ & $(2,3)$ & EEFF(b)\\
16 & 36 & 4 & $(3,5)$ & $(2,1)$ & EEFG\\
17 & 12 & 4 & $(2,3)$ & $(3,1)$ & EEEG\\
\hline
\end{tabular}
\label{table:maxcones}
\end{table}
Given three points $X = (X_0,X_1,X_2)$,
$Y = (Y_0,Y_1,Y_2)$ and $Z = (Z_0,Z_1,Z_2)$ in the tropical projective plane ${\mathbb T}{\mathbb{P}}^2$, we
write $L_{X,Y,Z}$ for the tropical $2$-plane in ${\mathbb T}{\mathbb{P}}^5$ determined,
as in \cite[(3.44)]{ascb} or
\cite[\S 2]{Sp}, by the tropical Pl\"ucker vector
$\,{\rm trop}(\Phi)(X,Y,Z) \,\in \,{\rm trop}({\rm Gr}(3,6)) \,\subset \,{\mathbb T}{\mathbb{P}}^{19}$.
Geometrically, $L_{X,Y,Z}$ is the tropical plane whose points are the
tropical quadrics that pass through the points $X,Y,Z$.
Our picture of this in Figure~\ref{fig:plane1} is
reminiscent of \cite[Fig.~19]{RST}.
Our main result is the classification of the $17$ types
of configurations of triples of points.
\begin{theorem}\label{thm:IV_3}
Precisely $48$ of the $1005$ generic $2$-planes in ${\mathbb T}{\mathbb{P}}^5$ arise
as $L_{X,Y,Z}$ for some triple $X,Y,Z \in {\mathbb T}{\mathbb{P}}^2$.
This covers six of the seven symmetry classes.
Table~\ref{table:maxcones} summarizes
the correspondence between the $17$ types of triples and
the $6$ combinatorial types of $2$-planes.
\end{theorem}
We proved this theorem by explicit computations. We shall explain our method and how to read Table~\ref{table:maxcones}.
For each of the $17$, types we list a representative configuration.
Here we break the symmetry by setting $X_0 = Y_0 = Z_0 = 0$
and by fixing the third point to lie at the origin, i.e.~$Z = (Z_1,Z_2) = (0,0)$.
For each configuration, the first point $X= (X_1,X_2)$ is listed in the fourth column,
and the second point $Y=(Y_1,Y_2)$ is listed in the fifth column.
\begin{figure}
\begin{center}
\includegraphics[height=4.3in]{divisionplane.eps}\\
\caption{Partition of the $(Z_1,Z_2)$-plane obtained
by $X = (0,0)$ and $Y= (2,3)$.}
\label{DivisionPlane}
\end{center}
\end{figure}
The second column, ``Orbit Size'', lists of the cardinality of the orbit of the configuration
under permuting both points and coordinates. The sum of the $17$ orbit
sizes is $504$, the total number of vertices of ${\rm NP}(\Phi)$. The third column, ``Valency'', lists the number of edges of
the Newton polytope ${\rm NP}(\Phi)$ that are adjacent to the
given vertex. Equivalently, this is the number of linear inequalities needed to
characterize the configurations of the type in question.
For instance, there are
six such linear inequalities for type 1:
\begin{equation}
\label{eq:sixineq}
X_1 \geq X_2, \,\, X_1 \geq Y_1,\,\,
Y_2 \leq 0 ,\,\,
X_2+2 Y_1 \geq 2 X_1, \,\,
2 X_2+Y_1 \geq 2 X_1,\,\,
X_2+Y_1+Y_2 \geq X_1.
\end{equation}
One solution is $\{X = (4,3), Y = (3,-1) \}$. A solution is equivalent,
in the sense that the plane
$L_{X,Y,Z}$ is in the same maximal cone
of ${\rm trop}({\rm Gr}(3,6))$, if and only if the six inequalities in
(\ref{eq:sixineq}) are satisfied.
The solution set of (\ref{eq:sixineq}) is the
cone over a bipyramid. The corresponding vertex figure of
${\rm NP}(\Phi)$ is a cube. The vertex figures for
types 2 and 3 are Egyptian pyramids.
All other $14$ types of vertices are simple, so those vertex
figures of ${\rm NP}(\Phi)$ are tetrahedra.
One way to visualize the partition of $\mathbb{R}^4$ into $504$ normal
cones to ${\rm NP}(\Phi)$ is to intersect this normal fan with the
$2$-dimensional affine space obtained by also fixing
the second point $Y = (Y_1,Y_2)$
at a particular location. For instance, in Figure \ref{DivisionPlane}
we fix $Y=(2,3)$, and we allow $X = (X_1,X_2)$ to vary over
the plane. The regions of equivalence are convex polygons.
On each polygon, the plane of conics $L_{X,Y,Z}$ has a fixed
combinatorial type in ${\rm trop}({\rm Gr}(3,6))$.\
The tropicalization of ${\rm Gr}(3,6)$ has seven $\mathfrak{S}_6$-orbits of maximal cones. The interior of each corresponds to a distinct type
of plane in ${\mathbb T}{\mathbb{P}}^5$. These types were classified in \cite[\S 5]{SS} and given the names EEEE, EEFF(a), EEFF(b), EFFG, EEEG, EEFG, and FFFGG.
See \cite[Figure 1]{HJJS} for a diagram that shows these seven planes.
In the last column of Table~\ref{table:maxcones}, we see that
EEEE is the unique type that does not arise as $L_{X,Y,Z}$
for any triple $X,Y,Z$ in ${\mathbb T}{\mathbb{P}}^2$. The other six types arise
as planes of conics through three points, as also seen
in Figure \ref{DivisionPlane}.
Each tropical plane consists of bounded and unbounded faces.
These are dual to the interior cells and boundary cells of
a matroid subdivision of the second hypersimplex \cite{Sp}.
Each cell is indexed by a matroid of rank $3$ on
the ground set $\{1,2,3,4,5,6\}$. Intersecting a generic tropical plane
in ${\mathbb T}{\mathbb{P}}^5$ with the six tropical hyperplanes at infinity gives a tree arrangement consisting of six trees with five leaves each. A detailed description of the correspondence between tropical planes
and tree arrangements is given in \cite[\S 4]{HJJS}. For our planes $L_{X,Y,Z}$ that arise from triples $X,Y,Z \in {\mathbb T}{\mathbb{P}}^2$,
we can construct the corresponding arrangement of six trees by removing in turn each of the
six monomials $\{U_0^2, U_1^2, U_2^2, U_0 U_1, U_0 U_2, U_1 U_2 \}$
from the defining equation of the conic. In other words, each of the
six trees represents a line in ${\mathbb T}{\mathbb{P}}^4$. That line is a tree which parameterizes
conics with five fixed terms that pass through the three given points.
These trees are similar to \cite[Fig.~19]{RST} but have only five taxa.\
Now we explain how we constructed Table \ref{table:maxcones}.
We first computed the Newton polytope ${\rm NP}(\Phi)$
using {\tt Gfan}, and we picked a representative in each
maximal cone of its normal fan. Up to symmetries, these
are the $17$ displayed configurations $X = (X_1,X_2), Y= (Y_1,Y_2) , Z = (0,0)$.
For these we calculated the corresponding vectors
of tropical Pl\"ucker coordinates
$$ ({\rm trop}(\Phi))(X,Y,Z) \,\,= \,\,
\bigl( P_{012}, P_{013}, P_{014}, \ldots, P_{245}, P_{345} \bigr) \,\, \in \,\, \mathbb{R}^{20}. $$
For each of the six indices, we consider the restricted vector
of Pl\"ucker coordinates involving that index. For instance, for index
``$0$'', corresponding to the monomial $U_0^2$, this is the vector
$$(P_{012}, P_{013}, P_{014}, P_{023}, P_{024}, P_{034}) .$$
This vector represents the pairwise distances in a phylogenetic
tree with taxa $\{1,2,3,4,5\}$. This is the first among the six trees
that represent the plane $L_{X,Y,Z}$, in its guise as
a $3$-tree \cite[(3.44)]{ascb}. At this point, the last column in
Table \ref{table:maxcones} can simply read off from
\cite[Table 2]{HJJS}.
\bigskip
{\bf Acknowledgements.}
We are grateful to Anders Jensen for helping us with our {\tt Gfan} computation.
This work is based on the undergraduate senior thesis of the first author.
The second author was partially supported by the NSF
(DMS-0456960 and DMS-0757207).
\bigskip
|
\section{Introduction}
The {\it Gamma-ray Large Area Space Telescope} was launched on 2008 June 11.
It began its scientific operations two months later, and shortly thereafter,
it was renamed the {\it Fermi Gamma-ray Space Telescope}. Its primary
instrument is the Large Area Telescope \citep[LAT;][]{atwood}, the successor
to the Energetic Gamma-Ray Experiment Telescope (EGRET) on board the
{\it Compton Gamma-Ray Observatory} \citep{egret}. The LAT offers a major
increase in sensitivity over EGRET and the Italian Space Agency's
{\it Astro-rivelatore Gamma a Immagini Leggero} \citep[{\it AGILE};][]{agile},
allowing it to study the \mbox{$\gamma$-ray} sky in unprecedented detail. In sky
survey mode ({\it Fermi}'s main observing mode), the LAT observes the entire
sky every 3~hours, providing effectively uniform exposure on the timescale of
days.
One of the major scientific goals of the {\it Fermi} mission is to investigate
high-energy emission in active galactic nuclei (AGNs). Although it is
generally accepted that the \mbox{$\gamma$-rays} detected from blazars are emitted
from collimated jets of charged particles moving at relativistic speeds
\citep{blandrees,maraschi}, open questions remain. The mechanisms by which
the particles are accelerated, the precise site of the \mbox{$\gamma$-ray} emission,
and the origin of AGN variability and the \mbox{$\gamma$-ray} duty cycle of blazars
are not well understood. The physical reasons for the observational
differences between radio-loud and radio-quiet AGNs and between FSRQs and
BL~Lacs are also unclear. LAT observations of blazars and other AGNs are
already helping to address these and other issues. Several in-depth spectral
and/or multiwavelength studies of specific blazars
\citep[e.g.,][]{3c454, pks1454, pks2155} and of non-blazar radio galaxies
\citep{ngc1275,M87,pmnj0948} have been performed.
The high sensitivity and nearly uniform sky coverage of the LAT make it a
powerful tool for investigating the properties of large populations. The first
list of bright AGNs detected by the LAT, the LAT Bright AGN Sample
\citep[LBAS;][]{LBAS} included bright AGNs at high Galactic latitude
($|b|>10\arcdeg$) detected with high significance ($TS>100$, or
$\ga$$10\sigma$) during the first three months of scientific operation. This
list comprised 58 FSRQs, 42 BL~Lacs, two radio galaxies, and four AGNs of
unknown type. Following the models used to describe the \mbox{$\gamma$-ray} spectra
obtained with previous \mbox{$\gamma$-ray} observatories \citep[e.g.,][]{mat96}, the
early analysis reported in the LBAS was carried out by fitting the \mbox{$\gamma$-ray}
spectra at energies above 200 MeV using a simple power-law model. This
analysis revealed a fairly distinct spectral separation between FSRQs and
BL~Lacs, with FSRQs having significantly softer spectra. The division between
the two classes was found to be at power law index $\Gamma \approx 2.2$. It
has been suggested \citep{Ghi09} that this separation results from different
radiative cooling of the electrons due to distinct accretion regimes in the two
blazar classes. The \mbox{$\gamma$-ray} spectral properties and time-resolved
multifrequency spectral energy distributions of LBAS sources were further
investigated in \citet{Spe09,SEDpaper}.
Here, we report on a larger AGN sample detected after 11 months of scientific
operations. The LAT first-year catalog \citep[1FGL;][]{Catalog} contains a
total of 1451 sources detected with $TS > 25$, and 1043 of these are
at high Galactic latitudes ($|b| > 10\arcdeg$). We present a catalog of the
high-latitude 1FGL sources that are associated with blazars and other AGNs. We
refer to this as the First LAT AGN Catalog (1LAC). In addition to the 1LAC, we
also provide, where possible, AGN associations for low-latitude LAT sources and
AGN ``affiliations''---candidate counterparts for which a quantitative
association probability could not yet be computed---for unassociated
high-latitude sources.
In Section 2, we describe the observations by the LAT and the analysis that led
to the first-year catalog. In Section 3, we explain the method for associating
\mbox{$\gamma$-ray} sources with AGN counterparts in a statistically meaningful
way, present the results of this method, and describe the two schemes for
classifying 1LAC AGNs. Section 4 provides a brief census of the 1LAC sample.
Section 5 summarizes some of the properties of the 1LAC, including the
\mbox{$\gamma$-ray} flux distribution, the \mbox{$\gamma$-ray} photon spectral
index distribution, the \mbox{$\gamma$-ray} variability properties, the
redshift distribution, and the \mbox{$\gamma$-ray} luminosity distribution. In
Section 6, we describe the multiwavelength properties, from radio to TeV, of
the 1LAC AGNs. We discuss the implications of the 1LAC results in Section 7
and conclude in Section 8.
In the following, we use a $\Lambda$CDM cosmology with values within $1\sigma$ of
the {\it Wilkinson Microwave Anisotropy Probe} ({\it WMAP}) results
\citep{komatsu}; in particular, we use $h = 0.71$, $\Omega_m = 0.27$, and
$\Omega_\Lambda = 0.73$, where the Hubble constant
$H_0=100h$~km~s$^{-1}$~Mpc$^{-1}$. We also define the radio spectral indices
such that $S(\nu) \propto \nu^{-\alpha}$.
\section{\label{obs}Observations with the Large Area Telescope --- Analysis Procedures}
The \mbox{$\gamma$-ray} sources in the 1LAC are a subset of those in the 1FGL catalog;
we summarize here the procedures used in producing the 1FGL catalog. The data
were collected from 2008 August 4 to 2009 July 4, primarily with standard
sky-survey observations. Only photons in the ``Diffuse'' event
class\footnote{See http://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone/Cicerone\_Data/LAT\_DP.html .}
with energies in the range 0.1-100~GeV were considered in this analysis in
order to minimize contamination from misclassified cosmic rays \citep{atwood}.
This photon class is described in more detail (comparison to the class used in
the Bright Source List \citep[BSL;][]{BSL} paper, systematic uncertainties) in
\citet{Catalog}. To minimize contamination from \mbox{$\gamma$-rays} from the
Earth's limb, photons with incident directions greater than $105\arcdeg$ from
the local zenith were removed. In addition, time ranges during which the
rocking angle of the LAT was greater than $43\arcdeg$ were excluded from the
data set because the bright limb of the Earth entered the field of view. This
rocking angle limit removed only a small fraction of the data, with the
excluded time intervals occurring during occasional 5-hour pointed observations
at larger rocking angles (\mbox{$\gamma$-ray} burst [GRB] afterglow searches)
and during even briefer intervals related to Sun avoidance during survey mode
observations. A few minutes were excised around two bright GRBs (GRB~080916C
and GRB~090510). The few time intervals with poor data quality, flagged as
anything other than ``Good'' in the pointing/live time history (FT2) files,
were also excluded. The resulting data set includes 245.6~d of live time. The
standard {\it Fermi}-LAT {\it ScienceTools} software
package\footnote{http://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone/}
(version v9r15p2) was used with the ``P6\_V3\_DIFFUSE'' set of instrument
response functions.
The source detection step made use of two wavelet algorithms, {\it mr\_filter}
\citep{sp98} and {\it PGWAVE} \citep{ciprini07}, as well as tools that maximize
a simplified likelihood function \citep[{\it pointfind};][]{Catalog} and that
implement a minimum spanning tree algorithm \citep{cmg08}. The intention in
using a variety of algorithms to generate a list of ``seed'' positions for
sources was to keep the source-detection step from being a limiting factor in
the analysis. As described in the 1FGL paper, the algorithms were run
independently for different energy ranges to find both soft- and hard-spectrum
sources. Yet more seeds were introduced from the Roma-BZCAT \citep{bzcat}
and {\it WMAP} \citep{Hin07,Gio09} catalogs if no nearby LAT seeds were
present. The seeds were essentially candidate sources, and each was evaluated
in detail in the subsequent steps of the 1FGL catalog analysis using the
standard {\it gtlike} tool to arrive at the final list.
The Galactic diffuse background model consistently employed throughout the
analysis is the currently recommended version (gll\_iem\_v02), publicly
released through the {\it Fermi} Science Support
Center\footnote{http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html}.
The isotropic background (including the \mbox{$\gamma$-ray} diffuse and residual
instrumental backgrounds) model was derived from an overall fit of the diffuse
component over the $|b|>30\arcdeg$ sky. The Galactic diffuse model and
corresponding isotropic spectrum are described in more detail in documentation
available from the {\it Fermi} Science Support Center.
To evaluate the source significance, we used the maximum-likelihood algorithm
implemented in {\it gtlike}. For the 1FGL catalog, a threshold of 25 was
adopted for the test statistic ($TS$) from the {\it gtlike} likelihood
analysis. Sources found to have $TS > 25$ were included in the 1FGL catalog.
This corresponds approximately to a minimum significance of $4.1\sigma$.
Figure~\ref{fig:sens} displays a sky map, in Galactic coordinates, of the flux
limit for photon spectral index $\Gamma=2.2$ and $TS=25$. The anisotropy
(about a factor of 2) is due to the non-uniform Galactic diffuse background and
non-uniform exposure (mostly arising from the passage of the {\it Fermi}
satellite through the South Atlantic Anomaly). For soft sources, source
confusion decreases the sensitivity to some extent \citep{Catalog}.
The flux, photon spectral index ($\Gamma$), and test statistic of each source
in the energy range 0.1-100~GeV were determined by analyzing regions of
interest (ROI) typically $12\arcdeg$ in radius. The model of the ROI used to
fit the data was built taking into account all the sources detected within a
given ROI. The fluxes in five bands (0.1-0.3, 0.3-1, 1-3, 3-10, and
10-100~GeV) were also evaluated, with the photon spectral index held fixed to
the best fit over the whole interval. The energy flux was determined from
these fluxes, resulting in better accuracy than would be obtained from the
power-law fitted function. For hard sources, the test statistics provided by
the 10-100~GeV band can be appreciable, which represents a notable difference
with respect to EGRET, for which the acceptance dropped sharply in this range.
The departure of the spectrum from a power-law shape can be estimated via a
simple $\chi^2$ test on the five fluxes, referred to as the curvature index
\citep{Catalog}.
A TS map---a 2-dimensional array of likelihood TS values evaluated at a finely
spaced grid centered on the direction of a \mbox{$\gamma$-ray} source---was generated
for each source using {\it pointfit}. TS values are determined for each grid
position independently by maximum-likelihood fitting of a test point source.
From the TS maps, elliptical fits to the 95\% confidence source location
contours were derived; for this, the decrease of the TS away from the
maximum-likelihood position of a source is interpreted in terms of the $\chi^2$
distribution with two degrees of freedom (the coordinates of the test source).
The semimajor and semiminor axes of these ellipses were multiplied by 1.1 to
account for systematic errors, which were evaluated by comparing the measured
positions of bright sources to the known positions. The fiducial 95\% error
radius (the geometric mean of the semimajor and semiminor axes) is plotted
against TS (derived as described above) in Figure~\ref{fig:r95}. For each
source, a simple variability index was derived from the $\chi^2$ value of the
monthly flux distribution with a 3\% systematic uncertainty included.
\section{Source Associations}
Any procedure for associating a \mbox{$\gamma$-ray} source with a lower-energy
counterpart necessarily relies on a spatial coincidence between the two. In
the EGRET era, the problem of counterpart associations was made very difficult
by the large \mbox{$\gamma$-ray} localization contours:\ the mean 95\% error radius
for sources in the 18-month EGRET sky survey \citep{1eg} was $0\fdg62$. The
LAT offers a great improvement in source localization, resulting in a mean 95\%
error radius of $0\fdg15$ for the high-latitude 1FGL sources. (For comparison
to LBAS, which included only sources with $TS > 100$ and had a mean 95\% error
radius of $0\fdg14$, the corresponding value for high-latitude 1FGL sources
with $TS > 100$ is $0\fdg09$.)
However, the LAT localization accuracy is not good enough to permit the
determination of a lower-energy counterpart based only on positional
coincidence. A firm counterpart identification is asserted only if the
variability detected by the LAT corresponds with variability at longer
wavelengths. In practice, such identifications are made only for a few sources
(see Table~7 in \citealt{Catalog}). For the rest, we use a method for finding
associations between LAT sources and AGNs based on the calculation of
association probabilities using a Bayesian approach implemented in the
{\it gtsrcid} tool included in the LAT {\it ScienceTools} package and described
in the BSL paper.
\subsection{\label{sec:gtsrcid}The Bayesian Association Method}
The Bayesian method \citep{deRuiter77,ss92}, implemented by the {\it gtsrcid}
tool in the LAT {\it ScienceTools}, is similar to that used by \citet{mhr01} to
associate EGRET sources with flat-spectrum radio sources. An earlier version
of the method was used and described in the BSL and LBAS papers. A more
complete description is given in \citet{Catalog}, but we provide a basic
summary here. The method uses Bayes's theorem to calculate the posterior
probability that a source from a catalog of candidate counterparts is truly an
emitter of \mbox{$\gamma$-rays} detected by the LAT. The significance of a
spatial coincidence between a candidate counterpart from a catalog $C$ and a
LAT-detected \mbox{$\gamma$-ray} source is evaluated by examining the local
density of counterparts from $C$ in the vicinity of the LAT source. We can
then estimate the likelihood that such a coincidence is due to random chance
and establish whether the association is likely to be real. To each catalog
$C$, we assign a prior probability, assumed for simplicity to be the same for
all sources in $C$, for detection by the LAT. The prior probability for each
catalog can be tuned to give the desired number of false positive associations
for a given threshold on the posterior probability, above which the
associations are considered reliable (see Section~\ref{sec:cat}). We use a
slightly different configuration of {\it gtsrcid} from that used for the 1FGL
catalog. This allowed us to assign multiple associations to a single LAT
source and to find associations with probabilities above 50\% (compared with a
threshold of 80\% for the 1FGL associations).
Candidate counterparts were drawn from a number of source catalogs. The most
important ones are the Combined Radio All-sky Targeted Eight GHz Survey
\citep[CRATES;][]{crates}, the Candidate Gamma-Ray Blazar Survey
\citep[CGRaBS;][]{cgrabs}, and the Roma-BZCAT \citep{bzcat}. The CRATES
catalog contains precise positions, 8.4~GHz flux densities, and radio spectral
indices for over 11,000 flat-spectrum sources over the entire $|b| > 10\arcdeg$
sky. CGRaBS, a sample of the 1,625 CRATES sources with radio and \mbox{X-ray}
properties most similar to the blazars in the Third EGRET Catalog
\citep[3EG;][]{3EGcatalog}, provides optical magnitudes, optical
classifications, and spectroscopic redshifts. Roma-BZCAT is a list of blazars
compiled based on an accurate examination of data from the literature and
currently includes over 2800 sources, all observed at radio and optical
frequencies and showing the observational characteristics of blazars. A
complete list of the source catalogs used by {\it gtsrcid} can be found in
\citet{Catalog}.
The same association method can be used at low latitudes; most of the candidate
counterparts in this region are drawn from the VLBA Calibrator Survey
\citep{vcs1,vcs2,vcs3,vcs4,vcs5,vcs6}. These associations are discussed in
Section~\ref{sec:lowlat}.
\subsection{Association Results}
\subsubsection{\label{sec:cat}The First LAT AGN Catalog (1LAC)}
The 1LAC comprises all high-latitude ($|b| > 10\arcdeg$) sources with an
association from {\it gtsrcid}; the full catalog includes 709 AGN associations
for 671 distinct 1FGL sources and is shown in Table~1. An AGN is in the
``high-confidence'' sample if and only if its association probability $P$ is at
least 80\%; this sample contains 663 AGNs. An AGN is in the ``clean'' sample
if and only if it has $P \ge 80$\%, it is the sole AGN associated with the
corresponding 1FGL \mbox{$\gamma$-ray} source (as indicated by an ``S'' in the
last column of Table~1), and it is not ``flagged'' in the 1FGL catalog as
exhibiting some problem or anomaly that casts doubt on its detection. This
last criterion eliminates 12 sources from the clean sample:\
1FGL~J0217.8$+$7353, 1FGL~J0258.0$+$2033, 1FGL~J0407.5$+$0749,
1FGL~J0433.5$+$3230, 1FGL~J0539.4$-$0400, 1FGL~J0540.9$-$0547, 1FGL~J1424.5$-$7847, 1FGL~J1702.7$-$6217,
\linebreak
1FGL~J1727.9$+$5010,
1FGL~J1938.2$-$3957, 1FGL~J2212.9$+$0654, and 1FGL~J2343.6$+$3437. All figures
presented here are for the clean sample, which contains 599 AGNs.
The LBAS associations included one LAT source that was associated with two
radio counterparts. \citet{LBAS} noted that, as the number of LAT detections
increased, source confusion was likely to be more of a problem, and indeed, the
1LAC includes 35 LAT sources that are associated with more than one AGN (for a
total of 73 such associations). In cases of multiple associations, we list
each counterpart separately in Table~1 and indicate them in the last column of
the table.
The prefix ``FRBA'' in the column of AGN names refers to sources observed at
8.4~GHz as part of VLA program AH996 (``Finding and Rejecting Blazar
Associations for {\it Fermi}-LAT \mbox{$\gamma$-ray} Sources''). The prefix
``CLASS'' refers to sources from the Cosmic Lens All-Sky Survey \citep{class1,
class2}.
Figure~\ref{fig:separation} shows the distribution of normalized angular
separations between the \mbox{$\gamma$-ray} sources and their AGN counterparts. The
solid curve corresponds to the expected distribution for true associations
while the dashed curve represents the expected distribution for purely random
associations. These results provide confidence that most of the associations
found are real.
The association probabilities can be used to estimate the number of false
positive associations. In a sample of $k$ sources with association
probabilities $P_i$, the number of false positives is
\mbox{$N_\mathrm{false} \approx \sum_{i=1}^k (1-P_i)$}. Among the 709
associations in the entire 1LAC, $\sim$30 are false, but of the 663 sources
in the high-confidence list, there are only $\sim$14 false positives, and only
$\sim$11 of the sources in the clean sample are falsely associated.
Additionally, there should be less than one false positive among the 363
most likely associations in the whole catalog.
\subsubsection{\label{sec:lowlat}Low-Latitude AGNs}
A simple extrapolation, based on the global density of 1LAC sources on the sky
and the solid angle subtended by the Galactic plane region ($|b| < 10\arcdeg$),
indicates that the LAT should be detecting $\sim$150 AGNs at low Galactic
latitudes. Diffuse radio emission, interloping Galactic point sources, and
heavy optical extinction make the low-latitude sky a difficult region for AGN
studies, and catalogues of AGNs and AGN candidates often avoid it partially or
entirely. However, we are able to make associations with 51 low-latitude AGNs;
these are presented in Table~\ref{tbl-lowlat}. Although the associations are
considered valid, these sources have, in general, been studied much less
uniformly and much less thoroughly than the high-latitude sources at virtually
all wavelengths, so we do not include them as part of the 1LAC in order to keep
them from skewing any of our analyses of the overall \mbox{$\gamma$-ray} AGN
population.
\subsubsection{\label{sec:affil}AGN ``Affiliations''}
For many of the 1FGL sources that are not formally associated with AGNs in the
1LAC, it is still possible to find nearby AGNs or AGN candidates for which
reliable association probabilities can not (yet) be computed but which show
some indication that they may be the correct counterpart. For example, using
the ASDC multifrequency tools\footnote{These tools are available from the ASDC
web site at the following URLs:\ http://tools.asdc.asi.it/ and
http://www.asdc.asi.it/ .}, we performed a visual inspection of the vicinity of
each unassociated 1FGL source. We considered objects inside the 95\% error
ellipse that showed hints of any blazar properties, such as coincident radio
and \mbox{X-ray} emission and indications in the literature of variability,
polarization, etc. For some sources, an optical spectrum, often from the
Sloan Digital Sky Survey \citep[SDSS;][]{sdss}, was available, allowing us to
classify them as BL~Lacs or FSRQs. These sources have been evaluated by a
method based on the known $\log N - \log S$ relationships of several types of
AGNs (FSRQs, BL~Lacs, flat-spectrum radio sources, etc.). We also used the
figure of merit methodology developed by \citet{srm03} and employed in the
assembly of the LBAS source list; although we have discovered some problems
with the calibration of the probabilities calculated by this approach, most of
the resulting AGN associations seem likely to be legitimate. Both of these
methods are being studied more carefully, but we list all of the unquantified
correspondences (which we call ``affiliations'') derived by them, along with
some of the properties of the affiliated AGNs, in Table~\ref{tbl-extras}. In
all, we find 109 AGN affiliations for 104 high-latitude LAT sources. We
expect that future refinements to our association methods will allow us to turn
many of these into true, quantitative associations.
\subsection{\label{sec:classif}Source Classification}
\subsubsection{\label{sec:optclass}Optical Classification}
We classify each AGN according to its optical spectrum where available.
Blazars are assigned optical classifications either as flat-spectrum radio
quasars (FSRQs) or BL~Lacertae objects (BL~Lacs) using the same scheme as for
CGRaBS. In particular, following \citet{stocke91}, \citet{up95}, and
\citet{marcha96}, we classify an object as a BL~Lac if the equivalent width
(EW) of the strongest optical emission line is $< 5$~\AA, the optical spectrum
shows a \ion{Ca}{2} H/K break ratio $C < 0.4$, and the wavelength coverage of
the spectrum satisfies
$(\lambda_\mathrm{max} - \lambda_\mathrm{min})/\lambda_\mathrm{max} > 1.7$ in
order to ensure that at least one strong emission line would have been detected
if it were present. Although other definitions of BL~Lac objects are sometimes
applied (e.g., using the EWs of the [\ion{O}{2}]~$\lambda$3727 and
[\ion{O}{3}]~$\lambda$5007 lines and/or different limits on $C$; see, e.g.,
\citealt{landt04}), the definition used here can be applied over a large
redshift range, with the caveat that high-redshift blazars may be classified as
BL~Lacs or FSRQs using different emission lines from those used for
low-redshift objects. The classification of higher-redshift sources will
preferentially use lines at shorter wavelengths (e.g., Ly$\alpha$~$\lambda$1216
and \ion{C}{4}~$\lambda$1549) than for low-redshift sources (e.g.,
\ion{Mg}{2}~$\lambda$2798 and H$\alpha$~$\lambda$6563). A few of the 1LAC
sources (e.g., radio galaxies such as Centaurus~A and NGC~1275) that are not
considered blazars are listed in Table 1 simply as ``AGNs''; these objects are
discussed individually in greater detail in the following section. Sources for
which no optical spectrum was available or for which the optical spectrum was
of insufficient quality to determine the optical classification are listed as
being of unknown type. Some redshifts and source types given in Table~1 may
differ from previously published results (generally from the NASA Extragalactic
Database\footnote{http://nedwww.ipac.caltech.edu/} [NED], SDSS, and/or
\citealt{vcv}). We thoroughly examined the data in the literature for accuracy
and compatibility with our classification scheme, and our reevaluations of
these results are reflected in Table~1.
A substantial number of the redshifts and optical classifications presented in
Table~1 are from our own optical follow-up campaigns. Most of these come from
spectroscopic observations conducted with the Marcario Low-Resolution
Spectrograph \citep{lrs} on the 9.2~m Hobby-Eberly Telescope (HET) at McDonald
Observatory. Other facilities that have contributed optical results include
the 3.6~m New Technology Telescope at La~Silla, the 5~m Hale Telescope at
Palomar, the 8.2~m Very Large Telescope at Paranal, and the 10~m Keck~I
Telescope at Mauna Kea. These spectra will be examined in greater detail in a
subsequent paper \citep{Shaw10}. A number of the HET redshifts were confirmed,
and new results were obtained, with the 3.6~m Telescopio Nazionale Galileo at
La~Palma. This work will also be detailed in an upcoming publication
\citep{tng}.
\subsubsection{\label{sec:sedclass}SED Classification}
The 1LAC blazars are also classified based on the peak of the synchrotron
component of the broadband SED. In most cases, it is not possible to build a
complete SED with simultaneous data, so we follow the scheme outlined by
\citet{SEDpaper}. This scheme is an extension to all blazars of a standard
classification system introduced by \citet{pg95} for BL~Lacs. We estimate the
frequency of the synchrotron peak, $\nu_\mathrm{peak}^\mathrm{S}$, using the
broadband spectral indices $\alpha_\mathrm{ro}$ (between 5~GHz and 5000~\AA)
and $\alpha_\mathrm{ox}$ (between 5000~\AA\ and 1~keV). This method uses an
analytic relationship calibrated on the $\nu_\mathrm{peak}^\mathrm{S}$ values
directly measured from the SEDs of the 48 sources studied by \citet{SEDpaper},
who confirm that these sources are representative of the global sample. This
relationship is as follows:
$$ \log \nu_\mathrm{peak}^\mathrm{S} = \left\{ \begin{array}{ll}
13.85 + 2.30X & \mbox{ if $X<0$ and $Y<0.3$,} \\
13.15 + 6.58Y & \mbox{ otherwise,}
\end{array} \right. $$
where $X = 0.565 - 1.433\alpha_\mathrm{ro} + 0.155\alpha_\mathrm{ox}$ and
$Y = 1.000 - 0.661\alpha_\mathrm{ro} - 0.339\alpha_\mathrm{ox}$. We use the
estimated value of $\nu_\mathrm{peak}^\mathrm{S}$ to classify the source as a
low-synchrotron-peaked, or LSP, blazar (for sources with
$\nu_\mathrm{peak}^\mathrm{S} < 10^{14}$~Hz), an
intermediate-synchrotron-peaked, or ISP, blazar (for
$10^{14}$~Hz~$< \nu_\mathrm{peak}^\mathrm{S} < 10^{15}$~Hz), or a
high-synchrotron-peaked, or HSP, blazar (for sources with
$\nu_\mathrm{peak}^\mathrm{S} > 10^{15}$~Hz). Figure~\ref{fig:alpha} displays
the locations of BL~Lacs from the clean sample in the ($\alpha_\mathrm{ox}$,
$\alpha_\mathrm{ro}$) plane. The data used for calculating the broadband
spectral indices were obtained mainly from {\it Swift}-XRT observations,
the NRAO VLA Sky Survey \citep[NVSS;][]{nvss}, and the USNO-B1.0 catalog
\citep{usno} and completed with data from Roma-BZCAT. It is evident from
Figure~\ref{fig:alpha} that \mbox{$\gamma$-ray}--selected BL~Lacs cover the
region in this plane typically occupied by blazars selected at other
frequencies, such as radio and \mbox{X-rays}. Figure~\ref{fig:Fr_Fx} shows the
soft \mbox{X-ray} flux vs.\ the radio flux density at 1.4~GHz for the blazars
in the clean sample. As in LBAS, we find that the broadband properties of the
1LAC blazars are consistent with those of the parent populations of BL~Lacs and
FSRQs.
We note that the SED classification method assumes that the optical and \mbox{X-ray}
fluxes come from the non-thermal emission without contamination from the disk
or accretion. For blazars in which the thermal components are non-negligible,
this method may lead to a significant overestimation of the position of
$\nu_\mathrm{peak}^\mathrm{S}$.
\section{\label{sec:census}1LAC Population Census}
Table~\ref{tbl-census} summarizes the breakdown of 1LAC sources by type for
the full catalog, the high-confidence sample, and the clean sample.
The fraction of blazars (BL~Lacs and FSRQs) that are BL~Lacs (51\% for the
high-confidence sample) is higher than in LBAS (39\%) and much higher than in
3EG (21\%). As discussed in Section~\ref{sec:flux}, for a given significance,
BL~Lacs can typically be detected by the LAT at a lower flux than can FSRQs,
which have softer spectra. The presence of one ISP-FSRQ and one HSP-FSRQ
warrants comment. The existence of higher-peaked FSRQs has not been
definitively proven yet. We stress that the SED classifications given here can
be improved with simultaneous data, and for blazars in which the thermal
components are non-negligible, it may lead to a significant overestimation of
$\nu_\mathrm{peak}^\mathrm{S}$. Moreover, the average error on the
\mbox{$\gamma$-ray} spectral indices is 0.14, so it can not be ruled out that the
candidate ISP- and HSP-FSRQs actually have lower synchrotron peak frequencies
and are consistent with the rest of the sample.
The distributions of synchrotron peak positions $\nu_\mathrm{peak}^\mathrm{S}$
are shown in Figure~\ref{fig:syn_hist}. The large fraction of HSPs in the 1LAC
sample contrasts sharply with the results from the 3EG sample, in which most
BL~Lacs ($>$80\%) were of the LSP type. This is also a consequence of the
dependence of the limiting flux on the spectral index (see
Section~\ref{sec:flux} and Figure~\ref{fig:index_flux}).
Figure~\ref{fig:sky_map} shows the sky locations of the sources in the clean
sample. It is clear that the distribution is not isotropic; there are more
sources in the northern galactic hemisphere than in the southern one. The
Galactic latitude distributions for FSRQs and BL~Lacs are shown in
Figure~\ref{fig:sky_lat} along with the corresponding distributions for the
Roma-BZCAT and 1FGL. The anisotropy, defined as $(N_+ - N_-)/(N_+ + N_-)$,
where $N_+$ and $N_-$ are the number of sources at $b > 10\arcdeg$ and
$b < -10\arcdeg$ respectively, amounts to $-4$\% ($N_+ = 135$, $N_- = 146$) for
high-confidence FSRQs and 18\% ($N_+ = 171$, $N_- = 120$) for high-confidence
BL~Lacs. For comparison, the sources in the 1FGL catalog are very evenly
distributed over the two hemispheres:\ 559 sources with $b > 10\arcdeg$ and 550
sources with $b < -10\arcdeg$.
\subsection{\label{sec:misaligned}Misaligned AGNs}
We have performed a search for possible associations between LAT detections and
non-blazar radio-loud sources (i.e., radio galaxies and steep-spectrum radio
quasars) using three main low-frequency surveys:\ the 3CR catalog \citep{3cr,
3cr2}; its revised version, the 3CRR catalog \citep{3crr}; and the Molonglo
Southern 4~Jy Sample \citep[MS4;][]{ms4}. These catalogs are flux limited
(3CR:\ 9~Jy; 3CRR:\ 10.9~Jy; MS4:\ 4~Jy) and cover large portions of the
northern (3CR, 3CRR) and southern (MS4) sky. In addition, because these
surveys were conducted at low frequencies (3CR and 3CRR:\ 178~MHz;
MS4:\ 408~MHz), they detect radio sources primarily in the relatively steep
part of the synchrotron emission spectrum, which is generally associated with
the extended lobes rather than the compact cores.
As a result, these catalogs are particularly appropriate for selecting AGNs
with jets misaligned with respect to the line of sight. The sources in these
surveys exhibit radio maps with resolved and possibly symmetrical structures
and steep radio spectra ($\alpha_\mathrm{r}>0.5$). At higher radio
frequencies, the distinction between blazars and misaligned radio sources can
be less sharp as the compact core emission (characterized by a flat or inverted
spectrum) can emerge and dominate the optically thin synchrotron emission from
extended regions. In addition to Cen~A \citep{LBAS}, NGC~1275
\citep{LBAS,ngc1275}, and M~87 \citep{M87}, this search has resulted in the
classification of five LAT-detected sources as misaligned AGNs:\ three radio
galaxies (NGC~1218~=~3C~78, PKS~0625$-$35, and NGC~6251) and two steep-spectrum
radio quasars (3C~207 and 3C~380). The properties of steep-spectrum radio
sources detected by the LAT will be examined further in a future publication
\citep{ssrs}.
Five other 1LAC sources have flat or nearly flat radio spectra between
$\sim$1~GHz and 8.4~GHz and strong emission lines in their optical spectra, but
these lines are narrow, in contrast to the broad emission features seen in
FSRQs. Under the standard AGN unification paradigm \citep{antonucci,up95},
this is interpreted as an indication that the jet axis is at a larger angle to
the line of sight than for a typical FSRQ. This causes the broad-line region
to be obscured by the dusty torus surrounding the central black hole, leaving
only narrow lines in the resulting spectrum, and such sources are commonly
described as narrow-line radio galaxies (NLRGs). The NLRGs appearing in the
1LAC are 4C~+15.05, PKS~1106+023, CGRaBS~J1330+5202, 4C~+15.54, and
CGRaBS~J2250$-$2806.
\subsection{\label{sec:sy}Radio-Quiet AGNs}
Radio-loud narrow-line Seyfert~1 galaxies have previously been shown to be
emitters of \mbox{$\gamma$-rays} (PMN J0948$+$0022, \citealt{pmnj0948}; B2~0321$+$33B,
PKS~1502+036, and PKS~2004$-$447, \citealt{rlnlsy1}). There is a tentative
indication that the LAT may have detected radio-quiet Seyfert galaxies or
radio-quiet quasars. Indeed, the 1LAC includes 10 radio-quiet AGN associations
for LAT sources (RX~J0008.0$+$1450, 1WGA~0405.6$-$1313,
\mbox{CXOMP}~J084045.2$+$131617,
\linebreak
\mbox{CXOMP}~J084054.3$+$131456, SDSS~J112042.47$+$071311.5,
\mbox{CXOCY}~J113008.8$-$144737,
\linebreak
SDSS~J125257.95$+$525925.6, SDSS~J143847.94$+$371342.7,
QSO~J150255.20$-$415430.2, and
\linebreak
SDSS~J155140.52$+$085226.1). In seven of these cases, however, the
LAT source is also associated with at least one radio-loud AGN (a blazar or a
radio galaxy) with higher probability, so the \mbox{$\gamma$-ray} emission is very
likely attributable to the radio-loud source. 1FGL~J1120.4$+$0710 is
associated both with SDSS~J112042.47$+$071311.5, a Seyfert~1 galaxy, and
CRATES~J1120$+$0704; although the latter is of unknown type, it is a radio-loud
source (probably a blazar) and is the likely source of the \mbox{$\gamma$-ray}
emission. The remaining two radio-quiet 1LAC associations, RX~J0008.0$+$1450
and QSO~J150255.20$-$415430.2, have association probabilities under 70\%. More
data and further study will be necessary to establish whether these objects are
indeed emitters of \mbox{$\gamma$-rays}.
\subsection{\label{sec:indiv}Notes on Individual Sources}
\noindent {\bf 1FGL~J0047.3$-$2512:} This source is associated with NGC~253,
a starburst galaxy previously detected by the LAT \citep{starbursts}. It has
for some time been designated a Seyfert galaxy in the V{\'e}ron catalog (most
recently in \citealt{vcv}), but this classification is questionable
\citep[see, e.g.,][]{foerster}. \citet{engelbracht} determine that the
LINER-like characteristics of NGC~253 can be attributed entirely to starburst
activity.
\noindent {\bf 1FGL~J0339.1$-$1734:} This source is associated with
PKS~0336-177. The optical spectrum from the final release of 6dFGS
\citep{6dF04,6dF09} shows no strong emission lines, but the \ion{Ca}{2} H/K
break is too large to satisfy the criterion $C < 0.4$ for classification as a
BL~Lac.
\noindent {\bf 1FGL~J0405.6$-$1309:} This source is associated with
PKS~0403$-$13 (an FSRQ) and with
\linebreak
1WGA~J0405.6$-$1313. \citet{rixos} classify the latter simply as an ``AGN''
but provide the redshift ($z = 0.226$) based on the detection of three emission
lines, including broad H$\beta$. The source is radio-faint (undetected in
NVSS) and likely a Seyfert galaxy.
\noindent {\bf 1FGL~J0627.3$-$3530:} This source is associated with
PKS~0625$-$35, which shows an FR-I radio morphology. However, \citet{wills2}
suggest a BL~Lac classification for this object. They also note that, in
contrast with the other genuine FR-I sources in their sample, an adequate fit
to the optical continuum of PKS~0625$-$35 required a nonthermal power law in
addition to emission from the host galaxy.
\noindent {\bf 1FGL~J0645.5$+$6033:} This source is associated with
BZU~J0645+6024, a source with broad emission lines but with a steep radio
spectrum.
\noindent {\bf 1FGL~J0923.2$+$4121:} This source is associated with
B3~0920+416, which \citet{falco} characterize as a late-type galaxy.
\noindent {\bf 1FGL~J0956.5$+$6938:} This source is associated with M~82, also
known as NGC~3034 and 3C~231, a starburst galaxy previously detected by the LAT
\citep{starbursts}. There are indications that it may host a weak AGN
\citep{wills}, though this is not confirmed.
\noindent {\bf 1FGL~J1202.9$+$6032:} This source is associated with
CRATES~J1203+6031. \citet{falco} describes it as an early-type galaxy; it is
listed in NED as a LINER.
\noindent {\bf 1FGL~J1305.4$-$4928:} This source is associated with NGC~4945,
which is both a Seyfert~2 galaxy and a starburst galaxy. It is the third
starburst galaxy seen by the LAT, and its detection is reported here for the
first time. A more careful analysis will be required to determine whether
the \mbox{$\gamma$-ray} emission comes from the starburst activity or the AGN
or both.
\noindent {\bf 1FGL~J1307.0$-$4030:} This source is associated with
\mbox{ESO~323-G77}, a nearby ($z=0.015$) Seyfert 1.2 galaxy detected also by
{\it Swift}-BAT (see, e.g., \citealt{ajello09} and Section~\ref{sec:hardX}).
It exhibits a hard \mbox{X-ray} spectrum typical of radio-quiet AGN. There is
little indication that the source is a starburst galaxy. If the
\mbox{$\gamma$-ray} emission from 1FGL~J1307.0$-$4030 truly comes from
\mbox{ESO~323-G77}, then its origin might be connected to the central AGN; this
hypothesis is being examined more closely.
\noindent {\bf 1FGL~J1641.0$+$1143:} This source is associated with
CRATES~J1640+1144, described by \citet{mitton} only as a ``galaxy.'' Since
the authors do classify other sources as BL~Lacs, it seems likely that
CRATES~J1640+1144 is not a BL~Lac.
\noindent {\bf 1FGL~J1647.4$+$4948:} This source is associated with
CGRaBS~J1647+4950, which \citet{falco} characterize as a late-type galaxy.
\noindent {\bf 1FGL~J1724.0$+$4002:} This source is associated with B2~1722+40.
The optical spectrum from \citet{vermeulen} shows narrow forbidden emission
lines and absorption features from the host galaxy.
\noindent {\bf 1FGL~J1756.6$+$5524:} This source is associated with
BZB~J1756$+$5522 (a BL~Lac) and with CRATES~J1757+5523. The optical spectrum
of the latter from \citet{caccianiga} shows strong absorption features from the
host galaxy and a large \ion{Ca}{2} H/K break. The authors describe it as a
passive elliptical galaxy.
\noindent {\bf 1FGL~J2008.6$-$0419:} This source is associated with 3C~407, a
source with broad emission lines but with a fairly steep radio spectrum.
\noindent {\bf 1FGL~J2038.1$+$6552:} This source is associated with NGC~6951,
which has been classified as a Seyfert~2 galaxy and a LINER. It also has a
circumnuclear ring of starburst activity.
\noindent {\bf 1FGL~J2204.6$+$0442:} This source is associated with
4C~$+$04.77, which is often called a BL~Lac in the literature. A spectrum
from \citet{vcv1993}, however, shows that the object exhibits strong broad
emission lines characteristic of Seyfert~1 galaxies.
\section{Properties of the 1LAC Sources}
\subsection{\label{sec:flux}Flux and Photon Spectral Index Distributions}
As demonstrated in \cite{Spe09}, many bright LAT blazars (most notably FSRQs
and some LSP-BL~Lacs) exhibit breaks in their \mbox{$\gamma$-ray} spectra. Despite
this caveat, determining a photon spectral index fitted over the whole band is
useful as it does reflect the source's spectral hardness and can be obtained
with reasonable accuracy even for fairly faint sources.
The 11-month average photon spectral index is plotted against the flux
($E>100$~MeV) estimated from a power law fit in Figure~\ref{fig:index_flux} for
blazars in the clean sample. As the figure shows, the limiting flux
corresponding to $TS=25$ depends fairly strongly on the photon spectral index;
the solid curve corresponds to a simple analytical estimate \citep{Lot07}.
This effect is discussed in greater detail in \citet{Catalog}. FSRQs (red
circles in Figure~\ref{fig:index_flux}) mostly cluster in the soft spectral
index region, while BL~Lacs (blue circles) primarily occupy the hard spectral
index region, confirming the trend seen for the LBAS sources \citep{LBAS}.
This implies that the limiting flux is different for FSRQs and BL~Lacs. The
respective flux distributions are compared in Figure~\ref{fig:flux}. The mean
fluxes are $8.5 \times 10^{-8}$ \pflux{} and $2.9 \times 10^{-8}$ \pflux{} for
FSRQs and BL~Lacs respectively. The faintest BL~Lacs are about a factor of 3
fainter than the faintest FSRQs.
The overall flux distribution (of FSRQs and BL~Lacs combined) is compared to
that measured by EGRET in Figure~\ref{fig:flux_lat_egret}a. The high-flux ends
of the two distributions are in reasonable agreement. The peak flux
distributions (maximum flux in a $\approx$15-day viewing period for EGRET vs.\
maximum monthly flux for the LAT) are compared in
Figure~\ref{fig:flux_lat_egret}b. The peak fluxes observed by EGRET are
substantially higher than those observed for the brightest 1LAC sources; this
is illustrated in Figures~\ref{fig:flux_lat_egret}c and
\ref{fig:flux_lat_egret}d, which show the peak flux vs.\ mean flux and the peak
flux/mean flux ratio respectively. The effect of using two different time
binnings to determine the peak flux (15~days for EGRET vs.\ 1~month for the
LAT) has been studied for the bright LBAS sources; the 15-day peak flux is only
13\% higher, on average, than the 1-month peak flux. Hence, the main reason
for the different peak fluxes observed by EGRET and the LAT is probably the
different time spans over which the observations were conducted (4.5~years for
EGRET vs.\ 11~months for the LAT), enabling the sources to explore a wider
range of different states during the EGRET era. Note also that the EGRET
observations were often triggered by flaring alerts provided by other
facilities, causing a bias that is not present for the LAT.
The photon spectral index distributions for the clean sample, shown in
Figure~\ref{fig:index}, confirm the trend found for the LBAS sources
\citep{LBAS} and already mentioned above, with a clear difference between FSRQs
and BL~Lacs. Only 7\% of FSRQs (17/231) have a photon spectral index
$\Gamma < 2.2$, and only two FSRQs have $\Gamma < 2$ (with uncertainties of
$\sim$0.2). Thus, it is quite clearly established that LAT-detected FSRQs
exhibit soft \mbox{$\gamma$-ray} spectra. The 1LAC BL~Lac distribution is broader
than that observed for LBAS; it shows a larger overlap with the FSRQ
distribution. In \cite{Spe09}, the BL~Lacs with $\Gamma > 2.2$ were found to
be mostly LSP sources. This trend is also confirmed for the 1LAC sources, as
illustrated in Figure~\ref{fig:index_log_nu_syn}, which shows $\Gamma$ vs.
$\nu_\mathrm{peak}^\mathrm{S}$. \cite{Spe09} also point out that the BL~Lac
photon spectral index distribution suffers from a bias since the
high-spectral-index (i.e., soft) end is cut off somewhat due to the TS
threshold, as seen in Figure~\ref{fig:index_flux}. Figure~\ref{fig:index_c}
shows the photon spectral index distributions for the different blazar classes
for sources with $F$[$E>100$~MeV] $>3\times10^{-8}$ \pflux{}, above which the
sample is essentially complete (see Figure~\ref{fig:index_flux}). The
completeness of the sample will be examined more carefully in an upcoming paper
on the LAT blazar population \citep{blazpop}.
The presence of a few very hard ($\Gamma < 1.7$) sources warrants comment.
These are fairly low-significance sources ($TS<150$), as is evident from
Figure~\ref{fig:TS_index}. Note that typical statistical error bars reach
0.1--0.2 for all sources with TS in that range. Figure~\ref{fig:index} also
displays the photon spectral index distribution for sources with unknown types.
This distribution overlaps with both those of FSRQs and BL~Lacs, which seems to
rule out that these sources belong preferentially to either class.
Since the flux of each source is evaluated in five different bands, the
departure of the spectrum from a power-law (PL) shape can be estimated via a
simple $\chi^2$ test, referred to as the curvature index \citep{Catalog}.
Figure~\ref{fig:chi2_flux} displays the curvature index vs.\ total flux for
FSRQs and for the three BL~Lac subclasses. The trend corroborates the findings
reported in \cite{Spe09}, namely, that the spectra of bright FSRQs and
LSP-BL~Lacs show strong departures from a PL shape while those of HSP-BL~Lacs
are essentially compatible with PLs. Note that due to their harder spectra,
any deviation from a PL behavior would be more easily visible for HSP-BL~Lacs.
Because of larger statistical uncertainties, no conclusion can be drawn from
the $\chi^2$ test for fainter sources (a similar limitation affects the
variability index and is described below).
\subsection{Variability}
One of the defining characteristics of AGNs is their variability, measured at
all time scales and at all wavelengths. In the LBAS sample, 46 sources (35
FSRQs, 10 BL~Lacs, and one blazar of unknown type) were flagged as variable
based on the results of a $\chi^2$ test applied to weekly light curves covering
the first three months of the LAT sky survey. Recently, \citet{LBASvar}
detected variability in 64 LBAS sources by analyzing weekly light curves over a
period of 11 months.
The electronic version of the 1FGL catalog provides the light curves for all
sources over 11 time intervals of 30.37 days each in the energy range
100~MeV--100~GeV. The procedure used to make these light curves is described
in \citet{Catalog}. Our analysis of the light curves of the sources in the
clean sample shows that only six sources (3C~454.3, PKS~1510$-$08, 3C~279,
PKS~1502$+$106, 3C~273, and PKS~0235$+$164) have a maximum monthy flux
($E > 100$~MeV) greater than $10^{-6}$ \pflux{} (see also
Figure~\ref{fig:flux_lat_egret}b). The number of sources with peak flux values
above $10^{-6}$ \pflux{} increases if we consider shorter time intervals, as
illustrated by the 41 Astronomer's Telegrams issued by the LAT collaboration
(see below), mostly related to sources that showed short flares that reached
this flux level on the time scale of days.
The 1FGL catalog also lists a variability index $V$, obtained using a simple
$\chi^2$ test, that can be used to determine the probability that a source is
variable. Sources for which $V > 23.21$ have a 99\% probability (for 10
degrees of freedom) of being variable. Figure~\ref{fig:varind} shows the
distribution of the variability index for blazars (BL~Lacs and FSRQs) in the
clean sample. 189 blazars in the clean sample are found to be variable (to the
right of the vertical line in Figure~\ref{fig:varind}); they comprise 129 FSRQs
(68\%), 46 BL~Lacs (24\%), five AGNs of other types (NGC~1275, B2~0321$+$33B,
4C~$+$15.54, B2~1722$+$40, and CGRaBS~J2250$-$2806), and nine blazars of
unknown type. The 46 variable BL~Lacs include 24 LSPs, nine ISPs, 10 HSPs, and
three BL~Lacs (CRATES~J0058$+$3311, CGRaBS~J0211$+$1051, and
CRATES~J1303$+$2433) for which there are not enough multiwavelength data to
permit an SED classification. Figure~\ref{fig:varph} shows the distribution of
the photon spectral index for the variable blazars. This distribution is
dominated by sources with $\Gamma > 2.2$. These are quite bright sources
observed at energies higher than the peak energies of their SEDs, where the
amplitude of the variability is generally larger.
It is important to bear in mind that for a source to be labeled as variable on
the basis of its variability index, it must be both intrinsically variable and
sufficiently bright. Larger statistical flux uncertainties obtained for
fainter sources lead to a reduction of the variability index for a given
fractional flux variation. To illustrate this effect, the variability index is
plotted against the flux in Figure~\ref{fig:varind_flux} for FSRQs and BL~Lacs.
The curves display the evolution of the variability index for two sample
sources (the FSRQ 3C~454.3 and the BL~Lac object AO~0235$+$164) that would be
observed for the same temporal variation but lower mean fluxes. As a result of
this effect, fainter sources appear less variable than brighter sources simply
because we can not measure their variability as well. This may also explain
why only $\sim$17\% of the 1LAC BL~Lacs, which are generally fainter and harder
than FSRQs, are found to be variable by this criterion. The histograms in
Figure~\ref{fig:varind} are artificially broadened by the flux dependence of
the variability index shown in Figure~\ref{fig:varind_flux}. Allowing for the
behavior shown by the the trends for 3C~454.3 and AO~0235$+$164, there is no
strong evidence for a significant difference in the intrinsic variability of
FSRQs and BL~Lacs.
The LAT Collaboration routinely issues Astronomer's Telegrams (ATels) to alert
the community to transient sources or sources that are exhibiting flaring
states in order to encourage simultaneous multiwavelength follow-up. Through
2010 Jan 20, 43 1LAC sources (and one at low latitude) have been the subjects
of ATels from the LAT team. They comprise 32 FSRQs (with redshifts as high as
$z = 2.534$, for B3~1343+451), 10 BL~Lacs, and one AGN of another type
(CGRaBS~J2250$-$2806, a narrow-line radio galaxy).
\subsection{\label{sec:z}Redshift Distributions}
The redshift distributions for FSRQs and BL~Lacs in the clean sample are
presented in Figure~\ref{fig:redshift} along with the corresponding ones for
the {\it WMAP} blazars \citep{Hin07,Gio09}, which constitute a flux-limited
all-sky sample above 1~Jy in the central {\it WMAP} observing band (Q band,
41~GHz). Note that only 121 out of 291 (42\%) high-confidence 1LAC BL~Lacs
have measured redshifts. This is notably worse than for LBAS, in which 29 out
of 42 (69\%) BL~Lacs had measured redshifts. The redshift distributions of
\mbox{$\gamma$-ray}--detected blazars---both for FSRQs and for BL~Lacs---are
fairly similar to those for {\it WMAP}, peaking around $z=1$ for FSRQs and at a
lower redshift for BL~Lacs (the lowest-redshift 1LAC BL~Lac has $z = 0.030$).
The highest redshift for a high-confidence 1LAC FSRQ is $z=3.10$.
The photon spectral index is plotted against redshift in
Figure~\ref{fig:index_redshift}. For FSRQs, no significant evolution is
visible. This behavior is compatible with what was previously observed for
LBAS. The attenuation effect of the extragalactic background light (EBL) would
tend to introduce spurious evidence of evolution \citep{Chen04}, but the soft
spectra of FSRQs and the common presence of spectral breaks at a few GeV
\citep{Spe09} both minimize this effect. A stronger evolution is seen for
BL~Lacs:\ hard sources are mostly located at low redshifts, while most
high-redshift sources are softer than average (though it is important to bear
in mind that most BL~Lacs do not have measured redshifts). This trend was less
clear for the LBAS BL~Lacs \citep{LBAS} due to lower statistics.
Figure~\ref{fig:bll_z} displays the photon spectral index distributions for
sources with $z<0.5$ (top) and $z>0.5$ (middle), which are clearly different.
This provides some insight into the properties of BL~Lacs without measured
redshifts. The photon spectral index distributions of BL~Lacs with measured
redshifts and with unknown redshifts are shown in the bottom panel of
Figure~\ref{fig:bll_z}. The distribution of BL~Lacs with unknown redshifts
includes notably fewer hard sources than that for BL~Lacs with known redshifts.
A Kolmogorov-Smirnov (K-S) test yields a probability of $6 \times 10^{-3}$ that
the two distributions are drawn from the same underlying population. The
similarity of the distribution of BL~Lacs with unknown redshifts to that of
BL~Lacs with $z > 0.5$ (K-S probability of 0.54 that the two distributions are
drawn from the same underlying population) supports the idea of some bias
toward higher redshifts for this class of objects.
\subsection{Luminosity Distributions}
The \mbox{$\gamma$-ray} luminosity $L_\gamma$ is calculated as in \cite{Ghi09}:\
$$
L_\gamma = 4 \pi d_L^2 \frac{S (E_1, E_2)}{(1+z)^{2-\Gamma}}
$$
where $d_L$ is the luminosity distance, $\Gamma$ is the photon spectral index,
$S$ is the energy flux, and $E_1$ and $E_2$ are the lower and upper energy
bounds (taken here to be 100~MeV and 100~GeV) respectively. Implicit in this
derivation is the assumption that $E_2\gg E_1$, which is satisfied for our
calculations. Figure~\ref{fig:Lum_redshift} shows $L_\gamma$ plotted against
the source redshift. The curves represent approximate instrumental limits
calculated for two photon indices, $\Gamma = 1.8$ and $\Gamma = 2.2$. The
low-redshift FSRQs with $L_\gamma > 10^{48}$ erg s$^{-1}$ could potentially
still be detected at redshifts $z > 3.1$.
Figure~\ref{fig:index_lum} shows the photon spectral index plotted against the
\mbox{$\gamma$-ray} luminosity. The Pearson correlation coefficient for the
two parameters is 0.17. It is important to bear in mind two issues when
interpreting this correlation:\ the higher flux limit for soft sources (see
Figures~\ref{fig:index_flux} and \ref{fig:flux}) and the difference in redshift
distributions between FSRQs and BL~Lacs (see Figure~\ref{fig:redshift}). Given
their relative softness (and thus, high flux limit), FSRQs are excluded from
the soft-faint region. BL~Lacs are also partly excluded because of their
relatively low fluxes. Luminosity limits calculated for different redshifts
(with the same approach as for the flux limit shown in
Figure~\ref{fig:index_flux}) are displayed for reference in
Figure~\ref{fig:index_lum} and illustrate somewhat the effect of the Malmquist
bias\footnote{In a flux-limited sample, sources located at larger distances
appear more luminous than closer ones since fainter sources are below the
detection threshold \citep{Mal20}.}.
\section{Multiwavelength Properties of 1LAC Sources}
\subsection{Sources Detected at TeV Energies}
Over the last two decades, ground-based \mbox{$\gamma$-ray} instruments operating in
the ``TeV'' or very-high-energy (VHE; $E\gtrsim100$~GeV) regime have detected
32 AGNs, with the pace of discovery increasing significantly\footnote{TeVCat
(\url{http://tevcat.uchicago.edu/}) presents an up-to-date catalog of TeV
sources.} as the latest generation of instruments---CANGAROO, H.~E.~S.~S.,
MAGIC and VERITAS---has been commissioned. Of these 32 TeV AGNs, the majority
are BL~Lacs (23 HSPs, three ISPs, and two LSPs), with the remainder comprising
one FSRQ, two FR-I radio galaxies, and one AGN of unknown type. A detailed
analysis of the TeV AGNs detected with {\it Fermi} during the first 5.5~months
of operation is given by \citet[][and see references to the TeV detections
therein]{FermiTeV}. Five new TeV AGNs have since been detected
(1ES~0414$+$009, \citealt{ATelHofmann}; PKS~0447$-$437, \citealt{ATelRaue};
RBS~0413, \citealt{ATelOng1}; 1ES~0502$+$675, \citealt{ATelOng2};
VER~J0521$+$211, \citealt{ATelOng3}). Two of these detections and one previous
detection \citep[PKS~1424$+$240;][]{pks1424} were motivated directly by the
detection of GeV emission with {\it Fermi}.
Of the TeV AGNs, the 28 listed in Table~\ref{tbl-tev} are detected by
\textit{Fermi} as 1FGL sources, with a mean photon spectral index of
$\langle\Gamma_\mathrm{GeV}\rangle=2.02\pm0.01$. Taking only the subsample of
25 GeV--TeV BL~Lacs, the mean index is
$\langle\Gamma_\mathrm{GeV}\rangle=1.92\pm0.01$ ($\sigma_\Gamma=0.26$),
compared with $2.07\pm0.01$ ($\sigma_\Gamma=0.28$) for the larger sample of
{\it Fermi} BL~Lacs (see Figure~\ref{fig:index}b), illustrating that the TeV
sources are among the hardest BL~Lacs in the GeV regime. The measured spectra
of the majority of the GeV--TeV BL~Lacs are well described by power laws in
both regimes. For many of the sources, the photon spectral index of the GeV
emission differs significantly from that of the TeV emission, which may be an
indication of the presence of a break in the \mbox{$\gamma$-ray} spectrum between the
two regimes. However, most of the TeV spectra have not been measured
simultaneously with the GeV spectra, and caution is advised when comparing the
spectra in detail. The largest such break, consistent with
$\Delta\Gamma \equiv \Gamma_\mathrm{TeV}-\Gamma_\mathrm{GeV}\sim2$, is evident
in the spectra of 1ES~1101$+$496, H~1426$+$428 and PG~1553$+$113. By contrast,
the spectra of the nearby radio galaxies M~87 and Cen~A show no evidence of a
spectral break. The mean break index is $\langle\Delta\Gamma\rangle=1.3$. The
values of $\Delta\Gamma$ for the GeV--TeV sources are shown in
Figure~\ref{fig:break}, plotted against the redshift of the source. Among the
possible explanations for the apparent deficit of sources with small
$\Delta\Gamma$ at high redshift is the effects of pair production with the EBL,
which is expected to introduce a redshift-dependent steepening into the TeV
spectra of extragalactic objects.
The four TeV AGNs not detected thus far by {\it Fermi} are RGB~J0152$+$017,
1ES~0229$+$200, 1ES~0347$-$121 and PKS~0548$-$322, all HSP-BL~Lacs.
Extrapolating the measured power-law TeV spectra from all the TeV AGNs down to
200~GeV, it is evident that these four have among the smallest fluxes of all
TeV AGN at this energy.
\subsection{Sources Detected Previously at GeV Energies}
Of the 709 sources in the 1LAC, 114 sources were included in the BSL and were
thus significantly detected (at $10\sigma$ or greater) during the first three
months of LAT observations. All of those sources were previously associated
with AGNs as part of LBAS. Three low-confidence LBAS sources
(0FGL~J0909.7$+$0145, 0FGL~J1248.7$+$5811, and 0FGL~J1641.4$+$3939) are not
confirmed in the 1LAC sample.
Ten years after EGRET, it is interesting to look at the fraction of the AGNs
that were active in the EGRET era and are detected again by the LAT with a
comparable flux. We consider two sources to be ``positionally coincident''
when the separation between their positions is less than the quadratic sum of
their 95\% error radii. In the 1LAC sample, 63 AGNs are positionally
coincident with 3EG sources. Of these, the 3EG catalog lists 51 sources as AGN
identifications and four as AGN associations. The 3EG catalog listed a total
of 62 sources as AGN identifications, so there are 11 identified EGRET AGN that
are not positionally coincident with 1LAC sources. 45 sources from the 1LAC
have positions compatible with the revised EGRET catalog \citep[EGR;][]{EGR},
while 22 sources are positionally coincident with sources from the high-energy
EGRET catalog \citep[GEV;][]{Lamb1997}. In all, 75 1LAC sources have
coincident detections from one or more of these three EGRET catalogs.
11 of the sources in this catalog are included in the first year {\it AGILE}
catalog \citep[1AGL;][]{AGILEcatalog}. All 11 are identified as blazars by
{\it AGILE}, and indeed, these 11 comprise the total sample of blazars in the
1AGL catalog. Only two of these {\it AGILE}-detected sources do not have
similar EGRET detections. In all, 77 1LAC sources have been cataloged by other
GeV instruments.
These 77 sources are listed in Table \ref{tbl-fmrdetect} along with the mean
fluxes and photon indices measured by the LAT and by EGRET. These AGNs are
composed of 49 FSRQs, 21 BL~Lacs, four AGNs of other types, and three AGNs of
unknown types.
During its 4.5-year mission, EGRET found few sources with flux
(${E > 100}$~MeV) less than ${10 \times 10^{-8}}$~\pflux{}. A large number
of the 1LAC sources have fluxes well below this value; such sources would not
have been visible to EGRET. Some sources, such as 1FGL~J0428.6$-$3756
(associated with PKS 0426$-$380) and 1FGL~J2229.7$-$0832
(associated with PKS 2227$-$08), have flux values well above the EGRET
threshold but were not seen by EGRET and yet are not noted as being variable in
the 1FGL data. These examples continue to demonstrate that long-duration
variability is evident in AGN \mbox{$\gamma$-ray} flux.
All of the EGRET sources seen at $10\sigma$ significance and associated with
flaring blazars have been detected by the LAT and appear with blazar
associations.
\subsection{\label{sec:hardX}Sources Detected in the Hard X-ray Band}
In recent years, a new generation of hard \mbox{X-ray} telescopes has drastically
improved our knowledge of the source populations in this energy band. The
{\it Swift}-BAT and {\it INTEGRAL}-IBIS instruments have been conducting
surveys at hard \mbox{X-ray} energies; the most recent catalogs are the fourth IBIS
catalog \citep{bird} of 723 sources detected at 17-100 keV and the 54-month
Palermo BAT catalog \citep{swiftCat} of 1049 sources\footnote{\citet{swiftCat}
describes the 39-month catalog; the 54-month catalog is available at:\
http://www.ifc.inaf.it/cgi-bin/INAF/pub.cgi?href=activities/bat/index.html .}
detected at 14-150 keV. These two telescopes perform in a complementary
fashion; while IBIS is the more sensitive instrument with better angular
resolution, it has a smaller field of view and has concentrated exposure in the
Galactic plane. Conversely, the BAT instrument has lower instantaneous
sensitivity but has a much larger field of view and a much more uniform
exposure across the whole sky. An initial comparison between the LBAS sources
and the fourth IBIS catalog by \citet{ubertini09} showed that only a small
subset of the $>$250 {\it INTEGRAL} AGNs, of which 19 are blazars, was detected
by the LAT.
Restricting the catalogs to $|b|>10\arcdeg$, there are 291 IBIS sources and
736 BAT sources. If we consider all of the 1LAC sources with associated
counterparts that fall within the error circles of the hard \mbox{X-ray} catalogs, we
find that 50 of the 1LAC sources can be associated with known hard \mbox{X-ray}
sources. Of these 50 sources, one appears only in the fourth IBIS catalog,
16 appear in both the fourth IBIS catalog and the 54-month Palermo BAT catalog,
and the remaining 35 are detected only in the 54-month Palermo BAT catalog.
This is still a very small subset of AGNs despite the depth of the LAT
catalog and the sky coverage of the BAT catalog.
Table~\ref{tab:hardXray} lists the high-confidence 1LAC AGNs detected in hard
\mbox{X-rays}. This sample contains 27 FSRQs, 16 BL~Lacs, and seven AGNs of
other types. Most of the AGNs (54\%) detected in hard \mbox{X-rays} are FSRQs;
this is expected since the external Compton peak of these high-luminosity
objects reaches a maximum at MeV energies and declines in the hard
\mbox{X-rays}, maintaining a substantial fraction of the emitting power. The
difference in the spectral shapes of BL~Lacs and FSRQs observed in hard
\mbox{X-rays} indicates that this energy range probes the high-energy tail of
the synchrotron peak in BL~Lacs and the ascending part of the Compton peak in
FSRQs.
Finally, there is evidence that the hard \mbox{X-ray} emission from most bright LAT
blazars is, apparently, missed in spite of the sub-milliCrab sensitivity
reached with {\it INTEGRAL} and {\it Swift} in the deepest fields.
\subsection{Radio Properties}
The high-latitude 1FGL sources associated with AGNs and presented here in the
1LAC are all radio sources at some level. The large number of sources found in
CRATES can, by definition, be characterized as relatively bright, flat-spectrum
radio sources. However, since a significant number of associations are found
from other catalogs, it is interesting to contemplate the distributions of the
flux densities, luminosities, and spectral indices.
In Figure~\ref{fig:radioflux}, we plot the distribution of the radio flux
density at 8.4~GHz for the clean sample and separately for the FSRQs and the
BL~Lacs in this sample. The overall distribution, including also the blazars
of unknown type and the other AGNs, has one main peak (corresponding to
$\langle S_\mathrm{8\,GHz,\,1LAC}\rangle \sim 900$~mJy). However, the FSRQs
are on average brighter than the BL~Lacs, with
$\langle S_\mathrm{8\,GHz,\,FSRQ}\rangle \sim 1200$~mJy and
$\langle S_\mathrm{8\,GHz,\,BLL}\rangle \sim 400$~mJy.
In Figure~\ref{fig:radiolum}, we plot the corresponding radio luminosities,
calculated and $K$-corrected for the sources with measured redshifts. The
distributions for FSRQ and BL~Lacs are clearly different, with the FSRQs
concentrated at higher radio luminosities
($\log$~$(L_{r,\,\mathrm{FSRQ}}$~[erg~s$^{-1}$]$)=44.1\pm0.7$~erg~s$^{-1}$) and
the BL~Lacs distributed over a broad range
($\log$~$L_{r,\,\mathrm{BLL}}$~[erg~s$^{-1}$]$)=42.3\pm1.1$~erg~s$^{-1}$). This is
similar to the distribution found in the LBAS, and the only sources with radio
luminosities below $L_r = 10^{40}$~erg~s$^{-1}$ are a few nearby AGNs (Cen~A,
NGC~253, and NGC~4945).
Finally, in Figure~\ref{fig:radiospecind}, we plot the spectral index
distribution, calculated between either 1.4~GHz from NVSS for sources with
$\delta > -40\arcdeg$ or 0.84~GHz from the Sydney University Molonglo Sky
Survey \citep[SUMSS;][]{sumss} for sources with $\delta < -40\arcdeg$ and
8.4~GHz from CRATES (or other measurements from NED). The overall distribution
is consistent with a flat spectral index ($\alpha=0.08 \pm 0.32$). No
difference is found between FSRQs and BL~Lacs. Interestingly, a tail of
sources with steeper spectral index is found, and it is mostly composed of
non-blazars, such as radio galaxies. This suggests that \mbox{$\gamma$-ray}
sources can also be associated with radio sources that are less closely aligned
with the line of sight than blazars (see Section~\ref{sec:misaligned}).
Since the different classes have different properties in the \mbox{$\gamma$-ray} band,
the relationship between the radio emission and the high-energy emission merits
a deeper discussion. A paper dedicated to this subject is in preparation, with
a focus on the correlation between radio and \mbox{$\gamma$-ray} fluxes.
\section{Discussion}
Of the 1043 high-latitude ($|b| > 10\arcdeg$) \mbox{$\gamma$-ray} sources in
the 1FGL catalog, 671 are associated with 709 AGNs; these constitute the 1LAC.
The 663 high-confidence AGN sources detected by {\it Fermi} in 11~months of
data can be compared with the 66 high-confidence AGNs detected by EGRET at
$>5 \sigma$ significance over its lifetime \citep{3EGcatalog}. We estimate
that a large fraction of the other 372 high-latitude sources are likely AGNs on
the basis of their spectral and variability properties; indeed, we have found
plausible AGN candidates for $\sim$100 of them (see Section~\ref{sec:affil}.
In addition, $\sim$20 known millisecond pulsars, and possibly a comparable
number of yet unidentified millisecond pulsars, are among the remaining
high-latitude 1FGL sources.
The 1LAC represents a $\sim$5-fold increase in the number of sources associated
with \mbox{$\gamma$-ray} blazars over previously published source lists (in
particular, the 106 high-confidence blazars in the LBAS catalog). It is likely
that other blazars remain unidentified, as implied by the observed north-south
anisotropy of the associated sources, which reflects the incompleteness of the
counterpart AGN catalogs. This is partly supported by the similarity of the
\mbox{$\gamma$-ray} properties, including flux and photon spectral index, of
the unidentified sources to those of the 1LAC blazars \citep{unID}. There are
also indications from our early radio follow-up of the unidentified sources
that the LAT is detecting blazars that are fainter in the radio than those
appearing in the catalogs of flat-spectrum radio sources from which we draw
candidate associations.
The overall properties of the 1LAC sources are generally consistent with those
of the LBAS \citep{LBAS,Spe09,SEDpaper}, including observations of:
\begin{enumerate}
\item A small number of non-blazar sources.
The misaligned AGNs include six radio galaxies, of which only three (NGC~1275,
Cen~A, and M~87) were previously reported. Five of the six radio galaxies are
low-power FR-I galaxies associated with LAT sources with very high
probabilities ($P > 98$\%). The sixth, NGC~6251, is a moderately powerful
radio galaxy \citep{Perley84} and has a lower association probability. The
misaligned FR-II radio galaxy 3C~111 is associated with a LAT source (see
Table~\ref{tbl-lowlat}), but it is not a member of the 1LAC since it resides at
low Galactic latitude. In addition, two steep-spectrum radio quasars and five
NLRGs are among the other non-blazar sources, though many of these have
prominent radio cores and some lack extended radio emission.
\item Redshift distributions peaking at $z \approx 1$ for 1LAC FSRQs and at low
redshift for 1LAC BL~Lacs with known redshifts.
The maximum redshift of the high-confidence 1LAC FSRQs is $z=3.10$, larger than
the maximum redshift ($z=2.286$) for an EGRET blazar. For comparison, the
maximum redshift in both CGRaBS and Roma-BZCAT is for an FSRQ with $z > 5$.
The redshift distribution for BL~Lacs in the 1LAC with known redshift displays
a broad low-redshift peak from the lowest known redshift (for Mkn~421) at
$z=0.030$ to $z \approx 0.4$. The good agreement with the {\it WMAP} redshift
distribution for FSRQs, with a comparable number of sources, supports the idea
that similar populations are being sampled by {\it WMAP} and the LAT. Both the
FSRQ and BL~Lac redshift distributions are compatible with the corresponding
distributions from LBAS; the K-S probabilities that they derive from the same
parent distributions are 0.99 and 0.89 for FSRQs and BL~Lacs respectively.
No strong evidence for a new population of misaligned FSRQs emerging at lower
redshifts is found. Other than NGC~6251, just mentioned, notably absent are
detections, at least at the 1LAC significance level of $TS > 25$, of nearby
powerful FR-II radio galaxies such as Cyg~A and Pic~A, though the low-latitude
FR-II radio galaxy 3C~111 is detected by the LAT.
\item A high BL~Lac/FSRQ ratio, close to unity.
This ratio is even higher than that found for the LBAS sources, which comprise
42 BL~Lacs and 57 FSRQs. It stands in sharp contrast to the population of 3EG
blazars, in which FSRQs outnumber BL~Lacs by a factor of 3.
\item A high HSP/LSP ratio among BL~Lacs.
The large HSP/LSP radio for BL~Lacs is a result of the fact that the sample is
significance-limited. The LAT detects hard-spectrum sources at higher
significance than soft-spectrum sources with comparable photon number fluxes.
Moreover, the LAT is far more sensitive to multi-GeV photons than was EGRET,
which lost sensitivity above several GeV due to self-vetoing effects.
Consequently, HSP-BL~Lacs, with harder spectra than LSP-BL~Lacs, are more
prevalent in the 1LAC than in the 3EG catalog.
\item Little evidence for different variability properties for FSRQs and
BL~Lacs.
No significant differences in the variability of FSRQs and BL~Lacs are observed
for sources at comparable fluxes. Because HSP blazars are generally observed
at low flux levels, however, the comparison of variability properties primarily
concerns LSP/ISP-BL~Lacs and FSRQs.
\item A fairly strong correlation between photon spectral index and blazar
class for the detected sources.
FSRQs, which are almost all LSP blazars, are found to be soft, while BL~Lacs,
whether of the LSP, ISP, or HSP type, represent on average a population of
harder-spectrum sources. The average photon spectral index
$\langle \Gamma \rangle$ continuously shifts to lower values (i.e., harder
spectra) as the class varies from FSRQs ($\langle \Gamma \rangle = 2.48$) to
LSP-BL~Lacs ($\langle \Gamma \rangle = 2.28$), ISP-BL~Lacs
($\langle \Gamma \rangle = 2.13$), and HSP-BL~Lacs
($\langle \Gamma \rangle = 1.96$).
\end{enumerate}
Furthermore, as noted by \citet{Ghi09} and seen in Figure~\ref{fig:index_lum},
the \mbox{$\gamma$-ray} blazars detected with the {\it Fermi}-LAT separate into
hard-spectrum, low-luminosity sources, primarily consisting of BL~Lac objects,
and high-luminosity, soft-spectrum sources, primarily consisting of FSRQs and
LSP-BL~Lacs. Compared with the LBAS sample, the blazar divide now displays
an extension of the sample toward low-luminosity objects in each of the blazar
subclasses. In addition, we identify a third branch consisting of radio
galaxies, which are distinguished from the aligned blazars by their soft
spectra and very low \mbox{$\gamma$-ray} luminosities.
The finding that FSRQs in the 1LAC almost all have soft ($\Gamma \ga 2.2$)
\mbox{$\gamma$-ray} spectra suggests a connection between the presence of strong
emission lines and the nonthermal electron maximum Lorentz factor or spectral
shape. It is consistent with scenarios \citep{Ghi98,Boe02} in which the strong
ambient radiation fields, as revealed by the presence of emission lines,
affects the acceleration and cooling of particles and controls the formation of
the blazar SED. The luminosity dependence of the \mbox{$\gamma$-ray} spectral
hardness is also consistent with the blazar sequence \citep{Ghi98,Fos98,
GhiTav08}, in which the frequencies of the synchrotron and Compton peaks are
inversely correlated with the apparent isotropic luminosities of the blazars.
This behavior has been interpreted in terms of an evolution of FSRQs into
BL~Lac objects as the supply of matter fueling the accretion disk declines, the
surrounding external radiation fields become less intense, and the jet weakens
\citep{Boe02,Cav02,GhiTav08}.
Important caveats must, however, be considered before drawing firm conclusions
about the blazar sequence from the 1LAC sources. First, identification of
the blazars used to construct the blazar sequence is biased by the flux limits
and frequency ranges of the underlying radio and \mbox{X-ray} samples
\citep{Gio99,Pad03}, and different conclusions about its validity can be drawn
depending on the catalogs used to examine the sequence \citep{Niep06}. Second,
specific sources that do not conform to the sequence have been identified
\citep{caccianiga2,j0810}. Third, the 1LAC is significance-limited rather than
flux-limited, which results in preferential detection of lower-flux,
harder-spectrum sources. This means that there is a strong bias against
detecting weak, soft-spectrum \mbox{$\gamma$-ray} sources from low-luminosity
LSP blazars that might violate the sequence. On the other hand, hard-spectrum
HSP sources should be more easily detected, but there are no high-luminosity
HSP sources detected in the LAT band (see Figure~\ref{fig:index_lum}), as all
high-luminosity objects have SEDs of the LSP type. This may, however, reflect
the difficulty of measuring the redshifts of BL~Lacs, despite efforts even with
10~m-class telescopes \citep{Shaw10}. Even for sources associated with AGNs
with high probabilities, a large fraction lack redshift measurements, including
58\% of the high-confidence BL~Lacs and $\sim$30 BL~Lacs with $\Gamma \la 1.9$.
Considering the recent results of \citet{Plot10}, who find that many BL~Lacs
might have redshifts larger than 2 (up to $z=4.9$), and the indication that
BL~Lacs with unknown redshifts may be found preferentially at higher redshift
(see Section~\ref{sec:z}), it is possible that a number of distant,
high-luminosity HSP-BL~Lacs are included in the 1LAC sample. According to our
SED classification, $\sim$20 BL~Lacs with unknown redshifts are already
classified as HSPs. If they were all at high redshift (e.g., $z=2$), and
excluding those already detected at very high energy (which indicates a low
redshift due to a lack of EBL absorption), $\sim$7 objects could have
\mbox{$\gamma$-ray} luminosities $L_\gamma \ga 10^{48}$~erg~s$^{-1}$.
The existence of such sources, though quite rare, would be contrary to
expectations from the blazar sequence \citep{Ghi98,Fos98,GhiTav08}. In the
absence of redshift measurements of possible high-luminosity, hard-spectrum
sources, we can only conclude that the {\it Fermi} results seem compatible
with the sequence. Identifying hard-spectrum blazar sources at high redshifts
is also important for studies of the cosmological evolution of the
extragalactic background light \citep{EBL}, which are hampered by their
absence.
Related to this, it is interesting to compare the redshift distribution of the
LAT blazars (in particular, the FSRQs) with that of the objects detected by BAT
on {\it Swift} \citep{ajello09}. Despite a fourfold increase in exposure time
with respect to the LBAS sample and a consideration of \mbox{$\gamma$-ray} sources
with $TS > 25$ in the 1LAC compared with $TS > 100$ in LBAS, the number of
LAT-detected blazars still drops sharply at $z \sim 2.5$, and only two of the
high-confidence 1LAC blazars have $z > 3$ (keeping in mind biases from the lack
of redshift measurements). This is very different from the BAT survey, in
which half of all FSRQs have $z > 2$, and the distribution extends to
$z \sim 4$. This behavior may be indicative of a shift in the SED peak
frequencies toward lower values (i.e., a ``redder'' SED) for blazars at high
redshifts. Indeed, the jets of the high-redshift BAT blazars are found to be
more powerful than those of the LAT blazars and are among the most powerful
known \citep{Ghi10}. The peak of the inverse Compton flux for these objects,
estimated to be in the MeV or even sub-MeV range, is located closer to the BAT
band than to the LAT band, and the LAT instead samples the cutoff region of the
inverse Compton spectrum.
Comparison of {\it Fermi} results with pre-launch {\it GLAST} expectations
depends on how the 1LAC results are characterized. Except for a few dozen
millisecond pulsars, nearly all of the 1043 high-latitude \mbox{$\gamma$-ray}
sources in the 1FGL are likely to be AGNs, but only 671 high-latitude 1FGL
sources are associated with AGNs. Adopting a radio/\mbox{$\gamma$-ray}
correlation, \citet{ss96} and \citet{cm98} predicted many thousands of blazars
to a flux limit of $2 \times 10^{-9}$ \pflux{} (for $E > 100$~MeV) reached
after five years for a source with $\Gamma = 2.1$, as did \citet{nt06}, who
treated the problem using radio and \mbox{X-ray} luminosity functions.
\citet{mp00}, using a physical blazar model matched to the EGRET data,
predicted an increased fraction of BL~Lac objects at fainter flux levels, with
nearly 2000 BL~Lacs and 1000 FSRQs at the one-year flux limit of
$4 \times 10^{-9}$ \pflux{} (for $E > 100$~MeV). \citet{Der07} predicted that
$\sim$800 FSRQs and $\sim$200 BL~Lacs would be detected after one year, also
employing a physical blazar model. Assuming the validity of the blazar
sequence, \citet{it09} predicted $\sim$600-1200 blazars to a flux limit of
$2 \times 10^{-9}$ \pflux{} (for $E > 100$~MeV). However, predictions based on
a minimum flux depend sensitively on the source spectral index, as shown by
Figure~\ref{fig:index_flux}, so more reliable predictions needed to take this
effect into account.
The populations of AGNs that can be studied jointly at GeV and at TeV energies
has increased dramatically beyond what could have been expected based on the
experience of the EGRET era. There is now a large and increasing population of
GeV--TeV AGNs that can be best studied through observations with both
{\it Fermi} and the ground-based \mbox{$\gamma$-ray} observatories. Since its
launch, the number of TeV AGNs detected also by {\it Fermi} has steadily
increased as its exposure on these sources grows. Complementing this, the TeV
observatories have started to observe the more promising {\it Fermi} AGNs, and
a number of detections have resulted. One promising feature of joint GeV--TeV
observations is the power to infer the redshifts of BL~Lacs from spectral
modeling, as in the case of PG~1553$+$113 \citep{pg1553}.
\section{Conclusions}
We have presented the First LAT AGN Catalog (1LAC), consisting of 671
high-latitude \mbox{$\gamma$-ray} sources with $TS > 25$ measured in the first
11~months of the {\it Fermi}-LAT sky survey. Due to multiple associations,
these sources are associated with 709 AGNs. Besides a growing number of
blazars and six radio galaxies, the extragalactic \mbox{$\gamma$-ray} sky now
includes a small number of steep-spectrum quasars and narrow-line radio
galaxies. Confirming many of the LBAS results, we find a strong correlation
between the 0.1-100~GeV \mbox{$\gamma$-ray} spectral index and the blazar type,
whether defined in terms of optical emission line strength or peak frequency of
the synchrotron component of the SED. The relationship between the
\mbox{$\gamma$-ray} spectral properties and the spectral luminosity is
complicated by selection biases for source identification and redshift
determination, but the {\it Fermi} results seem compatible with the blazar
sequence.
The First LAT AGN Catalog represents significant progress in terms of the
number of detected sources, the diversity of source types, and the accuracy of
measured \mbox{$\gamma$-ray} properties. This sample is likely to evolve in the
future as more unidentified sources can be associated with AGNs thanks to
better counterpart catalogs and continued correlated variability studies with
more LAT observations. Bearing in mind the instrumental limitations described
above, the 1LAC provides an essential foundation on which to build a much
better understanding of the population of \mbox{$\gamma$-ray}--emitting AGNs.
\acknowledgments
\section*{Acknowledgments}
The \textit{Fermi} LAT Collaboration acknowledges generous ongoing support from
a number of agencies and institutes that have supported both the development
and the operation of the LAT as well as scientific data analysis. These
include the National Aeronautics and Space Administration and the Department of
Energy in the United States; the Commissariat \`a l'Energie Atomique and the
Centre National de la Recherche Scientifique / Institut National de Physique
Nucl\'eaire et de Physique des Particules in France; the Agenzia Spaziale
Italiana and the Istituto Nazionale di Fisica Nucleare in Italy; the Ministry
of Education, Culture, Sports, Science and Technology (MEXT), High Energy
Accelerator Research Organization (KEK) and Japan Aerospace Exploration Agency
(JAXA) in Japan; and the K.~A.~Wallenberg Foundation, the Swedish Research
Council and the Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is
gratefully acknowledged from the Istituto Nazionale di Astrofisica in Italy and
the Centre National d'\'Etudes Spatiales in France.
This research has made us of the NASA/IPAC Extragalactic Database (NED) which
is operated by the Jet Propulsion Laboratory, California Institute of
Technology, under contract with the National Aeronautics and Space
Administration. Part of this work is based on archival data, software, or
online services provided by the ASI Science Data Center (ASDC).
Some of the results presented in this paper are based on observations obtained
with the Hobby-Eberly Telescope (HET), a joint project of the University of
Texas at Austin, the Pennsylvania State University, Stanford University,
Ludwig-Maximilians-Universit{\"a}t M{\"u}nchen, and
Georg-August-Universit{\"a}t G{\"o}ttingen. The HET is named in honor of its
principal benefactors, William P.\ Hobby and Robert E.\ Eberly. The Marcario
Low Resolution Spectrograph (LRS) is named for Mike Marcario of High Lonesome
Optics, who fabricated several optics for the instrument but died before its
completion. The LRS is a joint project of the Hobby-Eberly Telescope
partnership and the Instituto de Astronom{\'i}a de la Universidad Nacional
Aut{\'o}noma de M{\'e}xico.
This work is also partly based on optical spectroscopy performed at the
Telescopio Nazionale Galileo (TNG), La Palma, Canary Islands (proposal
AOT20/09B). These observations confirm some of the redshifts and
classifications found with the HET. We thank the HET and TNG personnel for
their assistance during the observing runs.
The data in this paper are based partly on observations obtained at the Hale
Telescope, Palomar Observatory, as part of a collaborative agreement between
the California Institute of Technology, its divisions Caltech Optical
Observatories and the Jet Propulsion Laboratory (operated for NASA), and
Cornell University.
Some of the data presented herein were obtained at the W.~M.~Keck Observatory,
which is operated as a scientific partnership among the California Institute of
Technology, the University of California, and the National Aeronautics and
Space Administration. The Observatory was made possible by the generous
financial support of the W.~M.~Keck Foundation.
The authors wish to recognize and acknowledge the very significant cultural
role and reverence that the summit of Mauna Kea has always had within the
indigenous Hawaiian community. We are most fortunate to have the opportunity
to conduct observations from this mountain.
Some of the data in this paper are based on observations made with ESO
telescopes at the La Silla Observatory under programs \mbox{083.B-0460(B)} and
\mbox{084.B-0711(B).}
The National Radio Astronomy Observatory is a facility of the National Science
Foundation operated under cooperative agreement by Associated Universities, Inc.
Funding for the Sloan Digital Sky Survey (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/.
The SDSS is managed by the Astrophysical Research Consortium (ARC) for the
Participating Institutions. The Participating Institutions are the American
Museum of Natural History, Astrophysical Institute Potsdam, University of
Basel, University of Cambridge, Case Western Reserve University, The University
of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the
Japan Participation Group, The Johns Hopkins University, the Joint Institute
for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and
Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences
(LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for
Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico
State University, Ohio State University, University of Pittsburgh, University
of Portsmouth, Princeton University, the United States Naval Observatory, and
the University of Washington.
{\it Facilities:} \facility{{\it Fermi} LAT}.
|
\section{Introduction}
A classical question in stochastic process theory is to understand the asymptotic behavior of a given stochastic process $X=(X_t)_{t\geq 0}$ on the level of paths. In the present work we consider general L\'evy processes and find Chung type LIL (laws of the iterated logarithm) at zero, that is, given the L\'evy process $X$, we aim at characterizing a norming function $b$ satisfying
\begin{align}\label{LIL}
\liminf_{t\rightarrow 0}\frac{||X||_t}{b(t)}=1,\qquad \text{where}\qquad ||X||_t:=\sup_{0\leq s\leq t}|X_s|.
\end{align}
The topic of large and small time fluctuations of L\'evy processes has been studied extensively in the past (see for instance \cite{D} for an overview and \cite{B96,sato}).
It is well-known that, via the Borel-Cantelli Lemma, Chung type LIL for a general stochastic process are connected to the so-called small deviation rate of the process, i.e.\
\begin{align}\label{small}
-\log \P(||X||_t\leq \eps),\qquad \text{as $\eps\to 0$ and $t\to0$}.
\end{align}
The main motivation for this paper originates from the recent work \cite{AD09}, where a framework for obtaining the small deviation rate (\ref{small}) for general L\'evy processes is provided. The difficulty in passing over from the small deviation estimate to the respective LIL concerns circumventing the independence assumption of the Borel-Cantelli lemma.
\medskip
In this paper we show how the asymptotics of (\ref{small}) imply explicit LIL. \medskip
Small deviation problems are studied independently of LIL and have connections to other fields, such as the approximation of stochastic processes, coding problems, the path regularity of the process, limit laws in statistics, and entropy numbers of linear operators. We refer to the surveys \cite{lishao,lif} for an overview of the field and to \cite{sdbib} for a regularly updated list of references, which also includes references to laws of the iterated logarithm of Chung type. The papers \cite{taylor,mogulskii,bormog,simon01,simonpvar,lindeshi,ls,lindezipfel,elena,elena2} provide a good source for earlier results on small deviations of L\'evy processes.\medskip
We now discuss LIL for special L\'evy processes that have already appeared in the literature. The norming function $b(t)=\sqrt{\pi^2 t/ (8\log|\log t|)}$ for a standard Brownian motion can be derived from the large time LIL, proved by Chung \cite{chung}, via time inversion. \\ For any L\'evy process with non-trivial Brownian component the recent result of \cite{BM09} shows that (\ref{LIL}) holds with the norming function for a standard Brownian motion. If $X$ is an $\alpha$-stable L\'evy process (\ref{LIL}) holds with norming function $b(t)= ( c_\alpha t/\log|\log t|)^{1/\alpha}$, which goes back to \cite{taylor}.\medskip
Of course it is natural to ask for the general structure of the norming function for arbitrary L\'evy processes not having the special features of the examples mentioned so far.
LIL for more general L\'evy processes were obtained by Wee in \cite{W88} (see \cite{wee2} for more examples). It was shown that if for some positive constant $\theta$
\begin{align} \label{eqn:weecondition}
\P(X_t>0)\geq \theta\quad \text{and}\quad \P(X_t<0)\geq \theta,\qquad\text{for all $t$ sufficiently small,}
\end{align}
holds, then upper and lower bounds in the LIL hold in the following sense: for $\lambda_1$ sufficiently small and $\lambda_2$ sufficiently large,
\begin{align*}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_t}{b_{\lambda_1}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_t}{b_{\lambda_2}(t)}\leq 1
\end{align*}
for norming functions $b_{\lambda}$ given by
\begin{align*}
b_\lambda(t):=f^{-1}\left(\frac{\log|\log t|}{\lambda t} \right),
\end{align*}
where $f$ is given by some explicit, but complicated expression depending on the L\'evy triplet.\medskip
Although the results of Wee are quite general, there are some drawbacks which we aim to overcome in the present work.\\
First, the proofs given in \cite{W88} and \cite{wee2} are rather obscure. This might be due to the use of the old-fashioned notation for the L\'evy measure. In particular, it does not become clear that actually the LIL follow from small deviation estimates of type (\ref{small}) and which behavior of the process is actually responsible for the correct norming function.
Secondly, even in the symmetric case, if the norming function $b_{\lambda}$ is not regularly varying at zero, the unspecified (and suboptimal) constants $\lambda_1$ and $\lambda_2$ do not only appear in a weaker limiting constant, but they influence the norming function essentially (see (\ref{eqn:constinexp}) below for an example of influence on the exponential level). In our approach, we keep track of the appearing constants in an optimal way.
Thirdly, although condition (\ref{eqn:weecondition}) looks innocent at a first glance it turns out to be quite delicate. It is certainly fulfilled for symmetric L\'evy processes. Unfortunately, only given the L\'evy triplet it seems to be unknown how (\ref{eqn:weecondition}) can be checked. On the contrary, our conditions are explicit in terms of the L\'evy triplet. We give a couple of examples showing that our conditions can be checked easily.
\medskip
This paper is structured as follows. In Section~\ref{sec:results}, we give the main results that manage the transfer between small deviations and LIL. Several examples of LIL for concrete L\'evy processes are collected in Section~\ref{sec:examples}. The proofs are given in Section~\ref{sec:proofs}.
\medskip
Let us finally fix some notation. In this paper we let $X$ be a L\'evy process with characteristic triplet $(\gamma, \sigma^2, \Pi)$, where $\gamma\in \R$, $\sigma^2\geq 0$, and the L\'evy measure $\Pi$ has no atom at zero and satisfies
\begin{align*}
\int (1\wedge x^2)\Pi(\d x)<\infty.
\end{align*}
For basic definitions and properties of L\'evy processes we refer to \cite{B96,sato}. As we are interested only in the behavior for small times we may truncate large jumps. In particular, we restrict ourselves to L\'evy processes involving only jumps of absolute value at most $1$. Hence, the characteristic exponent, $\E e^{i z X_t } =: e^{ t \psi(z)}$, has the form
\begin{align*}
\psi(z)=i\gamma z -\frac{\sigma^2z^2}{2}+\int_{-1}^{1}(e^{iz x}-1-iz x)\Pi(\d x), \qquad z\in\R.
\end{align*}
For later use we denote by $\Phi$ the Laplace exponent of a subordinator $A$, $\E e^{-u A_1}=e^{-\Phi(u)}$,
\begin{align*}
\Phi(u)= u \gamma_A + \int_0^\infty (1-e^{-ux}) \Pi_A(\d x).
\end{align*}
Further, we use the standard notation $\bar\Pi(\eps):=\Pi([-\eps,\eps]^c)$ for the tail of the L\'evy measure.
In the following, we denote by $f \sim g$ the strong asymptotic equivalence, i.e.\ $\lim f/g=1$, and by $f\approx g$ the weak asymptotic equivalence, i.e.\ $0<\liminf f/g \leq \limsup f/g < \infty$.
\section{Main results} \label{sec:results}
Our first theorem manages the transfer from small deviation rates to LIL under minimal loss of constants.
\begin{theorem}\label{t3}
Let $X$ be a L\'evy process and $F$ be a function increasing to infinity at zero such that with some $0<\lambda_1\leq \lambda_2<\infty$
\begin{align}\label{eqn:sdestmain}
\lambda_1 F(\eps)t\leq -\log \P(||X||_{t}<\eps) \leq \lambda_2 F(\eps)t, \qquad \text{for all $\eps<\eps_0$ and $t<t_0$.}
\end{align}
Further, define
\begin{align*}
b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)
\end{align*}
for $\lambda>0$ and assume that
\begin{align} \label{eqn:conditionM}
(n+1)^{-(n+1)^\beta} \left| \int_{|x|>b_{\lambda_2'}(n^{-n^\beta})} x \Pi(\d x) - \gamma\right|=o\big(b_{\lambda_2'}\big(n^{-n^\beta}\big)\big),\qquad \text{for all $\beta>1$ and $\lambda_2'>\lambda_2$.}
\end{align}
Then the LIL
\begin{align*}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_1'}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_2'}(t)}\leq 1
\end{align*}
hold almost surely for any $\lambda_1'<\lambda_1$ and $\lambda_2'>\lambda_2$.
\end{theorem}
Unfortunately, our proof forces us to assume condition (\ref{eqn:conditionM}) in order to prove the more delicate upper bound. This condition is clearly satisfied for symmetric processes and can be checked readily from the L\'evy triplet.
It is crucial that there is almost no loss of constants in the transfer from the small deviations to the LIL as in cases when $b_{\lambda}$ is not regularly varying, the constants $\lambda_1', \lambda_2'$ may influence the rate function drastically (see (\ref{eqn:constinexp}) for an extreme example).\\
If instead $b_{\lambda}$ only depends on $\lambda$ via a multiplicative constant, our approach allows to strengthen the previous theorem to the optimal limiting constants. Such examples occur for instance if the small deviation rate function $F$ is regularly varying.
\begin{corollary}\label{cor:t1corollary}
In the setting of Theorem~\ref{t3} assume additionally that $F$ is regularly varying at zero with non-positive exponent. Then the following LIL hold almost surely:
\begin{align} \label{eqn:lilnonre}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_1}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_2}(t)}\leq 1.
\end{align}
In particular, if there is $\lambda>0$ such that (\ref{eqn:sdestmain}) holds for all $\lambda_1<\lambda$ and all $\lambda_2>\lambda$ then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda}(t)}=1\qquad\text{a.s.}
\end{align*}
\end{corollary}
In the setting of regularly varying rate function, say $F$ is regularly varying at zero with exponent $-\alpha$, $\alpha>0$, one can express (\ref{eqn:lilnonre}) as
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{1}(t)}\in \big[\lambda_1^{1/\alpha},\lambda_2^{1/\alpha}\big],\qquad\text{a.s.}
\end{align*}
This shows that only the quality of the small deviation estimate (\ref{eqn:sdestmain}) matters in order to obtain the limiting constant in the LIL. Recall that the Blumenthal zero-one law implies that the limit is almost surely equal to a deterministic constant, which in this case can be specified.\medskip
Theorem~\ref{t3} reduces the question of the right norming function for the LIL to the question of small deviations which is known precisely for many examples.
For general L\'evy processes those have been obtained in \cite{AD09} (their results were stated for $t=1$ only but hold in general as we discuss in Proposition~\ref{prop:ad} below). In particular, for symmetric L\'evy processes their main result states that the rate function is given by
\begin{align}\label{a2}
F(\eps)=\eps^{-2}U(\eps),
\end{align}
where $U(\eps)$ is the variance of $X$ with jumps larger than $\eps$ replaced by jumps of size $\eps$:
\begin{align}\label{eqn:defnU}
U(\eps):=\eps^2\bar\Pi(\eps)+\sigma^2+\int_{-\eps}^{\eps}x^2\Pi(\d x).
\end{align}
From these specific small deviations we can deduce the following corollary for symmetric processes.
\begin{corollary} \label{cor:sddirectsymmetric}
Let $X$ be a symmetric L\'evy process, then there are $0<\lambda_1\leq \lambda_2<\infty$ such that
almost surely,
\begin{align*}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_1}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_2}(t)}\leq 1,
\end{align*}
with
\begin{align*}
b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)
\end{align*}
and $F$ defined in (\ref{a2}). If additionally $F$ is regularly varying at zero with exponent $-\alpha$, $\alpha>0$, then the following general bounds hold:
\begin{align*}
\frac{1}{12}\frac{1}{2^{\alpha}}\leq \lambda_1\leq \lambda_2\leq 3^{\alpha}10.
\end{align*}
\end{corollary}
The loss of constants in the corollary is only due to the general formulation. For some examples we will see below that the small deviations are known in the strong asymptotic sense so that Theorem~\ref{t3} gives the precise law.\medskip
For strongly non-symmetric L\'evy processes we have to proceed differently, since here condition (\ref{eqn:conditionM}) does not hold. For this case, we provide another, different link between small deviation rates and LIL; we keep track of the constants in the norming function in an optimal way and only lose the limiting constant.
\begin{theorem}\label{t}
Let $X$ be a L\'evy process and $F$ be a function increasing to infinity at zero such that for $0<\lambda_1\leq \lambda_2<\infty$
\begin{align} \label{eqn:yetanothersdestimate}
\lambda_1 F(\eps)t\leq -\log \P(||X||_{t}<\eps) \leq \lambda_2 F(\eps)t, \qquad \text{for all $\eps<\eps_0$ and $t<t_0$.}
\end{align}
Furthermore, set
\begin{align} \label{eqn:quantitythebee}
b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)
\end{align}
and suppose that there is a constant $C>0$ such that
\begin{align} \label{eqn:regularityofb}
C b_{\lambda}(t)\leq b_{\lambda}(t/2),\qquad 0<t\leq t_0, \lambda\in(\lambda_1/2,2\lambda_2).
\end{align}
Then the LIL
\begin{align*}
0<\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_1'}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_2'}(t)}<\infty
\end{align*}
hold almost surely for all $\lambda_1'<\lambda_1$ and $\lambda_2<\lambda_2'$.
\end{theorem}
Again, if the rate function $F$ is regularly varying, then we can strengthen the result.
\begin{corollary}\label{cor:t1corollary2}
In the setting of Theorem~\ref{t} assume additionally that $F$ is regularly varying at zero with negative exponent. Then the following LIL hold almost surely:
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{1}(t)} \in (0,\infty).
\end{align*}
\end{corollary}
The theorems listed so far manage the transfer between small deviation order and LIL. Similarly to Corollary~\ref{cor:sddirectsymmetric}, we can combine them with the main results of \cite{AD09}. This looks more technical in the present case. We give an explanation of the role of the different terms after the result.
\begin{theorem}\label{t2}
Let $X$ be a L\'evy process with triplet $(\gamma, \sigma^2, \Pi)$. Assume that $u_\eps$ is the solution of the equation $\Lambda_\eps'(u)=0$, where $\Lambda_\eps$ is the following log Laplace transform:
\begin{align} \label{eqn:quantitylambda}
\Lambda_\eps(u)=\frac{\sigma^2}{2}\, u^2 + \left(\gamma-\int_{[-1,1]\setminus [-\eps,\eps]} x \Pi(\d x)\right) u + \int_{-\eps}^{\eps} (e^{u x}-1 - u x) \Pi(\d x).
\end{align}
Set
\begin{align} \label{eqn:quantitythef}
F(\eps):=\eps^{-2} U_\eps(\eps)-\Lambda_\eps(u_\eps),\qquad U_\eps(\eps):=\eps^2\bar\Pi(\eps)+\sigma^2+\int_{-\eps}^{\eps}x^2 e^{-u_\eps x} \Pi(\d x),
\end{align}
define $b$ as in (\ref{eqn:quantitythebee}) and assume that $b$ satisfies (\ref{eqn:regularityofb}).
If furthermore
\begin{align} \label{eqn:cond-esschervanishes}
\eps |u_\eps| = o( \log \log F(\eps) ), \qquad \text{as $\eps\to 0$,}
\end{align}
is satisfied then we have for some $\lambda_1, \lambda_2>0$
\begin{align*}
0< \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_1}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{\lambda_2}(t)}<\infty\qquad \text{a.s.}
\end{align*}
\end{theorem}
Let us explain the quantities occurring in Theorem~\ref{t2} in more detail. The main observation is that the proof for the small deviation estimates in \cite{AD09} (Theorem~1.5) can be used directly for any $t>0$ to obtain the following proposition.
\begin{proposition} \label{prop:ad}
Let $\Lambda_\eps$ be as defined in (\ref{eqn:quantitylambda}) and assume that $u_\eps$ is the solution of $\Lambda_\eps'(u_\eps)=0$. Then, with $F$ as in (\ref{eqn:quantitythef}), we have for all $t>0$ and all $\eps<1$
\begin{align} \label{eqn:sdquantityadstrengthend}
\frac{1}{12}\, t\, F(2\eps) - \eps |u_{2\eps}| -1 \leq - \log \P\left( ||X||_t \leq \eps \right) \leq 10 t\, F\left(\frac{\eps}{3}\right) + \eps |u_{\eps/3}|+3.
\end{align}
\end{proposition}
The term $\bar\Pi(\eps)$ in (\ref{eqn:sdquantityadstrengthend}) (included in the $F$ term) comes from the requirement that there should be no jumps larger than $\eps$. After removing these jumps, the process may drift out of the interval $[-\eps,\eps]$, which is prevented by applying an Esscher transform to the process, whose `price' is given by the term $-\Lambda_\eps(u_\eps)$. The quantity $u_\eps$ is the drift that has to be subtracted in order to make the process a martingale. Then the remaining process is treated as in the symmetric case, and the same term $\eps^{-2} U_\eps(\eps)$ appears as in (\ref{a2}), but this time with respect to the L\'evy measure transformed by the change of measure.
Note that (\ref{eqn:sdquantityadstrengthend}) is almost the required estimate in (\ref{eqn:yetanothersdestimate}), except for the term $\eps |u_\eps|$, which may spoil the estimate. It is exactly condition (\ref{eqn:cond-esschervanishes}) that ensures that the term $\eps |u_\eps|$ can be neglected.\\
We stress that in some cases $\eps |u_\eps|$ does give an order that is larger than $t F(\eps)$ so that the function $b$ from (\ref{eqn:quantitythebee}) is not the right norming function. This effect can be observed in some examples below. In particular, this happens for processes of bounded variation with non-zero drift.
\begin{proposition} \label{prop:bvnodrift}
Let $X$ be a L\'evy process with bounded variation and non-vanishing effective drift, i.e.\ $\int_{[-1,1]} |x| \Pi(\d x)< \infty$ and $c:=\gamma-\int_{-1}^1 x\,\Pi (\d x)\neq 0$. Then
\begin{align*}
\lim_{t\rightarrow 0}\frac{||X||_t}{t} =|c|\qquad\text{a.s.}
\end{align*}
\end{proposition}
The proof of this proposition is based on classical arguments rather than any connection to small deviations. \medskip
\section{Explicit LIL for L\'{e}vy processes}\label{sec:examples}
In this section we collect concrete L\'evy processes for which we can transform small deviation results to an LIL. As we have seen, understanding the small deviation rates is crucial. \medskip
The first corollary gives us a useful variance domination principle for LIL that works for many examples.
\begin{corollary}\label{cor:domination}
Suppose $X^1$ and $X^2$ are independent symmetric L\'evy processes, then $X^1+X^2$ and $X^2$ fulfill precisely the same LIL if
\begin{align*}
\lim_{\eps\to0}\frac{U_{X^1}(\eps)}{U_{X^2}(\eps)}=0.
\end{align*}
\end{corollary}
\begin{proof}
This follows directly from Corollary~\ref{cor:sddirectsymmetric} noticing that $U_{X^1+X^2}=U_{X^1}+U_{X^2}$.
\end{proof}
In the same spirit the following corollary (recovering (3.2) in \cite{BM09}) displays the intuitive fact that a non-zero Brownian component dominates the jumps of a L\'evy process.
\begin{corollary}\label{cor:withbrownian}
If $X$ is a L\'evy process with $\sigma\neq 0$, then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{\sqrt{t/\log| \log t|}}=\frac{\pi\sigma}{\sqrt{8}}\qquad\text{ a.s.}
\end{align*}
\end{corollary}
\begin{proof}
Following precisely the proof of Corollary~2.6 of \cite{AD09} one can show that the small deviation rates of L\'evy processes with non-zero Brownian component is given by
\begin{align*}
-\log \P(||X||_{t}<\eps)\sim \frac{\pi^2\sigma^2}{8}\eps^{-2}t, \qquad\text{as $\eps\to 0$ and $t\to 0$.}
\end{align*}
Hence, the norming function follows from Theorem~\ref{t3}. As the process is not necessarily symmetric, condition (\ref{eqn:conditionM}) has to be checked: Since $b(t)=\sqrt{t\pi^2/ (8\log |\log t|)}$ and $\int_{|x|>\eps} |x| \Pi(\d x)=o(\eps^{-1})$, it remains to be seen that
$$ a_{n+1} \leq c b(a_n)^2= a_n / \log |\log a_n|$$ for $a_n=n^{-n^\beta}$ and $\beta>1$. This can be verified by simple computations.
\end{proof}
Similarly to L\'evy processes with non-zero Brownian component, symmetric processes of smaller small deviation order (e.g.\ stable processes of smaller index) are dominated by stable L\'evy processes.
\begin{corollary}\label{cor:stables}
Let $X$ be a symmetric $\alpha$-stable L\'evy process with $\alpha\in (0,2]$ and let $Y$ be symmetric with $U_Y(x)=o(x^{2-\alpha})$. Then there is a constant $0<c_\alpha<\infty$ such that
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X+Y||_t}{(t/\log|\log t|)^{1/\alpha}}=\liminf_{t\rightarrow 0}\frac{||X||_t}{(t/\log|\log t|)^{1/\alpha}}=c_\alpha^{1/\alpha}\qquad\text{a.s.}
\end{align*}
\end{corollary}
\begin{proof}
The small deviation rate is given by
\begin{align*}
-\log \P(||X||_{t}<\eps)\sim c_{\alpha} \eps^{-\alpha}t, \qquad\text{as $\eps\to 0$ and $t\to 0$},
\end{align*}
for some constant $c_{\alpha}>0$ (see e.g.\ page 220 in \cite{B96}). Hence, the LIL follow from Corollary~\ref{cor:t1corollary} and Corollary~\ref{cor:domination}.
\end{proof}
\begin{rem}
The constant $c_\alpha$ in the LIL of stable L\'evy processes is the unknown constant of the small deviations for respective $\alpha$-stable L\'evy processes (see \cite{taylor} and Proposition~3 and Theorem~6 in Chapter VIII of \cite{B96}). The results of \cite{AD09} entail the following concrete bounds:
\begin{align*}
\frac{2 C }{2^{\alpha}} \left( \frac{1}{\alpha} + \frac{1}{12(2-\alpha)}\right) < c_\alpha < 3^\alpha\cdot 2 C \left( \frac{1}{\alpha} + \frac{10}{2-\alpha}\right),
\end{align*}
where $C$ is the constant in the L\'evy measure: $\Pi(\d x)=C |x|^{-(1+\alpha)}\d x$. This implies $c_\alpha\sim 2 C/\alpha$, as $\alpha\to 0$. We remark that, contrary to the symmetric case, the constant $c_\alpha$ is known explicitly for completely asymmetric stable L\'evy processes, see \cite{bertoin96}.
\end{rem}
\medskip
If $\Pi$ behaves as a regularly varying function at zero and is symmetric the following LIL are satisfied.
\begin{corollary}
Let $X$ be a L\'evy process with triplet $(0,0,\Pi)$ with $\Pi$ being symmetric and
$$\bar\Pi(\eps)\approx \eps^{-\alpha} |\log\eps|^{-\gamma},\qquad\text{as $\eps\to 0$,}$$
with $0<\alpha<2$ or $\alpha=2, \gamma>1$. Then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_{t}}{b(t)}\in(0,\infty)\qquad{a.s.}
\end{align*}
with
\begin{align*}
b(t)=\begin{cases}
\left(\frac{t|\log t|^{-\gamma}}{\log|\log t|}\right)^{1/\alpha}&:0<\alpha<2,\\
\left(\frac{t|\log t|^{1-\gamma}}{\log|\log t|}\right)^{1/2}&:\alpha=2, \gamma>1.
\end{cases}
\end{align*}
\end{corollary}
\begin{proof}
The corollary follows from Theorem~\ref{t2}. The required small deviation estimate,
\begin{align*}
-\log \P(||X||_{t}<\eps)\approx \begin{cases}
\eps^{-\alpha}|\log \eps|^{-\gamma} t&:0<\alpha<2,\\
\eps^{-2}|\log \eps|^{1-\gamma}t&:\alpha=2, \gamma>1,
\end{cases}
\end{align*}
as $\eps\to 0$ and $t\to 0$, is obtained from Proposition~\ref{prop:ad} (cf.\ Example~2.2 in \cite{AD09} for $t=1$). Since we deal with a symmetric process, condition (\ref{eqn:cond-esschervanishes}) is trivially satisfied due to $u_{\eps}=0$.
\end{proof}
Having discussed the $\alpha$-stable like cases, we now consider L\'evy processes with polynomial tails near zero of {\it different} exponents. The technique used for this example can be extended to any case with essentially regularly varying L\'{e}vy measure at zero. Let $X$ be a L\'evy process with triplet $(\gamma,0,\Pi)$, where $\Pi$ is given by \begin{equation} \frac{\Pi(\d x)}{\d x} = \frac{C_1 \ind_{(0,1]}(x)}{x^{1+\alpha_1}} + \frac{C_2 \ind_{[-1,0)}(x)}{(-x)^{1+\alpha_2}}, \label{eq:regularlm} \end{equation} with $2>\alpha_1\geq\alpha_2$ and $C_1,C_2\geq 0$, $C_1+C_2\neq 0$. We now analyze the pathwise behavior at zero in the cases when $\alpha_1>1$, $\alpha_1=1$, and $0<\alpha_1<1$, respectively. The second exponent $\alpha_2$ can be even negative.
\begin{corollary}\label{pol} Let $X$ be a L\'evy process with triplet $(\gamma,0,\Pi)$ with $\Pi$ as in (\ref{eq:regularlm}). Then the following holds:
\begin{enumerate}
\item If $\alpha_1\geq \alpha_2$ and $\alpha_1>1$, then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{(t / \log |\log t|)^{1/\alpha_1}}\in (0,\infty)\qquad\text{a.s.}
\end{align*}
\item If $\alpha_1=\alpha_2=1$ and $C_1=C_2$ then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{t / \log |\log t|}\in (0,\infty)\qquad\text{a.s.}
\end{align*}
\item If $1>\alpha_1\geq \alpha_2$ and the effective drift $c=\gamma-\int_{-1}^1 x\,\Pi(\d x)$ vanishes, then
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{(t / \log |\log t|)^{1/\alpha_1}}\in (0,\infty)\qquad\text{a.s.}
\end{align*}
\item If $1>\alpha_1\geq \alpha_2$ and the effective drift does not vanish, then
\begin{align*}
\lim_{t\rightarrow 0}\frac{||X||_t}{t} = |c|\qquad\text{a.s.}
\end{align*}
\end{enumerate}
\end{corollary}
\begin{proof}
Parts (1), (2), and (3) follow from Theorem~\ref{t2}. The required small deviation estimates,
\begin{align*}
-\log \P(||X||_{t}<\eps)\approx \eps^{-\alpha_1}t
\end{align*}
for $\eps\to 0$ and $t\to 0$, are obtained from Proposition~\ref{prop:ad} (cf.\ Corollary~2.7,~2.8, and~2.9 of \cite{AD09} for $t=1$; note that $u_\eps\approx \eps^{-1}$ in all cases). One can easily check condition (\ref{eqn:cond-esschervanishes}).\\
In part (4), the process is of bounded variation, so that the claim is included in Proposition~\ref{prop:bvnodrift}.
\end{proof}
\medskip
We now come to L\'evy processes obtained from Brownian motion by subordination, i.e.\ $X_t=\sigma B_{A_t}$, where $B$ is a Brownian motion independent of the subordinator $A$. In this case, the resulting L\'evy process is symmetric and the small deviation asymptotics is governed by the truncated variance $U$ from (\ref{eqn:defnU}).
\begin{corollary}
Let $B$ be a Brownian motion independent of the subordinator $A$, where $A$ has Laplace exponent $\Phi$. For $\lambda>0$ we set $b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)$ with
\begin{align*}
F(\eps):= \Phi(\sigma^2\eps^{-2}) + \gamma_A \sigma^2 \eps^{-2}.
\end{align*}
Then for some $\lambda_1, \lambda_2>0$
\begin{align*}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_t}{b_{\lambda_1}(t)}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_t}{b_{\lambda_2}(t)}\leq 1\qquad\text{a.s.}
\end{align*}
In particular, if $\gamma_A=0$ and $\Phi$ is regularly varying with positive exponent, we have
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{(\Phi^{-1}\left(\log | \log t| / t\right))^{-1/2}}\in(0,\infty)\qquad\text{a.s.}
\end{align*}
\end{corollary}
\begin{proof}
The corollary follows from Theorem~\ref{t3} with small deviation estimate from Proposition~\ref{prop:ad}
$$
-\log \P( ||X||_t \leq \eps) \approx (\Phi(\sigma^2\eps^{-2}) + \gamma_A \sigma^2 \eps^{-2})t,
$$
as $\eps\to 0$ and $t\to 0$ (cf.\ Example~2.13 of \cite{AD09} for $t=1$ and note the misprint there). Condition (\ref{eqn:conditionM}) is trivially fulfilled as the process is symmetric.
\end{proof}
For a more specific example, in particular exhibiting exotic small time behavior, we choose the subordinator $A$ to be a Gamma process. Then one defines the so called Variance-Gamma process as
\begin{align*}
X_t=\sigma B_{A_t}+\mu A_t,
\end{align*}
for some constants $\sigma\neq 0$ and $\mu\in\R$.
\begin{corollary
Let $X$ be a Variance-Gamma process, then for $\mu=0$ there are some constants $0<\lambda_1\leq \lambda_2<\infty$ such that
\begin{align} \label{eqn:constinexp}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_t}{ e^{ - \lambda_1 \log |\log t|/t}}\qquad\text{and}\qquad\liminf_{t\rightarrow 0}\frac{||X||_t}{ e^{- \lambda_2 \log |\log t|/t}}\leq 1\qquad\text{ a.s.},
\end{align}
whereas for $\mu\neq 0$
\begin{align*}
\liminf_{t\rightarrow 0}\frac{||X||_t}{t}=|\mu|\,\E A_1\qquad\text{ a.s.}
\end{align*}
\end{corollary}
\begin{proof}
The second part is included in Proposition~\ref{prop:bvnodrift}, since the process is of bounded variation with non-zero effective drift. In the first part, the effective drift is zero, and the claim follows from Theorem~\ref{t3}. The small deviation estimate,
\begin{align*}
-\log \P\left( ||X||_t \leq \eps\right) \approx t |\log \eps|, \qquad \text{as $\eps\to 0$ and $t\to 0$}
\end{align*}
follows from Proposition~\ref{prop:ad} (cf.\ Example~2.12 of \cite{AD09} for $t=1$).
\end{proof}
In the first case of the previous corollary the dependence of good small deviation estimates and good LIL becomes transperant. The fact that we cannot specify the constants $\lambda_1, \lambda_2$ in (\ref{eqn:constinexp}) is only caused by the weak asymptotics for the small deviation estimate as we do not lose any further constants in the transfer of small deviations to the LIL. If one does not have more control on the constants $\lambda_1, \lambda_2$, the understanding of the precise small time behavior of $X$ is far from optimal as the error enters exponentially.
\section{Proofs} \label{sec:proofs}
We start with a lemma which shows that the small deviation order is at least as large as the term induced by the variance, defined in (\ref{eqn:defnU}).
\begin{lemma}
Let $\eps>0$ and let $X$ be a L\'evy process with L\'evy measure concentrated on $[-\eps,\eps]$, then
\begin{align*}
\P( ||X||_t \leq \eps/2 )\leq e^{- \eps^{-2} \left(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2\right) t/12+1}, \qquad\text{for $t\geq 0$.}
\end{align*}
\end{lemma}
\begin{proof}
We proceed similarly to Lemma~4.2 in~\cite{AD09}. Let $\tau$ be the first exit time of $X$ out of $[-\eps,\eps]$. Then, by Wald's identity,
\begin{align*}
4 \eps^2 & \geq \limsup_{t\to\infty} \E[ X^2_{t\wedge \tau}] \geq \limsup_{t\to\infty} {\rm var}[ X_{t\wedge \tau}] \\ &= \limsup_{t\to\infty} \left(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2\right) \E[ t\wedge \tau] = \left(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2\right) \E [ \tau].
\end{align*}
Therefore,
$$
\P\left(\tau\geq 8 \eps^2 / \left(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2\right)\right) \leq \frac{\left(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2\right) \E [\tau]}{8 \eps^2} \leq \frac12.
$$
Let $n:=\lfloor t(\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2)/(8 \eps^2) \rfloor$ and set $t_i := 8i \eps^2 / (\int_{-\eps}^\eps x^2 \Pi(\d x) + \sigma^2)$, $i=0,\ldots, n$. Then
$$
\P( ||X||_t \leq \eps/2) \leq \P\left( \forall i=0, \ldots, n-1 : \sup_{s\in[t_i,t_{i+1})} |X_s -X_{t_i}|\leq \eps\right) = \P( \tau \geq t_1)^n \leq 2^{-n}.
$$
\end{proof}
This shows that the small deviation order is always at least as large as the term induced by the truncated variance process. This fact will be needed later on.
\begin{lemma}\label{lem:flargeru}
Let $F$ be a function that increases to infinity at zero. If for some L\'evy process $X$ for $t\leq t_0$ and $\eps<\eps_0$
\begin{align*}
-\log \P( ||X||_t \leq \eps) \leq F(\eps) t
\end{align*}
then, for some absolute constant $c>0$ and all $\eps>0$ small enough,
\begin{align*}
\eps^{-2} U(\eps) \leq c (F(\eps)+1).
\end{align*}
\end{lemma}
\begin{proof}
We use the assumption together with the fact that if $||X||_t\leq \eps$ then $X$ must not have jumps larger than $2\eps$ and the previous lemma:
\begin{align*}
e^{-F(\eps)t} \leq \P(||X||_t\leq \eps) = e^{-\bar\Pi(2\eps) t} \P(||X'||_t\leq \eps)
\leq e^{-\bar\Pi(2\eps) t} e^{- (2\eps)^{-2} \left(\int_{-2\eps}^{2\eps} x^2 \Pi(\d x) + \sigma^2\right) t/12+1},
\end{align*}
where $X'$ has L\'evy measure $\Pi$ restricted to $[-2\eps,2\eps]$. Noting that Lemma~5.1 of \cite{AD09} implies that $U(\eps)/\eps^2 \approx U(2\eps)/(2\eps)^2$, the statement of the lemma is proved.
\end{proof}
The lower bound in the LIL comes from the following lemma.
\begin{lemma}\label{lem:lower}
Let $F$ be a function that increases to infinity at zero such that for all $t\leq t_0$ and $\eps\leq \eps_0$
\begin{align*}
\lambda F(\eps)t \leq -\log \P(||X||_t\leq \eps)
\end{align*}
and, for $\lambda>0$, we set $b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)$. Then, for any $\lambda'<\lambda$,
\begin{align*}
1\leq \liminf_{t\rightarrow 0}\frac{||X||_t}{b_{\lambda'}(t)}\qquad \text{a.s.}
\end{align*}
\end{lemma}
\begin{proof}
For any $\lambda'<\lambda$, we can find $0<r<1$ such that $1< \lambda r/\lambda'$. Note that
\begin{align*}
\sum_{n}\P\big(||X||_{r^{n+1}}\leq b_{\lambda'}(r^n)\big)<\infty
\end{align*}
since
\begin{align}\label{est}
-\log \P\big(||X||_{r^{n+1}}\leq b_{\lambda'}(r^n)\big)\geq \lambda F( b_{\lambda'}(r^n)) r^{n} r = \lambda \frac{r}{\lambda'} \log |\log r^n| = \log n^{r \lambda/\lambda'} + {\rm const.}
\end{align}
Hence, by the Borel-Cantelli lemma,
\begin{align*}
\big\{n ~:~||X||_{r^{n+1}}\leq b_{\lambda'}(r^{n})\big\}
\end{align*}
is almost surely a finite set. Thus, for each path $\omega$, we have that for any $n\geq n_{0}(\omega)$ and any $t\in[r^{n+1},r^{n})$
\begin{align*}
&\frac{||X||_{t}}{b_{\lambda'}(t)}\geq\frac{||X||_{r^{n+1}}}{b_{\lambda'}(r^{n})}\geq 1,
\end{align*}
as $b_{\lambda'}$ is an increasing function. We take $\liminf_{t\to 0}$ to obtain the statement.
\end{proof}
The proof of the upper bound in the LIL requires the following lemma.
\begin{lemma}\label{L2}
Let $F$ be a function that increases to infinity at zero such that for all $t\leq t_0$ and $\eps\leq \eps_0$
\begin{align*}
-\log \P(||X||_t\leq \eps) \leq \lambda F(\eps) t
\end{align*}
and, for $\lambda>0$, set $b_\lambda(t):=F^{-1}\left(\frac{\log|\log t|}{\lambda t} \right)$. Assume that
\begin{align}\label{mladen}
\limsup_{n\to\infty}\frac{|X_{(n+1)^{-(n+1)^{\beta}}}|}{b_{\lambda}\big(n^{-n^{\beta}}\big)}=0\qquad \text{ a.s.},
\end{align}
for all $\beta>1$. Then, for any $\lambda'>\lambda$,
\begin{align} \label{eqn:upper1ml}
\liminf_{t\to 0}\frac{||X||_{t}}{b_{\lambda'}(t)}\leq 1\qquad\text{a.s.}
\end{align}
\end{lemma}
\begin{proof}
For $\lambda'>\lambda$, we choose $\beta>1$ such that $\lambda'>\lambda \beta$. First note that (\ref{mladen}) implies
\begin{align} \label{eqn:suparg1}
\limsup_{n\to\infty}\frac{||X||_{(n+1)^{-(n+1)^{\beta}}}}{b_{\lambda'}\big(n^{-n^{\beta}}\big)}=0\quad\text{ a.s.},
\end{align}
as $b_{\lambda}(t)$ is an increasing function in $\lambda$ for fixed $t\geq 0$.
Using the L\'evy property we see the following:
\begin{align*}
&\,\,\,\,\,\,\,\sum_{n}\P\Big(\sup_{(n+1)^{-(n+1)^{\beta}}\leq t < n^{-n^{\beta}}}|X_{t}-X_{(n+1)^{-(n+1)^{\beta}}}|\leq b_{\lambda'}\big(n^{-n^{\beta}}\big)\Big)\\
&=\sum_{n}\P\big(||X||_{n^{-n^{\beta}}-(n+1)^{-(n+1)^{\beta}}}\leq b_{\lambda'}\big(n^{-n^{\beta}}\big)\big)\\
&\geq \sum_{n}\P\big(||X||_{n^{-n^{\beta}}}\leq b_{\lambda'}\big(n^{-n^{\beta}}\big)\big)=\infty.
\end{align*}
The last step follows as in (\ref{est}) since now $\lambda \beta/\lambda'<1$. The Borel-Cantelli lemma shows that the sequence of independent events
\begin{align*}
A_{n}=\Big\{\sup_{(n+1)^{-(n+1)^{\beta}}\leq t < n^{-n^{\beta}}}|X_{t}-X_{(n+1)^{-(n+1)^{\beta}}}|\leq b_{\lambda'}\big(n^{-n^{\beta}}\big)\Big\}
\end{align*}
satisfies $\P(A_{n}\,\,\text{i.o.})=1$. To reduce to the supremum note that
\begin{align*}
\frac{||X||_{n^{-n^{\beta}}}}{b_{\lambda'}(n^{-n^{\beta}})}\leq \frac{\sup_{(n+1)^{-(n+1)^{\beta}}\leq t < n^{-n^{\beta}}}|X_{t}-X_{(n+1)^{-(n+1)^{\beta}}}|}{b_{\lambda'}(n^{-n^{\beta}})}+\frac{2||X||_{(n+1)^{-(n+1)^{\beta}}}}{b_{\lambda'}(n^{-n^{\beta}})}
\end{align*}
and therefore by (\ref{eqn:suparg1})
\begin{align*}
\liminf_{n\to\infty}\frac{||X||_{n^{-n^{\beta}}}}{b_{\lambda'}(n^{-n^{\beta}})}\leq\liminf_{n\to\infty}\frac{\sup_{(n+1)^{-(n+1)^{\beta}}\leq t < n^{-n^{\beta}}}|X_{t}-X_{n^{-(n+1)^{\beta}}}|}{b_{\lambda'}(n^{-n^{\beta}})}\leq 1.
\end{align*}
This shows (\ref{eqn:upper1ml}).
\end{proof}
Now we are in position to prove Theorem~\ref{t3}. For a detailed analysis of the $\limsup$ case, we refer to~\cite{S09}.
\begin{proof}[Proof of Theorem~\ref{t3}:]
The claim follows from Lemmas~\ref{lem:lower} and~\ref{L2}. To verify the use of Lemma~\ref{L2} we still need to check that condition (\ref{mladen}) holds for all $\beta>1$.
We fix $\beta>1$ and $\lambda'_2>\lambda_2$. Since $\lambda'_2$ is fixed, we suppress the subscript $\lambda'_2$ in the definition of $b$ in order to increase readability. We define the auxiliary function
\begin{align*
h(t)=b(\phi(t)),
\end{align*}
where $\phi(t)$ is chosen such that $\phi\big((\frac{t}{t+1})^{(\frac{t+1}{t})^{\beta}}\big)=t^{\frac{1}{t^{\beta}}}$ and $\phi(0)=0$. Note that $\phi$ is increasing and that $\phi(s^{-s^\beta})=(s-1)^{-(s-1)^\beta}$. We also do not mark that $\phi$ and $h$ depend on $\beta$ and $\lambda_2'$.
{\it Step 1:} We show that
\begin{align}\label{int}
\int_0^{1/2} \bar \Pi( h(t))\,\d t<\infty.
\end{align}
First, by the definition of $h$ and a change of variables we obtain
\begin{align*}
&\,\,\,\,\,\,\,\int_0^{1/2} \bar \Pi( h(t))\,\d t\\
&=\int_0^{ C(\beta)} \bar \Pi\big( b\big(s^{s^{-\beta}}\big)\big)\frac{d\big(\frac{s}{s+1}\big)^{(\frac{s+1}{s})^{\beta}}}{\d s}\\
&=\int_0^{ C(\beta)} \bar \Pi\big( b\big(s^{s^{-\beta}}\big)\big)\big(\frac{s}{s+1}\big)^{(\frac{s+1}{s})^{\beta}}(\frac{s+1}{s})^{\beta-1}s^{-2}(1-\beta\log{(1-(s+1)^{-1}}))\,\d s,
\end{align*}
which can be estimated from above by
\begin{align*}
&\,\,\,\,\,\,\,\,C\int_0^{ C(\beta)} \frac{b^{2}(s^{s^{-\beta}})\bar\Pi( b(s^{s^{-\beta}}))}{b^{2}(s^{s^{-\beta}})}\Big(\frac{s}{s+1}\Big)^{(\frac{s+1}{s})^{\beta}}s^{-1-\beta} |\log s|\d s\\
&\leq C\int_0^{ C(\beta)} \frac{U( b(s^{s^{-\beta}}))}{b^{2}(s^{s^{-\beta}})}\Big(\frac{s}{s+1}\Big)^{(\frac{s+1}{s})^{\beta}}s^{-1-\beta} |\log s| \d s\\
&\leq C'\int_0^{ C(\beta)} F( b(s^{s^{-\beta}}))\Big(\frac{s}{s+1}\Big)^{(\frac{s+1}{s})^{\beta}}s^{-1-\beta}|\log s| \d s\\
&=\frac{C'}{\lambda}\int_0^{ C(\beta)}\frac{\log\Big|\log s^{s^{-\beta}}\Big|}{s^{s^{-\beta}}}\Big(\frac{s}{s+1}\Big)^{(\frac{s+1}{s})^{\beta}}s^{-1-\beta}|\log s| \d s\\
&\leq \frac{C'}{\lambda}\int_0^{ C(\beta)}s^{-1-\beta} \left( \log\Big|\log s^{s^{-\beta}}\Big| \right) s^{(\frac{s+1}{s})^{\beta}-\frac{1}{s^{\beta}}} |\log s| \d s<\infty,
\end{align*}
where we have used $x^{2}\bar\Pi(x)\leq x^{2}\bar\Pi(x)+\int_{-x}^{x}y^{2}\Pi(\d y) + \sigma^2=U(x)\leq c x^2F(x)$ for some absolute $c>0$ by Lemma~\ref{lem:flargeru} and the definition of $b$.
{\it Step 2:} We denote by
\begin{align} \label{eqn:defnAn}
A_{n}:=\Big\{\text{ there are jumps with modulus larger than $b(n^{-n^{\beta}})$ up to time $(n+1)^{-(n+1)^{\beta}}$}\Big\}
\end{align}
and show that
\begin{align}\label{ba}
\sum_n \P\big(A_{n}\big)<\infty.
\end{align}
This comes from (\ref{int}). Indeed, note that $h$ inherits the monotonicity of $b$ and $\phi$ and hence (\ref{int}) implies that
\begin{align} \label{eqn:nozerosu}
&\sum_{n} \big((n+1)^{-(n+1)^{\beta}}-(n+2)^{-(n+2)^{\beta}}\big) \bar\Pi\big( h\big((n+1)^{-(n+1)^{\beta}}\big)\big)\leq\sum_{n}\int_{(n+2)^{-(n+2)^{\beta}}}^{(n+1)^{-(n+1)^{\beta}}}\bar\Pi(h(t))\d t<\infty.
\end{align}
Using
\begin{align*}
(n+1)^{-(n+1)^{\beta}}-(n+2)^{-(n+2)^{\beta}}&\sim(n+1)^{-(n+1)^{\beta}},\\
b(n^{-n^{\beta}})&=h((n+1)^{-(n+1)^{\beta}}),
\end{align*}
and that the sequence $(n+1)^{-(n+1)^{\beta}}\bar\Pi( h((n+1)^{-(n+1)^{\beta}}))$ tends to zero by (\ref{eqn:nozerosu}), we obtain that
\begin{align*}
\P\big(A_{n}\big)=1-e^{-(n+1)^{-(n+1)^{\beta}}\bar\Pi( b(n^{-n^{\beta}}))}&\sim (n+1)^{-(n+1)^{\beta}}\bar\Pi\big(h\big((n+1)^{-(n+1)^{\beta}}\big)\big)
\end{align*}
is summable. Therefore (\ref{ba}) is proved.
{\it Step 3:} Let us now show how to use (\ref{ba}) to deduce (\ref{mladen}). Apparently, it suffices to show that
\begin{align*}
\limsup_{n\rightarrow \infty}\frac{|X_{(n+1)^{-(n+1)^{\beta}}}|}{b(n^{-n^{\beta}})}<\eps\,\,\,\text{a.s.},
\end{align*}
for any $\eps>0$ and, hence, by the Borel-Cantelli lemma it suffices to show that
\begin{align*}
\sum_n \P\big({|X_{(n+1)^{-(n+1)^{\beta}}}|}>\eps{b\big(n^{-n^{\beta}}\big)}\big)<\infty.
\end{align*}
Separating jumps of absolute value larger or smaller than $ b\big(n^{-n^{\beta}}\big)$ and using the definition of $A_n$ in (\ref{eqn:defnAn}), we obtain that
\begin{align*}
&\,\,\,\,\,\,\,\sum_n \P\big({|X_{(n+1)^{-(n+1)^{\beta}}}|}>\eps{b(n^{-n^{\beta}})}\big)\\
&=\sum_n \P\big({|X_{(n+1)^{-(n+1)^{\beta}}}|}>\eps{b(n^{-n^{\beta}})}\,;\,A^{c}_{n}\big) + \sum_n \P\big({|X_{(n+1)^{-(n+1)^{\beta}}}|}>\eps{b(n^{-n^{\beta}})}\,;\, A_{n}\big),
\end{align*}
which is bounded from above by
\begin{align*}
&\sum_n \P\left(\left.{|X_{(n+1)^{-(n+1)^{\beta}}}|}>\eps{b(n^{-n^{\beta}})}\,\right|A^{c}_{n}\right)\cdot \P\big(A^{c}_{n}\big)+\sum_n \P\big(A_{n}\big).
\end{align*}
The second term is finite by (\ref{ba}); and the first term is bounded by
\begin{align} \label{eqn:gaptoolarge}
\sum_n \P\left(\left.\left|X_{(n+1)^{-(n+1)^{\beta}}}\right|>\eps b(n^{-n^{\beta}})\,\right|A^{c}_{n}\right).
\end{align}
To estimate this sum note that conditionally on $A^{c}_{n}$, $X_t\stackrel{d}=X_{t}(n)$, where $X(n)$ differs from $X$ only by removing jumps of size larger than $|b(n^{-n^{\beta}})|$. Clearly, by Wald's identity,
\begin{align*}
{\rm var}(X_{t}(n))=t\left(\int_{-b(n^{-n^{\beta}})}^{b(n^{-n^{\beta}})}y^{2}\Pi(\d y)+\sigma^{2}\right) \leq t U(b(n^{-n^{\beta}})).
\end{align*}
Note that
\begin{align*}
\E X_{(n+1)^{-(n+1)^{\beta}}}(n) = (n+1)^{-(n+1)^{\beta}} \left| \int_{|x|>b(n^{-n^\beta})} x \Pi(\d x) - \gamma\right|.
\end{align*}
Therefore, by assumption (\ref{eqn:conditionM}),
\begin{align*}
\E X_{(n+1)^{-(n+1)^{\beta}}}(n) = o( b(n^{-n^{\beta}}) ).
\end{align*}
Using this (first step), Chebychev's inequality (second step), Lemma~\ref{lem:flargeru} (third step), and the definition of $b$ (fourth step), we are led to the upper bound of the term in (\ref{eqn:gaptoolarge}):
\begin{align*}
&\sum_n \P\left({|X_{(n+1)^{-(n+1)^{\beta}}}(n)|}> \eps{b(n^{-n^{\beta}})}\right)\\
&\leq \sum_n \P\left({|X_{(n+1)^{-(n+1)^{\beta}}}(n) - \E X_{(n+1)^{-(n+1)^{\beta}}}(n)|}> \frac{1}{2} \,\eps{b(n^{-n^{\beta}})}\right)\\
&\leq \sum_n\frac{{(n+1)^{-(n+1)^{\beta}}}U\big(b(n^{-n^{\beta}})\big)}{(\eps/2)^2b(n^{-n^{\beta}})^2}\\
&\leq \sum_n\frac{{(n+1)^{-(n+1)^{\beta}}}C\cdot F\big(b(n^{-n^{\beta}})\big)}{(\eps/2)^2}\\
&=\frac{C'}{\lambda \eps^2}\sum_n\frac{{(n+1)^{-(n+1)^{\beta}}}\log |\log n^{-n^{\beta}}|}{n^{-n^{\beta}}}<\infty,
\end{align*}
where we used the definition of $b$ in the last step. Thus, the term in (\ref{eqn:gaptoolarge}) is finite, as required.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{cor:t1corollary}]
If $F$ is regularly varying so is $b_{\lambda}$, see \cite{bgt}, Proposition~1.5.7. Now note that if $F$ is regularly varying with exponent $-\alpha<0$, we have
\begin{align*}
b_{\lambda}(t)&=F^{-1}(\log|\log t|/\lambda t)\\
&\sim \lambda^{{1/\alpha}} F^{-1}(\log|\log t|/ t)\\
&=\lambda^{{1/\alpha}}b_1(t).
\end{align*}
Hence, the statement of Theorem~\ref{t3} reads
\begin{align*}
(\lambda'_1)^{{1/\alpha}}\leq \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{1}(t)}\leq (\lambda'_2)^{{1/\alpha}}\qquad\text{ a.s.}
\end{align*}
for all $\lambda_1'<\lambda_1$ and $\lambda_2'>\lambda_2$. Taking the limits on both sides we obtain
\begin{align*}
(\lambda_1)^{{1/\alpha}}\leq \liminf_{t\rightarrow 0}\frac{||X||_{t}}{b_{1}(t)}\leq (\lambda_2)^{{1/\alpha}}\qquad\text{ a.s.}
\end{align*}
Applying the regular variation argument in the reverse direction yields the claim.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{cor:sddirectsymmetric}]
This follows directly from Theorem~\ref{t3}. The bounds on the constants can be obtained from the absolute constants in Proposition~\ref{prop:ad}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{t}]
Lemma~\ref{lem:lower} gives the lower LIL of the theorem. Unfortunately, the arguments for the proof of Theorem~\ref{t3} do not apply here. Hence, for the reverse direction we show more directly that the given norming function of the LIL implies the rate function of the small deviations. The following arguments go back to Kesten. The proof is via contradiction assuming that
\begin{align}\label{eqn:ass}
\liminf_{t\to 0}\frac{||X||_{t}}{b_{\lambda_2'}(t)}>\frac{2}{C}+\delta
\end{align}
for some $\delta>0$ and $\lambda_2'>\lambda_2$. We show that under this assumption we can derive the estimates
\begin{align}
1&\geq\sum_{n\geq l}\P\bigg(\frac{||X||_{r^{j}-r^{n}}}{b_{\lambda_2'}(r^{j}-r^{n})}>\frac{2}{C};\text{for all $l\leq j\leq n-1$}\bigg)\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big)\label{eqn:schonwiedera}\\
&\geq\frac{1}{2}\sum_{n\geq l}\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big)\label{eqn:schonwiederb}
\end{align}
which is a contradiction as, by the choice of $b_{\lambda_2'}$ and the small deviation rate (\ref{eqn:yetanothersdestimate}), the sum in (\ref{eqn:schonwiederb}) is infinite. First, let us derive estimate (\ref{eqn:schonwiedera}) for which Assumption (\ref{eqn:ass}) is not needed. For any fixed integer $l$ partitioning the probability space we obtain
\begin{align*}
1&\geq \sum_{n\geq l}\P\big(||X||_{r^{j}}>b_{\lambda_2'}(r^{j})\text{ for all $l\leq j\leq n-1$};||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big)\\
&\geq\sum_{n\geq l}\P\Big(\sup_{r^{n}\leq s<r^{j}}|X_{s}|>b_{\lambda_2'}(r^{j})\text{ for all $l\leq j\leq n-1$};||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\Big).
\end{align*}
In order to employ the independence of increments of $X$ we estimate from below by
\begin{align*}
\sum_{n\geq l}\P\Big(\sup_{r^{n}\leq s<r^{j}}|X_{s}-X_{r^{n}}|>2 b_{\lambda_2'}(r^{j})\text{ for all $l\leq j\leq n-1$};||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\Big)
\end{align*}
which equals
\begin{align*}
&\,\,\,\,\,\,\,\sum_{n\geq l}\P\big(||X||_{r^{j}-r^{n}}>2 b_{\lambda_2'}(r^{j})\text{ for all $l\leq j\leq n-1$}\big)\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big)\\
&=\sum_{n\geq l}\P\Bigg(\frac{||X||_{r^{j}-r^{n}}}{b_{\lambda_2'}(r^{j}-r^{n})}>2 \frac{b_{\lambda_2'}(r^{j})}{b_{\lambda_2'}(r^{j}-r^{n})}\text{ for all $l\leq j\leq n-1$}\Bigg)\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big).
\end{align*}
By the monotonicity of $b_{\lambda_2'}$ this yields the lower bound
\begin{align*}
\sum_{n\geq l}\P\bigg(\frac{||X||_{r^{j}-r^{n}}}{b_{\lambda_2'}(r^{j}-r^{n})}>2 \frac{b_{\lambda_2'}(r^{j})}{b_{\lambda_2'}(r^{j}-r^{j+1})};\text{for all $l\leq j\leq n-1$}\bigg)\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big).
\end{align*}
Finally, we utilize the regularity of $b_{\lambda_2'}$ from (\ref{eqn:regularityofb}) to obtain the lower bound
\begin{align*}
\sum_{n\geq l}\P\bigg(\frac{||X||_{r^{j}-r^{n}}}{b_{\lambda_2'}(r^{j}-r^{n})}>\frac{2}{C};\text{for all $l\leq j\leq n-1$}\bigg)\P\big(||X||_{r^{n}}\leq b_{\lambda_2'}(r^{n})\big).
\end{align*}
As required we derived Estimate (\ref{eqn:schonwiedera}).
Assuming (\ref{eqn:ass}) we now derive Estimate (\ref{eqn:schonwiederb}). The assumption directly shows that
\begin{align*}
\lim_{t\to 0}\P\Big(\bigcap_{s\leq t}\big\{||X||_{s}\geq 2C^{-1} b_{\lambda_2'}(s)\big\}\Big)=1
\end{align*}
which implies that we may choose $l$ large enough such that
\begin{align*}
\P\Bigg(\frac{||X||_{r^{j}-r^{n}}}{b_{\lambda_2'}(r^{j}-r^{n})}>\frac{2}{C};\text{for all $l\leq j\leq n-1$}\Bigg)\geq\P\Big(\bigcap_{s\leq r^{l}}\big\{||X||_{s}\geq 2C^{-1} b_{\lambda_2'}(s)\big\}\Big)\geq \frac{1}{2}.
\end{align*}
Hence, we derived estimate (\ref{eqn:schonwiederb}) so that the proof is complete.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{cor:t1corollary2}] This is completely analogous to the proof of Corollary~\ref{cor:t1corollary}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{t2}] We use Proposition~\ref{prop:ad} and Theorem~\ref{t3}. In order to do so, we have to see that the term $\eps u_\eps$ in (\ref{eqn:sdquantityadstrengthend}) has no influence on the order. We apply Lemma~\ref{lem:lower} and the proof of Theorem~\ref{t} with the scaling
$$t=r^n \qquad\text{and}\qquad \eps=b(r^n)$$
and with the sequence $n^{-n^\beta}$, respectively. Therefore, it is sufficient to show that
$$\eps u_\eps =o( t F(\eps))$$
with the above scalings of $t$ and $\eps$. Since $\eps=b(t)$ and thus $t\sim F(\eps)^{-1} \log \log F(\eps)$, we need to show that
$$\eps u_\eps =o( \log \log F(\eps)).$$
As this is precisely what we stated in condition (\ref{eqn:cond-esschervanishes}), the proof is complete.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:bvnodrift}]
As $X$ is of bounded variation, the representation
\begin{align*}
X_t=A^1_t-A^2_t+ct
\end{align*}
holds with two independent pure jump subordinators $A^1, A^2$. Next, we use the simple observation
\begin{align*}
\frac{|X_t|}{t}\leq \frac{||X||_t}{t}\leq \frac{||A^1||_t+||A^2||_t+|c|t}{t}=\frac{A^1_t}{t}+\frac{A^2_t}{t}+|c|
\end{align*}
to conclude the proof. The left hand side converges to $|c|$ as $X$ has bounded variation (see Theorem~39 of \cite{D}). Finally, the right hand side converges to $|c|$ as $|A^i_t|/t$ converge at zero almost surely to their drift (see Proposition~5 of \cite{D}).
\end{proof}
\section*{Acknowledgment}
We thank Thomas Simon (Lille) for pointing out the reference \cite{W88} to the authors.
|
\section{Introduction}
\input{section_intro.tex}
\section{The CMS detector}
\input{section_detector.tex}
\section{Event selection}
\input{section_evtSel.tex}
\section{Reconstruction algorithms}
\input{section_reco.tex}
\subsection{Primary vertex reconstruction}
\input{section_reco_vertex.tex}
\subsection{Pixel cluster counting method}
\input{section_reco_clusters.tex}
\subsection{Pixel-tracklet method}
\input{section_reco_tracklets.tex}
\subsection{Tracking method}
\input{section_reco_tracks.tex}
\section{Results}
\input{section_results.tex}
\section{Discussion}
\input{section_discuss.tex}
\section{Summary}
\input{section_summary.tex}
\section*{Acknowledgements}
\input{section_acknowledge.tex}
\subsection{Charged hadron transverse-momentum distributions}
Tracks with $|\eta|<2.4$ and $\pt>0.1$~\GeVc\ were used for the
measurement of \dnchdpt. The measured average charged-hadron
yields per NSD event are shown in Fig.~\ref{fig:spectra}a, as a function
of \pt\ in bins of $|\eta|$. The yields were fit by the Tsallis function
(Eq.~\ref{eq:tsallis}), which empirically describes both the low-\pt\
exponential and the high-\pt\ power-law behaviours
\cite{Tsallis:1987eu,Wilk:2008ue}:
\begin{equation}
E \frac{d^3N_{\mathrm{ch}}}{d p^3} =
\frac{1}{2\pi p_T} \frac{E}{p} \frac{d^2N_{\mathrm{ch}}}{d\eta dp_T} =
C(n,T,m) \frac{dN_{\mathrm{ch}}}{dy}\left(1 + \frac{E_T}{nT}\right)^{-n} ,
\label{eq:tsallis}
\end{equation}
\noindent where $y=0.5\ln[(E+p_z)/(E-p_z)]$ is the rapidity;
$C(n,T,m)$ is a normalization constant that depends on $n$, $T$ and $m$;
$\ET = \sqrt{m^2 + \pt^2} - m$ and $m$ is the charged
pion mass. This function provides both the inverse slope parameter $T$,
characteristic for low \pt, and the exponent $n$, which parameterizes the
high-\pt\ power-law tail. These fit parameters change by less than 5\%
with $\eta$, thus a fit to the whole region $|\eta|<2.4$ was performed.
The \pt\ spectrum of charged hadrons,
$1/(2\pi \pt)d^2N_{\rm ch}/d\eta d \pt$, in the region $|\eta| < 2.4$,
was also fit with the empirical function (Eq.~\ref{eq:tsallis}) and is
shown in Fig.~\ref{fig:spectra}b.
The \pt resolution of the CMS tracker was found to have a negligible
effect on the measured spectral shape and was therefore ignored in the fit function.
For the 0.9~TeV data, the inverse slope parameter and the exponent
were found to be $T = 0.13 \pm 0.01$~\GeV and $n = 7.7 \pm 0.2$.
For the 2.36~TeV data, the values were $T = 0.14 \pm 0.01$~\GeV and
$n = 6.7 \pm 0.2$.
The average transverse momentum, calculated from the measured data points
adding the low- and high-\pt\ extrapolations from the fit is
$\langle \pt \rangle = 0.46 \pm 0.01$~(stat.)~$\pm$~0.01~(syst.)~\GeVc
for the 0.9~TeV and
$0.50 \pm 0.01$~(stat.)~$\pm$~0.01~(syst.)~\GeVc for the 2.36~TeV data.
The \dnchdeta\ spectrum was obtained by summing the measured differential
yields for $0.1 < \pt < 3.5$~\GeVc\ and adding the result to the integral
of the fit function for $\pt < 0.1$~\GeVc and $\pt > 3.5$~\GeVc.
The latter term amounts to 5\% of the total.
\begin{figure}
\begin{flushleft}
\subfigure{\label{fig:spectra_a}
\includegraphics[width=0.48\textwidth,height=0.5\textwidth]{figures/differential_hap_ham_combined_A}}
\hspace{0cm}
\subfigure{\label{fig:spectra_b}
\includegraphics[width=0.48\textwidth,height=0.5\textwidth]{figures/invariant_all_B}}
\end{flushleft}
\caption{(a) Measured differential yield of charged hadrons in the
range $|\eta| < 2.4$ in $0.2$-unit-wide bins of $|\eta|$ for the 2.36~TeV
data. The measured values with systematic uncertainties (symbols) and the
fit functions (Eq.~\ref{eq:tsallis}) are displayed. The values with
increasing $\eta$ are successively shifted by four units along the
vertical axis.
(b) Measured yield of charged hadrons for $|\eta| < 2.4$ with systematic
uncertainties (symbols), fit with the empirical function
(Eq.~\ref{eq:tsallis}).
}
\label{fig:spectra}
\end{figure}
\subsection{Charged hadron pseudorapidity density}
The summary of results on the pseudorapidity density distribution of
charged hadrons is shown in Fig.~\ref{dndeta_results}.
The \dnchdeta\ results for the three layers in the cluster-counting method
and the three layer-pairs in the pixel-tracklet method are consistent
within 3\%. These results from the various layers and from the different
layer pairs were combined to provide one set of data from each analysis
method, as shown in Fig.~\ref{dndeta_results}a.
The error bars include the systematic uncertainties of about
2.4--4.4\% specific to each method, estimated from the variations of
model parameters in the simulation used for corrections and the
uncertainties in the data-driven corrections. The systematic uncertainties
common to all the three methods, which amount to 3.2\%, are not shown.
The results from the three analysis methods are in agreement.
The larger fraction of background hits in the data
compared to simulation affects the cluster-counting method differently
from the other two, which results in a small difference at high $\eta$,
well accounted for by the systematic uncertainty of the measurement.
|
\section{Introduction}
The object of our study is the massless field on $D_n = D \cap \tfrac{1}{n} \mathbf{Z}^2$, with Hamiltonian $\mathcal {H}(h) = \sum_{b \in D_n^*} \mathcal {V}(\nabla h(b))$. Here, $D \subseteq \mathbf{R}^2$ is a bounded domain with smooth boundary. The sum is over all edges in the induced subgraph of $\tfrac{1}{n} \mathbf{Z}^2$ with vertices in $D$ and $\nabla h(b) = h(y) - h(x)$ denotes the discrete gradient of $h$ across the oriented bond $b=(x,y)$. We take our boundary conditions to be a continuous perturbation of a macroscopic tilt: $h(x) = nu \cdot x + f(x)$ when $x \in \partial D_n$ for $u \in \mathbf{R}^2$ and $f \colon \mathbf{R}^2 \to \mathbf{R}$ is a continuous function. We consider a general interaction $\mathcal {V} \in C^2(\mathbf{R})$ which is assumed only to satisfy:
\begin{enumerate}
\item $\mathcal {V}(x) = \mathcal {V}(-x)$ (symmetry),
\item $0 < a_\mathcal {V} \leq \mathcal {V}''(x) \leq A_\mathcal {V} < \infty$ (uniform convexity), and
\item $\mathcal {V}''$ is $L$-Lipschitz.
\end{enumerate}
The purpose of the first condition is merely to simplify the notation since the symmetrization of a non-symmetric potential does not change its behavior. Note that we can assume without loss of generality that $\mathcal {V}(0) = 0$. This is the so-called \emph{Ginzburg-Landau $\nabla \phi$ effective interface (GL) model}, also known as the \emph{anharmonic crystal}. The variables $h(x)$ represent the heights of a random surface which serves as a model for an interface separating two pure phases.
The macroscopic behavior of the GL model has been the subject of much recent study. An important step in this development is the construction and classification of \emph{Gibbs states}, which are infinite volume versions of the model. In two-dimensions, it turns out that the height variable $h(x)$ diverges as the size of the domain tends to infinity, so in order to construct a Gibbs state one must first pass to the gradient field $\nabla h$. The existence and uniqueness of \emph{gradient Gibbs states} is proved by Funaki and Spohn in \cite{FS97}, where they also study macroscopic dynamics. Deuschel, Giacomin, and Ioffe in \cite{DGI00} establish a large deviations principle for the surface shape with zero boundary conditions but in the presence of a chemical potential and Funaki and Sakagawa in \cite{FS04} extend this result to the case of non-zero boundary conditions using the contraction principle. The behavior of the maximum is studied by Deuschel and Giacomin in \cite{DG00} and by Deuschel and Nishikawa in \cite{DN07} in the case of Langevin dynamics. The central limit theorem for linear functionals of the infinite gradient Gibbs states was proved first by Naddaf and Spencer for the static model with zero tilt in \cite{NS97} and Giacomin, Olla, and Spohn handle the time-varying case with general tilt evolving under Langevin dynamics in \cite{GOS01}.
The special situation in which the interaction is quadratic, i.e. $\mathcal {V}(x) = \tfrac{1}{2} x^2$, corresponds to the so-called \emph{discrete Gaussian free field} (DGFF) or \emph{harmonic crystal}. Many of the results in \cite{DG00, DN07, FS04, FS97, GOS01, NS97} have been established separately for the quadratic case and often the results in this setting are more refined. The reason for the latter is that its Gaussian structure greatly simplifies its analysis; we will discuss this point in more detail later. For the DGFF, large deviations principles for the surface shape as well as a central limit theorem for the height variable was proved by Ben Arous and Deuschel in \cite{BAD96}, the behavior of the maximum studied by Bolthausen, Deuschel, and Giacomin in \cite{BDG01} and by Daviaud in \cite{DAV06}. In a particularly impressive and difficult work, Schramm and Sheffield in \cite{SS06} show that the macroscopic level sets are described by a family of conformally invariant random curves which are variants of $SLE(4)$.
Beyond having a Gaussian distribution, the main feature that makes the analysis of the DGFF tractable is that its mean and covariance are completely described in terms of the harmonic measure and Green's function associated with a simple random walk. These objects are very well understood in the planar case \cite{LAW91}, which often allows for very precise estimates. The mean and covariance of the more general GL model also admit a representation in terms of a random walk (Helffer-Sj\"ostrand representation or HS random walk \cite{HS94, DGI00}). The general situation, however, is much more difficult because in addition to being non-Guassian, the corresponding representations involve a \emph{random walk in a dynamic random environment} whose behavior depends non-trivially on the boundary data. The hypothesis that $\mathcal {V}$ is uniformly convex is helpful in assuring that the jump rates of the HS random walk are uniformly bounded from above and below. This implies that its Green's function is comparable to that of a simple random walk, which in turn allows for rough (up to multiplicative constants) variance estimates and, more generally, centered moments in terms of the corresponding moments for the DGFF (Brascamp-Lieb inequalities). Furthermore, that the jump rates are bounded means that the Nash-Aronson and Nash continuity estimates apply, which give some rough control of the off-diagonal covariance structure.
\subsection{Main Results}
The main result of this article is the following central limit theorem for linear functionals of the height. Let $D \subseteq \mathbf{R}^2$ be a bounded domain with smooth boundary. For $\kappa \geq 0$, let $H^\kappa(D)$ be the Sobolev space of degree $\kappa$ and $H^{-\kappa}(D)$ its Banach space dual.
\begin{theorem}[Central Limit Theorem]
\label{thm::clt}
Let $f \colon \mathbf{R}^2 \to \mathbf{R}$ be a continuous function. Suppose that $D_n = D \cap \tfrac{1}{n} \mathbf{Z}^2$ and that $h^n$ is distributed according to the GL model on $D_n$ with $h^n(x) = \varphi_n(x) + f(x)$ for all $x \in \partial D_n$ where $\varphi_n(x) = n u \cdot x$. Let $a_u(b) = \mathbf{E}[ \mathcal {V}''(\eta(b))]$ where $\eta$ has the law of the Funaki-Spohn state with tilt $u$. Define the linear functional
\[ \xi_\nabla^{n,D}(g) = \sum_{b \in D_n^*} a_u(b) \nabla g(b) \nabla (h^n - \varphi_n)(b) \text{ for } g \in H^{\kappa}(D).\]
For every $\kappa > 4$, the law of $\xi_\nabla^{n,D}$ on $H^{-\kappa}(D)$ converges weakly with respect to the weak topology of $H^{-\kappa}(D)$ to a Gaussian free field on $D$, the standard Gaussian with respect to the weighted Dirichlet inner product $(g_1,g_2)_\nabla^\beta = \int_D \sum_i \beta_i \partial_i g_1 \partial_i g_2$, with boundary condition $f$, where $(\beta_1,\beta_2)$ is proportional to $(\mathbf{E}[\mathcal {V}''( \eta(0,e_1))], \mathbf{E}[\mathcal {V}''( \eta(0,e_2))])$. In particular, $\beta_1 = \beta_2 > 0$ when $u = 0$.
\end{theorem}
\noindent We will recall both the notion of a Sobolev space and the GFF in subsection \ref{subsec::gff}. We will also review the notion of a Funaki-Spohn state (as well as a new construction for zero-tilt) in subsection \ref{subsec::shift_invariant}.
As we mentioned earlier, central limit theorems have already been established by Naddaf and Spencer in \cite{NS97} and Giacomin, Olla, and Spohn in \cite{GOS01} for linear functionals of infinite gradient Gibbs states of the GL model. Both articles are based on the beautiful observation that the CLT can be reduced to a homogenization problem using the HS representation. The proof in \cite{GOS01} has more of a probabilistic flavor while the approach in \cite{NS97} is to use PDE techniques. The reason that these results are restricted to infinite gradient Gibbs states is that the homogenization techniques they employ fail to carry over to the finite case. In particular, the main step in \cite{GOS01} is a proof that the macroscopic covariance structure of the gradient Gibbs state for the GL model is the same as that in the GFF by showing that the HS random walk converges in the limit to a Brownian motion. The key tool here is the so-called Kipnis-Varadhan method \cite{KV86}, which is to represent the random walk as an additive functional of the environment from the perspective of the walker. When the environment is an infinite, stationary, ergodic Markov process then it remains so when viewed from the walker, thus the convergence to Brownian motion is a consequence of Corollary 1.5 of \cite{KV86}. If the environment is finite and, in particular, \emph{not ergodic with respect to shifts}, this approach can no longer be used since \emph{the environment viewed from the particle is not ergodic}.
The covariance matrix of the limiting Gaussian in \cite{GOS01} is given in terms of a complicated variational formula. It is therefore not explicit, except in the case of zero-tilt where it is possible to argue that it is proportional to the identity using rotational invariance. The proof of Theorem \ref{thm::clt} gives the covariance matrix explicitly, up to a multiplicative constant, which is another new result for gradient Gibbs states.
The careful reader may note that the convergence in Corollary 2.2. of \cite{GOS01} is in $H^{-\kappa}(D)$ for $\kappa > 3$ while we require $\kappa > 4$. The reason for the distinction is that we assume only continuity of the boundary condition $f$. This forces us to perform an extra integration by parts, which in turn puts an extra derivative on the test function. In the more restrictive setting of $C^1$ boundary conditions, our proof also gives convergence in $H^{-\kappa}(D)$ for $\kappa > 3$.
The main step in our proof is motivated by the Markovian structure enjoyed by the quadratic case: the law of a DGFF on $D_n$ with boundary condition $f_n$ is equal in law to that of a \emph{zero-boundary} DGFF on $D_n$ plus the discrete harmonic extension of $f_n$ to $D_n$. This property is a higher dimensional analog of the fact that a random walk $X_t$ on $\mathbf{Z}$ conditioned to satisfy $X_{t_1} = x_1$ and $X_{t_2} = x_2$ for $t_1 < t_2$ has the law of $Y_t + H_t$ where $Y_t$ is a random walk on $I = [t_1,t_2]$ conditioned to vanish at $t_1,t_2$ and $H$ is the discrete harmonic function on $I$ with boundary values $H(t_1) = X_{t_1}$ and $H_{t_2} = X_{t_2}$. Our next theorem is a quantitative estimate of the degree to which this property approximately holds for the GL model. Although we state it as our second theorem, it is the key step in the proof of Theorem \ref{thm::clt} and much of the article is dedicated to its proof. In order to give a precise statement of this result, we first need to setup some notation.
Suppose that $D \subseteq \mathbf{Z}^2$ is a bounded subset of diameter $R > 0$. Fix $\overline{\Lambda} > 0$ and let $\mathbf{B}_{\overline{\Lambda}}^u(D)$ be the set of functions $\phi \colon \partial D \to \mathbf{R}$ satisfying $\max_{x \in \partial D} | \phi(x) - u \cdot x| \leq \overline{\Lambda} (\log R)^{\overline{\Lambda}}$. For $r > 0$, let $D(r) = \{ x \in D : {\rm dist}(x,\partial D) \geq r\}$. With $\phi \in \mathbf{B}_{\overline{\Lambda}}^u(D)$, let $\mathbf{P}_D^\phi$ denote the law of the GL model on $D$ with boundary condition $\phi$. In other words, $\mathbf{P}_D^\phi$ is the measure on functions $h \colon D \to \mathbf{R}$ with density
\[ \frac{1}{\mathcal {Z}} \exp\left( - \sum_{b \in D^*} \mathcal {V}(\nabla (h \vee \phi)(b)) \right)\]
with respect to Lebesgue measure on $\mathbf{R}^{|D|}$. Here, $h \vee \phi$ is used to denote the function
\begin{equation}
\label{intro::eqn::vee} h \vee \phi (x) = \begin{cases} h(x) \text{ if } x \in D,\\
\phi(x) \text{ if } x \in \partial D. \end{cases}
\end{equation}
We will write $O_{\overline{\Lambda}}(f(x))$ to denote the set of functions $g$ for which there exists a constant $c_{\overline{\Lambda}}$ depending on $\overline{\Lambda}$ but independent of $R$ so that $|g(x)| \leq c_{\overline{\Lambda}}|f(x)|$. For $\beta = (\beta_1,\beta_2)$, let
\[ (\Delta^\beta f)(x) =
\beta_1 ( f(x+e_1) + f(x-e_1) - 2 f(x)) +
\beta_2 ( f(x+e_2) + f(x-e_2) - 2 f(x))\]
where $e_1 = (1,0)$ and $e_2 = (0,1)$.
Note that $\Delta^\beta$ is the usual discrete Laplacian for $\beta = (1,1)$.
\begin{theorem}
\label{harm::thm::coupling}
Suppose that $\psi,\widetilde{\psi} \in \mathbf{B}_{\overline{\Lambda}}^u(D)$ and $\beta = \beta(u)$ as in the statement of Theorem \ref{thm::clt}. There exists $C,\epsilon,\delta > 0$ depending only on $\mathcal {V}$ such that if $r \geq CR^{1-\epsilon}$ then the following holds. There exists a coupling $(h^\psi, h^{\widetilde{\psi}})$ of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$ such that if $\widehat{h} \colon D(r) \to \mathbf{R}$ solves the elliptic problem $\Delta^\beta \widehat{h} = 0$ with $\widehat{h}|_{\partial D(r)} = \overline{h} = h^\psi - h^{\widetilde{\psi}}$ then
\[ \mathbf{P}[ \overline{h} \neq \widehat{h} \text{ in } D(r)] = O_{\overline{\Lambda}}(R^{-\delta}).\]
When $u = 0$, we can take $\beta = (1,1)$ so that $\Delta^\beta$ is the usual Laplacian.
\end{theorem}
One of the main challenges in the analysis of the GL model is the lack of useful comparison inequalities for its mean. The difficulty is that the only explicit formula is given in terms of the annealed first exit distribution of the HS walk \cite{DGI00}. It is not possible to extract any sort of asymptotic contiguity of this measure with respect to the harmonic measure of simple random walk using only that the HS walk jumps with bounded rates, which is all that is required to prove comparability of the corresponding Green's functions hence also of centered moments with DGFF. Indeed, examples have been worked out in the continuum setting of diffusions in which the two measures are absolutely singular and that the support of the former has a fractal structure. The situation is further complicated in the setting of the HS walk since in addition to being dynamic, its jump rates also depend on the boundary conditions, hence it seems difficult to rule out pathological behavior whenever the walk gets close to the boundary and the boundary conditions are rough.
Applying Theorem \ref{harm::thm::coupling} to the special case $\widetilde{\psi}(x) = u \cdot x$ gives the following estimate of the mean, which we believe to be sufficiently important that we state it as a separate theorem.
\begin{theorem}
\label{harm::thm::mean_harmonic}
Suppose that $\psi \in \mathbf{B}_{\overline{\Lambda}}^u(D)$. There exists $C, \epsilon,\delta > 0$ such that if $r \geq CR^{1-\epsilon}$ and $\beta = \beta(u)$ as in the statement of Theorem \ref{thm::clt} then the following holds. If $\widehat{h} \colon D(r) \to \mathbf{R}$ is the $\Delta^\beta$-harmonic extension of $\mathbf{E}^\psi h$ from $\partial D(r)$ to $D(r)$ then
\[ \max_{x \in D(r)} |\mathbf{E}^\psi h(x) - \widehat{h}(x)| = O_{\overline{\Lambda}}(R^{-\delta}).\]
When $u = 0$, we can take $\beta = (1,1)$ so that $\widehat{h}$ is harmonic with respect to the usual discrete Laplacian.
\end{theorem}
It is worth pointing out that both of these theorems place no restrictions on the regularity of the boundary conditions $\psi, \widetilde{\psi}$ nor the regularity of $\partial D$.
\begin{comment}
Recalling the relationship between the Green's function of the HS random walk and the covariance structure of the underlying field, Theorem \ref{thm::clt} implies that the Green's function of the HS random walk is the same as that of a Brownian motion. Using this observation, we are able to prove the homogenization of the HS random walk.
\begin{theorem}
\label{thm::hs_homogenization}
Supposing we are in the setting of Theorem \ref{thm::clt} and that $X_t^n$ denotes the HS random walk associated with the dynamics of $h^n$. Then $X_t^n$ stopped on its first exit from $D_n$ converges weakly to a standard Brownian motion on $D$ stopped on its first exit from $D$.
\end{theorem}
\end{comment}
\subsection{Sequel} This article the first in a series of two and will be a prerequisite for the second. In the sequel, we will make use of many of the estimates developed here in order to resolve a conjecture made by Sheffield (Problem 10.1.3 in \cite{SHE05}) that the macroscopic level lines of the GL model converge in the limit to $SLE(4)$; the case of quadratic potentials is proved by Schramm and Sheffield in \cite{SS06}. The two papers together are meant to be fairly self-contained.
\subsection{Outline} The remainder of the article is structured as follows. The second section is a short discussion of discrete and continuum Gaussian free fields. We chose to include the former part of this section since the special Markovian structure of the DGFF is the inspiration for Theorem \ref{harm::thm::coupling} and also to serve as an illustration of the complications associated with non-quadratic interaction. In the latter part, we provide a brief description of the GFF, the standard Gaussian law on $H_0^1(D)$. A much more thorough introduction can be found in \cite{SHE06}. In Section \ref{sec::gl}, we will give a formal introduction to the GL model, its Langevin dynamics as well as the HS representation, and the Brascamp-Lieb inequalities. In Section \ref{sec::dyn}, we will explain how the Langevin dynamics can be used to construct couplings of the GL model and prove an energy inequality for the discrete Dirichlet energy of such a coupling. This section is concluded with an equivalence of ensembles result: the Funaki-Spohn shift-ergodic gradient Gibbs state can be realized as an infinite volume limit of models on finite domains. In Section \ref{sec::harm}, we will prove Theorems \ref{harm::thm::coupling} and \ref{harm::thm::mean_harmonic} using an entropy estimate which is based on technical estimates from Section \ref{sec::correlation_decay}. Finally, Section \ref{sec::clt} is relatively short and deduces the CLT from Theorem \ref{harm::thm::coupling}. We conclude the article with two appendices containing useful estimates on discrete harmonic functions and symmetric random walks.
\section{Gaussian Free Fields}
In this section we will introduce the discrete and continuum Gaussian free fields (DGFF and GFF). The reason that we include a discussion of the latter separate from the general case of the GL model is to emphasize its special Markovian structure, which is the motivation behind the ideas used in Section \ref{sec::harm}.
\subsection{Discrete Gaussian Free Field}
\label{subsec::dgff_construction}
Suppose that $G = (V \cup \partial V,E)$ is a finite, undirected, connected graph with distinguished subset $\partial V \neq \emptyset$ and edge weights $\omega > 0$. The zero-boundary discrete Gaussian free field (DGFF) is the measure on functions $h \colon V \cup \partial V \to \mathbf{R}$ vanishing on $\partial V$ with density
\[ \frac{1}{\mathcal {Z}_G} \exp\left(-\frac{1}{2} \sum_{b \in E} \omega(b) (\nabla (h \vee 0)(b))^2\right)\]
with respect to Lebesgue measure. Here, $h \vee 0$ has the same meaning as in \eqref{intro::eqn::vee} and $\mathcal {Z}_G$ is a normalizing constant so that the above has unit mass. Equivalently, the DGFF is the standard Gaussian associated with the Hilbert space $H_0^1(V)$ of real-valued functions $h$ on $V$ vanishing on $\partial V$ with weighted Dirichlet inner product
\[ (f,g)_\nabla^\omega = \sum_{b \in E} \omega(b) \nabla f(b) \nabla g(b).\]
This means that the DGFF $h$ can be thought of as a family of Gaussian random variables $(h,f)_\nabla^\omega$ indexed by elements $f \in H_0^1(V)$ with mean zero and covariance
\begin{equation}
{\rm Cov}((h,f)_\nabla^\omega,(h,g)_\nabla^\omega) = (f,g)_\nabla^\omega,\ f,g \in H_0^1(V) \label{dgff::covariance}.
\end{equation}
Although perhaps non-standard since our Hilbert space is finite dimensional, this representation is convenient since it allows for a simple derivation of the mean and covariance of $h$. Let $\Delta^\omega \colon V \to \mathbf{R}$ denote the discrete Laplacian on $V$, i.e.
\[ \Delta^\omega f(x) = \sum_{b \ni x} \omega(b) \nabla f(b)\]
and let $G^\omega(x,y) = (\Delta^\omega)^{-1} \mathbf{1}_{\{x\}}(y)$ be the discrete Green's function on $V$. Summation by parts gives that
\[ (f,g)_\nabla^\omega = -\sum_{x \in V} f(x) \Delta^\omega g(x) = -\sum_{x \in V} \Delta^\omega f(x) g(x) \text{ for } f,g \in H_0^1(V).\]
Thus
\[h(x) = (h,\mathbf{1}_{\{x\}}(\cdot))_{L^2} = -(h,G^\omega(x,\cdot))_\nabla,\]
hence
\[ {\rm Cov}(h(x),h(y)) = (G^\omega(x,\cdot),G^\omega(y,\cdot))_\nabla = G^\omega(x,y).\]
Suppose that $W \subseteq V$. Then $H_0^1(V)$ admits the orthogonal decomposition $H_0^1(V) = \mathcal {M}_I \oplus \mathcal {M}_B \oplus \mathcal {M}_O$ where $\mathcal {M}_I,\mathcal {M}_B,\mathcal {M}_O$ are the subspaces of $H_0^1(V)$ consisting of those functions that vanish on $V \setminus W$, are $\Delta^\omega$-harmonic off of $\partial W$, and vanish on $W$, respectively. It follows that we can write $h = h_I + h_B + h_O$ with $h_I \in \mathcal {M}_I, h_B \in \mathcal {M}_B, h_O \in \mathcal {M}_O$ where $h_I,h_B,h_O$ are independent. This implies that the DGFF possesses the following Markov property: the law of $h|_W$ conditional on $h|_{V \setminus W}$ is that of a zero boundary DGFF on $W$ plus the $\Delta^\omega$-harmonic extension of $h$ from $\partial W$ to $W$. In particular, the conditional mean of $h|_W$ given $h|_{V \setminus W}$ is the $\Delta^\omega$-harmonic extension of $h|_{\partial W}$ to $W$.
More generally, if $\phi \colon \partial V \to \mathbf{R}$, the DGFF with boundary condition $\phi$ is the measure on functions $h \colon V \to \mathbf{R}$ with $h|_{\partial V} = \phi$ with density
\[ \frac{1}{\mathcal {Z}_G} \exp\left( -\frac{1}{2} \sum_{b \in E} \omega(b)(\nabla (h \vee \phi)(b))^2 \right).\]
That is, $h$ has the law of a zero boundary DGFF on $V$ plus the $\Delta^\omega$-harmonic extension of $\phi$ from $\partial V$ to $V$.
\subsection{The Continuum Gaussian Free Field}
\label{subsec::gff}
The GFF is a $2$-time dimensional analog of the Brownian motion. Just as the Brownian motion can be realized as the scaling limit of many random curve ensembles, the GFF arises as the scaling limit of a number of random surface ensembles \cite{BAD96, GOS01, KEN01, NS97, RV08}, as well as the model under consideration in this article. In this subsection, we will describe the basic properties of the GFF necessary for our analysis. Let $D$ be a bounded domain in $\mathbf{R}^2$ with smooth boundary and let $C_0^\infty(D)$ denote the set of $C^\infty$ functions compactly supported in $D$. We begin with a short discussion of Sobolev spaces; the reader is referred to Chapter 5 of \cite{EVAN02} or Chapter 4 of \cite{TAY96} for a more thorough introduction. With $\mathbf{N}_0 = \{0,1,\ldots\}$ the non-negative integers, when $f \in C_0^\infty(D)$ and $\alpha = (\alpha_1,\alpha_2) \in \mathbf{N}_0^2$ we let $D^\alpha f = \partial_1^{\alpha_1} \partial_2^{\alpha_2} f$. For $k \in \mathbf{N}_0$ we define the $H^k(D)$-norm
\begin{equation}
\label{eqn::sobolev}
\|f\|_{H^k(D)}^2 = \sum_{|\alpha| \leq k} \int_D |D^\alpha f(x)|^2 dx
\end{equation}
where $|\alpha| = \alpha_1 + \alpha_2$. The Sobolev space $H_0^k(D)$ is the Banach space closure of $C_0^\infty(D)$ under $\Vert \cdot \Vert_{H^k(D)}$. If $s \geq 0$ is not necessarily an integer then $H_0^s(D)$ can be constructed via the complex interpolation of $H_0^0(D) = L^2(D)$ and $H_0^k(D)$ where $k \geq s$ is any positive integer (see Chapter 4 section 2 of \cite{TAY96} for more on this construction and also Chapter 4 of \cite{KAT04} for more on interpolation). A consequence of this is that if $T \colon C_0^\infty(D) \to C_0^\infty(D)$ is a linear map continuous with respect to the $L^2(D)$ and $H^k(D)$ topologies then it is also continuous with respect to $H^s(D)$ for all $0 \leq s \leq k$. For $s \geq 0$ we define $H^{-s}(D)$ to be the Banach space dual of $H_0^s(D)$ where the dual pairing of $f \in H^{-s}(D)$ and $g \in H_0^s(D)$ is given formally by the usual $L^2(D)$ inner product
\[ (f,g) = (f,g)_{L^2(D)} = \int_D f(x) g(x) dx.\]
More generally, for any $s \in \mathbf{R}$ the $H^s(D)$-topology can be constructed explicitly via the inner product
\begin{equation}
\label{gff::eqn::sobolev_inner_product}
(f,g)_{s} = \int (1-\overline{\Delta})^{s/2} f \cdot (1-\overline{\Delta})^{s/2} g;
\end{equation}
see the introduction of Chapter 4 of \cite{TAY96}.
We are using the notation $\overline{\Delta}$ for the Laplacian on $\mathbf{R}^2$ to keep the notation consistent since elsewhere in the article $\Delta$ refers to the discrete Laplacian.
Here,
\[ (1-\overline{\Delta})^p f = \mathcal {F}^{-1} [(1+\xi_1^2 + \xi_2^2)^p (\mathcal {F} f)] \text{ for } p \in \mathbf{R}\]
where
\[ \mathcal {F} f (\xi) = \int e^{-i \xi \cdot x} f(x) dx\]
is the Fourier transform of $f$. We will be most interested in the space $H_0^1(D)$. Fix a positive definite $2 \times 2$ real matrix $A$. An application of the Poincare inequality (Chapter 4, Proposition 5.2) gives that the norm induced by the weighted Dirichlet inner product
\[ (f,g)_\nabla^A \equiv \int_D \sum_{i,j} a_{ij} \partial_i f \partial_j g \text{ for } f,g \in C_0^\infty(D)\]
is equivalent to $\| \cdot \|_{H^1(D)}$. This choice of inner product is particularly convenient because it is invariant under precomposition by conformal transformations when $A$ is a multiple of the identity.
The $A$-GFF $h$ on $D$ can be expressed formally as a random linear combination of an $(\cdot,\cdot,)_\nabla^A$-orthonormal basis $(f_n)$ of $H_0^1(D)$
\[ h = \sum_n \alpha_n f_n\]
where $(\alpha_n)$ is an iid sequence of standard Gaussians. Although the sum defining $h$ does not converge in $H_0^1(D)$, for each $\epsilon > 0$ it does converge almost surely in $H^{-\epsilon}(D)$ (\cite[Proposition 2.7]{SHE06} and the discussion thereafter). If $f,g \in C_0^\infty(D)$ then an integration by parts gives $(f,g)_\nabla^A = -( f,\overline{\Delta}^A g)$. Here, $\overline{\Delta}^A = \overline{\nabla} A \overline{\nabla}$. Using this, we define
\[ (h,f)_\nabla^A = -(h, \overline{\Delta}^A f) \text{ for } f \in C_0^\infty(D).\]
Observe that $(h,f)_\nabla^A$ is a Gaussian random variable with mean zero and variance $(f,f)_\nabla^A$. Hence by polarization $h$ induces a map $C_0^\infty(D) \to \mathcal {G}$, $\mathcal {G}$ a Gaussian Hilbert space, that preserves the Dirichlet inner product. This map extends uniquely to $H_0^1(D)$ which allows us to make sense of $(h,f)_\nabla^A$ for all $f \in H_0^1(D)$. We are careful to point out, however, that while $(h, \cdot)_\nabla^A$ is well-defined off of a set of measure zero as a linear functional on $C_0^\infty(D)$ this is not the case for general $f \in H_0^1(D)$.
Suppose that $W \subseteq D$ is a smooth, open set. Then there is a natural inclusion of $H_0^1(W)$ into $H_0^1(D)$ given by the extension by value zero. If $f \in C_0^\infty(W)$ and $g \in C_0^\infty(D)$ then as $(f,g)_\nabla^A = -(f,\overline{\Delta}^A g)$ it is easy to see that $H_0^1(D)$ admits the $(\cdot,\cdot)_\nabla^A$-orthogonal decomposition $\mathcal {M} \oplus \mathcal {N}$ where $\mathcal {M} = H_0^1(W)$ and $\mathcal {N}$ is the set of functions in $H_0^1(D)$ that are $\overline{\Delta}^A$-harmonic on $W$. Thus we can write
\[ h = h_W + h_{W^c} = \sum_n \alpha_n f_n + \sum_n \beta_n g_n\]
where $(\alpha_n),(\beta_n)$ are independent iid sequences of standard Gaussians and $(f_n)$, $(g_n)$ are orthonormal bases of $\mathcal {M}$ and $\mathcal {N}$, respectively. Observe that $h_W$ has the law of the GFF on $W$, $h_{W^c}$ the $\overline{\Delta}^A$-harmonic extension of $h|_{\partial W}$ to $W$, and $h_W$ and $h_{W^c}$ are independent. We arrive at the following proposition:
\begin{proposition}[Markov Property]
\label{gff::prop::markov}
The conditional law of $h|_W$ given $h |_{ D \setminus W}$ is that of the $A$-GFF on $W$ plus the $\overline{\Delta}^A$-harmonic extension of the restriction of $h$ on $\partial W$ to $W$.
\end{proposition}
This proposition will be critical in the proof of Theorem \ref{thm::clt}. It also allows us to make sense of the $A$-GFF with non-zero boundary conditions: if $f \colon \partial D \to \mathbf{R}$ is a continuous function and $F$ is its $\overline{\Delta}^A$-harmonic extension from $\partial D$ to $D$ then the law of the $A$-GFF on $D$ with boundary condition $f$ is given by the law of $F + h$ where $h$ is a zero boundary $A$-GFF on $D$.
\section{The Ginzburg-Landau Model}
\label{sec::gl}
The Ginzburg-Landau $\nabla \phi$-interface (GL) model is a general effective interface model first studied by Funaki and Spohn in \cite{FS97} and Naddaf and Spencer in \cite{NS97}. Suppose that $G = (V \cup \partial V,E)$ is a finite, undirected, connected graph with a distinguished set of vertices $\partial V$. Let $\mathcal {V} \in C^2(\mathbf{R})$ satisfy:
\begin{enumerate}
\item $\mathcal {V}(x) = \mathcal {V}(-x)$ (symmetry),
\item $0 < a_\mathcal {V} \leq \mathcal {V}''(x) \leq A_\mathcal {V} < \infty$ (uniform convexity), and
\item $\mathcal {V}''$ is $L$-Lipschitz.
\end{enumerate}
The law of the GL model on $V$ with potential function $\mathcal {V}$ and boundary condition $\psi \colon \partial V \to \mathbf{R}$ is the measure on functions $h \colon V \to \mathbf{R}$ with $h|_{\partial V} = \psi$ described by the density
\[ \frac{1}{\mathcal {Z}_\mathcal {V}} \exp\left( - \sum_{b \in E} \mathcal {V}(\nabla (h \vee \psi)(b)) \right) \]
with respect to Lebesgue measure and $h \vee \psi$ is as in \eqref{intro::eqn::vee}.
\subsection{Langevin Dynamics}
\begin{figure}
\centering
\includegraphics[width=.7\textwidth]{figures/gl_picture2.pdf}
\caption{A typical realization of the GL model with zero boundary conditions on $\{0,\ldots,30\}^2$ and potential function $\mathcal {V}(x) = x^2 + \cos(x)$, sampled using a discretization of the SDS \eqref{gl::eqn::dynam}}
\end{figure}
Consider the stochastic differential system (SDS)
\begin{equation}
\label{gl::eqn::dynam}
dh_t^\psi(x) = \sum_{b \ni x} \mathcal {V}'(\nabla (h_t^\psi \vee \psi)(b))dt + \sqrt{2}dW_t(x) \text{ for } x \in D
\end{equation}
where $W_t(x)$, $x \in V$, is a family of independent standard Brownian motions. The generator for \eqref{gl::eqn::dynam} is given by
\begin{align*}
\mathcal {L} \varphi(h)
&= \sum_{x \in V} \left(\partial_{h(x)}^2 \varphi(h) + \sum_{b \ni x} \mathcal {V}'(\nabla h \vee \psi(b)) \partial_{h(x)}\varphi(h)\right)\\
&= \sum_{x \in V} e^{\mathcal {H}_\mathcal {V}^\psi(h)} \frac{\partial}{\partial h(x)} \left( e^{-\mathcal {H}_\mathcal {V}^\psi(h)} \frac{\partial}{\partial h(x)} \varphi(h) \right),
\end{align*}
where
\[ \mathcal {H}_\mathcal {V}^\psi(h) = \sum_{b \in E} \mathcal {V}( \nabla h \vee \psi(b))\]
is the Hamiltonian for the GL model.
Thus it is easy to see that $\mathcal {L}$ is self-adjoint in the space $L^2(e^{-\mathcal {H}(h)})$, hence the dynamics \eqref{gl::eqn::dynam} are reversible with respect to the law of the GL model. These are the \emph{Langevin dynamics}.
\subsection{The Helffer-Sj\"ostrand Representation}
\label{subsec::hs_representation}
We showed in subsection \ref{subsec::dgff_construction} that if $\mathcal {V}(x) = \tfrac{1}{2} x^2$ then the mean height is harmonic and that the covariance of heights is described by the discrete Green's function. Both of these quantities admit simple probabilistic representations: if $X_t$ is a continuous-time random walk (CTRW) on $G$ that jumps with uniform rate $1$ equally to its neighbors and $\tau$ is the time it first hits $\partial V$, then
\[ \mathbf{E} h(x) = \mathbf{E}_x h(X_\tau) \text{ and } {\rm Cov}(h(x),h(y)) = \mathbf{E}_x \int_0^\tau \mathbf{1}_{\{X_s = y\}} ds\]
where the subscript $x$ indicates $X_0 = x$.
The idea of the Helffer-Sj\"ostrand (HS) representation, originally developed in \cite{HS94} and reworked probabilistically in \cite{DGI00, GOS01}, is to give an expression for the corresponding quantities for the GL model in terms of the first exit distribution and occupation time of another CTRW. In contrast to the the quadratic case, the CTRW is rather complicated for non-quadratic $\mathcal {V}$ as its jump rates are not only \emph{random}, but additionally are \emph{time varying} and \emph{depend} on the boundary data. Nevertheless, the HS representation is a rather useful analytical tool due to comparison inequalities (Brascamp-Lieb and Nash-Aronson).
Specifically, let $h_t^\psi$ solve \eqref{gl::eqn::dynam} with boundary condition $\psi$. Conditional on the realization of the trajectory of the time-varying gradient field $(\nabla h_t^\psi(b) : b \in D^*)$, we let $X_t^\psi$ be the Markov process on $G$ with time-varying jump rates $\mathcal {V}''(\nabla h_t^\psi(b))$. Let $\tau = \inf\{ t \geq 0 : X_t^\psi \in \partial V\}$. Let $\mathbf{P}_x$ denote the joint law of $(h_t^\psi, X_t^\psi)$ given $X_0^\psi = x$ and $\mathbf{E}_x$ the expectation under $\mathbf{P}_x$.
\begin{lemma}
\label{gl::lem::hs_mean_cov}
The mean and covariances of $h^\psi$ admit the representation
\begin{align}
{\rm Cov}(h^\psi(x),h^\psi(y)) &= \mathbf{E}_x \int_0^\tau \mathbf{1}_{\{X_s^\psi = y\}} ds\\
\mathbf{E} h^\psi(x) &= \int_0^1 \mathbf{E}_x \psi(X_\tau^{r\psi}) dr.
\end{align}
\end{lemma}
We refer the reader to \cite[Section 2]{DGI00} for a proof and also a much more detailed discussion on the HS representation.
\subsection{Brascamp-Lieb Inequalities}
\label{subsec::bl}
For $\nu \in \mathbf{R}^{|V|}$, we let
\[ \langle \nu, h^\psi \rangle = \sum_{x \in V} \nu(x) h^\psi(x).\]
The following inequalities, first proved in \cite{BL76} and redeveloped probabilistically in \cite{DGI00}, bound from above the centered moments of $h^\psi$ with those of $h^*$, where $h^*$ is a zero-boundary DGFF on $G$. Recall that $a_\mathcal {V},A_\mathcal {V}$ are positive, finite constants so that $a_\mathcal {V} \leq \mathcal {V}'' \leq A_\mathcal {V}$.
\begin{lemma}[Brascamp-Lieb inequalities]
\label{bl::lem::bl_inequalities}
There exists a constant $C > 0$ depending only on $a_\mathcal {V},A_\mathcal {V}$ such that the following inequalities hold:
\begin{align}
& {\rm Var}(\langle \nu, h^\psi \rangle) \leq C {\rm Var}( \langle \nu, h^* \rangle ) \label{gl::eqn::bl_var}, \\
& \mathbf{E} \exp(\langle \nu, h^\psi \rangle - \mathbf{E} \langle \nu,h^\psi \rangle) \leq \mathbf{E} \exp( C\langle \nu, h^* \rangle) \label{gl::eqn::bl_exp}
\end{align}
for all $\nu \in \mathbf{R}^{|V|}$.
\end{lemma}
We again refer the reader to \cite[Section 2]{DGI00} for a proof. The Brascamp-Lieb inequalities allow for the following bound on the moments of the maximum which we will make use of many times throughout the rest of the article.
\begin{lemma}[Moments of the Maximum]
\label{gl::lem::mom_max}
Suppose that $F \subseteq \mathbf{Z}^2$ is bounded and connected with $R = {\rm diam}(F)$. Let $\zeta \in \mathbf{B}_{\overline{\Lambda}}^u(F)$, $h^\zeta \sim \mathbf{P}_F^\zeta$, and $M = \max_{x \in F} |h^\zeta(x) - u \cdot x|$. For every $\epsilon > 0$ and $p \geq 1$ we have that
\[ (\mathbf{E} M^p)^{1/p} = O_{\overline{\Lambda},p}(R^\epsilon).\]
\end{lemma}
\begin{proof}
We may assume without loss of generality that $\overline{\Lambda} \geq 1$. Combining the Brascamp-Lieb and Chebyshev inequalities, we have the tail bound
\[ \mathbf{P}[M \geq t] = O_{\overline{\Lambda}}(\exp( O_{\overline{\Lambda}}((\log R)^{\overline{\Lambda}}) - t)).\]
Furthermore, we have
\[ \mathbf{E} M^p \leq \sum_{x \in F} \mathbf{E} |h^\zeta(x)|^p \leq O_{\overline{\Lambda}}( R^2 (\log R)^{p\overline{\Lambda}}).\]
Consequently,
\begin{align*}
\mathbf{E} M^p &\leq R^{p\epsilon} + \mathbf{E} M^p \mathbf{1}_{\{ M \geq R^{\epsilon}\}}
\leq R^{p\epsilon} + (\mathbf{E} M^{2p})^{1/2} (\mathbf{P}[ M \geq R^{\epsilon}])^{1/2} \\
&\leq R^{p\epsilon} + O_{\overline{\Lambda}}( R (\log R)^{p\overline{\Lambda}} \exp(c_p- \tfrac{1}{2}R^{\epsilon})).
\end{align*}
Therefore $(\mathbf{E} M^p)^{1/p} = O_{\overline{\Lambda},p}(R^\epsilon)$,
as desired.
\end{proof}
\section{Dynamics}
\label{sec::dyn}
We now specialize to the case where $G$ is a bounded, connected subgraph of $\mathbf{Z}^2$. We will write $D$ for its vertices, $\partial D = \{ x \in \mathbf{Z}^2 : {\rm dist}(x,D) = 1\}$ for its boundary, and $D^* = \{ b = (x_b,y_b) \in (\mathbf{Z}^2)^* : x_b,y_b \in D\}$ for its edges, where $(\mathbf{Z}^2)^*$ denotes the set of edges of $\mathbf{Z}^2$. Finally, let $\partial D^*$ be the set of edges that are either contained in $\partial D$ or intersect both $\partial D$ and $D$. The Langevin dynamics are extremely useful for constructing couplings of instances of the GL model with either different boundary conditions, defined on different (though overlapping) domains, or both. Suppose that $h^\psi, h^{\widetilde{\psi}}$ are solutions of \eqref{gl::eqn::dynam} driven by the same Brownian motions with boundary conditions $\psi,\widetilde{\psi}$, respectively. Let $\overline{\psi} = \psi - \widetilde{\psi}$ and $\overline{h} = h^\psi - h^{\widetilde{\psi}}$. Observe that
\begin{equation}
\label{gl::eqn::diff_dynam}
d \overline{h}_t(x) = \sum_{b \ni x} [\mathcal {V}'(\nabla(h_t^\psi \vee \psi)(b)) - \mathcal {V}'(\nabla (h_t^{\widetilde{\psi}} \vee \widetilde{\psi})(b))] dt
\end{equation}
Let
\begin{equation}
\label{gl::eqn::stat_rate_def} c_t(b) = \int_0^1 \mathcal {V}''(\nabla (h_t^{\widetilde{\psi}} + s \overline{h}_t)(b)) ds \text{ and }
\mathcal {L}_t f(x) = \sum_{b \ni x} c_t(b) \nabla f(b).
\end{equation}
Then we can rewrite \eqref{gl::eqn::diff_dynam} more concisely as
\begin{equation}
\label{gl::eqn::diff_dynam_ell}
d \overline{h}_t(x) = \mathcal {L}_t \overline{h}_t(x) dt.
\end{equation}
The following is \cite[Lemma 2.3]{FS97}:
\begin{lemma}[Energy Inequality]
\label{gl::lem::ee}
For every $T > 0$ we have
\begin{align}
\label{gl::eqn::ee}
&\sum_{x \in D} |\overline{h}_T(x)|^2 + \int_0^T \sum_{b \in D^*} |\nabla \overline{h}_t(b)|^2 dt \notag\\
\leq& C \left(\sum_{x \in D} |\overline{h}_0(x)|^2 + \int_0^T \sum_{b \in \partial D^*} |\overline{\psi}(x_b)||\nabla \overline{h}_t(b)|dt\right)
\end{align}
for $C > 0$ depending only on $a_\mathcal {V},A_\mathcal {V}$.
\end{lemma}
More generally, if $f_t$ solves $\partial_t f = \mathcal {L}_t f_t$ then $f_t$ also satisfies \eqref{gl::eqn::ee}.
\subsection{Coupling Bounds}
\label{dyn::subsec::coupling_bounds}
The purpose of the next lemma is to show that $\lim_{T \to \infty} (h_{T}^\psi, h_{T}^{\widetilde{\psi}})$ gives the unique invariant measure of the Markov process $(h_t^\psi,h_t^{\widetilde{\psi}})$, i.e. where $h_t^\psi,h_t^{\widetilde{\psi}}$ both solve \eqref{gl::eqn::dynam} with the same driving Brownian motions.
\begin{lemma}$\ $
\label{gl::lem::ergodic}
\begin{enumerate}
\item \label{gl::lem::unique_invariant} The SDS \eqref{gl::eqn::dynam} is ergodic.
\item \label{gl::lem::unique_invariant_multiple} More generally, any finite collection $h^1,\ldots,h^n$ satisfying the SDS \eqref{gl::eqn::dynam} and driven by the same family of Brownian motions is ergodic.
\end{enumerate}
\end{lemma}
\begin{proof}
Part \eqref{gl::lem::unique_invariant} follows immediately from Lemma \ref{gl::lem::ee}. Indeed, the Poincar\'e inequality implies that there exists $c_D > 0$ such that for all functions $f \colon D \to \mathbf{R}$ with $f|_{\partial D} \equiv 0$ we have
\[ \sum_{x \in D} (f(x))^2 \leq c_D \sum_{b \in D^*} (\nabla f(b))^2.\]
Thus Lemma \ref{gl::lem::ee} implies that if $(h_t^\psi,\widetilde{h}_t^\psi)$ both solve \eqref{gl::eqn::dynam} with the same Brownian motions and boundary data (though with possibility different initial distributions) then with $\overline{h}_t = h_t^\psi - \widetilde{h}_t^\psi$ we have
\[ \sum_{x \in D} (\overline{h}_T(x))^2 + \frac{1}{c_D} \int_0^T \sum_{x \in D} (\overline{h}_t(x))^2 dt \leq C \sum_{x \in D^*} (\overline{h}_0(b))^2.\]
In particular, the integral is bounded as $T \to \infty$ which implies
\[ \lim_{T \to \infty} \int_T^\infty \sum_{x \in D^*} (\overline{h}_t(x))^2 dt = 0.\]
The energy inequality also gives
\[ \frac{1}{C}\sum_{x \in D} (\overline{h}_T(x))^2 \leq \sum_{x \in D} (\overline{h}_t(x))^2\]
for all $0 < t < T$. Therefore $\overline{h}_t \stackrel{d}{\to} 0$ as $t \to \infty$. This proves \eqref{gl::lem::unique_invariant}.
We will now prove part \eqref{gl::lem::unique_invariant_multiple}. In the interest of keeping the notation simple, we will prove the result in the special case $n=2$. Suppose that $(h_t^1,h_t^2)$ solve the SDS \eqref{gl::eqn::dynam} driven by the same Brownian motions with boundary conditions $\psi^1,\psi^2$ but with arbitrary initial conditions $(h_0^1, h_0^2)$. We know that $h_t^i$ converges in distribution $\mathbf{P}_D^{\psi^i}$ by part \eqref{gl::lem::unique_invariant}. Consequently, the pair $(h_t^1,h_t^2)$ is tight.
We will now prove the existence of the limit $\lim_{t \to \infty} (h_t^1,h_t^2)$. Suppose that $(T_k)$ and $(S_k)$ are arbitrary increasing sequences diverging to infinity. Fix $k$ and assume that $T_k \leq S_k$. Let $(\widetilde{h}_t^1, \widetilde{h}_t^2)$ solve \eqref{gl::eqn::dynam} where $\widetilde{h}_0^i = h_{S_k-T_k}^i$. Let $(\breve{h}_t^1,\breve{h}_t^2)$ be another pair of solutions to \eqref{gl::eqn::dynam} with the same boundary and initial conditions of $(h^1,h^2)$ but driven by the same Brownian motions as $(\widetilde{h}^1,\widetilde{h}^2)$. Then $(\widetilde{h}_{T_k}^1,\widetilde{h}_{T_k}^2) \stackrel{d}{=} (h_{S_k}^1, h_{S_k}^2)$ and $(\breve{h}_{T_k}^1, \breve{h}_{T_k}^2) \stackrel{d}{=} (h_{T_k}^1, h_{T_k}^2)$. By the energy inequality,
\begin{align*}
\frac{1}{T_k} \int_{0}^{T_k} \sum_{b \in D^*} |\nabla (\breve{h}_t^i - \widetilde{h}_t^i)(b)|^2 dt
&\leq \frac{C}{T_k} \sum_{x \in D} |\breve{h}_0^i(x) - \widetilde{h}_0^i(x)|^2\\
&\leq \frac{2C}{T_k} \sum_{x \in D} \big(|h_0^i(x)|^2 + |h_{S_k-T_k}^i(x)|^2 \big).
\end{align*}
As $T_k \to \infty$, the first term in the summation on the right hand side clearly converges to zero almost surely. The second term in the summation converges to zero in distribution since $h_t^i$ is tight. Consequently, for every $\delta > 0$ there exists $k_0$ sufficiently large so that for all $k \geq k_0$ with $S_k \geq T_k$ we have
\[ \mathbf{P}\left[ \frac{1}{T_k} \int_{0}^{T_k} \sum_{b \in D^*} |\nabla (\breve{h}_t^i - \widetilde{h}_t^i)(b)|^2 dt > \delta \right] < \delta.\]
Let $\epsilon > 0$ be arbitrary and fix $\delta = \epsilon / c_D$ so that if $f \colon D \to \mathbf{R}$ is an arbitrary function vanishing on $\partial D$ with $\sum_{b \in D^*} |\nabla f(b)|^2 \leq \delta$ then $\sum_{x \in D} |f(x)|^2 \leq \epsilon$. Assume that $k$ is sufficiently large so that with probability $1-\epsilon$ we have
\[\frac{1}{T_k} \int_{0}^{T_k} \sum_{b \in D^*} |\nabla (\breve{h}_t^i - \widetilde{h}_t^i)(b)|^2 dt \leq \delta.\]
Then there exists (random) $t_0 \in [0,T_k/2]$ such that $\sum_{b \in D^*} |\nabla (\breve{h}_{t_0}^i - \widetilde{h}_{t_0}^i)(b)|^2 \leq \delta$ hence $\sum_{x \in D} | \breve{h}_{t_0}^i(x) - \widetilde{h}_{t_0}^i(x)|^2 \leq \epsilon$ with probability $1-\epsilon$. Applying the energy inequality once again yields
\begin{align*}
\sum_{x \in D} | \breve{h}_{T_k}^i(x) - \widetilde{h}_{T_k}^i(x)|^2
&\leq C\sum_{x \in D} |\breve{h}_{t_0}^i(x) - \widetilde{h}_{t_0}^i(x)|^2
\leq C\epsilon
\end{align*}
with probability $1-\epsilon$. Of course, we can do the same if $T_k \geq S_k$. Therefore we conclude that the subsequential limits of $(h_t^1,h_t^2)$ are unique, hence $\mu \stackrel{d}{=} \lim_{t \to \infty} (h_t^1, h_t^2)$ exists. The same argument also implies that for any $s > 0$ we have $\lim_{t \to \infty} (h_{s+t}^1,h_{s+t}^2)$ exists and has the same distribution as $\mu$. Therefore $\mu$ is a stationary measure.
To finish proving the lemma, we just need to establish uniqueness. Suppose that each of the pairs $(h_t^1, h_t^2), (\widetilde{h}_t^1, \widetilde{h}_t^2)$ solve the SDS \eqref{gl::eqn::dynam}, $h_t^i, \widetilde{h}_t^i$ all driven by the same Brownian motions. Suppose further that both pairs have stationary initial conditions. Then we can use the energy inequality exactly in the same manner as in the proof of part \eqref{gl::lem::unique_invariant} to deduce that $|h_t^i - \widetilde{h}_t^i| \stackrel{d}{\to} 0$ as $t \to \infty$. Since $(h_t^1,h_t^2)$, $(\widetilde{h}_t^1, \widetilde{h}_t^2)$ are stationary, it therefore follows that $(h_0^1,h_0^2) \stackrel{d}{=} (\widetilde{h}_0^1, \widetilde{h}_0^2)$.
\end{proof}
Suppose that $(h_t^\psi,h_t^{\widetilde{\psi}})$ is the stationary coupling of instances of the model with boundary conditions $\psi,\widetilde{\psi} \in \mathbf{B}_{\overline{\Lambda}}^u(D)$, respectively. Letting $\overline{h} = h^\psi - h^{\widetilde{\psi}}$, the Caccioppoli inequality \eqref{symm_rw::eqn::caccioppoli} implies that
\begin{equation}
\label{gl::lem::cap_bound}
\int_{r^2}^{2r^2} \sum_{b \in B^*(x_0,r)} |\nabla \overline{h}_t(b)|^2 dt \leq \frac{C}{r^2} \int_0^{2r^2} \sum_{x \in B(x_0,r)} |\overline{h}_t(x)|^2 dt
\end{equation}
for $r > 0$ and $x_0 \in D$ with $B(x_0,2r) \subseteq D$. The maximum principle implies that $\overline{h}$ attains its maximum on $\partial D$, hence $\overline{h} = O_{\overline{\Lambda}}((\log R)^{\overline{\Lambda}})$. Consequently, taking expectations of both sides of \eqref{gl::lem::cap_bound} and using the stationarity of the dynamics yields
\[ \sum_{b \in B^*(x_0,r)} \mathbf{E} |\nabla \overline{h}_t(b)|^2 \leq O_{\overline{\Lambda}}( (\log R)^{2 \overline{\Lambda}}).\]
Recall that $D(r) = \{ x \in D : {\rm dist}(x,\partial D) \geq r\}$ and let $\epsilon > 0$ be arbitrary. Then $D(R^{1-\epsilon})$ can be covered by $O(R^2 / R^{2-2\epsilon}) = O(R^{2\epsilon})$ balls of radius $R^{1-\epsilon}$, all of which are contained in $D$. Therefore
\[ \sum_{b \in D^*(R^{1-\epsilon})} \mathbf{E} |\nabla \overline{h}_t(b)|^2 \leq O_{\overline{\Lambda}}( R^{3\epsilon}).\]
Let $D(R_1,R_2) = \{x \in D : R_1 \leq {\rm dist}(x,\partial D) < R_2\}$, $r_k = kR^{1-5\epsilon}$, and $D_k = D(R^{1-\epsilon}+r_k, R^{1-\epsilon} + r_{k+1})$. Note that we can write
\[ \sum_{b \in D^*(R^{1-\epsilon},2R^{1-\epsilon})} \mathbf{E} | \nabla \overline{h}_t(b)|^2 = \sum_{k=0}^{R^{4\epsilon}-1} \sum_{b \in D_k^*} \mathbf{E} | \nabla \overline{h}_t(b)|^2 = O_{\overline{\Lambda}}(R^{3\epsilon}).\]
This implies there exists $0 \leq k \leq R^{4\epsilon}-1$ such that
\[ \sum_{b \in D_k^*} \mathbf{E} |\nabla \overline{h}_t(b)|^2 \leq O_{\overline{\Lambda}}(R^{-\epsilon})\]
We have proven:
\begin{lemma}
\label{gl::lem::grad_error}
Suppose that $(h^\psi,h^{\widetilde{\psi}})$ is a stationary coupling of two solutions of the SDS \eqref{gl::eqn::dynam} driven by the same Brownian motions with $\psi,\widetilde{\psi} \in \mathbf{B}_{\overline{\Lambda}}^u(D)$. For every $\epsilon > 0$ there exists $R^{1-\epsilon} \leq R_1 \leq 2R^{1-\epsilon}$ such that with $R_2 = R_1 + R^{1-5\epsilon}$ we have that
\begin{align*}
\sum_{b \in D^*(R_1, R_2)} \mathbf{E} |\nabla \overline{h}_t(b)|^2 &= O_{\overline{\Lambda}}(R^{-\epsilon}) \text{ and }
\sum_{b \in D^*(R_1)} \mathbf{E} |\nabla \overline{h}_t(b)|^2 &= O_{\overline{\Lambda}}(R^{3\epsilon}).
\end{align*}
\end{lemma}
This lemma will be particularly useful for us in Section \ref{sec::harm} in order to construct an intermediate coupling of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$ exhibiting pointwise regularity near $\partial D$ with high probability.
\subsection{Gradient Gibbs States and Equivalence of Ensembles}
\label{subsec::shift_invariant}
By the reverse Brascamp-Lieb inequality \cite[Lemma 2.8]{DGI00}, it follows that if $D_n$ is any sequence of domains tending locally to the infinite lattice $\mathbf{Z}^2$ and, for each $n$, $h^n$ is an instance of the GL model on $D_n$ then ${\rm Var}(h^n(x)) \to \infty$ as $n \to \infty$. This holds regardless of the choice of boundary conditions, which suggests that it is not possible to take an infinite volume limit of the height field $h^n(x)$. However, the Brascamp-Lieb inequality (Lemma \ref{bl::lem::bl_inequalities}) gives that ${\rm Var}(\nabla h^n(b))$ remains uniformly bounded as $n \to \infty$, indicating that it should be possible to take an infinite volume limit of the \emph{gradient field}.
Working with gradient rather than height fields, though unnecessary for $d \geq 3$, is convenient since it allows for a unified treatment of Gibbs states for all dimensions. Let $\mathcal {X}$ be the set of functions $\eta \colon (\mathbf{Z}^d)^* \to \mathbf{R}$. Let $\mathcal {F} = \sigma(\eta(b) : b \in (\mathbf{Z}^d)^*)$ be the $\sigma$-algebra on $\mathcal {X}$ generated by the evaluation maps and, for each $D^* \subseteq (\mathbf{Z}^d)^*$, let $\mathcal {F}_{D^*} = \sigma(\eta(b) : b \in D^*)$ be the $\sigma$-algebra generated by the evaluation maps in $D^*$. Suppose that $D \subseteq \mathbf{Z}^d$ is bounded, $\varphi \in \mathcal {X}$, and $\nabla \phi = \varphi$. If $h$ is distributed according to $\mathbf{P}_D^\phi$, then the gradient field $\nabla h$ induces a measure $\mathbf{P}_{D^*}^\varphi$ on functions $D^* \to \mathbf{R}$. We call $\mathbf{P}_{D^*}^\varphi$ the law of the GL model on $D^*$ with \emph{Neumann boundary} conditions $\varphi$. Let $\mu$ be a measure on $\mathcal {X}$ and suppose that $\eta$ has the law $\mu$. We say that $\mu$ is a \emph{gradient Gibbs state} associated with the potential $\mathcal {V}$ if for every finite $D^* \subseteq (\mathbf{Z}^2)^*$,
\[ \mu(\cdot | \mathcal {F}_{(D^*)^c}) =\mathbf{P}_{D^*}^{\eta|\partial D^*} \text{ almost surely.}\]
Fix a vector $u \in \mathbf{R}^d$. A gradient Gibbs state $\mu$ is said to have \emph{tilt} $u$ if $\mathbf{E}^\mu \eta(x+ b_i) = u \cdot e_i$ where $b_i = (0,e_i)$, $e_i$ a generator of $\mathbf{Z}^d$, and $x \in \mathbf{Z}^d$ is arbitrary. We say that $\mu$ is \emph{shift invariant} if $\mu \circ \tau_x^{-1} = \mu$ for every $x \in \mathbf{Z}^d$ where $\tau_x \colon \mathbf{Z}^d \to \mathbf{Z}^d$ is translation by $x$. Finally, a shift-invariant $\mu$ is said to be shift-ergodic if whenever $f$ is a shift-invariant $\mathcal {F}$-measurable function, then $f$ is $\mu$-almost surely constant.
Funaki and Spohn in \cite{FS97} proved that the shift-ergodic Gibbs states are parameterized according to their tilt $u$; from now on we will refer to the law of such as $\texttt{SEGGS}_u$. The natural construction is to take an infinite volume limit of gradient measures $\mathbf{P}_{D^*}^\varphi$ as $D^*$ tends locally to $\mathbf{Z}^d$ with boundary conditions $\varphi(b) = u \cdot (y_b - x_b)$ \cite[Remark 4.3]{STF_Funaki}. The difficulties with this approach are that $\mathbf{P}_{D^*}^\varphi$ is itself not shift-invariant and it is not clear that the mean gradient field approximately has tilt $u$. This issue is cleverly circumvented in \cite{FS97} by instead considering the finite volume measure
\[ d\mu_n(\eta) = \frac{1}{\mathcal {Z}_n} \exp\left( -\sum_{b \in (\mathbf{Z}_n^d)^*} \mathcal {V}(\eta(b) - (y_b - x_b) \cdot u) \right) d\nu_n(\eta)\]
on gradient fields on the torus, where $\nu_n$ is the uniform measure on the set of functions $\eta \colon (\mathbf{Z}_n^d)^* \to \mathbf{R}$ which can be expressed as the gradient of a function $h \colon \mathbf{Z}_n^d \to \mathbf{R}$. By construction, $\mu_n$ is shift invariant, has tilt $u$ and both of these properties are preserved in the limit as $n \to \infty$.
We will now explain how to use the method of dynamic coupling to prove that the gradient field of $\mathbf{P}_{D^*}^\varphi$, $\varphi(b) = u \cdot (y_b - x_b)$ as before, converges to $\texttt{SEGGS}_u$. Our proof will also yield an alternative construction of the Funaki-Spohn state in the special case $u = 0$. We will include a statement of this result here as well as a short sketch of the proof since this will be important for the proof of Theorem \ref{thm::clt}. We note in passing that Theorem \ref{harm::thm::coupling} gives a much better coupling which will be critical for the proof of Theorem \ref{thm::clt}, but this result uses the convergence of the finite volume gradient fields hence we cannot simply apply it here.
Fix a tilt $u$. Suppose that $D_n$ is any sequence of bounded domains in $\mathbf{Z}^d$ converging locally to $\mathbf{Z}^d$. For each $n$, let $\eta^n \sim \mathbf{P}_{D_n^*}^{\varphi}$. Suppose that $\eta \sim \texttt{SEGGS}_u$. Fix $x_n \in \partial D_n$ and let $h^n, h^{n,S}$ be the height fields associated with the gradient fields $\eta^n, \eta$, respectively, both set to vanish at $x_n$.
By the Brascamp-Lieb inequality \eqref{gl::eqn::bl_var},
\begin{equation}
\label{eqn::var_bound}
{\rm Var}( h^{n,S}(x) - h^{n,S}(y)) \leq C \log(1+|x-y|).
\end{equation}
Here, $h^{n,S}(x) - h^{n,S}(y)$ is interpreted as $\sum_{i=1}^n \eta(b_i)$ where $b_1,\ldots,b_n$ is any sequence of bonds connecting $x$ to $y$. Of course the same is also true with $h^n$ in place of $h^{n,S}$. Let $R_n = {\rm diam}(D_n)$. As $h^n(x) = \mathbf{E} h^{n,S}(x)$ for $x \in \partial D_n$, the Brascamp-Lieb \eqref{gl::eqn::bl_exp} and Chebychev inequalities thus imply that
\begin{equation}
\label{eqn::gs_diff}
\mathbf{P}[ \max_{x \in \partial D_n} |h^{n,S}(x) - h^n(x)| \geq (\log R_n)^2] = O(R_n^{-100}).
\end{equation}
Assume that $(h_t^{n,S}, h_t^n)$ is the stationary coupling of $h^n$ and $h^{n,S}$ conditional on $h^{n,S} |_{\partial D_n}$. Then as $\overline{h}_t^n = h_t^{n,S} - h_t^n$ satisfies the parabolic equation \eqref{gl::eqn::diff_dynam_ell} and $\overline{h}^n$ is static on $\partial D_n$, the maximum principle implies
\[ \max_{x \in D_n} |\overline{h}_t^n(x)| \leq \max_{x \in \partial D_n} |\overline{h}_0^n(x)|.\]
Combining this with \eqref{eqn::gs_diff} implies
\[ \mathbf{P}[ \max_{x \in D_n} |\overline{h}_t^n(x)|\geq (\log R_n)^2] = O(R_n^{-100}).\]
The Nash continuity estimate (Lemma \ref{symm_rw::lem::nash_continuity_bounded}) applied to $\overline{h}_t^n$ thus implies that
\begin{equation}
\label{gl::eqn::quantitative_bound}
\mathbf{E} \big[ \max_{b \in D_n^*(R_n^\zeta)} |\nabla \overline{h}_t^n(b)| \big] = O_{\overline{\Lambda}}( R_n^{\epsilon-\zeta \xi_{\rm NC}})
\end{equation}
for $\epsilon, \zeta > 0$ fixed. This proves the desired convergence.
Existence in the special case of $u = 0$ can be proved in a very similar manner. The reason is that, in this case, $h^n \sim \mathbf{P}_{D_n}^0$ hence $\mathbf{E} h^n(x) = 0$. Thus it is clear that the subsequential limits of gradient fields, which exist by the Brascamp-Lieb inequalities, have zero tilt. The Brascamp-Lieb and Chebychev inequalities imply that the maximum of $h^n$ is with high probability $O(\log R_n)$. Thus if $\widetilde{h}^n$ denotes the law of the GL model on the domain $\widetilde{D}_n$ given by shifting $D_n$ by one unit, then using the argument of the previous paragraph we can couple $h^n$ and $\widetilde{h}^n$ such that $\nabla (h^n - \widetilde{h}^n)$ is with high probability $O(R_n^{\epsilon - \zeta \xi_{\rm NC}})$ at distance at least $R_n^{\zeta}$, $\zeta > 0$, from both $\partial D_n$ and $\partial \widetilde{D}_n$. Therefore the subsequential limits of $\mathbf{P}_{D_n}^0$ are shift-invariant which proves the existence of a zero-tilt shift-invariant Gibbs state. Uniqueness (and also existence of limits) follows by taking two such states $\eta, \eta'$, then applying the argument of the previous paragraph.
We have obtained:
\begin{theorem}[Equivalence of Ensembles] $\ $
\label{gl::thm::infinite_coupling}
\item If $(D_n)$ is any sequence of bounded domains in $\mathbf{Z}^d$ tending locally to $\mathbf{Z}^d$ and, for each $n$, $h^n$ is an instance of the GL model on $D_n$ with boundary conditions $\varphi_n \in \mathbf{B}_{\overline{\Lambda}}^u(D)$, then $\eta^n = \nabla h^n$ converges weakly to $\texttt{SEGGS}_u$ and we have
\[ \mathbf{E} \big[ \max_{b \in D_n^*(R_n^\zeta)} |\eta^n(b) - \eta(b)| \big] = O_{\overline{\Lambda}}(R^{\epsilon -\zeta \xi_{\rm NC}}).\]
\end{theorem}
\subsection{Invariance under Reflections}
\label{subsec::reflect_invariance}
The following proposition will be especially useful for us later in the next section when applied to $f = \mathcal {V}''$. We let $\varphi^h,\varphi^v \colon \mathbf{R}^2 \to \mathbf{R}^2$ be the maps which reflect about the horizontal and vertical axes, respectively, and $\varphi_x^h(b) = \varphi^h(b-x)$, $\varphi_x^v(b) = \varphi^v(b-x)$ for $x \in \mathbf{Z}^2$.
\begin{proposition}
\label{gl::prop::reflection_invariance}
Fix a tilt $u$, $x \in \mathbf{Z}^2$, and let $f \colon \mathbf{R} \to \mathbf{R}$ be an even function. Let $b^h = \varphi_x^h(b)$ for $b \in (\mathbf{Z}^2)^*$.
If $\eta_u \sim \texttt{SEGGS}_u$, then $( f(\eta_u(b)) : b \in (\mathbf{Z}^2)^*) \stackrel{d}{=} ( f(\eta_{u}(b^h)) : b \in (\mathbf{Z}^2)^*)$. The same is also true when the horizontal reflection is replaced with vertical reflection.
\end{proposition}
\begin{proof}
Let $w = \varphi^h(u)$ and $\eta_w \sim \texttt{SEGGS}_w$. Let $s_b = 1$ if $b$ is orientated vertically and $-1$ otherwise.
Since $f$ is even, we know that $f(\eta_u(b)) = f( s_b \eta_u(b))$. Since $(s_b \eta_u(b)) \stackrel{d}{=} (\eta_w(b))$, we thus have that
\[ (f(\eta_u(b))) \stackrel{d}{=} (f(\eta_w(b))) \stackrel{d}{=} f(\eta_u(b^h)).\]
The reason for the last inequality is that $( \eta_u(b^h) : b \in (\mathbf{Z}^2)^*)$ is still a shift-ergodic Gibbs state but with tilt $w$.
\end{proof}
\section{Correlation Decay}
\label{sec::correlation_decay}
Suppose that $F \subseteq \mathbf{Z}^2$ is a bounded, connected domain with $R = {\rm diam}(F)$. Let $\zeta, \widetilde{\zeta} \in \mathbf{B}_{\overline{\Lambda}}^u(F)$ and assume that $(h_t^\zeta,h_t^{\widetilde{\zeta}})$ is the stationary coupling of $\mathbf{P}_F^\zeta, \mathbf{P}_F^{\widetilde{\zeta}}$. Throughout, we let $\beta = \beta(u)$ as in the statement of Theorem \ref{thm::clt}. The main result of this section is that $\mathcal {V}''(\nabla h^\zeta(b))$ and $\nabla \overline{h}(b)$, $\overline{h} = h^\zeta - h^{\widetilde{\zeta}}$, are uncorrelated when averaged against a $\Delta^\beta$-harmonic function.
\begin{theorem}
\label{cd::thm::cd}
Suppose that $x_0 \in F$ with ${\rm dist}(x_0,\partial F) \geq R^{\alpha+\epsilon}$ for $\alpha,\epsilon > 0$, $E = B(x_0, R^\alpha)$, and let $g \colon F \to \mathbf{R}$ be $\Delta^\beta$-harmonic. We have that
\begin{align}
\label{cd::eqn::estimate}
\mathbf{E} \sum_{b \in E^*} \mathcal {V}''(\nabla h^\zeta(b)) \nabla \overline{h}(b) \nabla g(b) =& \sum_{b \in E^*} a_u(b) \mathbf{E}[\nabla \overline{h}(b)] \nabla g(b) +\\
&O_{\overline{\Lambda}}(R^{\epsilon+\alpha(1-\rho_{\rm CD}) } \| \nabla g \|_\infty) \notag
\end{align}
for $\rho = \rho_{\rm CD}(\mathcal {V}) > 0$ and $a_u(b) = \mathbf{E}[\mathcal {V}''(\eta(b))]$ with $\eta \sim \texttt{SEGGS}_u$.
\end{theorem}
Note that $a_u$ depends on $b$ only through its orientation (either vertical or horizontal) by the shift-invariance of $\eta$. Our typical choice of $g$ will have $\| \nabla g \|_\infty = O(R^{-\alpha})$, in which case the exponent in the error term is actually negative. The idea of the proof is to show that replacing $\overline{h}(x)$ by its average $\overline{h}^\rho(x)$ over the ball $B(x,R^{\rho})$ introduces a small amount of error. The advantage of this replacement is that the time-derivative of $\overline{h}_t^\rho$ possesses more regularity than that of $\overline{h}_t$. This allows us to replace the left hand side of \eqref{cd::eqn::estimate} with
\[ \mathbf{E} \sum_{b \in E^*} \mathcal {V}''(\nabla h_T^\zeta(b)) \nabla \overline{h}_0(b) \nabla g(b).\]
The proof is then completed by coupling $h_T^\zeta$ to $\eta \sim \texttt{SEGGS}_u$ conditional on $\overline{h}_0$, which can be accomplished at the cost of negligible error by the argument used to prove Theorem \ref{gl::thm::infinite_coupling}.
\subsection{Change of Environment}
Let $\eta_t$ follow the $\texttt{SEGGS}_u$ dynamics independent from $(h_t^\zeta,h_t^{\widetilde{\zeta}})$. That is, $\eta_t$ solves the infinite dimensional SDS
\[ d\eta_t((x,y)) = \left( \sum_{b \ni y} \mathcal {V}'(\eta_t(b)) - \sum_{b \ni x} \mathcal {V}'(\eta_t(b)) \right) dt + \sqrt{2}(dW_t(y) - dW_t(x))\]
for $b \in (\mathbf{Z}^2)^*$ where $W_t(x), x \in \mathbf{Z}^2,$ is a family of iid standard Brownian motions; see Section 9 of \cite{STF_Funaki} for a discussion of the existence and uniqueness of solutions to this equation.
Fix $T > 0$, let $\breve{c}_t(b) = \mathcal {V}''( \eta_t(b)),$
and let $\breve{p}$ be the transition kernel of the random walk jumping with rates $\breve{c}_{T-t}(b)$, $t \in [0,T]$, stopped on its first exit from $F$.
\begin{proposition}
\label{cd::prop::environment_change}
Suppose that we have the setup as Theorem \ref{cd::thm::cd}. Let $\gamma_1, \gamma_2 \in (0,\alpha]$ and $\delta_i = 4 \alpha - 4\gamma_i - \gamma_i \rho_{\rm EC}$. Let $S_2 = R^{2\gamma_2}$. There exists $\tfrac{3}{4} R^{2\gamma_1} \leq S_1 \leq R^{2\gamma_1}$ such that the following holds. Let $\breve{p}^1, \breve{p}^2 \sim \breve{p}$ associated with independent environments $\eta^1,\eta^2 \sim \texttt{SEGGS}_u$ which are in turn independent of $\overline{h}_0$. Let $S = S_1 + S_2$ and let
\[ \breve{h}_t(x) = \sum_{y,z} \breve{p}^1(S-t,S_1;x,y) \breve{p}^2(0,S_2;y,z) \overline{h}_0(z),\ \ S - S_2 \leq t \leq S.\]
There exists a coupling of $(\eta^1,\eta^2,\breve{h}_t)$ and $(\nabla h, \overline{h})$ such that
\begin{align*}
\mathbf{E} \sum_{b \in E^*} (\nabla \breve{h}_{S}(b) - \nabla \overline{h}_{S}(b))^2
= O_{\overline{\Lambda}}(R^{\epsilon+2\alpha +2\gamma_2- 4 \gamma_1 + \delta_2}) + O_{\overline{\Lambda}}(R^{\epsilon + \delta_1}),\\
\mathbf{E} \bigg[ \max_{b \in E^*} \sup_{S/2 \leq t \leq S} | \nabla h_{t}^\zeta(b) - \eta_{t}^1(b)| \bigg] = O_{\overline{\Lambda}}(R^{\epsilon-\gamma \xi_{\rm NC}}),\label{harm::eqn::ec_change_grad}
\end{align*}
for $\rho_{\rm EC} > 0$ depending only on $\mathcal {V}$.
\end{proposition}
The following heat kernel estimates are crucial ingredients in the proof of the proposition.
\begin{lemma}
\label{harm::lem::hc_time_space_deriv}
Suppose that $x_0 \in F$ with ${\rm dist}(x_0,\partial F) \geq R^{\gamma+\epsilon}$ for $\gamma,\epsilon > 0$ and let $E = B(x_0,R^{\gamma})$. For $T = R^{2\gamma}$ and $\tfrac{3}{4} T \leq t_1 < t_2 \leq T$, we have
\begin{align*}
\sum_{x \in E} \sum_{y \in F} | q(u,t_1;x,y) - q(u,t_2;x,y)|^2 &= O (|t_1-t_2|^2 R^{\epsilon - 2 \gamma \xi_{\rm NC}})
\end{align*}
for $0 \leq u \leq T/4$ and
\begin{align*}
\sum_{b \in E^*} \sum_{y \in F} \int_0^{T/4} |\nabla q(u,t_1;b,y) - \nabla q(u,t_2;b,y)|^2 du &= O (|t_1-t_2|^2 R^{\epsilon-2\gamma \xi_{\rm NC}})
\end{align*}
where $q$ is the transition kernel of a random walk on $F$ evolving with rates $a_\mathcal {V} \leq d_t(b) \leq A_\mathcal {V}$.
\end{lemma}
\begin{proof}
Using that $\partial_t q(u,t;x,y) = \sum_{b \ni y} d_t(b) \nabla q(u,t;x,b)$, we have
\begin{align*}
& \sum_{b \in E^*} \sum_{y \in F} \int_{0}^{T/4} |\nabla q(u,t_1;b,y) - \nabla q(u,t_2;b,y)|^2 du \\
\leq& |t_1 - t_2| \sum_{b \in E^*} \sum_{y \in F} \int_{t_1}^{t_2} \int_0^{T/4} |\partial_t \nabla q(u,t;b,y)|^2 du dt\\
\leq& C |t_1 - t_2| \sum_{b \in E^*}\sum_{b' \in F^*} \int_{t_1}^{t_2} \int_0^{T/4} |\nabla \nabla q(u,t;b,b')|^2 du dt,
\end{align*}
for $C > 0$ depending only on $\mathcal {V}$. Applying the Cacciopoli inequality \eqref{symm_rw::eqn::caccioppoli} to the first time and spatial coordinates, we see that this is bounded from above by
\begin{equation}
\label{harm::eqn::hc_mixed_deriv_est}
O(1) \frac{|t_1- t_2|}{R^{2\gamma}} \sum_{x \in E} \sum_{b' \in F^*} \int_{t_1}^{t_2} \int_{0}^{T/2} |\nabla q(u,t;x,b')|^2 du dt.
\end{equation}
The Nash-Aronson estimates (Lemma \ref{symm_rw::lem::nash_aronson}) imply that the contribution to the sum given by those $b' \in F^*$ with ${\rm dist}(b', E) \geq R^{\gamma+\epsilon/2}$ is negligible in comparison to the upper bound we seek to prove. For $b' \in F^*$ with ${\rm dist}(b',E) \leq R^{\gamma+\epsilon/2}$, the Nash continuity and Nash-Aronson estimates (Lemmas \ref{symm_rw::lem::nash_continuity_bounded}, \ref{symm_rw::lem::nash_aronson}) imply
\[ |\nabla q(u,t;x,b')| = O(R^{-\gamma \xi_{\rm NC}-2\gamma})\]
for $0 \leq u \leq T/2$ and $t_1 \leq t \leq t_2$.
Inserting this into \eqref{harm::eqn::hc_mixed_deriv_est}, we arrive at the bound
\begin{align*}
|t_1-t_2|^2 O(R^{4\gamma+\epsilon}) \cdot O(R^{- 2\gamma \xi_{\rm NC}-4\gamma})
= O(|t_1-t_2|^2 R^{\epsilon-2\gamma \xi_{\rm NC}}).
\end{align*}
This proves the second part of the lemma. The first is exactly the same except the application of the Cacciopoli inequality is unnecessary.
\end{proof}
\begin{lemma}
Suppose that $x_0 \in F$ with ${\rm dist}(x_0,\partial F) \geq R^{\gamma+\epsilon}$ for $\gamma,\epsilon > 0$ and let $E = B(x_0,R^\gamma)$, $E_0 = B(x_0,R^{\gamma+\epsilon/2})$, $E_1 = B(x_0,R^{\gamma+\epsilon})$. Let $q,q'$ be the transition kernels for two random walks on $F$ jumping with rates $a_\mathcal {V} \leq d_t(b), d_t'(b) \leq A_\mathcal {V}$, respectively. With $T = R^{2\gamma}$, assume that $d_t \equiv d_t'$ for all $0 \leq t \leq \tfrac{1}{2} T$ and let
\[ \overline{d}_\infty = \sup_{0 \leq t \leq T} \max_{b \in E_1^*} |d_t(b) - d_t'(b)|.\]
Uniformly in $x \in E_0$ we have that
\[ \sum_{y \in E_1^*} | \overline{q}(0,T;x,y)|^2 + \sum_{b \in E_1^*} \int_0^{T} |\nabla \overline{q}(0,t;x,b)|^2 dt =
O(\overline{d}_\infty^2 R^{-2\gamma})\]
where $\overline{q} = q-q'$.
\end{lemma}
\begin{proof}
By definition,
\[ \partial_t q(0,t;x,y) = [\mathcal {L}_t q(0,t;x,\cdot)](y) \text{ and } \partial_t q'(0,t;x,y) = [\mathcal {L}_t' q'(0,t;x,\cdot)](y)\]
where
\[ \mathcal {L}_t f(y) = \sum_{b \ni y} d_{t}(b) \nabla f(b) \text{ and } \mathcal {L}_t' f(y) = \sum_{b \ni y} d_{t}'(b) \nabla f(b).\]
Consequently,
\begin{align}
\partial_t \overline{q}^2
&= 2\overline{q}(\mathcal {L}_t \overline{q} +\overline{\mathcal {L}}_t q')
\label{harm::eqn::diff_square}
\end{align}
where $\overline{\mathcal {L}}_t = \mathcal {L}_t - \mathcal {L}_t'$. Using the same proof as the energy inequality (Lemma \ref{gl::lem::ee}), by integrating both sides of \eqref{harm::eqn::diff_square} from $0$ to $T$ then using summation by parts,
\begin{align*}
&\sum_{y \in E_1} |\overline{q}(0,T;x,y)|^2 + 2\sum_{b \in E_1^*} \int_0^T d_t(b) |\nabla \overline{q}(0,t;x,b)|^2 dt\\
\leq& 2\sum_{b \in \partial E_1^*} \int_0^T \bigg[ d_t(b) |\overline{q}(0,t;x,x_b) \nabla \overline{q}(0,t;x,b)| + |\overline{d}_t(b) \overline{q}(0,t;x,x_b)\nabla q'(0,t;x,b)|\bigg] dt+\\
& 2\sum_{b \in E_1^*} \int_0^T |\overline{d}_{t}(b) \nabla \overline{q}(0,t;x,b) \nabla q'(0,t;x,b)| dt.
\end{align*}
By the Nash-Aronson estimates (Lemma \ref{symm_rw::lem::nash_aronson}),
\[ \sum_{b \in \partial E_1^*} \int_0^T |\overline{q}(0,t;x,x_b) \nabla \overline{q}(0,t;x,b)| dt = O(R^{3\gamma+\epsilon}\exp(-c'R^{\epsilon})) = O(\exp(-cR^{\epsilon}))\]
for some $c, c' > 0$ depending only on $a_\mathcal {V},A_\mathcal {V}$. The other boundary term is similarly of order $O(\exp(-c R^{\epsilon}))$.
Consequently,
\begin{align}
&\sum_{y \in E_1} |\overline{q}(0,T;x,y)|^2 + \sum_{b \in E_1^*} \int_{T/2}^T |\nabla \overline{q}(0,t;x,b)|^2 dt
\leq O(\exp(-cR^{\epsilon})) + \notag\\
&
C\overline{d}_\infty \sum_{b \in E_1^*} \int_{T/2}^T |\nabla \overline{q}(0,t;x,b) \nabla q'(0,t;x,b)| dt \label{harm::eqn::ec_hc_upper_bound}.
\end{align}
The reason that the lower bound of integration is $T/2$ rather than $0$ is $d_t \equiv d_t'$ for all $t \leq T/2$.
If
\[\sum_{b \in E_1^*} \int_{T/2}^T |\nabla \overline{q}(0,t;x,b)|^2 dt = O(\exp(-cR^{\epsilon/2}))\]
then we are obviously done. If not, we apply Cauchy-Schwarz to the sum on the right hand side of \eqref{harm::eqn::ec_hc_upper_bound} and rearrange to arrive at
\begin{align*}
& \sum_{b \in E_1^*} \int_{T/2}^T |\nabla \overline{q}(0,t;x,b)|^2 dt
\leq C \overline{d}_\infty^2 \sum_{b \in E_1^*} \int_{T/2}^T |\nabla q'(0,t;x,b)|^2 dt,
\end{align*}
increasing $C$ if necessary.
By the Nash-Aronson estimates,
\begin{align*}
\sum_{y \in E_1^*} |q'(0,T;x,y)|^2 &= O(R^{-2\gamma}) \text{ and}\\
\sum_{b \in \partial E_1^*} \int_{T/2}^T |q'(0,t;x,x_b)| |\nabla q'(0,t;x,b)|dt &= O(\exp(-c R^{\epsilon})).
\end{align*}
Therefore, by the energy inequality (Lemma \ref{gl::lem::ee}),
\begin{equation}
\label{harm::eqn::hc_int_bound1}
\sum_{b \in E_1^*} \int_{T/2}^T |\nabla q'(0,t;x,b)|^2 dt = O(R^{-2\gamma}),
\end{equation}
which in turn implies
\begin{equation}
\label{harm::eqn::hc_int_bound2}
\sum_{b \in E_1^*} \int_{T/2}^T |\nabla \overline{q}(0,t;x,b)|^2 dt = O(\overline{d}_\infty^2 R^{-2\gamma}).
\end{equation}
Inserting \eqref{harm::eqn::hc_int_bound1}, \eqref{harm::eqn::hc_int_bound2} into \eqref{harm::eqn::ec_hc_upper_bound} proves the lemma.
\end{proof}
\begin{lemma}
Suppose that we have the same setup as Theorem \ref{cd::thm::cd} and let $\breve{p}, \eta$ be as in the introduction of this subsection. Let $S = R^{2\gamma}$ for $0 \leq \gamma \leq \alpha$ and
\[ \breve{h}_t(y) = \sum_{z} \breve{p}(S-t,S;y,z) \overline{h}_0(z) \text{ for } 0 \leq t \leq S.\]
There exists a coupling of $(\eta,\breve{h})$ and $(\nabla h^\zeta, \overline{h})$ such that
\begin{align}
\mathbf{E} \sum_{b \in E^*} \int_{3S/4}^{S} | \nabla \breve{h}_t(b) - \nabla \overline{h}_t(b)|^2 dt
= O_{\overline{\Lambda}}(R^{\epsilon+4\alpha-2\gamma - \gamma \rho_{\rm EC}}), \label{harm::eqn::ec_change_int}\\
\mathbf{E} \sum_{x \in E} | \breve{h}_S(x) - \overline{h}_S(x)|^2 = O_{\overline{\Lambda}}(R^{\epsilon+4\alpha-2\gamma -\gamma \rho_{\rm EC}}), \label{harm::eqn::ec_change_sum}\\
\mathbf{E} \bigg[ \max_{b \in E^*} \sup_{S/2 \leq t \leq S} | \nabla h_{t}^\zeta(b) - \eta_{t}(b)| \bigg] = O_{\overline{\Lambda}}(R^{\epsilon-\gamma \xi_{\rm NC}}),\label{harm::eqn::ec_change_grad_single}
\end{align}
where $\rho_{\rm EC} > 0$ depends only on $\mathcal {V}$ and $\eta$ is independent of $(\nabla h_0^\zeta, \overline{h}_0)$.
\end{lemma}
This constant $\rho_{\rm EC}$ from Proposition \ref{cd::prop::environment_change} is the same as that appearing in the statement of this lemma.
\begin{proof}
Let $T = R^{2\alpha}$ and let $\eta_0 \sim \texttt{SEGGS}_u$ independent of $(h_0^\zeta,h_0^{\widetilde{\zeta}})$. Assume further that the evolution of the Brownian motions driving $\eta_t$ in $F$ are independent from those of $(h_t^\zeta, h_t^{\widetilde{\zeta}})$ until time $T- S$, $S \equiv R^{2\gamma}$, after which they are the same. Let $\breve{c}_t(b) = \mathcal {V}''(\eta_{t}(b))$, $c_t(b)$ be as in \eqref{gl::eqn::stat_rate_def} with $h^\zeta, h^{\widetilde{\zeta}}$ in place of $h^\psi, h^{\widetilde{\psi}}$, and let $p$ be the transition kernel of a random walk in $F$ stopped on its first exit jumping with rates $c_{T-t}(b)$. Note that
\[ \overline{h}_t(x) = \sum_{y \in F} p(T-t,T;x,y) \overline{h}_0(y) \text{ for } 0 \leq t \leq T.\]
Set $S' = S - R^{2\sigma}$, $\sigma \in (0,\gamma)$ to be determined later, and define environments
\[ \widetilde{c}_t(b) = \begin{cases}
\breve{c}_t(b) \text{ for } T-S \leq t \leq T\\
c_t(b) \text{ for } 0 \leq t < T-S
\end{cases}, \ \ \
\widetilde{c}_t'(b) = \begin{cases}
\breve{c}_t(b) \text{ for } T-S' \leq t \leq T\\
c_t(b) \text{ for } 0 \leq t < T-S'.
\end{cases}\]
Let $\widetilde{p}, \widetilde{p}'$ be the transition kernels for the random walks in $F$ stopped on their first exit jumping with rates $\widetilde{c}_{T-t}(b), \widetilde{c}_{T-t}'(b)$, respectively. Finally, let $\overline{p} = p - \widetilde{p}$ and $\overline{p}' = p - \widetilde{p}'$.
For $S \leq t \leq T$, we have $p(S,t;x,y) = \widetilde{p}(S,t;x,y) = \widetilde{p}'(S,t;x,y)$, hence
\begin{align*}
& \sum_{b \in E^*} \sum_{y \in F} (\nabla \overline{p}(u,T;b,y))^2
= \sum_{b \in E^*} \sum_{y \in F} \left( \sum_{z \in F} \nabla \overline{p}(u,S;b,z) p(S, T;z,y)\right)^2.
\end{align*}
By Jensen's inequality, this is bounded from above by
\begin{align*}
& 4\sum_{b \in E^*} \sum_{z \in F} \bigg( (\nabla \overline{p}(u,S;b,z) - \nabla \overline{p}(u,S';b,z))^2\\
&+
(\nabla \overline{p}'(u,S';b,z) - \nabla \overline{p}'(u,S;b,z))^2
+(\nabla \overline{p}'(u,T;b,z))^2 \bigg) \equiv I_1 + I_2 + I_3.
\end{align*}
Fix a base point $a_0 \in \partial F$, set $h_t^S(a_0) = 0$, and let $h_t^S$ solve $\nabla h_t^S(b) = \eta_t(b)$. Applying the Nash continuity and Nash-Aronson estimate (Lemma \ref{symm_rw::lem::nash_continuity_bounded}) to $h_t^S - h_t^\zeta$ and $h_t^{\zeta} - h_t^{\widetilde{\zeta}}$, similar to the proof of Theorem \ref{gl::thm::infinite_coupling}, yields
\begin{equation}
\label{cd::eqn::env_change} \overline{d}_\infty \equiv \sup_{0 \leq t \leq S'} \sup_{b \in E^*} |c_{T-t}(b) - \breve{c}_{T-t}(b)| = O(M_0 R^{-\sigma \xi_{\rm NC}})
\end{equation}
where $M_t = \| \overline{h}_t \|_\infty + \| h_t^\zeta - h_t^S\|_\infty$, as $\mathcal {V}''$ is Lipschitz. Here, we are taking the maximum over $x \in F$. Let $q(s,t;x,y) = p(T-t,T-s;y,x)$ and $q'(s,t;x,y) = \widetilde{p}'(T-t,T-s;y,x)$. Then $q,q'$ are the transition kernels for random walks jumping with rates $c_t(b), \widetilde{c}_t'(b)$, respectively. The previous lemma thus yields
\[ \sum_{b \in E^*} \sum_{y \in F} \int_{0}^T (\nabla \overline{p}'(u,T;b,y))^2 du = O(M_0^2 R^{\epsilon/2-2\sigma \xi_{\rm NC}})\]
since the contribution to the sum given by those $y \in F$ with ${\rm dist}(y, E) \geq R^{\alpha+\epsilon/4}$ is negligible.
Since we can cover $E$ by $O(R^{2(\alpha-\gamma)})$ disks of radius $R^\gamma$, applying Lemma \ref{harm::lem::hc_time_space_deriv} to the terms corresponding to $I_1,I_2$ gives us the bound
\begin{align}
&\sum_{b \in E^*} \sum_{y \in F} \int_{0}^{S/4} (\nabla \overline{p}(u,T;b,y))^2 du \notag\\
=& O(R^{\epsilon/2+4\sigma + 2(\alpha-\gamma) - 2\gamma \xi_{\rm NC}}) + O(M_0^2 R^{\epsilon/2-2\sigma \xi_{\rm NC}}). \label{harm::eqn::hc_diff_bound}
\end{align}
Observe that $M$ is the only random quantity on the right hand side. Let
\[ \widetilde{h}_t(x) = \sum_{y \in F} \widetilde{p}(T-t,T;x,y) \overline{h}_0(y) \text{ for } 0 \leq t \leq T.\]
Note that
\begin{align*}
\widetilde{h}_t(x) &= \sum_{y,z \in F} \widetilde{p}(T-t,S;x,y) p(S,T;y,z) \overline{h}_0(z)\\
&= \sum_{y \in F} \widetilde{p}(T-t,S;x,y) \overline{h}_{T-S}(z) \text{ for } T-S \leq t \leq T.
\end{align*}
Hence as $t \mapsto \widetilde{p}(T-t,S;x,y)$, $T-S \leq t \leq T$, is independent from $\overline{h}_{T-S}$,
it follows that $\widetilde{h}_{t + (T-S)} \stackrel{d}{=} \breve{h}_t$, $0 \leq t \leq S$, with $\breve{h}$ as in the statement of the proposition. We can write
\begin{align*}
&\mathbf{E} \sum_{b \in E^*} \int_{T-S/4}^T (\nabla \overline{h}_t(b) - \nabla \widetilde{h}_t(b))^2 dt\\
= & \mathbf{E} \sum_{b \in E^*} \int_{T-S/4}^T \left( \sum_{y \in F} \nabla \overline{p}(T-t,T;b,y) \overline{h}_{T-t}(y) \right)^2 dt.
\end{align*}
The terms in the summation over $y \in F$ which are of distance at least $R^{\alpha+\epsilon/2}$ from $E$ make a negligible contribution to the summation by the Nash-Aronson estimate. Consequently, by making a change of variables and applying the Cauchy-Schwarz inequality, we see that it suffices to control
\begin{align}
\mathbf{E} \left[ R^{2\alpha+\epsilon} \sum_{b \in E^*} \sum_{y \in F} \int_0^{S/4} (\nabla \overline{p}(u,T;b,y))^2 M_u^2 du \right]. \label{harm::eqn::suff_bound}
\end{align}
Choosing $\sigma = \gamma \xi_{\rm NC}/4$ and applying \eqref{harm::eqn::hc_diff_bound} yields that the expression in \eqref{harm::eqn::suff_bound} is bounded by
\begin{align*}
& \sup_{0 \leq u \leq S/4} \bigg( \mathbf{E} O(M_u^2 R^{3\epsilon/2+ 4\alpha-2\gamma - \gamma \xi_{\rm NC}}) + \mathbf{E} O(M_0^2 M_{u}^2 R^{3\epsilon/2+2\alpha - 2\gamma \xi_{\rm NC}^2/4}) \bigg).
\end{align*}
Now, Lemma \ref{gl::lem::mom_max} and the Cauchy-Schwarz inequality imply $\mathbf{E} (M_u^2+M_0^2 M_{u}^2) = O_{\overline{\Lambda}}(R^{\epsilon/2})$ uniformly in $u$. Thus, we are left with the bound
\begin{align*}
O_{\overline{\Lambda}}(R^{2\epsilon+4\alpha - 2\gamma - \gamma \xi_{\rm NC}^2/2}).
\end{align*}
Taking $\rho_{\rm EC} = \xi_{\rm NC}^2/2$ gives \eqref{harm::eqn::ec_change_int}. Equation \eqref{harm::eqn::ec_change_sum} follows from exactly the same argument except using the first part of Lemma \ref{harm::lem::hc_time_space_deriv} rather than the second. The final part of the proposition is immediate from the construction and the Nash continuity estimate.
\end{proof}
We can now prove Proposition \ref{cd::prop::environment_change}.
\begin{proof}[Proof of Proposition \ref{cd::prop::environment_change}]
We now construct couplings as follows. First, we couple $(\eta^2, \breve{h}^2), (\nabla h, \overline{h})$ as in Proposition \ref{cd::prop::environment_change} for $\gamma = \gamma_2$. Equation \eqref{harm::eqn::ec_change_sum} implies that with $S_2 = R^{2\gamma_2}$ we have
\begin{align}
\mathbf{E} \sum_{x \in \widetilde{E}} |\overline{h}_{S_2}(x) - \breve{h}_{S_2}^2(x)|^2 &= O_{\overline{\Lambda}}(R^{\epsilon/2 + \delta_2+2\gamma_2}). \label{harm::eqn::ec_coupling_error1}
\end{align}
where $\widetilde{E} = B(x_0, 2R^{\alpha+\epsilon/100})$. Now we apply Proposition \ref{cd::prop::environment_change} a second time except with $\gamma = \gamma_1$ and starting at $S_2$ to yield a coupling $\big((\eta^1,\breve{h}^1), (\nabla h, \overline{h}) \big)$. Equation \eqref{harm::eqn::ec_change_int} implies the existence of $\tfrac{3}{4} R^{2\gamma_1} \leq S_1 \leq R^{2\gamma_1}$ such that with $S = S_1 + S_2$ we have
\begin{equation}
\label{harm::eqn::ec_coupling_error2}
\mathbf{E} \sum_{b \in E^*} |\nabla \overline{h}_{S}(b) - \nabla \breve{h}_{S}^1(b)|^2 =
O_{\overline{\Lambda}}(R^{\epsilon + \delta_1}).
\end{equation}
Let $\breve{p}^i$ be the kernel associated with $\eta^i$. In this coupling, $\overline{h}_{t}$ for $0 \leq t \leq S_2$ is independent from $\breve{p}^1$ and $\breve{p}^1,\breve{p}^2$ are independent. Let
\begin{align*}
\breve{h}_t(x)
&= \sum_{y,z \in F} \breve{p}^1(S-t,S_1;x,y)\breve{p}^2(0,S_2;y,z) \overline{h}_0(z)\\
&= \sum_{y \in F} \breve{p}^1(S-t,S_1;x,y) \breve{h}_{S_2}^{2}(y)
\end{align*}
for $S_2 \leq t \leq S$.
We have
\begin{align*}
& \mathbf{E} \sum_{b \in E^*} |\nabla \breve{h}_S(b) - \nabla \breve{h}_S^1(b)|^2
= \mathbf{E} \sum_{b \in E^*} \left(\sum_{y \in F} \nabla \breve{p}^1(0,S_1;b,y) (\breve{h}_{S_2}^2(y) - \overline{h}_{S_2}(y)) \right)^2\\
\leq& \left(\sum_{b \in E^*} \sum_{y \in \widetilde{E}} \mathbf{E} (\nabla \breve{p}^1(0,S_1;b,y))^2\right) \left(\sum_{y \in \widetilde{E}}\mathbf{E} (\breve{h}_{S_2}^2(y) - \overline{h}_{S_2}(y))^2\right) + O(\exp(-R^{10^{-5}\epsilon})),
\end{align*}
where the last inequality came from Cauchy-Schwarz and the Nash-Aronson estimates (Lemma \ref{symm_rw::lem::nash_aronson}).
It follows from equation (1.4) of Theorem 1.1 from \cite{DD05} and the Nash-Aronson estimates that
\[ \sum_{y \in \widetilde{E}} \sum_{b \in E^*} \mathbf{E} (\nabla \breve{p}^1(0,S_1;b,y))^2 = O(R^{2\alpha+\epsilon-4\gamma_1}).\]
Combining everything proves the proposition.
\end{proof}
\subsection{Approximation by the Average}
\begin{proposition}
\label{harm::prop::average_approx}
Suppose that we have the same setup as Theorem \ref{cd::thm::cd}. There exists $\rho_{\rm A} > 0$ depending only on $\mathcal {V}$ such that the following holds. If $\overline{h}^\rho(x)$ is the average of $\overline{h}(x)$ on the ball $B(x,R^\rho)$, then
\begin{align}
&\mathbf{E} \sum_{b \in E^*} \mathcal {V}''(\nabla h_{T_2}^\zeta(b)) (\nabla \overline{h}_{T_1}(b) - \nabla \overline{h}_{T_1}^\rho(b)) \nabla g(b) \notag\\
=& O_{\overline{\Lambda}}(R^{\epsilon+\rho+\alpha(1 -\rho_{\rm A})} \| \nabla g \|_\infty) \label{harm::eqn::average_approx}
\end{align}
for $T_1 \leq T_2$.
\end{proposition}
\begin{proof}
While the proof for $T_1 \neq T_2$ does not introduce any additional technical challenges, in the interest of keeping the notation simple we will only provide the proof for the case $T_1=T_2 =S$, with $S$ from Proposition \ref{cd::prop::environment_change}.
Let
\[ \gamma_1 = \alpha \left( \frac{ 1 + \tfrac{3}{16} \rho_{\rm EC}}{1+ \tfrac{1}{4} \rho_{\rm EC}} \right) \text{ and } \gamma_2 = \alpha.\]
Let $\delta_1,\delta_2$ be as in Proposition \ref{cd::prop::environment_change} with these choices of $\gamma_1,\gamma_2$. Note that then there exists $\rho_{\rm A} > 0$ depending only on $\mathcal {V}$ such that
\begin{align*}
\delta_1 < -\alpha \rho_{\rm A} \text{ and } 2\alpha - 4\gamma_1 + \delta_2 + 2 \gamma_2 < - \alpha \rho_{\rm A}.
\end{align*}
Consequently, by Proposition \ref{cd::prop::environment_change} it suffices to prove \eqref{harm::eqn::average_approx} with $\breve{h}, \breve{h}^\rho$ in place of $\overline{h}, \overline{h}^\rho$ where $\breve{h}_t^{\rho}(x)$ is the average of $\breve{h}_t(x)$ on $B(x,R^\rho)$. Moreover, \eqref{harm::eqn::ec_change_grad_single} combined with Lemma \ref{gl::lem::grad_error} imply that it suffices to prove \eqref{harm::eqn::average_approx} with $\mathcal {V}''(\nabla h_{S}^\zeta(b))$ replaced with $\breve{c}^1(b) = \mathcal {V}''(\eta_{S_1}^1)$. For $f \colon F \to \mathbf{R}$ let
\[ \mathcal {L} f(x) = \sum_{b \ni x} \breve{c}^1(b) \nabla f(b).\]
Summation by parts implies it suffices to bound
\begin{equation}
\label{harm::eqn::avg_err_approx_sbp}
\sum_{b \in \partial E^*} \breve{c}^1(b) (\breve{h}_{S}(x_b) - \breve{h}_{S}^\rho(x_b)) \nabla g(b) - \sum_{x \in E} [\breve{h}_{S}(x) - \breve{h}_{S}^\rho(x)] \mathcal {L} g(x).
\end{equation}
By the Nash continuity estimate (Lemma \ref{symm_rw::lem::nash_continuity_bounded}), we know that
\[ \mathbf{E} | \breve{h}_{S}(x_b) - \breve{h}_{S}^\rho(x_b)| = O_{\overline{\Lambda}}(R^{\epsilon+(\rho-\alpha)\xi_{\rm NC}}),\]
hence the boundary term in \eqref{harm::eqn::avg_err_approx_sbp} is of order $O_{\overline{\Lambda}}(R^{\epsilon+(\rho-\alpha)\xi_{\rm NC}+\alpha}\| \nabla g \|_\infty) = O_{\overline{\Lambda}}(R^{\epsilon + \alpha(1-\rho_{\rm A})} \| \nabla g \|_\infty)$, shrinking $\rho_{\rm A}$ if necessary.
We now deal with the interior term. For $y, \theta \in \mathbf{Z}^2$, let $y^\theta = y+\theta$. We are going to omit the times when referring to $\breve{p}^i$ and just write $\breve{p}^i(x,y)$ for $\breve{p}^i(0,S_i;x,y)$. Say that a bond $b$ is ``positively oriented'' if it points either in the positive hortzonal or vertical directions. For each triple $(x,y,b)$, let $y_b = y$ if $b$ is positively oriented and $2x-y$ otherwise. The latter is the reflection of $y$ about $x$. With $B_\rho = B(0,R^{\rho})$, we can rewrite the interior term of \eqref{harm::eqn::avg_err_approx_sbp} as
\begin{align*}
\frac{1}{|B_\rho|} \mathbf{E} \sum_{x \in E} \sum_{b \ni x} \sum_{y,z \in F} \sum_{\theta \in B_\rho} &\breve{c}^1(b)\big[ \breve{p}^1(x,y_b) \breve{p}^2(y_b,z) -\\
&\breve{p}^1(x_b^\theta,y_b)\breve{p}^2(y_b,z) \big] \overline{h}_0(z) \nabla g(b).
\end{align*}
Using the independence properties of $\eta^i, \overline{h}_0$ as well as the $\Delta^\beta$-harmonicity of $g$, we see that this is the same as
\begin{align*}
\frac{1}{|B_\rho|} \sum_{x ,b,y,z,\theta} &w(x,y,b,\theta) \mathbf{E} \big[\breve{p}^2(y_b,z) - \breve{p}^2(x,z)\big] \mathbf{E}[\overline{h}_0(z)] \nabla g(b)
\end{align*}
where the summation is over $x \in E$, $b \ni x$, $y,z \in F$, $\theta \in B_\rho$ and
\[ w(x,y,b,\theta) = \mathbf{E}\big[\breve{c}^1(b) (\breve{p}^1(x,y_b) - \breve{p}^1(x_b^\theta,y_b)) \big].\]
Here we are crucially using that $\breve{p}^2(x,z)$ does not depend on $y$.
For $b \ni x$ let $b' \ni x$ have the opposite orientation of $b$. By Proposition \ref{gl::prop::reflection_invariance}, we have
\[ w(x,y,b,\theta) = w(x,y,b',\theta)\]
Consequently, we can rewrite our sum as
\begin{align*}
&\frac{1}{|B_\rho|} \sum_{x ,b,y,z,\theta} w(x,y,b,\theta) \bigg( \mathbf{E} \big[(\breve{p}^2(y_b,z) - \breve{p}^2(x,z)) + (\breve{p}^2(y_{b'},z) - \breve{p}^2(x,z)) \big] \nabla g(b) \\
&+ \mathbf{E} \big[\breve{p}^2(y_{b'},z) - \breve{p}^2(x,z) \big] (\nabla g(b') - \nabla g(b))\bigg) \mathbf{E}[\overline{h}_0(z)]\\
\equiv& \frac{1}{|B_\rho|} \sum_{x ,b,y,z,\theta} w(x,y,b,\theta) \bigg( A(x,y,z,b) \nabla g(b) + B(x,y,z,b) (\nabla g(b') - \nabla g(b))\bigg) \mathbf{E}[\overline{h}_0(z)]
\end{align*}
where the summation is now only over positively oriented bonds.
We will deal with the term involving $A$ first. By the Nash-Aronson estimates (Lemma \ref{symm_rw::lem::nash_aronson}), the sum over $y$ of distance from $x$ more than $R^{\gamma_1+\epsilon}$ from $x$ is negligible. Similarly, we may ignore those $z$ with $|z-x| \geq R^{\gamma_2 + \epsilon}$. For $|y-x| \leq R^{\gamma_1+\epsilon}$, it is a consequence of Theorem 1.1 equation (1.5b) of \cite{DD05} and the Nash-Aronson estimates (Lemma \ref{symm_rw::lem::nash_aronson}) that
\begin{equation}
\label{harm::eqn::double_deriv}
|A(x,y,z,b)| \leq CR^{2\gamma_1+2\epsilon - 4 \gamma_2}.
\end{equation}
Indeed, this can be seen by rewriting the difference as a sum of $O(R^{2\gamma_1+2\epsilon})$ discrete second derivatives. Using that $\mathbf{E} \overline{h}_0(z) = O_{\overline{\Lambda}}(R^{\epsilon})$ from Lemma \ref{gl::lem::mom_max}, we thus have
\begin{align*}
&\frac{1}{|B_\rho|} \sum_{x ,b,y,z,\theta} w(x,y,b,\theta) A(x,y,z,b) \nabla g(b)\\
=& \frac{1}{|B_\rho|} \sum_{x,b,y,\theta} w(x,y,b,\theta) O_{\overline{\Lambda}}(R^{5\epsilon + 2 \gamma_1 - 2\gamma_2} \| \nabla g \|_\infty).
\end{align*}
Now, equation (1.4) of Theorem 1.1 of \cite{DD05} implies
\[ \mathbf{E} \sum_{y} |w(x,y,b,\theta)| = O(R^{\rho-\gamma_1})\]
uniformly in $x$ since $|x-x_b^\theta| \leq R^{\rho}$. Since the sum over $x$ includes $O(R^{2\gamma_2})$ terms, our total error is $O_{\overline{\Lambda}}(R^{10 \epsilon + \rho + \gamma_1} \| \nabla g \|_\infty)$.
We now turn to the term involving $B(x,y,z,b)$. Since $g$ is $\Delta^\beta$-harmonic so is $x \mapsto \nabla g((x,x+e_i))$ hence we have that $\nabla g(b') - \nabla g(b) = O(R^{-\gamma_2} \| \nabla g \|_\infty)$. Again applying equation (1.4) of Theorem 1.1 of \cite{DD05} we thus see
\[ B(x,y,z,b) (\nabla g(b') - \nabla g(b)) \mathbf{E}[\overline{h}_0(z)] = O_{\overline{\Lambda}}(R^{3\epsilon+\gamma_1 - 4\gamma_2} \| \nabla g \|_\infty).\]
This is of an even smaller magnitude than the corresponding term with $A$, so we also get an error of $O_{\overline{\Lambda}}(R^{10 \epsilon + \rho + \gamma_1} \| \nabla g \|_\infty)$. The proposition now follows from the explicit form of $\gamma_1$.
\end{proof}
\subsection{Change of Time}
\begin{proposition}
\label{cd::prop::change_of_time}
Suppose that we have the same setup as Theorem \ref{cd::thm::cd}. Let $T = R^{2\gamma}$ for $0 \leq \gamma \leq \alpha$ and fix $g \colon F \to \mathbf{R}$. We have that
\begin{align}
&\mathbf{E} \sum_{b \in E^*} \mathcal {V}''(\nabla h_{T}^\zeta(b)) (\nabla \overline{h}_{T}(b) - \nabla \overline{h}_0(b)) \nabla g(b) \label{harm::eqn::change_of_time}\\
=& O_{\overline{\Lambda}}(R^{\epsilon+\alpha(1-\rho_{\rm CoT}) + \gamma}\| \nabla g\|_\infty) \notag
\end{align}
where $\rho_{\rm CoT} > 0$ depends only on $\mathcal {V}$.
\end{proposition}
\begin{proof}
For $\rho > 0$, Proposition \ref{harm::prop::average_approx} implies that replacing $\overline{h}_T, \overline{h}_0$ by $\overline{h}_T^\rho, \overline{h}_0^\rho$, respectively, in \eqref{harm::eqn::change_of_time} introduces an error of
\begin{equation}
\label{harm::eqn::cotexp1}
O_{\overline{\Lambda}}(R^{\epsilon+\rho+(1-\rho_{\rm A})\alpha} \| \nabla g\|_\infty).
\end{equation}
We will now prove that
\begin{align}
&\mathbf{E} \sum_{b \in E^*} \mathcal {V}''( \nabla h_T^{\zeta}(b)) (\nabla \overline{h}_T^\rho(b) - \nabla \overline{h}_0^\rho(b)) \nabla g(b) \notag\\
= &O_{\overline{\Lambda}}(T R^{\epsilon+\alpha-\rho} \| \nabla g\|_\infty) \label{harm::eqn::cotexp2}.
\end{align}
Since $\sum_{b \in E^*} (\nabla g(b))^2 = O(R^{2\alpha} \| \nabla g \|_\infty^2)$, applying the Cauchy-Schwarz inequality to the expression on the left hand side implies that it suffices to show
\[ \mathbf{E} \sum_{b \in E^*} (\nabla \overline{h}_T^\rho(b) - \nabla \overline{h}_0^\rho(b))^2 = O_{\overline{\Lambda}}(T^2 R^{\epsilon-2\rho}).\] As
\[ |\partial_t \overline{h}_t^\rho(x)| \leq \frac{C_0}{R^{2\rho}} \sum_{b \in \partial B^*(x,R^{\rho})} |\nabla \overline{h}_t(b)|\]
we have that
\begin{align*}
& \mathbf{E} \sum_{b \in E^*} (\nabla \overline{h}_T^\rho(b) - \nabla \overline{h}_0^\rho(b))^2
\leq C_1T \mathbf{E} \int_0^T \sum_{b \in E^*} |\nabla \partial_t \overline{h}_t^\rho(b)|^2\\
\leq& \frac{C_2T}{R^{2\rho}} \mathbf{E} \int_0^T \sum_{b \in E_1^*} |\nabla \overline{h}_t(b)|^2
= \frac{C_2T^2}{R^{2\rho}} \mathbf{E} \sum_{b \in E_1^*} |\nabla \overline{h}_0(b)|^2,
\end{align*}
where $E_1 = B(x_0,R^{\alpha} + R^{\rho})$ and the final equality comes by stationarity. We know that the latter quantity is of order $O_{\overline{\Lambda}}(R^{\epsilon})$ by Lemma \ref{gl::lem::mom_max}. Equating the exponents in \eqref{harm::eqn::cotexp1}, \eqref{harm::eqn::cotexp2} gives the equation
\[ \rho + (1-\rho_{\rm A}) \alpha = 2\gamma + \alpha - \rho,\]
which leads to the choice
\[ \rho = \gamma + \frac{\rho_{\rm A}}{2} \alpha.\]
Combining everything gives the proposition, where $\rho_{\rm CoT} = \tfrac{1}{2} \rho_{\rm A}$.
\end{proof}
\subsection{Proof of Theorem \ref{cd::thm::cd}}
Assume that $T = R^{2\gamma}$ where $\gamma = \alpha \rho_{\rm CoT}/2$.
By Proposition \ref{cd::prop::change_of_time} it suffices to estimate
\begin{equation}
\label{cd::eqn::cd_reduction}
\sum_{b \in E^*} \mathbf{E}[\mathcal {V}''(\nabla h_{T}^\zeta(b)) \nabla \overline{h}_0(b))] \nabla g(b)
\end{equation}
provided we pay an error of $O_{\overline{\Lambda}}(R^{\epsilon+\alpha(1- \rho_{\rm CoT}/2)} \| \nabla g \|_\infty)$. Now the result follows from an argument similar to the proof of \eqref{cd::eqn::env_change} from the proof of Proposition \ref{cd::prop::environment_change}. Indeed, we let $\eta_t$ follow the $\texttt{SEGGS}_u$ dynamics driven by the same Brownian motions as $(h_t^\zeta, h_t^{\widetilde{\zeta}})$ but with $\eta_0$ independent from $(h_0^\zeta,h_0^{\widetilde{\zeta}})$, then use the Nash continuity estimate (Lemma \ref{symm_rw::lem::nash_continuity_bounded}) to argue that
\[ \mathbf{P}[ \max_{b \in E^*} |\nabla h_T^\zeta(b) - \eta_T(b)| \geq R^{\epsilon-\gamma \xi_{\rm NC}}] = O_{\overline{\Lambda}}(R^{-100}).\]
Thus applying the Cauchy-Schwarz inequality, we see that replacing $\mathcal {V}''(\nabla h_T^\zeta(b))$ with $\mathcal {V}''(\eta_T(b))$ in \eqref{cd::eqn::cd_reduction} introduces an error of order
\[ \left( \sum_{b \in E^*} \mathbf{E}[ (\nabla \overline{h}_0(b))^2]\right)^{1/2} \cdot O_{\overline{\Lambda}}(R^\alpha \| \nabla g \|_\infty) \cdot O_{\overline{\Lambda}}(R^{\epsilon - \gamma \xi_{\rm NC}}).\]
The result now follows as the first term is of order $O_{\overline{\Lambda}}(R^{\epsilon})$ by Lemma \ref{gl::lem::grad_error}.
\qed
\section{Harmonic Coupling}
\label{sec::harm}
Throughout this section we will make use of the following notation. For $F \subseteq \mathbf{Z}^2$ and $\phi \colon \partial F \to \mathbf{R}$, we let $\mathbf{P}_F^\phi$ be the law of the GL model on $F$ with boundary condition $\phi$. We will denote by $h^\phi$ a random variable distributed according to $\mathbf{P}_F^\phi$, where $F$ is understood through the domain of definition of $\phi$. Finally, for $g \colon F \to \mathbf{R}$ we let $\mathbf{Q}_F^{\phi,g}$ be the law of $(h^\phi - g)$. We will write $\mathbf{E}^\psi$ for the expectation under $\mathbf{P}_F^\psi$ if we want to emphasize $\psi$ and, similarly, $\mathbf{E}^{\psi,\widetilde{\psi}}$ for the expectation under a coupling of $\mathbf{P}_F^\psi, \mathbf{P}_F^{\widetilde{\psi}}$ if we want to emphasize both $\psi$ and $\widetilde{\psi}$. We also let $\beta = \beta(u)$ as in the statement of Theorem \ref{thm::clt} for a fixed tilt $u \in \mathbf{R}^2$.
Suppose that $\mu,\nu$ are measures with $\mu$ absolutely continuous with respect to $\nu$. Recall that the relative entropy of $\mu$ with respect to $\nu$ is the quantity
\[ \mathbf{H}(\mu|\nu) = \mathbf{E}_\mu \left[ \log \frac{d\mu}{d\nu} \right].\]
We begin by fixing $D \subseteq \mathbf{Z}^2$ with $R = {\rm diam}(D) < \infty$. Fix $\overline{\Lambda} > 0$ and let $\psi,\widetilde{\psi} \in \mathbf{B}_{\overline{\Lambda}}^u(D)$.
Morally, the idea of our proof is to get an explicit upper bound on the rate of decay of the symmetrized relative entropy
\[ \mathbf{H}(\mathbf{P}^{\widetilde{\psi}}|\mathbf{Q}^{\psi,\widehat{h}}) + \mathbf{H}(\mathbf{Q}^{\psi,\widehat{h}}|\mathbf{P}^{\widetilde{\psi}}),\]
where $\widehat{h}$ is the $\Delta^\beta$-harmonic extension of $\psi - \widetilde{\psi}$ from $\partial D \to D$, as $R \to \infty$, then invoke Pinsker's inequality, the well-known bound that the total variation distance of measures is bounded from above by the square-root of their relative entropy \cite{DZ98}:
\begin{equation}
\label{harm::eqn::pinsker} \| \mu - \nu \|_{TV}^2 \leq \tfrac{1}{2} \mathbf{H}(\mu|\nu) .
\end{equation}
We will show shortly that the symmetrized relative entropy takes the form
\begin{equation}
\label{harm::eqn::entropy_form}
\sum_{b \in D^*} \mathbf{E}^{\psi,\widetilde{\psi}} c(b) \nabla \widehat{h}(b) (\nabla \widehat{h}(b) - \nabla \overline{h}(b))
\end{equation}
where $\overline{h} = h^\psi - h^{\widetilde{\psi}}$ and $c(b)$ is a collection of conductances which are \emph{random} but uniformly bounded from above and below. In the Gaussian case, $c(b) \equiv c$ is constant, hence one can sum by parts, then use the harmonicity of $\widehat{h}$ to get that the entropy vanishes. The idea of our proof is to use Theorem \ref{cd::thm::cd} repeatedly to show that this approximately holds in expectation:
\begin{align}
\label{harm::eqn::entropy_approx_constant}
&\sum_{b \in D^*} \mathbf{E}^{\psi,\widetilde{\psi}} c(b) \nabla \widehat{h}(b)(\nabla \widehat{h}(b)-\nabla \overline{h}(b))\notag\\
=& \sum_{b \in D^*} a_u(b) \mathbf{E}^{\psi,\widetilde{\psi}} \nabla \widehat{h}(b) (\nabla\widehat{h}(b)-\nabla\overline{h}(b)) + O_{\overline{\Lambda}}(R^{-\delta})
\end{align}
for some $\delta > 0$, where $a_u(b) = \mathbf{E}[ \mathcal {V}''(\eta(b))]$ for $\eta \sim \texttt{SEGGS}_u$. Note that $a_u(b)$ depends only on $\mathcal {V},u,$ and the orientation of $b$.
Theorem \ref{cd::thm::cd} is only applicable if the distance of $b$ to $\partial D$ is $\Omega(R^\xi)$. This will force us to deal with a boundary term, the magnitude of which will in turn depend on the regularity of both $\nabla \overline{h}$ and $\widehat{h}$ near $\partial D$. Since we make no hypotheses on $\psi, \widetilde{\psi}$ other than being pointwise bounded it may very well be that neither $\overline{h}$ nor $\widehat{h}$ possess any regularity near $\partial D$.
We will resolve this issue by invoking the length-area comparison technique of Section \ref{sec::dyn}. This gives us that, for $\epsilon > 0$ fixed, there exists $R^{1-\epsilon} \leq R_D^1 \leq R_D^2 \equiv R_D^1 + R^{1-5\epsilon} \leq 2R^{1-\epsilon}$ such that
\[ \mathbf{E}^{\psi,\widetilde{\psi}} \sum_{b \in D^*(R_D^1, R_D^2)} |\nabla \overline{h}(b)|^2 = O_{\overline{\Lambda}}(R^{-\epsilon})\]
where $D(R_1,R_2) = \{ x \in D : R_1 \leq {\rm dist}(x,\partial D) < R_2\}$. Let $g$ be the $\Delta^\beta$-harmonic extension of $\overline{h}$ from $\partial D(R_D^1, R_D^2)$ to $D(R_D^1,R_D^2)$. Note that $g$ is the minimizer of the variational problem
\[ \widehat{g} \mapsto \sum_{b \in D^*(R_D^1, R_D^2)} a_u(b) (\nabla \widehat{g}(b))^2,\ \ \widehat{g}(x) = \overline{h}(x) \text{ for } x \in \partial D(R_D^1,R_D^2).\]
Indeed, the first order conditions for optimality are exactly that $\Delta^\beta \widehat{g}(x) = 0$ for $x \in D(R_D^1, R_D^2)$. Consequently,
\[ \mathbf{E}^{\psi,\widetilde{\psi}} \sum_{b \in D^*(R_D^1, R_D^2)} |\nabla g(b)|^2 = O_{\overline{\Lambda}}(R^{-\epsilon}).\]
Going back to \eqref{harm::eqn::entropy_form}, by invoking Pinsker's inequality, this implies that we can construct our initial coupling so that $\overline{h}$ is harmonic in $D(R_D^1,R_D^2)$ on an event $\mathcal {H}$ with $\mathbf{P}[\mathcal {H}] = 1-O_{\overline{\Lambda}}(R^{-\epsilon/2})$. On $\mathcal {H}$, we have that $\nabla \overline{h}(b) = O_{\overline{\Lambda}}(R^{6\epsilon-1})$ uniformly in $b \in \partial D(R_D)$ where $R_D = R_D^1 + \tfrac{1}{2} R^{1-5\epsilon}$. Thus with high probability $\overline{h}$ has plenty of regularity a bit away from $\partial D$ while $\psi - \widetilde{\psi}$ need not have any.
Moving to the subdomain $D(R_D)$ from $D$ is also useful since it possesses the $r$-exterior ball property for $r = R_D$. This means that for every $x \in \partial D(R_D)$ there exists $y \notin D(R_D)$ such that $B(y,R_D) \cap D(R_D) = \emptyset$ and $x \in \partial B(y,R_D)$. The importance of this property is that it implies pointwise regularity of harmonic functions in $D(R_D)$ near $\partial D(R_D)$, more so than one has for such functions in $D$ near $\partial D$ without further hypotheses. This is related to the notion of ``stochastic regularity'' \cite{KS98} and that random walk ``exits much more quickly'' from such domains.
\begin{figure}
\centering
\subfigure[Domain without the exterior ball property]{
\includegraphics[width=.30\textwidth]{figures/r_exterior.pdf}}
\hspace{0.1in}
\subfigure[Domain with the $r$-exterior ball property]{
\includegraphics[width=.30\textwidth]{figures/r_exterior_inner.pdf}}
\caption{The domain on the left hand side does not have the $r$-exterior ball property for due to the fjord in its upper right corner. However, the inner domain $D(r) = \{ x \in D: {\rm dist}(x,\partial D) \geq r\}$ shaded in light grey on the right hand side trivially does possess this property. The distinction is important since the $r$-exterior ball property is related to the regularity near the boundary of discrete harmonic functions.}
\end{figure}
The rest of this section is organized as follows. In subsection \ref{harm::subsec::entropy}, we will justify \eqref{harm::eqn::entropy_form}. The purpose of subsection \ref{subsec::entropy_est} is to prove a form of \eqref{harm::eqn::entropy_approx_constant}. Finally, in subsection \ref{harm::subsec::proof} we will put everything together to prove Theorems \ref{harm::thm::coupling} and \ref{harm::thm::mean_harmonic}.
Before we proceed, we would like to emphasize that in this section we will prove that a certain symmetrized relative entropy decays like a small, negative power of $R = {\rm diam}(D)$. Because of this, many of our estimates will be accurate only up to very small powers of $R$. In particular, we do not try to derive the ``best possible'' exponents and make many cautious choices in order to avoid carrying around overly complicated exponents.
\subsection{The Symmetrized Relative Entropy}
\label{harm::subsec::entropy}
\begin{lemma}
\label{harm::lem::entropy_form}
Suppose that $F \subseteq \mathbf{Z}^2$ is bounded and $\zeta,\widetilde{\zeta} \colon \partial F \to \mathbf{R}$ are given boundary conditions. If $g \colon F \to \mathbf{R}$ is any function such that $g|_{\partial F} = \zeta - \widetilde{\zeta}$ then
\begin{align*}
& \mathbf{H}(\mathbf{P}_F^{\widetilde{\zeta}}| \mathbf{Q}_F^{\zeta,g}) + \mathbf{H}(\mathbf{Q}_F^{\zeta,g}|\mathbf{P}_F^{\widetilde{\zeta}})\\
=& \sum_{b \in F^*} \mathbf{E}^{\zeta,\widetilde{\zeta}} \bigg[ \mathcal {V}''(\nabla h^{\zeta}(b)) \nabla g(b) \nabla(g-\overline{h})(b) + O( (|\nabla \overline{h}(b)|^2 + |\nabla g(b)|^2)|\nabla g(b)|) \bigg].
\end{align*}
where $\mathbf{E}^{\zeta,\widetilde{\zeta}}$ denotes the expectation under any coupling of $\mathbf{P}_F^{\zeta},\mathbf{P}_F^{\widetilde{\zeta}}$ and $\overline{h} = h^\zeta - h^{\widetilde{\zeta}}$.
\end{lemma}
\begin{proof}
The densities $p,q=q_g$ of $\mathbf{P}_F^{\widetilde{\zeta}}$ and $\mathbf{Q}_F^{\zeta,g}$ with respect to Lebesgue measure are given by
\begin{align*}
p(h) &= \frac{1}{\mathcal {Z}_p} \exp\left(-\sum_{b \in F^*} \mathcal {V}( \nabla (h \vee \widetilde{\zeta})(b))\right),\\
q(h) &= \frac{1}{\mathcal {Z}_q} \exp\left(-\sum_{b \in F^*} \mathcal {V}( \nabla [(h + g) \vee \zeta](b))\right).
\end{align*}
With $\mathbf{E}^{\zeta,g}$ the expectation under $\mathbf{Q}_F^{\zeta,g}$, we have
\begin{align*}
\mathbf{H}(\mathbf{Q}_F^{\zeta,g}|\mathbf{P}_F^{\widetilde{\zeta}}) + \log \frac{\mathcal {Z}_q}{\mathcal {Z}_p}
&= \sum_{b \in F^*} \mathbf{E}^{\zeta,g} \bigg[\mathcal {V}( (\nabla h \vee \widetilde{\zeta})(b)) - \mathcal {V}( \nabla ((h + g) \vee \zeta)(b)) \bigg]\\
&= -\sum_{b \in F^*} \left( \mathbf{E}^{\zeta} \int_0^1 \mathcal {V}'(\nabla [(h + (s-1)g) ](b))ds \right) \nabla g(b)
\end{align*}
Similarly, with $\mathbf{E}^{\widetilde{\zeta}}$ the expectation under $\mathbf{P}_F^{\widetilde{\zeta}}$,
\begin{align*}
\mathbf{H}(\mathbf{P}_F^{\widetilde{\zeta}}|\mathbf{Q}_F^{\zeta,g}) + \log \frac{\mathcal {Z}_p}{\mathcal {Z}_q}
&= \sum_{b \in F^*} \mathbf{E}^{\widetilde{\zeta}} \bigg[\mathcal {V}(\nabla ((h + g) \vee \zeta)(b)) - \mathcal {V}( (\nabla h \vee \widetilde{\zeta})(b)) \bigg]\\
&= \sum_{b \in F^*} \left( \mathbf{E}^{\widetilde{\zeta}} \int_0^1 \mathcal {V}'(\nabla (h + s g)(b)) ds \right) \nabla g(b)
\end{align*}
Thus
\begin{align*}
& \mathbf{H}(\mathbf{P}_F^{\widetilde{\zeta}}|\mathbf{Q}_F^{\zeta,g}) + \mathbf{H}(\mathbf{Q}_F^{\zeta,g}|\mathbf{P}_F^{\widetilde{\zeta}})\\
=&\sum_{b \in F^*} \left( \mathbf{E}^{\widetilde{\zeta}} \int_0^1 \mathcal {V}'(\nabla (h + s g)(b)) ds - \mathbf{E}^{\zeta} \int_0^1 \mathcal {V}'(\nabla [(h + (s-1)g) ](b))ds\right) \nabla g(b)\\
=& \sum_{b \in F^*} \left( \mathbf{E}^{\zeta,\widetilde{\zeta}} \int_0^1 \int_0^1 \mathcal {V}''(\nabla (h^{\zeta} + (s-1) g)(b) + r \nabla (g-\overline{h})(b)) dr ds\right)\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \nabla g(b) \nabla (g-\overline{h})(b).
\end{align*}
As $\mathcal {V}''$ is Lipschitz,
\begin{align*}
& \int_0^1 \int_0^1 \mathcal {V}''(\nabla (h^{\zeta} + (s-1) g)(b) + r \nabla (g-\overline{h})(b)) dr ds\\
=& \mathcal {V}''(\nabla h^{\zeta}(b)) + O(\nabla \overline{h}(b)) + O(\nabla g(b)).
\end{align*}
The lemma now follows by an application of Cauchy-Schwarz.
\end{proof}
\subsection{Estimating the Entropy}
\label{subsec::entropy_est}
Suppose that $E \subseteq \mathbf{Z}^2$ with ${\rm diam}(E) = R$ and $\zeta,\widetilde{\zeta} \in \mathbf{B}_{\overline{\Lambda}}^u(E)$. We are now going to give a general estimate of the symmetrized relative entropy of the previous lemma when $g \colon E \to \mathbf{R}$ is the $\Delta^\beta$-harmonic extension of $\overline{\zeta} = \zeta - \widetilde{\zeta}$ from $\partial E$ to $E$. The error is going to be a function of the regularity of $g$, $\overline{\zeta}$, and the number of balls required to cover annuli near $\partial E$. When all of the boundary data is smooth, the error is actually negligible.
To this end, we let
\[ \| \overline{\zeta}\|_{\nabla}^E = \max_{x,y \in \partial E} \frac{|\overline{\zeta}(x) - \overline{\zeta}(y)|}{|x-y|}.\]
Fix $\epsilon > 0$, let $\gamma_k = \epsilon k$, $M$ be the largest integer so that $\gamma_M < 1$, and $N_k$ be the number of balls of radius $R^{\gamma_k}$ necessary to cover $E(R^{\gamma_k}, R^{\gamma_{k+1}}) = \{x \in E : R^{\gamma_k} \leq {\rm dist}(x, \partial E) < R^{\gamma_{k+1}}\}$. Finally, with $1 \leq \ell \leq M$ fixed, let
\begin{align*}
\mathcal {E}_E &= |E^*| \| \nabla g \|_\infty^3 + \mathcal {E} \| \nabla g \|_\infty,\\
\mathcal {E}_B^\ell &= |(E')^*(R^{\gamma_\ell})| \| \nabla g \|_\infty \mathcal {E}^\ell,\\
\mathcal {E}_I^\ell &= \sum_{k=\ell+1}^M N_k R^{\epsilon + \gamma_k(1-\rho_{\rm CD})} \| \nabla g \|_\infty,
\end{align*}
where
\[ \mathcal {E}^\ell = ( R^{1/2} \| \overline{\zeta}\|_\nabla^E + R^{(\gamma_\ell - 1/2)\rho_{\rm B}}\| \overline{\zeta}\|_\infty ) \text{ and } \mathcal {E} = |\partial E^*| \| \overline{\zeta}\|_\infty \mathcal {E}^1.\]
\begin{proposition}
\label{harm::prop::ent_est}
We have that
\[ \mathbf{H}(\mathbf{P}_E^\zeta | \mathbf{P}_E^{\widetilde{\zeta}}) + \mathbf{H}( \mathbf{P}_E^{\widetilde{\zeta}} | \mathbf{P}_E^\zeta) = O_{\overline{\Lambda}}(\mathcal {E}_E + \mathcal {E}_I + \mathcal {E}_B).\]
\end{proposition}
\begin{proof}
Let $(h^\zeta,h^{\widetilde{\zeta}})$ be the stationary coupling of $\mathbf{P}_E^\zeta, \mathbf{P}_E^{\widetilde{\zeta}}$ and $\overline{h} = h^\zeta - h^{\widetilde{\zeta}}$. We begin with an \emph{a priori} estimate on the Dirichlet energy of $\overline{h}$. First fix $b \in \partial E^*$. Lemma \ref{symm_rw::lem::beurling} implies that $\nabla \overline{h}(b) = O(R^{1/2} \| \overline{\zeta}\|_\nabla^E + R^{-\rho_{\rm B}/2} \| \overline{\zeta}\|_\infty)$.
Applying Lemma \ref{gl::lem::ee} to the stationary dynamics $(h_t^\zeta,h_t^{\widetilde{\zeta}})$, we have that
\begin{align*}
& \sum_{b \in E^*} |\nabla \overline{h}(b)|^2
\leq C \sum_{b \in \partial E^*} |\overline{\zeta}(x_b)| |\nabla \overline{h}(b)|.
\end{align*}
Consequently, we have
\begin{align}
\sum_{b \in E^*} |\nabla \overline{h}(b)|^2
\leq C\mathcal {E}. \label{harm::eqn::dir_en_bound}
\end{align}
We now break the right hand side in the statement of Lemma \ref{harm::lem::entropy_form} into three terms:
\begin{align}
\sum_{b \in E^*} &\mathbf{E}^{\zeta,\widetilde{\zeta}} (|\nabla g(b)|^2 + |\nabla \overline{h}(b)|^2)|\nabla g(b)| \label{harm::eqn::ent_extra},\\
\sum_{b \in (E')^*(R^{\gamma_\ell})} &\mathbf{E}^{\zeta,\widetilde{\zeta}} \bigg[ \mathcal {V}''(\nabla h^{\zeta}(b)) \nabla g(b) \nabla(g-\overline{h})(b) \bigg] \label{harm::eqn::ent_boundary},\\
\sum_{b \in E^*(R^{\gamma_\ell})} &\mathbf{E}^{\zeta,\widetilde{\zeta}} \bigg[ \mathcal {V}''(\nabla h^{\zeta}(b)) \nabla g(b) \nabla(g-\overline{h})(b) \bigg] \label{harm::eqn::ent_interior}.
\end{align}
\noindent{\it Estimate of \eqref{harm::eqn::ent_extra}}
The first term in the summation is trivially bounded by $|E^*| \| \nabla g \|_\infty^3$. By \eqref{harm::eqn::dir_en_bound} the second is at most $\|\nabla g \|_\infty \mathcal {E}.$ This gives an error of $O(\mathcal {E}_E)$.
\noindent{\it Estimate of \eqref{harm::eqn::ent_boundary}}
The error is easily seen to be $|(E')^*(R^{\gamma_\ell})| \|\nabla g\|_\infty \delta$ where $\delta = \max_{x \in E'(R^{\gamma_\ell})} | g(x) - \overline{h}(x)|$, so we just need to estimate $\delta$. Fix $x \in E'(R^{\gamma_\ell})$. We know that we can write $g(x) = \mathbf{E} \overline{\zeta} (X_\tau)$ where $X$ is a random walk initialized at $x$ with bounded rates and $\tau$ its first exit from $E$. By Lemma \ref{symm_rw::lem::beurling}, the probability that $X_\tau$ exits at distance less than $R^{1/2}$ from $x \in E'(R^{\gamma_\ell})$ is $1-O(R^{ (\gamma_\ell - 1/2) \rho_{\rm B}})$. Now, $\overline{h}$ admits a similar representation though the random walk has time-varying rates. Nevertheless, the same statement still holds. Consequently, $|\overline{h}(x) - g(x)| \leq \mathcal {E}^\ell$, which leads to the desired bound. This gives an error of $O(\mathcal {E}_B^\ell)$.
\noindent{\it Estimate of \eqref{harm::eqn::ent_interior}}
We break the summation over $E^*(R^{\gamma_\ell})$ into the annuli $E_k = E^*(R^{\gamma_k}, R^{\gamma_{k+1}})$. Each annulus can be covered by $N_k$ balls of radius $R^{\gamma_k}$ by hypothesis. Let $B$ be such a ball. Applying Theorem \ref{cd::thm::cd} on $B$ yields an error of $O_{\overline{\Lambda}}(R^{\epsilon + \gamma_k(1-\rho_{\rm CD})} \| \nabla g \|_\infty)$. Repeating this on all $N_k$ balls for $\ell+1 \leq k \leq M$ implies that \eqref{harm::eqn::ent_interior} is equal to
\begin{align*}
\sum_{b \in E^*(R^{\gamma_\ell})} a_u(b) \mathbf{E}[ \nabla(\overline{h} - g)(b)] \nabla g(b) +
\sum_{k=\ell}^M N_k O_{\overline{\Lambda}}(R^{\epsilon + \gamma_k(1-\rho_{\rm CD})} \| \nabla g \|_\infty)
\end{align*}
where $a_u(b) = \mathbf{E}[ \mathcal {V}''(\eta(b))]$, $\eta \sim \texttt{SEGGS}_u$. Note that the error term is precisely $\mathcal {E}_I^\ell$. By our bound of \eqref{harm::eqn::ent_boundary}, we can rewrite the above expression as
\begin{align*}
\sum_{b \in E^*} a_u(b) \mathbf{E}[ \nabla(\overline{h} - g)(b)] \nabla g(b) +
O_{\overline{\Lambda}}(\mathcal {E}_E + \mathcal {E}_I).
\end{align*}
By summation by parts, we see that this is equal to
\begin{align*}
\sum_{b \in E^*} \mathbf{E}[ (\overline{h} - g)(x)] \Delta^\beta g(x) +
O_{\overline{\Lambda}}(\mathcal {E}_E + \mathcal {E}_I) = O_{\overline{\Lambda}}(\mathcal {E}_E + \mathcal {E}_I),
\end{align*}
where $\beta = \beta(u)$ as $g$ is $\Delta^\beta$-harmonic. This gives an error of $O(\mathcal {E}_I^\ell)$.
\end{proof}
\subsection{Proof of Theorems \ref{harm::thm::coupling} and \ref{harm::thm::mean_harmonic}}
\label{harm::subsec::proof}
\subsubsection{The Initial Coupling}
\begin{figure}
\centering
\subfigure[Stage 1: Langevin Dynamics in $D$]{
\includegraphics[width=.28\textwidth]{figures/stages_of_coupling1.pdf}}
\hspace{0.1in}
\subfigure[Stage 2: Harmonic in $D(R_D^1, R_D^2)$]{
\includegraphics[width=.28\textwidth]{figures/stages_of_coupling2.pdf}}
\hspace{0.1in}
\subfigure[The domain $E = D(R_D)$]{
\includegraphics[width=.28\textwidth]{figures/stages_of_coupling3.pdf}}
\caption{The entropy estimate consists of two stages of coupling, indicated by the images above. The annulus surrounded by a dashed line is $D(R_D^1, R_D^2)$, and inner light grey region in (c) is $E = D(R_D)$.}
\end{figure}
In this subsection we are going to show that we can construct a coupling of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$ so that with high probability the error terms from Proposition \ref{harm::prop::ent_est} are with high probability negligible when applied to $\mathbf{P}_E^\zeta, \mathbf{P}_E^{\widetilde{\zeta}}$ where $E = D(R_D)$, $R_D = cR^{1-\epsilon_D}$, some $\epsilon_D > 0$, and $(\zeta,\widetilde{\zeta}) = (h^\psi,h^{\widetilde{\psi}})|_{(\partial E)^2}.$ We will accomplish this using the following steps:
\begin{enumerate}
\item Take the stationary coupling $(h^\psi,h^{\widetilde{\psi}})$ of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$.
\item Invoke Lemma \ref{gl::lem::grad_error} to find an annulus $D(R_D^1, R_D^2)$ on which the Dirichlet energy of $\overline{h}$ is controlled.
\item Use Lemma \ref{harm::lem::entropy_form} to show that we can recouple the laws on $D(R_D^1, R_D^2)$ so that with high probability $\overline{h}$ is $\Delta^\beta$-harmonic. On this event we will have all of the regularity that we need on $\partial D(R_D)$, where $R_D^1 < R_D < R_D^2$.
\end{enumerate}
Fix $\epsilon_D > 0$ so small that $ \epsilon_D < (10^{-100}\rho_{\rm CD} \wedge \rho_{\rm B} \wedge \xi_{{\rm NC}})^2$, where $\xi_{{\rm NC}}$ is the exponent from the Nash continuity estimate (Lemma \ref{symm_rw::lem::nash_continuity_bounded}), $\rho_{\rm B}$ is the exponent from Lemma \ref{symm_rw::lem::beurling}, and $\rho_{\rm CD}$ is from Theorem \ref{cd::thm::cd}. We assume $R^{1-\epsilon_D} \leq R_D^1 \leq 2R^{1-\epsilon_D}$ has been chosen such that with $R_D^2 = R_D^1 + R^{1-5\epsilon_D}$ and $R_D = R_D^1 + \tfrac{1}{2} R^{1-5\epsilon_D}$:
\begin{align}
\label{harm::eqn::h_reg_assump_annulus}
&\mathbf{E}^{\psi,\widetilde{\psi}} \sum_{b \in D^*(R_D^1, R_D^2)} |\nabla \overline{h}(b)|^2 = O_{\overline{\Lambda}}(R^{-\epsilon_D}),\\
\label{harm::eqn::h_reg_assump_interior}
&\mathbf{E}^{\psi,\widetilde{\psi}} \sum_{b \in D^*(R_D)} |\nabla \overline{h}(b)|^2 = O_{\overline{\Lambda}}(R^{3\epsilon_D}).
\end{align}
That such a choice is possible is ensured by Lemma \ref{gl::lem::grad_error}. Let $E = D(R_D)$.
In order to apply Proposition \ref{harm::prop::ent_est} we need to make sure that we can arrange for the number of balls required to cover annuli near the boundary is not too large. Such estimates would come for free if $D$ was a lattice approximation of a smooth domain in $\mathbf{R}^2$. As we want Theorem \ref{harm::thm::coupling} to hold for general bounded subsets of $\mathbf{Z}^2$, such estimates do not necessarily hold uniformly but only on the average provided we are far enough from $\partial D$. Let $\gamma_k = k \epsilon_D$ and let $N_k$ be as in Proposition \ref{harm::prop::ent_est}. In particular, by using the averaging technique of Lemma \ref{gl::lem::grad_error} we can arrange for $R_D$ to be such that
\begin{align}
N_{k} &= O(R^{1+\gamma_2 - \gamma_k}),\ \
|E(R^{\gamma_k})| = O(R^{1+\gamma_{k+2}}),\ \
|\partial E| = O(R^{1+\gamma_1}) \label{harm::eqn::boundary_growth}
\end{align}
for all $k \leq M$, $M$ the largest integer such that $\gamma_M < 1$.
\begin{lemma}[Harmonic Coupling at the Boundary]
\label{harm::lem::harmonic_coupling} There exists a coupling $(h^{\psi},h^{\widetilde{\psi}})$ of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$ such that
\[ \mathcal {H} = \{ \overline{h} = h^{\psi} - h^{\widetilde{\psi}} \text{ is $\Delta^\beta$-harmonic in }D(R_D^1, R_D^2)\}\]
occurs with probability $1- O_{\overline{\Lambda}}(R^{-\epsilon_D/2})$.
\end{lemma}
\begin{proof}
Let $\overline{h} = h^{\psi} - h^{\widetilde{\psi}}$, $F = D(R_D^1, R_D^2)$, and let $g \colon F \to \mathbf{R}$ be $\Delta^\beta$-harmonic in $F$ with boundary values $\overline{h}$ on $\partial F$. By our choice of $R_D$,
\[ \sum_{b \in F^*} \mathbf{E}^{\psi,\widetilde{\psi}} (\nabla g(b))^2 = O_{\overline{\Lambda}}(R^{-\epsilon_D})\]
as $g$ is harmonic in $F$, has the same boundary values as $\overline{h}$, and $\overline{h}$ satisfies the same estimate.
Let $\zeta, \widetilde{\zeta} = (h^{\psi},h^{\widetilde{\psi}})|_{ \partial F \times \partial F}$. Conditional on $(\zeta,\widetilde{\zeta})$, let $\mathbf{P}_F^\zeta, \mathbf{P}_F^{\widetilde{\zeta}}$ have the laws of the GL model on $F$ with boundary conditions $\zeta,\widetilde{\zeta}$, respectively, and let $\mathbf{Q}_F^{\zeta,g}$ have the law of $h^{\zeta} - g$ where $h^{\zeta} \sim \mathbf{P}_F^{\zeta}$. It follows from the Cauchy-Schwarz inequality and Lemma \ref{harm::lem::entropy_form} that
\[ \mathbf{E}^{\psi,\widetilde{\psi}}[ \mathbf{H}(\mathbf{P}_F^{\widetilde{\zeta}}|\mathbf{Q}_F^{\zeta,g}) + \mathbf{H}(\mathbf{Q}_F^{\zeta,g}|\mathbf{P}_F^{\widetilde{\zeta}})] = O_{\overline{\Lambda}}(R^{-\epsilon_D}).\]
The lemma follows by invoking Pinsker's inequality \eqref{harm::eqn::pinsker}.
\end{proof}
\subsubsection{Regularity Estimate}
In the following lemma we will use $(h^{\psi},h^{\widetilde{\psi}})$ to indicate a random variable with joint law given by the coupling of $\mathbf{P}_D^\psi,\mathbf{P}_D^{\widetilde{\psi}}$ from Lemma \ref{harm::lem::harmonic_coupling} and $\mathcal {H}$ the corresponding event. Let $(\zeta,\widetilde{\zeta}) = (h^\psi,h^{\widetilde{\psi}}) |_{ \partial E \times \partial E}$. Let $g \colon E \to \mathbf{R}$ be the $\Delta^\beta$-harmonic extension of $\overline{\zeta} = \zeta - \widetilde{\zeta}$ from $\partial E$ to $E$. Recall the definition of $\|\overline{\zeta}\|_\nabla^E$ just before the statement of Proposition \ref{harm::prop::ent_est}.
\begin{lemma}
\label{harm::lem::harmonic_gradient}
There exists $1 \leq c_D \leq 10$ so that
\begin{align}
\left(\mathbf{E}^{\psi,\widetilde{\psi}}(\| \overline{\zeta}\|_{\nabla}^E)^p \mathbf{1}_\mathcal {H}\right)^{1/p} &= O_{\overline{\Lambda},p}(R^{c_D \epsilon_D-1}), \label{harm::eqn::boundary_reg}\\
\left( \mathbf{E}^{\psi,\widetilde{\psi}}[ \max_{b \in E^*} |\nabla g(b)|^p \mathbf{1}_{\mathcal {H}}] \right)^{1/p} &= O_{\overline{\Lambda},p}(R^{c_D \epsilon_D-1}) \label{harm::eqn::harm_reg}
\end{align}
for every $p \geq 1$.
\end{lemma}
\begin{proof}
By construction, $E$ has the $r$-exterior ball property for $r = R^{1-\epsilon_D}$. Furthermore, if $x,y \in \partial E$ with $|x-y| \leq \tfrac{1}{8}R^{1-5\epsilon_D}$ then the shortest path connecting $x$ to $y$ in $\mathbf{Z}^2$ is contained in $B(x,\tfrac{1}{4} R^{1-5\epsilon_D})$. Consequently,
\begin{equation}
\label{harm::eqn::boundary_reg1}
|\overline{\zeta}(x) - \overline{\zeta}(y)| \leq \breve{g}|x-y|
\end{equation}
where
\[ \breve{g} = \max\{ |\nabla \overline{h}(b)| : b \in D^*(R_D^1 + \tfrac{1}{4} R^{1-5\epsilon_D}, R_D^1 + \tfrac{3}{4} R^{1-5\epsilon_D})\}.\]
Let $M = \max\{ |\overline{h}(x)| : x \in D\}$. Since $\overline{h}$ is harmonic in $D(R_D^1,R_D^2)$ on $\mathcal {H}$, we have that
\begin{equation}
\label{harm::eqn::g_bound}
\breve{g} \leq \overline{g} \equiv \frac{C M}{R^{1-5\epsilon_D}}.
\end{equation}
We may assume without loss of generality that $C \geq 100$.
If $|x-y| \geq \tfrac{1}{8}R^{1-5\epsilon_D}$ then
\[ \overline{g} |x-y| \geq \tfrac{1}{8} CM \geq 2M \geq |\overline{\zeta}(x) - \overline{\zeta}(y)|.\]
Therefore Lemma \ref{dhf::lem::exterior_regularity} implies, by possibly increasing $C > 0$, that
\begin{equation}
\label{harm::eqn::max_gradient}
\max_{b \in E^*} |\nabla g(b)| \leq \frac{CM}{R^{1-5\epsilon_D}} \left[ \log R + \frac{R}{R^{1-\epsilon_D}} \right] \leq \frac{C M}{R^{1-6\epsilon_D}}
\end{equation}
on $\mathcal {H}$. Trivially,
\[ M \leq \max_{x \in D} |h^\psi(x)| + \max_{x \in D} |h^{\widetilde{\psi}}(x)|\]
and by Lemma \ref{gl::lem::mom_max} we know that $(\mathbf{E} M^p)^{1/p} = O_{\overline{\Lambda},p}(R^{\epsilon})$. This clearly gives \eqref{harm::eqn::harm_reg}. Combining \eqref{harm::eqn::boundary_reg1} with \eqref{harm::eqn::g_bound} gives \eqref{harm::eqn::boundary_reg}.
\end{proof}
\subsubsection{Putting Everything Together}
\begin{proof}[Proof of Theorem \ref{harm::thm::coupling}]
To prove the theorem we just have to estimate the error terms from Proposition \ref{harm::prop::ent_est} on $\mathcal {H}$. First of all, by Lemmas \ref{gl::lem::mom_max}, \ref{harm::lem::harmonic_gradient} we observe
\begin{equation}
(\mathbf{E}[ (\mathcal {E})^p \mathbf{1}_\mathcal {H}])^{1/p} = O_{\overline{\Lambda},p}(R^{1+2\epsilon_D - a_1\rho_{\rm B}})
\end{equation}
for $a_1 = 1/10$. Consequently,
\begin{align}
\mathbf{E}[\mathcal {E}_E \mathbf{1}_{\mathcal {H}}] =& O_{\overline{\Lambda}}(R^2 \cdot R^{3c_D \epsilon_D - 3} + R^{1+2\epsilon_D - a_1\rho_{\rm B} + c_D \epsilon_D - 1}) \notag\\
=& O_{\overline{\Lambda}}(R^{c_1\epsilon_D - a_1\rho_{\rm B}}) \label{harm::eqn::exp1}
\end{align}
for $c_1 < 100$. Using Lemmas \ref{gl::lem::mom_max}, \ref{harm::lem::harmonic_gradient} again, we see that
\begin{align}
\mathbf{E}[\mathcal {E}_B^\ell \mathbf{1}_{\mathcal {H}}] =& O_{\overline{\Lambda}}(R^{1+2\epsilon_D + \gamma_{\ell+2} - 1}) O_{\overline{\Lambda}}(R^{1/2+c_D\epsilon_D - 1} + R^{\epsilon_D+(\gamma_\ell - 1/2) \rho_{\rm B}}) \notag\\
=&
O_{\overline{\Lambda}}(R^{c_1\epsilon_D + 2\gamma_{\ell+2} - a_1\rho_{\rm B}}) \label{harm::eqn::exp2},
\end{align}
the last line coming as $\rho_{\rm B} \leq 1$.
Finally, Lemma \ref{harm::lem::harmonic_gradient} clearly implies
\begin{align}
\mathbf{E}[ \mathcal {E}_I^\ell \mathbf{1}_\mathcal {H}] &= \sum_{k=\ell+1}^M O_{\overline{\Lambda}}(R^{1+\gamma_2-\gamma_k} R^{\epsilon_D + \gamma_k(1-\rho_{\rm CD})} R^{c_2 \epsilon_D - 1}) \notag \\
&= \sum_{k=\ell+1}^M O_{\overline{\Lambda}}(R^{c_3 \epsilon_D - \gamma_\ell \rho_{\rm CD}}) \label{harm::eqn::exp3}
\end{align}
for $c_3 \leq 1000$. The exponents from \eqref{harm::eqn::exp1}, \eqref{harm::eqn::exp2}, \eqref{harm::eqn::exp3} are
\[ c_1 \epsilon_D - a_1 \rho_{\rm B},\ \ c_1 \epsilon_D + 2 \gamma_{\ell+2} - a_1 \rho_{\rm B},\ \ c_3 \epsilon_D - \gamma_{\ell} \rho_{\rm CD}.\]
Choosing $\ell > 10^5$ we see that the second and third exponents are negative and the first is clearly negative.
\end{proof}
We finish this section with the short proof of Theorem \ref{harm::thm::mean_harmonic}.
\begin{proof}[Proof of Theorem \ref{harm::thm::mean_harmonic}]
Assume that we still have the setup of the previous theorem except $\widetilde{\psi} = 0$. Then there exists $\delta > 0$ so that we can find a coupling of $\mathbf{P}_D^\psi, \mathbf{P}_D^{\widetilde{\psi}}$ so that $\overline{h}$ is harmonic in $E$ on an event $\widetilde{\mathcal {H}}$ (this is different from $\mathcal {H}$ in the proof of Theorem \ref{harm::thm::coupling}) with probability $1-O_{\overline{\Lambda}}(R^{-\delta})$. Let $g,\widehat{h}$ be the harmonic extensions of $\overline{h}$, $\mathbf{E} h^\psi(x)$ from $\partial E$ to $E$, respectively. For $x \in E$ we have that
\begin{align*}
\mathbf{E} h^\psi(x) &= \mathbf{E} \overline{h}(x) = \mathbf{E} g(x) (1-\mathbf{1}_{\widetilde{\mathcal {H}}^c}) + \mathbf{E} \overline{h}(x) \mathbf{1}_{\widetilde{\mathcal {H}}^c}\\
&= \widehat{h}(x) + \mathbf{E} (\overline{h}(x) - g(x)) \mathbf{1}_{\widetilde{\mathcal {H}}^c}
\end{align*}
since $\mathbf{E} h^{\widetilde{\psi}}(x) = 0$.
Since $\psi \in \mathbf{B}_{\overline{\Lambda}}^u(D)$, Lemma \ref{gl::lem::mom_max} implies both $\mathbf{E} \overline{h}^2(x) = O_{\overline{\Lambda}}(R^{\delta/2})$ and $\mathbf{E} g^2(x) = O_{\overline{\Lambda}}(R^{\delta/2})$. Consequently, an application of Cauchy-Schwarz yields
\[ \mathbf{E}\big[ |\overline{h}(x)| + |g(x)| \big] \mathbf{1}_{\widetilde{\mathcal {H}}^c} = O_{\overline{\Lambda}}(R^{-\delta/4}),\]
from which the theorem follows.
\end{proof}
\section{The Central Limit Theorem}
\label{sec::clt}
We will prove Theorem \ref{thm::clt} in this section. The primary inputs are Theorem \ref{harm::thm::coupling} and the main result of either \cite{GOS01} or \cite{NS97}. Throughout, we let $D \subseteq \mathbf{R}^2$ be a connected, bounded, smooth domain and for each $n$ let $D_n = D \cap \tfrac{1}{n} \mathbf{Z}^2$. We will be dealing with both discrete and continuum derivatives, so to keep the notation consistent with the rest of the article we will still let $\nabla, \Delta$ denote the discrete gradient and Laplacian, respectively, and use $\overline{\nabla}$ and $\overline{\Delta}$ for their continuum counterpart. We begin with a simple analysis lemma. Let $\beta = \beta(u)$ be as in the statement of Theorem \ref{thm::clt}.
\begin{lemma}
\label{clt::lem::discrete_de}
For each $\epsilon > 0$ there exists $c > 0$ so that for all $g \in C^\infty(D)$ we have that
\begin{align*}
\sum_{b \in D_n^*} |\nabla g(b)|^2 &\leq c \sup_{x \in D} \|\overline{\nabla} g(x)\|^2,\ \
\sum_{b \in D_n^*} |\nabla g(b)|^2 \leq c \| g \|_{H^{2+\epsilon}(D)}^2,\\
\sum_{b \in \partial D_n^*} |\nabla g(b)| &\leq c \| g \|_{H^{2+\epsilon}(D)},\ \
\sum_{x \in D_n} | \Delta^\beta g(x)| \leq c \| g\|_{H^{3+\epsilon}(D)}.
\end{align*}
\end{lemma}
\begin{proof}
The first claim is obvious since $|\nabla g(b)| \leq \tfrac{1}{n} \sup_{x \in D} \|\overline{\nabla} g(x)\|$.
As for the second claim, we note that the Sobolev embedding theorem implies that for each $\epsilon > 0$ there exists $c,c' > 0$ so that
\[ \sup_{x \in D} \|\overline{\nabla} g(x) \| \leq c' \| \overline{\nabla} g \|_{H^{1+\epsilon}(D)} \leq c \| g\|_{H^{2+\epsilon}(D)}.\]
See, for example, from Proposition 1.3 in Chapter 4 of \cite{TAY96}. The final two claims are proved similarly.
\end{proof}
Let $\varphi_n(x) = nu \cdot x$. Fix a continuous function $f \colon \mathbf{R}^2 \to \mathbf{R}$ and let $h^n$ have the law of the GL model on $D_n$ with $h^n(x) = f(x) + \varphi_n(x)$ for $x \in \partial D_n$, $\eta^{n,D} = \nabla h^n$, and for $g \in H^{3+\epsilon}(D)$ define
\begin{align*}
\xi_\nabla^{n,D}(g) &= \sum_{b \in D_n^*} a_u(b) \nabla g(b) ( \eta^{n,D}(b) - \nabla \varphi_n(b))
\end{align*}
where $a_u(b) = \mathbf{E}[ \mathcal {V}''(\eta(b))]$ for $\eta \sim \texttt{SEGGS}_u$. Note that $a_u(b)$ depends only on the orientation of $b$.
\begin{lemma}
\label{clt::lem::exp_moments}
For each $\epsilon > 0$ there exists $c > 0$ so that for all $g \in H^{3+\epsilon}(D)$ we have
\[ \mathbf{E} \exp( \xi_\nabla^{n,D}(g)) \leq \exp\left[ c \left( \|f\|_\infty^2 + c \| g \|_{H^{3+\epsilon}(D)}^2\right) \right].\]
\end{lemma}
\begin{proof}
Using summation by parts, we have
\begin{align*}
|\mathbf{E} \xi_\nabla^{n,D}(g)|
\leq& \sum_{x \in D_n} |\mathbf{E} (h^n(x) - \varphi_n(x)) \Delta^\beta g(x)| + c_1\sum_{b \in \partial D_n^*} |f(x_b) \nabla g(b)|\\
\leq& c_2 \| f \|_{\infty} \| g \|_{H^{3+\epsilon}(D)}.
\end{align*}
In the final inequality we used that $|\mathbf{E}(h^n(x) - \varphi_n(x))| \leq \|f \|_\infty$, which is a consequence of Lemma \ref{gl::lem::hs_mean_cov}, in addition to Lemma \ref{clt::lem::discrete_de}.
The exponential Brascamp-Lieb inequality (Lemma \ref{bl::lem::bl_inequalities}) combined with \eqref{dgff::covariance} and the previous lemma implies there exists $c_3,c_4 > 0$ so that
\begin{align*}
\mathbf{E} \exp( \xi_\nabla^{n,D}(g))
&\leq \exp\left[ c_3 \left(\|f\|_\infty \| g \|_{H^{3+\epsilon}(D)} + \sum_{b \in D_n^*} (\nabla g(b))^2 \right) \right]\\
&\leq \exp\left[ c_4 \left(\|f\|_\infty^2 +\| g \|_{H^{3+\epsilon}(D)}^2 \right) \right].
\end{align*}
\end{proof}
\begin{lemma}
For each $\kappa > 4$, the law of $\xi_\nabla^{n,D}$ induces a tight sequence on $H^{-\kappa}(D)$ equipped with the weak topology.
\end{lemma}
\begin{proof}
Fix $\kappa > 4$. It suffices to show that for each $\delta > 0$ there exists $M = M(\delta)$ such that
\[ \mathbf{P}[ \| \xi_\nabla^{n,D} \|_{H^{-\kappa}(D)} \geq M] \leq \delta\]
as the Banach-Alaoglu theorem yields that the ball $\{ \| g \|_{H^{-\kappa}(D)} \leq M\}$ is compact in the weak topology of $H^{-\kappa}(D)$.
Let $\epsilon = \tfrac{1}{2}(\kappa-4)$ and $\kappa' = \kappa-1-\epsilon > 3$. Let $(\widetilde{g}_k)$ be the eigenvectors of $\overline{\Delta}$ on $D$, normalized to be orthonormal in $L^2(D)$, with negative eigenvalues $(\lambda_k)$ ordered to be non-increasing in $k$. Let $g_k = (1-\overline{\Delta})^{-\kappa/2} \widetilde{g}_k$. By \eqref{gff::eqn::sobolev_inner_product}, $(g_k)$ is an orthonormal basis of $H^{\kappa}(D)$. As $g_k / (1-\lambda_k)^{(1+\epsilon)/2} = (1-\overline{\Delta})^{-(1+\epsilon)/2} g_k$ we have that
\[ \| g_k \|_{H^{\kappa'}(D)} = \| (1-\overline{\Delta})^{-(1+\epsilon)/2} g_k \|_{H^{\kappa}(D)} = \frac{1}{(1-\lambda_k)^{(1+\epsilon)/2}}.\]
The Weyl formula implies that $k/(-\lambda_k)$ tends to a constant $c = c_D$ depending only on $D$ as $k \to \infty$. Therefore there exists $c_D' \geq c_D$ so that
\[ \| k^{(1+\epsilon)/2} g_k \|_{H^{\kappa'}(D)} \leq c_D' \text{ for all } k \in \mathbf{N}.\]
Combining the above with Chebychev's inequality yields
\[ \mathbf{P}[ |\xi_\nabla^{n,D}(g_k)| \geq M/k^{(1+\epsilon/2)/2} ] = \mathbf{P}[ |\xi_\nabla^{n,D}(k^{(1+\epsilon)/2} g_k)| \geq M k^{\epsilon/4}] \leq \exp(c - M k^{\epsilon/4} )\]
where $c = c(\epsilon,D,f)$.
Consequently, letting $A_M^n = \cap_k \{ |\xi_\nabla^{n,D}(g_k)| \leq M/k^{(1+\epsilon/2)/2}\}$, a union bound yields $\mathbf{P}[ A_M^n] \to 1$ as $M \to \infty$.
Note that if $g \in H^{\kappa}(D)$, $g = \sum_k \alpha_k g_k$ for $(\alpha_k) \in \ell^2$, we have
\[ \xi_\nabla^{n,D}(g) = \sum_k \alpha_k \xi_\nabla^{n,D}(g_k)\]
since $H^{\kappa}(D)$-convergence implies uniform convergence as $\kappa > 4$ and $\xi_\nabla^{n,D}$ is obviously continuous in the uniform topology on functions $D_n \to \mathbf{R}$.
Let
\[ N_\epsilon = \sum_{k} \frac{1}{k^{1+\epsilon/2}}.\]
On $A_M^n$, note that
\[ |\xi_{\nabla}^{n,D}(g)| \leq \left(\sum_k \alpha_k^2\right)^{1/2} \left( \sum_k (\xi_\nabla^{n,D}(g_k))^2\right)^{1/2} \leq \| g\|_{H^{\kappa}(D)} M \sqrt{N_\epsilon}.\]
Consequently,
\begin{align*}
\mathbf{P}\bigg[ \sup_{ \|g \|_{H^{\kappa}(D) }\leq 1} |\xi_\nabla^{n,D}(g)| \geq M \sqrt{N_\epsilon} \bigg]
&\leq \mathbf{P}[(A_M^n)^c].
\end{align*}
This proves for every $\delta > 0$ there exists $\widetilde{M} = \widetilde{M}(\delta)$ sufficiently large so that
\[ \mathbf{P}[ \| \xi_\nabla^{n,D}\|_{H^{-\kappa}(D)} \geq \widetilde{M}] \leq \delta\]
for every $n \in \mathbf{N}$.
\end{proof}
Let $\eta \sim \texttt{SEGGS}_u$, but thought of as a random gradient field on $(\tfrac{1}{n} \mathbf{Z}^2)^*$. Fix a base point $x^* \in \partial D$ and let $x_n$ be a point in $\partial D_n$ with minimal distance to $x^*$. Set $h^{n,0}(x_n) = 0$ and let $h^{n,0}$ be the function satisfying $\nabla h^{n,0} = \eta$.
Let
\begin{align*}
\xi_\nabla^{n}(g) &= \sum_{b} a_u(b) \nabla g(b)(\eta^{n}(b) - \nabla \varphi_n(b)) \text{ for } g \in C_0^\infty(\mathbf{R}^2).
\end{align*}
Corollary 2.2 of \cite{GOS01} implies that for any $g_1,\ldots,g_k \in C_0^\infty(\mathbf{R}^2)$ fixed, the random vector $(\xi_\nabla^n(g_1),\ldots, \xi_\nabla^n(g_k))$ converges in distribution to a zero-mean Gaussian vector $(Z_1,\ldots,Z_k)$ with covariance ${\rm Cov}(Z_i, Z_j) = (g_i,g_j)_\nabla^A$ for $A = A(u,\mathcal {V})$ depending only on the tilt $u$ and $\mathcal {V}$. Note that our definition of $\xi_\nabla^n$ differs from $\xi^\epsilon$ in \cite{GOS01} in that we do not have a normalization of $n^{-1}$. The reason is that $\xi_\nabla^n$ operates on discrete gradients of $C_0^\infty(\mathbf{R}^2)$ functions, which themselves are of order $n^{-1}$.
\begin{lemma}
\label{clt::lem::bc_harm}
There exists $\overline{\Lambda}$ depending only on $\mathcal {V}$ such that
\[ \mathbf{P}[ h^{n,0} |_{\partial D_n} \in \mathbf{B}_{\overline{\Lambda}}^u(D_n)] = 1- O(n^{-8}).\]
\end{lemma}
\begin{proof}
Let $x \in D_n$ and $\overline{x} = x-x_n$. Combining the exponential Brascamp-Lieb inequality with \eqref{eqn::var_bound} yields for $x \in \partial D_n$ that
\[ \mathbf{E} \exp( h^{n,0}(x) - \varphi_n(\overline{x})) = \mathbf{E} \exp( h^{n,0}(x) - h^{n,0}(x_n) - \varphi_n( \overline{x})) \leq \exp( C \log n) = n^C.\]
Assume without loss of generality that $C \geq 1$. By Chebychev's inequality,
\[ \mathbf{P}[ |h^{n,0}(x) - \varphi_n(\overline{x})| \geq 10 C \log n] \leq n^{-9}.\]
Using a union bound we thus have
\[ \mathbf{P}[ \max_{x \in \partial D_n} |h^{n,0}(x) - \varphi_n(\overline{x})| \geq 10 C \log n] \leq O(n^{-8}).\]
Consequently, taking $\overline{\Lambda} = 10C$ we have that $\mathbf{P}[ h^{n,0}|_{\partial D_n} \in \mathbf{B}_{\overline{\Lambda}}^u(D_n)] = 1-O(n^{-8})$.
\end{proof}
By the same proof as Lemma \ref{clt::lem::exp_moments} we have
\begin{equation}
\label{clt::eqn::seggs_exp_moment}
\mathbf{E} \exp( \xi_\nabla^n(g)) \leq \exp(c \| g\|_{H^{3+\epsilon}(D)}^2).
\end{equation}
\begin{proof}[Proof of Theorem \ref{thm::clt}]
Fix $\kappa > 4$ and $f \colon \mathbf{R}^2 \to \mathbf{R}$ continuous. Let $h^0,h^f$ be GFFs on $D$ with respect to $(\cdot,\cdot)_\nabla^A$, $A = A(u,\mathcal {V})$ as before, where $h^0$ has zero boundary conditions and the boundary condition of $h^f$ is given by $f|_{\partial D}$. Let $\xi_\nabla^D$ be a weak-$H^{-\kappa}(D)$ subsequential limit of $(\xi_\nabla^{n,D})$. We will prove for any $g_1,\ldots,g_k \in C^\infty(D)$ that
\[ (\xi_\nabla^D(g_1),\ldots,\xi_\nabla^D(g_k)) \stackrel{d}{=} ( (h^f,g_1)_\nabla^A,\ldots,(h^f,g_k)_\nabla^A)\]
since the continuity of $\xi_\nabla^D$ implies that its law is determined by its projections onto $C^\infty(D)$, a dense subset of $H^\kappa(D)$. We will identify $A$ at the end of the proof. To establish this, it suffices to show that
\[ (\xi_\nabla^{n,D}(g_1),\ldots,\xi_\nabla^{n,D}(g_k)) \stackrel{d}{\to} ((h^f,g_1)_\nabla^A,\ldots,(h^f,g_k)_\nabla^A).\]
We will first prove the result for $C_0^\infty(D)$, then using an approximation argument generalize to $C^\infty(D)$.
By Lemma \ref{clt::lem::bc_harm}, with probability $1-O(n^{-8})$ we can apply Theorem \ref{harm::thm::coupling} to $\mathbf{P}_{D_n}^\psi, \mathbf{P}_{D_n}^{\widetilde{\psi}}$ where $\psi = f + \varphi_n$ and $\widetilde{\psi} = h^{n,0}|_{\partial D_n}$. This implies the existence of $\epsilon,\delta > 0$ independent of $n$ such that we can couple $h^n, h^{n,0}$ so that with $\widehat{h}^n$ the $\Delta^\beta$-harmonic extension of $h^n - h^{n,0}$ from $\partial D_n(n^{-\epsilon})$ to $D_n(n^{-\epsilon})$, we have
\[ \mathbf{P}[ \mathcal {H}^c ] = O(n^{-\delta}) \text{ where } \mathcal {H} = \{\overline{h}^n = \widehat{h}^n \text{ in } D_n(n^{-\epsilon})\}\]
for all $n$ large enough. The reason that we see $n^{-\epsilon}$ rather than $n^{1-\epsilon}$ as in the statement of Theorem \ref{harm::thm::coupling} is that $D_n = D \cap \tfrac{1}{n} \mathbf{Z}^2$, so all of our distances need to be scaled by a factor of $n^{-1}$.
Fix $g_1,\ldots,g_k \in C_0^\infty(D)$ and assume that $n$ is sufficiently large so that ${\rm supp }(g_1),\ldots,{\rm supp }(g_k) \subseteq D_n(n^{-\epsilon})$. On $\mathcal {H}$, for each $1 \leq i \leq k$ we have that
\begin{equation}
\label{clt::eqn::sum_by_parts}
\xi_\nabla^{n,D}(g_i) = \xi_\nabla^{n}(g_i) + \sum_{b \in D_n^*} a_u(b) \nabla g_i(b) \nabla \overline{h}^n(b) = \xi_\nabla^{n}(g_i).
\end{equation}
The second equality follows from summation by parts and the $\Delta^\beta$-harmonicity of $\widehat{h}^n$; there is no boundary term since $g_i$ vanishes near $\partial D_n(n^{-\epsilon})$.
Combining Lemma \ref{clt::lem::exp_moments}, \eqref{clt::eqn::seggs_exp_moment}, and the Cauchy-Schwarz inequality yields
\begin{align*}
& \mathbf{E} |\xi_\nabla^{n,D}(g_i) - \xi_\nabla^{n}(g_i)|
= \mathbf{E} |\xi_\nabla^{n,D}(g_i) - \xi_\nabla^n(g_i)| \mathbf{1}_{\mathcal {H}^c}\\
\leq& \big[ O(1) O(n^{-\delta}) \big]^{1/2}
\leq O(n^{-\delta/2}).
\end{align*}
Therefore $(\xi_\nabla^D(g_1),\ldots,\xi_\nabla^D(g_k))$ is a Gaussian vector with ${\rm Cov}( \xi_\nabla^D(g_i), \xi_\nabla^D(g_j)) = {\rm Cov}( (h,g_i)_\nabla^A, (h,g_j)_\nabla^A)$ where $h$ is an $A$-GFF on $\mathbf{R}^2$. Proposition \ref{gff::prop::markov} implies that $h$ restricted to $D$ admits the decomposition $h = h^0 + \widehat{h}$ where $h^0$ is a zero-boundary $A$-GFF on $D$ and $\widehat{h}$ is a $\overline{\Delta}^A$-harmonic function. Integration by parts implies that $(\widehat{h},g_i)_\nabla^A \equiv 0$ for all $i$, consequently the covariance structure of $(\xi_\nabla^D(g_1),\ldots,\xi_\nabla^D(g_k))$ is the same as $( (h^0,g_1)_\nabla^A,\ldots,(h^0,g_k)_\nabla^A)$.
We now turn to the general case that $g_1,\ldots,g_k \in C^\infty(D)$ do not necessarily have compact support in $D$. Note that we can write
\[ g_i = (g_i - \widetilde{g}_i - \widehat{g}_i) + \widetilde{g}_i + \widehat{g}\]
where $\widehat{g}_i$ is the $\overline{\Delta}^\beta$-harmonic extension of $g_i$ from $\partial D$ to $D$ and $\widetilde{g}_i \in C_0^\infty$ satisfies $\|g_i - \widehat{g}_i - \widetilde{g}_i(x) \|_{H^1(D)} \leq \delta_1$. Note that such an approximation exists since $g_i - \widehat{g}_i \in H_0^1(D)$.
Since $\widehat{g}_i$ is harmonic with smooth boundary conditions, we have $\Delta^\beta \widehat{g}_i = o(1) n^{-2}$ uniformly in $D$. Thus summing by parts twice and using that $h^n(x) = f(x) + \varphi_n(x)$ on $\partial D_n$, with $f^n$ the $\Delta^\beta$-harmonic extension of $f$ from $\partial D_n$ to $D_n$ we have
\begin{align*}
\mathbf{E} \xi_\nabla^{n,D}(\widehat{g}_i) = \sum_{b \in \partial D_n^*} a_u(b) f(x_b) \nabla \widehat{g}_i(b) + o(1) =
\sum_{b \in D_n^*} a_u(b) \nabla f^n(b) \nabla \widehat{g}_i(b) + o(1).
\end{align*}
Thus it is not difficult to see that if $F$ denotes the $\overline{\Delta}^\beta$-harmonic extension of $f$ from $\partial D$ to $D$ then
\[ \lim_{n \to \infty} \mathbf{E} \xi_\nabla^{n,D}(\widehat{g}_i) = \int_D \overline{\nabla} F A_u \overline{\nabla} \widehat{g}_i = \int_D \overline{\nabla} F A_u \overline{\nabla} g_i\]
where $A_u$ is the diagonal matrix with entries $\beta = (\beta_1,\beta_2)$.
Applying summation by parts,
\[ \xi_{\nabla}^{n,D}(\widehat{g}_i) = \sum_{b \in \partial D_n^*} a_u(b) f(x_b) \nabla \widehat{g}_i(b) - \sum_{x \in D_n} f(x) \Delta^\beta \widehat{g}_i(x).\]
Note that the first summation on the right hand side is deterministic. Consequently, combining the Brascamp-Lieb inequalities with \eqref{dgff::covariance} implies
\begin{align*}
& {\rm Var}\left( \xi_\nabla^{n,D}(\widehat{g}_i) \right)
= O(1) \sum_{x,y \in D_n} |\Delta^\beta \widehat{g}_i(x) \Delta^\beta \widehat{g}_i(y)| G_n(x,y)\\
&= o(1) \sum_{x \in D_n} n^{-2} \sum_{y \in D_n} n^{-2} G_n(x,y)
= o(1)
\end{align*}
where $G_n$ is the discrete Green's function on $D_n$. This takes care of $\widehat{g}_i$. We already know that the limiting behavior of $\xi_\nabla^{n,D}(\widetilde{g}_i)$, which leaves us to deal with $g_i-\widehat{g}_i-\widetilde{g}_i$. Invoking Lemma \ref{clt::lem::discrete_de} and the Brascamp-Lieb inequality, we have
\begin{align}
\limsup_{n \to \infty} \mathbf{E} ( \xi_\nabla^{n,D} (g_i-\widehat{g}_i-\widetilde{g}_i))^2 &\leq
C \limsup_{n \to \infty} \sum_{b \in D_n^*} |\nabla (g_i - \widehat{g}_i - \widetilde{g}_i)(b)|^2 \notag\\
&= C \| g_i - \widehat{g}_i - \widetilde{g}_i \|_{H^1(D)}^2 \leq c \delta_1^2 \label{clt::eqn::approx_bound}.
\end{align}
In the equality, we are using that $g_i - \widehat{g}_i - \widetilde{g}_i \in C^\infty(D)$. We also have
\[ \mathbf{E} |(h^0, g_i-\widehat{g}_i - \widetilde{g}_i)_\nabla^A|^2 \leq c\| g_i - \widehat{g}_i - \widetilde{g}_i \|_{H^1(D)}^2 \leq c\delta_1^2.\]
Assume that $(\xi_\nabla^{n,D}(\widetilde{g}_1),\ldots,\xi_\nabla^{n,D}(\widetilde{g}_k))$ and $((h^0,\widetilde{g}_1)_\nabla^A,\ldots,(h^0,\widetilde{g}_k)_\nabla^A)$ have been embedded into a common probability space so that $\lim_n \xi_\nabla^{n,D}(\widetilde{g}_i) = (h^0,\widetilde{g}_i)_\nabla^A$ almost surely for each $1 \leq i \leq k$. By \eqref{clt::eqn::approx_bound},
\begin{align*}
&\mathbf{E} \limsup_{n \to \infty} |\xi_\nabla^{n,D}(g_i) - (h^0,g_i)_\nabla^A - \int_D \overline{\nabla} F A_u \overline{\nabla} g_i|\\
=& \mathbf{E} \limsup_{n \to \infty} |\xi_\nabla^{n,D}(g_i-\widehat{g}_i-\widetilde{g}_i) - (h^0,g_i - \widehat{g}_i -\widetilde{g}_i)|\\ \leq& 2\delta_1.
\end{align*}
Since $\delta_1 > 0$ was arbitrary, we therefore have that $(\xi_\nabla(g_1),\ldots,\xi_\nabla(g_k))$ has the same distribution as
\[\left((h^0,g_i)_\nabla^A + \int \overline{\nabla} F A_u \overline{\nabla} g_i : 1 \leq i \leq k \right).\]
We will now explain why $A = A_u$, which will complete the proof. We will not spell out all of the details exactly in order to avoid repetition. Suppose that $U \subseteq D$ is a smooth open subset, $U_n = U \cap \tfrac{1}{n} \mathbf{Z}^2$, and let $\psi_n$ be the function which is $\Delta^\beta$ harmonic in $U_n$ and is equal to $h^n(x) - \varphi_n(x)$ in $D_n \setminus U_n$. Consider the auxiliary functional
\[ \xi_\nabla^{n,U}(g) = \sum_{b \in D^*} a_u(b) \nabla (h^n - \varphi_n - \psi_n)(b) \nabla g(b).\]
Exactly the same argument implies that $\xi_\nabla^{n,U}$ converges to a zero-boundary $A$-GFF on $U$, say $\xi_\nabla^U$. Since $\psi_n = \xi_\nabla^{n,D} - \xi_\nabla^{n,U}$, as linear functionals, we also know that $\psi_n$ has a limit, say $\psi = \xi_\nabla^D - \xi_\nabla^U$. It is not difficult to see that $\psi$ is $\overline{\Delta}^\beta$ harmonic and depends on $\xi_\nabla^D$ only through its values on $D \setminus U$. Since $\xi_\nabla^D$ is an $A$-GFF on $D$, we know that it admits the decomposition $\xi_\nabla^D = \widetilde{\xi}_\nabla^U + \widetilde{\psi}$ where $\widetilde{\psi}$ is $\overline{\Delta}^A$-harmonic and $\widetilde{\xi}_\nabla^U$ is a zero-boundary $A$-GFF on $U$ independent of $\xi_\nabla^U$. Therefore we have that
\[ \xi_\nabla^U = \widetilde{\xi}_\nabla^U + (\widetilde{\psi}-\psi).\]
This implies that $\widetilde{\psi} = \psi$ almost surely since a zero-boundary $A$-GFF plus an independent function does not have the law of a zero-boundary $A$-GFF. This finishes the proof of the theorem.
\end{proof}
|
\section{INTRODUCTION}
The primary window into the formation and disruption of star clusters comes from their mass and age distributions. Some physical processes, such as the evaporation of stars due to internal relaxation, are known to disrupt low-mass clusters earlier than high-mass clusters. Any such mass-dependent disruption process will imprint features (such as a bend) in the mass and age distributions, $\psi(M)\propto dN/dM$ and $\chi(\tau)\propto dN/d\tau$,
and will show up as correlations in the joint distribution of cluster masses and ages $g(M,\tau)$. We have recently derived new formulae for $g(M,\tau)$ for three idealized models, two in which the rate of disruption depends on the mass of a cluster, and one in which it does not:
(1)~sudden mass-dependent disruption,
(2)~gradual mass-dependent disruption, and
(3)~gradual mass-{\em in}dependent disruption.
A comparison of the predictions from these models with the observations of star clusters in the merging Antennae galaxies shows that gradual mass-independent disruption (Model~3) provides a good, first-order description of these clusters, with $g(M,\tau) \propto \psi(M) \chi(\tau) \propto M^{-2}\,\tau^{-1}$ over the studied range $\tau \lea 10^7 (M/10^4\,M_{\odot})^{1.3}$~yr (Fall, Chandar, \& Whitmore 2009; hereafter FCW09).
Results plotted in Lada \& Lada (2003) for young clusters in the solar neighborhood suggest $\psi(M)$ and $\chi(\tau)$ distributions that are similar to those for the Antennae, although they are noisier and pertain to clusters with lower masses than in the Antennae.
The main motivation for the present work is to test whether or not any of the disruption models mentioned above also describe young ($\tau \lea 10^9$~yr) star clusters in more normal galaxies than the Antennae. To accomplish this, we apply the same methodology that we used in the Antennae, to the star cluster systems in the Large and Small Magellanic Clouds (LMC and SMC). These neighboring galaxies provide an important test case because their clusters are relatively easy to identify and study, and they are representative of low-mass, late-type galaxies commonly found
throughout the nearby universe. The Magellanic Clouds are also in a more quiescent phase than the Antennae, and have had a nearly constant rate of star formation (to within a factor of two) for the last few Gyr (Harris \& Zaritsky 2004; 2009).
We use catalogs of star clusters in the Magellanic Clouds provided by Hunter et~al.\ (2003), which are the most comprehensive and photometrically uniform samples currently available, to derive the mass and age for each cluster.\footnote{Hunter et~al.\ selected clusters to be highly concentrated aggregates of stars with a density higher than that of the surrounding stellar field, regardless of whether the clusters might be classified as open, populous or globular, and without assessing if the clusters are gravitationally bound, since this is virtually impossible to determine for clusters younger than $\sim$10 internal crossing times.} The mass and age distributions of the clusters are then compared with $g(M,\tau)$ predictions from the three disruption models mentioned above. We follow the methodology developed in FCW09, which plots relatively narrow projections of $g(M,\tau)$ averaged over $\tau$ and $M$ and which lends itself to easy graphical interpretation. This simple but {\em direct} method simplifies the search for possible features that change over time, such as a bend in the mass function, as required when mass-dependent disruption affects a cluster system. Previously, an {\em indirect} approach developed by Boutloukos \& Lamers (2003) was used to infer values for the disruption timescale for star clusters in the LMC and SMC. Their method relies on mass and age distributions averaged over wide ranges of $M$ and $\tau$, and makes the {\em a priori} assumption that clusters are disrupted in a mass-dependent fashion. Here, we revisit the disruption of star clusters in the Magellanic Clouds using our more direct approach, and assess the validity of the previous results.
As mentioned above, there have been several earlier studies aimed at determining the mass and age distributions of star clusters in the LMC and SMC, and interpreting these distributions in terms of the formation and
disruption of the clusters. The recent literature on this subject is somewhat confusing and full of contradictory claims. We have examined these papers carefully, and find that despite the apparent conflicts, there is reasonably good agreement on the shape of the cluster mass and age distributions among many previous works, and also with the results presented here.
This is particularly true for the LMC. We summarize the results from these previous works and how they relate to ours in an accompanying Appendix.
The remainder of this paper is organized as follows: Section~2 provides a brief summary of the three idealized, disruption models considered here; Section~3 describes the observations and the resulting mass, age, and luminosity distributions; and Section~4 compares the \mbox{$M$--$\tau$} properties for clusters in the LMC with the model predictions. Section~5 discusses the physical processes that disrupt star clusters throughout their lives. We summarize our main conclusions in Section~6.
\section{MODELS}
In FCW09, we derived new analytical formulae for the bivariate mass-age distribution $g(M,\tau)$, defined such that $g(M, \tau) dMd\tau$ is the number of clusters with masses between $M$ and $M+dM$ and ages between $\tau$ and $\tau+ d\tau$, for three idealized models for the disruption of star clusters. We compared predictions from these models with the mass-age distribution of star clusters in the Antennae, and found that
predictions from one model nicely reproduce the data, while the other two do not. The main goal of this paper is to determine whether or not any of these models describes the star cluster systems in the LMC and SMC. In the first two models, proposed by Boutloukos \& Lamers (2003), clusters are disrupted on a timescale that depends on their masses, with clusters disrupted either suddenly (Model~1) or gradually (Model~2). In the third model, clusters are disrupted suddenly or gradually, on a timescale that is independent of their mass, as indicated by our earlier studies of clusters in
the Antennae (Zhang \& Fall 1999; Fall et~al. 2005, hereafter FCW05; Fall 2006; Whitmore, Chandar, \& Fall 2007, hereafter WCF07; FCW09). For all models, we assume that clusters form with an initial power-law mass function $\psi_0(M_0) \propto M_0^{\beta}$ at a constant rate.\footnote{Our assumption that the cluster formation rate in the LMC and SMC is nearly constant when averaged over large portions of the galaxies is supported by analysis of multi-band imaging of millions of field stars, showing that there have been only minor variations (at approximately the factor of two level) in the rate of star formation over the last few Gyr (Harris \& Zaritsky 2004; 2009). Variations at this level have a negligible impact on the results presented in this work.} Here, we briefly summarize the three disruption models and relevant formulae from FCW09.
\begin{itemize}
\item {\em Model~1. Sudden Mass-Dependent Disruption.}
This model assumes that clusters retain all of their initial mass until they are destroyed suddenly at an age $\tau_d(M_0)=\tau_{*}(M_0/M_{*})^{k}$, where $k$ is the disruption index, and $\tau_*$ is the characteristic time it takes to disrupt an $M_*=10^4\,M_{\odot}$ cluster (Boutloukos \& Lamers 2003). The relevant equation from FCW09 for $g(M,\tau)$ is Equation~(6).
It has been claimed that Model~1 provides good fits to the mass and age distributions of clusters in both the LMC (de Grijs \& Anders 2006) and the SMC (Boutloukos \& Lamers 2003; Lamers et~al.\ 2005), with the same parameters $k=0.6$ and $\tau_{*}=8\times10^9$~yr for both galaxies.
We test these claims directly in Section~4.
\item {\em Model~2. Gradual Mass-Dependent Disruption.}
In this model, we assume that the mass $M$ of a cluster evolves with time $\tau$ according to the equations
\begin{equation}
dM/d\tau = - M/\tau_d(M),
\end{equation}
\begin{equation}
\tau_d(M) = \tau_* (M/M_*)^k,
\end{equation}
where the exponent $k$ and characteristic disruption timescale $\tau_*$ are adjustable parameters, while $M_*=10^4\,M_{\odot}$ is a fiducial mass scale as before. This second model is an analog of the first, but with continuous evolution (as also noted by Boutloukos \& Lamers). The value $k = 0$ corresponds to a special case of mass-independent disruption, which describes the disruption of clusters by external gravitational shocks (Spitzer 1987). The case $k = 1$, which implies that $M(\tau)$ has a linear dependence on $\tau$, is appropriate for the standard treatment of stellar evaporation driven by two-body relaxation in clusters of constant mean
density, as regulated by the smooth tidal field of their host galaxies (e.g., Spitzer 1987; Fall \& Zhang 2001). Under some circumstances the evaporation rate is modified when stars scattered to zero or positive energies are subsequently scattered back to negative energies before they can
escape from the cluster, a process known as retarded evaporation. Baumgardt \& Makino (2003) find that a linear $M(\tau)$ relation also provides
a good approximation for retarded evaporation. While $k=1$ is therefore appropriate for both standard and retarded evaporation, the coefficient in the relationship, which is related to the timescale $\tau_*$, has different dependencies on the internal density of the clusters in the two cases. See McLaughlin \& Fall (2008; hereafter MF08) for a thorough discussion of this topic.
\item {\em Model~3. Gradual Mass-Independent Disruption.}
In this case, both the mass and age distributions are approximated by power-laws (i.e., they have no preferred scales), and are independent of one another, as given by
\begin{equation}
g(M,\tau) \propto (M/M_*)^{\beta} (\tau/\tau_*)^{\gamma}.
\end{equation}
Here, $M_*$ and $\tau_*$ are both arbitrary scale factors that merely specify the units in which masses and ages are measured (in contrast to Models~1 and 2, where $\tau_*$ has physical significance). Equation~(3) is valid if the masses of all clusters decline gradually with the same power-law dependence on age. However this interpretation is not unique. Equation~(3) is also valid if clusters are disrupted suddenly with an age-dependent but a mass-independent probability such that the number of surviving clusters
declines as a power-law in age. Intermediate cases, with different combinations of age-dependent masses and survival probabilities are also possible. Moreover, some clusters may lose only part of their mass, while others are completely destroyed. We refer to all of these situations as gradual mass-independent disruption, because they lead to a gradual decline in the number of clusters at each mass with age, at a fractional rate that is independent of mass.
\end{itemize}
One of the simplest and most informative approaches for comparing observations and predictions is to use averages of $g(M,\tau)$ over several adjacent intervals of age and mass. This approach makes it easy to see features that change over time or with mass, as imprinted by mass-dependent
disruption. These averages over age and mass respectively, are:
\begin{equation}
\bar{g}(M) \equiv \frac{1}{(\tau_2 - \tau_1)}
\int_{\tau_1}^{\tau_2} g(M, \tau) d\tau,
\end{equation}
\begin{equation}
\bar{g}(\tau) \equiv \frac{1}{(M_2 - M_1)}
\int_{M_1}^{M_2} g(M, \tau) dM.
\end{equation}
We have performed the analytical integrations for $\bar{g}(M)$ and $\bar{g}(\tau)$ for each of the three models described above, and give the formulae in Appendix~B of FCW09. This approach allows us to focus on the large-scale shapes of the distributions, rather than relying on a statistical method (e.g., $\chi^2$ minimization) that can easily be overwhelmed by relatively minor systematic errors that result from the dating procedure, such as those which arise during the red supergiant phase (see Section~3.1).
These $\bar{g}(M)$ and $\bar{g}(\tau)$ functions will be our main tool for comparing observations and predictions in Section~4.
\section{OBSERVATIONS}
\subsection{Data and Determination of Cluster Masses and Ages}
We use the catalogs and integrated photometry published by Hunter et~al.\ (2003) to estimate the masses and ages of star clusters in the Magellanic Clouds. Hunter et~al.\ carefully inspected candidate star clusters (listed in previous surveys) on ground-based \textit{UBVR} images taken at the Michigan Curtis Schmidt telescope at CTIO (Massey 2002), which provide large (but partial) coverage of the LMC ($11~\mbox{kpc}^{2}$) and the SMC ($8.3~\mbox{kpc}^{2}$). They retained objects that could be visually distinguished from the surrounding stellar field and resolved with respect to an isolated star, resulting in a total of 854 clusters in the LMC and 239 clusters in the SMC. These are the most extensive and photometrically homogeneous catalogs currently available, and thus a good dataset to use for a comprehensive study of the cluster populations in the Magellanic Clouds.
Hunter et~al.\ did not, however, quantify any biases that result from incompleteness in their sample. Therefore in Section~3.3 we perform several tests to assess whether or not biases in the sample significantly impacts our results.
Hunter et~al.\ measured the integrated light for each cluster within an aperture selected to minimize contamination from foreground/background stars, finding good agreement with the integrated colors published in other
catalogs. They did not correct for cluster light beyond the selected aperture.
We find that the $V$-band magnitudes measured by Hunter et~al. are fainter by $0.6$~mag on average, when compared with the total $V$-band magnitudes determined from King profile fits for a subset of clusters (McLaughlin \& van der Marel 2005), with an RMS scatter of $\sim0.5$~mag between the two works. This lack of aperture corrections affects the mass
estimates of the clusters, and is discussed further below.
We estimate the age $\tau$ and extinction $A_V$ for each cluster by performing a $\chi^2$ fit comparing observed magnitudes with predictions from Bruzual \& Charlot (2003) stellar population models with metallicity $Z=0.008$ for the LMC ($Z=0.004$ for the SMC), a Salpeter (1955) initial mass
function (IMF), and a Galactic-type extinction law (Fitzpatrick 1999). The best-fit values of $\tau$ and $A_V$ are those that minimize
\begin{equation}
\chi^2(\tau,A_V) = \sum_{\lambda}
W_{\lambda}~(m_{\lambda}^{\mbox{obs}}
- m_{\lambda}^{\mbox{mod}})^2,
\end{equation}
where $m_{\lambda}^{\mbox{obs}}$ and $m_{\lambda}^{\mbox{mod}}$ are the observed and model magnitudes respectively, and the sum runs over
all four bands, $\lambda=U,B,V,R$. The weight factors in the formula for
$\chi^2$ are taken to be $W_{\lambda} = [\sigma_{\lambda}^2 +
(0.05)^2]^{-1}$, where $\sigma_{\lambda}$ is the photometric uncertainty
determined by Hunter et~al.\ for each band. The mass of each cluster is estimated from the observed $V$-band luminosity (corrected for extinction but not for aperture) and the mass-to-light ratios ($M/L_V$) at the fitted $\tau$ predicted by the models, assuming a distance modulus $\Delta(m-M)=18.50$ for the LMC (Alves 2004), and $18.89$ for the SMC (Harries
et~al.\ 2003). We used the same procedure to estimate the ages and masses of clusters in the Antennae galaxies (FCW05 and FCW09), except that for the Magellanic Cloud clusters we have $R$, rather than $I$-band photometry, and no $H\alpha$ measurements.
We have assumed a Salpeter rather than the (more modern) Chabrier stellar IMF mainly to facilitate comparison with our study of clusters in the Antennae, where we made a similar assumption. Another approach would be to assume a Chabrier IMF and to correct the luminosities for the average (0.6~mag) offset found above. We have repeated our dating analysis using this second approach, and find that all the mass and age estimates remain virtually the same. The reason for this can be understood from Figure~1a, which compares predictions from the Bruzual \& Charlot (2003) models
for Salpeter and Chabrier IMFs, and shows that the adopted IMF makes little
difference in the predicted colors, and hence ages, of the clusters.
Figure~1b shows that the $M/L_V$, and hence the masses, are reduced by a near constant (age-independent) 40\% for a Chabrier IMF relative to the Salpeter IMF; this reduction in the predicted $M/L_V$ is approximately compensated for by the increased luminosity of the clusters due to the aperture correction. For the Antennae clusters, we made self-consistent aperture corrections based on our $HST$ observations (a procedure not possible for the LMC and SMC clusters), and also adopted the Salpeter IMF. The resulting offset in the mass scale between the Antennae and the Magellanic Clouds is $\approx0.6$~mag ($\approx40$\%) or $\Delta\mbox{log}~M\approx0.24$ (on average). The ages however, are on the same scale, since they are based on the same IMF, and aperture corrections do not affect the colors.
We tested our method by also estimating ages and masses from our $\chi^2$ analysis for different assumptions regarding extinction (using the Calzetti
et~al.\ 1994 obscuration curve, assuming foreground extinction only, and assuming no extinction) and different sets of stellar population models (Bruzual \& Charlot and GALEV). We discuss the impact that these assumptions have on the mass and age distributions in Section~3.2. Hunter et~al.\ (2003) also provide an independent estimate of cluster ages, which they determined by comparing their integrated \textit{UBVR} measurements with predictions from the Starburst99 stellar population models (Leitherer et~al.\ 1999) for young ($\tau\leq10^9$~yr) clusters, and with predictions from Searle, Sargent, \& Bagnuolo (1973) and Reed (1985) for older clusters ($\tau > 10^9$~yr). Hunter et~al.\ did not provide mass estimates, so we used the $M/L_V$ predicted by the Bruzual \& Charlot models for the age determined by Hunter et~al.\ for each cluster, and then calculated the mass in the same manner as in our analysis (we refer to these as the ``Hunter et~al.\ masses'' for brevity). In Section~3.2 we repeat our analysis with these independent age and mass estimates, which gives an indication of how sensitive the results are to different assumptions, dating techniques, and models.
We estimate the accuracy of our age determinations as follows. We find that
$\sim$60\% of the clusters have good fits with $\chi^2\leq1$; slightly more than 80\% of the clusters have $\chi^2\leq3$. This results in typical 1$\sigma$ (internal) uncertainties of $\sim$0.3 in $\log\tau$, corresponding to a factor of $\sim$2 in $\tau$. These internal uncertainties agree well with the external uncertainties of $\approx0.3$--0.4, corresponding to a factor of 2.0--2.5 in $\tau$, based on a comparison with ages determined from absorption-line strengths or from main sequence turnoff fitting for $\approx50$ clusters (e.g., Geisler et~al.\ 1997; Santos et~al.\ 2006; Mackey \& Gilmore 2003a,b; Kerber et~al.\ 2007; Nota et~al.\ 2006; Sirianni et~al.\ 2002). A similar uncertainty was determined for Magellanic Cloud clusters
by Elson \& Fall (1988) and by de Grijs \& Anders (2006), who also used integrated colors to estimate ages.
We now assess the uncertainties in our mass estimates. The random uncertainties in log~$\tau$ translate to $1\sigma$ uncertainties of $\approx0.3$ in log~$M$, or a factor of two in $M$. We have already mentioned that the derived masses of the clusters, but not their ages, depend on the assumed IMF in the stellar population models. We find that the mass estimates for the clusters are virtually the same if we adopt a Salpeter IMF
but make no correction for aperture, rather than adopting the Chabrier (2003) IMF, and include a 0.6~mag aperture correction for each cluster.
Similarly, adopting a shorter (longer) distance to the Magellanic Clouds would reduce (increase) the derived masses of the clusters. It is important to note that none of these systematic uncertainties impacts the {\em ratios} of cluster masses or the shape of the mass function presented in Section~3.2.
Figure~2 shows the resulting luminosity-age ($L$--$\tau$) distribution for star clusters in the LMC, and Figure~3 shows the corresponding mass-age ($M$--$\tau$) representation. The solid line in Figure~3 represents $M_V=-4.0$,
corresponding approximately to the limit\footnote{This comes from the change in the mass-to-light ratio of clusters with age, which can be
approximated by $M/L_V \propto \tau^{0.8}$ for $\tau \ga 10^7$~yr.} $\tau \lea 10^7$ $(M/10^2\,M_{\odot})^{1.3}$~yr and shows that the sample does not contain clusters over the same range of mass at all ages, because clusters fade over time. We use this $M_V=-4.0$ limit to select clusters for the analysis in Section~3.2. Figures~4 and 5 show the corresponding diagrams for clusters in the SMC. Figures~2--5 contain most of the statistical information about clusters in the Magellanic Clouds, and can be projected in different directions to construct the cluster age, mass, and luminosity distributions. When compared with the Antennae, where we restricted our sample to $M_V \lea -9$ (a conservative limit for distinguishing clusters from individual stars), the present study extends the age range by more than an order of magnitude at a fixed mass, and the mass range by more than an order of magnitude at a fixed age.
For some ages, the errors in log~$\tau$ are approximately symmetric and introduce little bias in the age distribution. This is not true, however, in the range $7.0 \lea \mbox{log}(\tau/\mbox{yr}) \lea 7.5$, where the optical
emission from massive clusters is dominated by red supergiant (RSG) stars.
The integrated colors during this phase loop back on themselves, and the fitted ages become degenerate, tending to avoid values $\approx10^7$~yr
and to prefer somewhat higher values. As a result of this type of bias, there likely are features in the unbinned age distribution that are not present in the real age distribution. The relatively empty stripes and similar features in Figures~2--5, while visually prominent, do not affect our conclusions. For comparison, the bottom panel of Figure~3 shows that cluster ages in the Hunter et~al.\ analysis also bunch up at a few specific ages (which are different from ours), with almost no clusters having derived ages $\tau \lea 3\times10^6$~yr. These types of small-scale features occur regardless of the specific dating analysis that is used, and can vary in location and prominence from one technique to another. We deal with this issue simply by binning on scales over which we believe that any apparent features are real,
$\Delta \mbox{log} \tau \sim 0.6$--0.8.
\subsection{Cluster Age, Mass, and Luminosity Distributions}
All basic results of this paper can be discerned, at least qualitatively, in the \mbox{$\log L$--$\log\tau$} and $\log M$--$\log\tau$ diagrams. When we look down the vertical axis in Figures~3 and 5, we see that the density of clusters increases steadily with decreasing mass. This indicates that clusters have an approximate power-law distribution of masses $\psi(M) \propto M^{\beta}$, which must be steeper than $M^{-1}$, with no turnover or other obvious feature. If we look along the horizontal direction, we see that the number of clusters is approximately even or increases only gradually in equal bins of $\log\tau$ above a given mass, without piling up or dropping off sharply at older ages. This means that the distribution of cluster ages can be described approximately as $\chi(\tau) \propto \tau^{\gamma}$, with $\gamma\approx-1$. We give a more quantitative treatment below.
The mass function $\psi(M)$, the integral of $g(M,\tau)$ over $\tau$, provides important information about the dynamical evolution of the star clusters. It is obtained by projecting diagonally along the stellar population tracks in Figures~2 and 4 or equivalently projecting horizontally in Figures~3 and 5, and counting clusters in different mass bins. We include clusters brighter than $M_V=-4.0$ (LMC) and $M_V=-4.5$ (SMC), although in the $10^8$--$10^9$~yr interval for the LMC we do include one bin that extends slightly below $M_V=-4.0$, since older, lower mass clusters provide strong constraints on the evolution of clusters. The specific $M$--$\tau$ intervals used to construct the mass function of clusters in the LMC are shown in the
top panel of Figure~6. In the following, we multiply the mass function by a factor $(\tau_2-\tau_1)^{-1}$ to convert it to $\bar{g}(M)$ as defined by Equation~(3). The {\em amplitudes} of the $\bar{g}(M)$ distributions, in
addition to their {\em shapes}, provide critical information on the evolution of the cluster system.
In Figures~7 and 8, we show $\bar{g}(M)$ for star clusters in the LMC (SMC),
using a typical bin width of $0.5$ ($0.4$) in $\log M$ as a compromise between observational uncertainties in the masses and adequately sampling the mass function. We find that over the plotted mass-age domain, {\em the mass function declines monotonically with increasing mass, with no obvious breaks or bends}, and can be described by a power law\footnote{The mass functions shown in Figures~7 and 8 can be described by power laws up to $M_\mathrm{max}\approx3\times10^5\,M_{\odot}$ in the LMC and to $M_\mathrm{max}\approx10^5\,M_{\odot}$ in the SMC. The expected number of clusters more massive than $M_\mathrm{max}$ is $\approx3$--4 in the LMC (Chandar et~al.\ 2010) and $\approx1$--2 in the SMC, based on extrapolations of these power laws. Thus, we cannot make any definite statements about whether there are or are not physical cutoffs in the mass functions above $M_\mathrm{max}$ in the Magellanic Clouds.}, $\bar{g}(M) \propto M^{\beta}$. The distributions have approximately the same shape for all plotted ages, and are therefore roughly independent of age up to $\tau \approx10^9$~yr. We find $\beta \approx-1.8$ (our masses) and $\beta\approx-1.7$ (Hunter et~al.\ masses) for clusters in the LMC, and $\beta\approx-1.8$ (our masses) and $\beta\approx-2.0$ (Hunter et~al.\ masses) for the SMC. These values of $\beta$ are based on least square fits of the form: $\log\bar{g}(M)=\beta\log M+{\rm const}$. Any incompleteness in the cluster sample at lower masses would flatten the observed $\bar{g}(M)$ distribution relative to its true shape. We find, however, that $\beta$ hardly changes if we exclude the one or two lowest mass bins from the fit, suggesting that any incompleteness at lower masses does not strongly affect our results. An assessment of how incompleteness impacts the shape of the mass function is discussed more thoroughly in Section~3.3. Based on our extensive experiments with different stellar population models, assumptions about extinction, and binning (both variable and constant), as well as our analysis of the independent age determinations from Hunter et~al., we believe that realistic uncertainties in the exponent $\beta$ are of the order $\Delta\beta \approx 0.2$, and that therefore $\bar{g}(M) \propto M^{\beta}$ with $\beta=-1.8\pm0.2$ describes clusters with ages $\tau \lea 10^9$~yr in both the LMC and SMC.
The age distribution $\chi(\tau)$, the integral of $g(M,\tau)$ over $M$,
provides important information about the disruption of the clusters. We multiply the age distribution by a factor $(M_2-M_1)^{-1}$ to convert it to $\bar{g}(\tau)$ as defined by Equation~(4). We show $\bar{g}(\tau)$ for star clusters in the LMC (Figure~9) and SMC (Figure~10) based on our age and mass estimates, and also based on the ages determined by Hunter et~al.\ These $\bar{g}(\tau)$ distributions were constructed for {\em mass-limited} samples by counting the number of clusters ($N$) in different age bins, shown as the dashed lines in the bottom panel of Figure~6 for the LMC, and restricted to cover the mass-age domain above the fading lines. The $1\sigma$ uncertainties were taken to be $\pm\sqrt{N}$. In Section~3.1 we estimated $1\sigma$ uncertainties of $\sim0.3$--0.4 in $\log\tau$, and here we use bins twice this wide. Most of {\em the distributions decline starting at very young ages, with no obvious bends or breaks}. Each distribution can be described approximately by a single power law of the form $\bar{g}(\tau) \propto \tau^{\gamma}$. The $\bar{g}(\tau)$ distributions at different masses have the same slope within the uncertainties, and are therefore at least approximately independent of mass. We find typical values of
$\gamma\approx-0.7$ (our ages) and $\gamma\approx-1.0$ (Hunter et~al.\ ages) for clusters in the LMC, based on simple, linear least-square fits.
These results from two independent dating analyses, plus experiments with different binning and stellar evolution models, suggest $\gamma=-0.8\pm0.2$ for clusters in the LMC. In the SMC, we find $\gamma\approx-0.7$ (our ages) and $\gamma \approx-0.9$ (Hunter et~al.\ ages), where the
youngest bin has been excluded from the fits with the Hunter et~al.\ ages, since these flatten significantly relative to the older data points. These results for clusters in the SMC are similar to those we found previously (Chandar
et~al.\ 2006) based on the independent cluster sample and age determinations by Rafelski \& Zaritsky (2005).
We have performed extensive experiments with our dating procedure to assess the reliability of these results, as described in Section~3.1. Our experiments suggest that the largest uncertainties in the $\bar{g}(\tau)$ distributions are for the youngest ($\tau \lea 10^7$~yr) clusters, which are most sensitive to the specific assumptions made for extinction when dating the clusters. In general, observations of young clusters in the Magellanic Clouds (e.g., Hunter et~al.\ 1997), and in other galaxies (e.g., M51, Bastian et~al.\ 2005; NGC~5253 and NGC~3077, Harris et~al.\ 2004; Antennae, Whitmore et~al.\ 1999) are affected by extinction in both the Milky Way and in the host galaxy, and it is standard practice to correct for both. However, if we use more extreme assumptions, making {\em no} correction for extinction, or correcting only for extinction in the Milky Way, the youngest data point in the $\chi(\tau)$ distributions move to lower values in both the LMC and SMC, since these assumptions force reddened, young clusters to be assigned ages that are older than their true ages. For ages $\tau \gea10^7$~yr, the $\bar{g}(\tau)$ distributions in the LMC are fairly stable regardless of the detailed assumptions made concerning extinction. This is not true, however, in the SMC, where different assumptions about extinction
can significantly flatten $\bar{g}(\tau)$ relative to that shown in Figures~9 and 10, with clusters piling up at older ages when no correction for extinction is made. To summarize, we find that the shape of $\bar{g}(\tau)$ is robust for $\tau \gea10^7$~yr in the LMC regardless of the specific assumptions we make, while the results for the SMC are more sensitive to details of our analysis, particularly whether or not we correct for extinction.
However, the $\bar{g}(\tau)$ distributions in both galaxies decline progressively starting at very young ages when standard assumptions are used.
The shape of the luminosity function provides information on the relationship between cluster ages and masses. In the Antennae, we demonstrated that the cluster luminosity and mass functions have the same power-law form (FCW09), even though this is not generally expected for a population of young clusters with a large range in age and hence $M/L$.
Fall (2006) showed analytically that if the cluster age distribution is independent of mass, and the cluster mass function is a power law and independent of age, then the cluster luminosity and mass functions must
have the same power-law exponents. These conditions are automatically satisfied by $g(M,\tau)$ for Model~3 (but not Models~1 and 2), which provides a good description in the LMC, and possibly in the SMC. In Figure~11, we show that the extinction corrected luminosity functions for clusters in the LMC and SMC can be approximated by a single power law, $\phi(L) \propto L^{\alpha}$, with $\alpha=-1.8\pm0.2$. These have shapes identical to the cluster mass functions shown in Figures~7 and 8, within the uncertainties, as expected from the independence of $M$ and $\tau$, i.e., $g(M,\tau) \propto \psi(M) \chi(\tau)$.
In the Appendix, we show that results from previous works are generally in
good agreement with those presented here for the shape of the mass and age distributions of star clusters in the LMC (although this point is obscured
in the published descriptions of these results). Results from several works in the SMC are also consistent with ours, although there is more disparity for this galaxy. We have determined the shape of the age distribution for clusters brighter than a given luminosity rather than from mass-limited samples as done here, using the Monte Carlo simulations described in WCF07. Age distributions constructed from luminosity-limited
samples are used below to assess whether or not there are strong biases in either mass or age in the Hunter et~al.\ samples. In the Appendix, we show that the $\bar{g}(\tau)$ distributions presented here for mass-limited samples
are consistent with the age distribution expected from luminosity-limited samples, giving additional confidence that our $\bar{g}(\tau)$ distributions
reflect the actual age distribution of star clusters in the Magellanic Clouds.
We discuss the physical implications of these results for the formation and disruption of clusters in the Magellanic Clouds in Section~5.1.
\subsection{Tests for Biases Due to Incompleteness}
The Hunter et~al.\ samples of Magellanic Cloud clusters are known to be incomplete in some fashion, since they do not include clusters fainter than
$m_V\approx15$ and do not cover the Magellanic Clouds in their entirety.
What is critical for our work is not that the samples be complete, but that any incompleteness not be a strong function of either mass or age, and hence not significantly affect the shape of the $\bar{g}(M)$ and $\bar{g}(\tau)$ (mass and age) distributions over the plotted ranges. In other words, we do not require a complete sample, as long as it is unbiased (i.e., representative) with respect to mass and age. Here, we perform three (necessarily) indirect tests on various projections of the $g(M,\tau)$ distribution, which suggest that biases in the sample do not significantly impact the shape of the mass and age distributions presented in Section~3.2.
Several previous studies, described in the Appendix, have used the same dataset to study the formation and disruption of clusters in the Magellanic Clouds, and therefore would be affected by any incompleteness in the same way as in our study.
First, we check whether or not the shape of the mass and age distributions change if we exclude the data points most likely to be affected by incompleteness, namely the lowest mass bin in each $\bar{g}(M)$ distribution, and the oldest age bin in each $\bar{g}(\tau)$ distribution.
Incompleteness would cause $\bar{g}(M)$ and $\bar{g}(\tau)$ in these bins to be lower than their true values. However, we find that they do not deviate significantly from power-law extrapolations to the rest of the data points in the $\bar{g}(M)$ and $\bar{g}(\tau)$ distributions, suggesting that any incompleteness in these bins does not have a strong impact on the shapes of the mass and age distributions.
Next, we check that the $M_V$ limits used here ($M_V=-4.0$ for the LMC and $M_V=-4.5$ for the SMC) are not so faint that we are missing significant numbers of clusters. If the shape of the age distribution constructed from clusters brighter than an $M_V$ limit, i.e., a luminosity-limited sample, changes (flattens) when brighter limits are used, then our $M_V$ limits are too faint. We find that the age distribution barely
changes shape when constructed for clusters at brighter limits of $M_V$ (e.g., $-4.5$, $-5.0$, etc.\ for the LMC), indicating that the limits used in Section~3.2 are not too faint. This result also suggests that the Hunter et~al.\ samples are not limited in a bluer band such as $U$~band rather than
$V$, which could potentially lead to an artificially steep age distribution since clusters fade faster in $U$ than in $V$, with age. In Appendix~A we show that distributions constructed from ages estimated in previous works, but based on different samples of Magellanic Cloud clusters and different limits for $M_V$, are similar to those determined in this work.
Finally, we compare the age distribution constructed in two different ways, by counting clusters above a mass limit and above a luminosity limit. The latter should decline more steeply than the former because some older clusters have faded below the $M_V$ limit. More quantitatively, Monte Carlo simulations show that both distributions can be modeled by power laws, and
that the luminosity-limited distribution has an index that is lower by
$\approx0.4$--0.6. If incompleteness affects the mass-limited but not the luminosity-limited sample, the index for the mass-limited age distribution will be artificially low, and closer to the value determined for the luminosity-limited distribution than the predictions. We find that the age distribution constructed from LMC (SMC) clusters brighter than $M_V=-4.0$ ($M_V=-4.5$) has an index that is always lower by $\approx0.4$--0.5 than the age distribution constructed from the mass-limited samples shown in Figures~9 and 10. Hence the observed and predicted differences between the two distributions are very similar. Based on the results from the three tests
described above, we conclude that the Hunter et~al.\ samples do not have strong biases over the plotted ranges in mass and age,\footnote{The one exception is clusters in the SMC with ages $\tau \lea 10^7$~yr, which likely are under-represented in the Hunter et~al.\ sample; see the Appendix.} and therefore are sufficiently representative of the cluster populations in the Magellanic Clouds for our purposes, i.e., sample incompleteness does not significantly affect the shapes of the $\bar{g}(M)$ and $\bar{g}(\tau)$ distributions presented in Section~3.2. This should be checked in future studies that include artificial cluster experiments during the cluster selection process.
\section{COMPARISON WITH MODELS}
Here, we compare predictions from the three disruption models described in Section~2 with the observed $\bar{g}(M)$ and $\bar{g}(\tau)$ distributions for star clusters in the Magellanic Clouds presented in Section~3. Overall, the $\bar{g}(\tau)$ and $\bar{g}(M)$ distributions for star clusters in the
SMC appear to be similar to those in the LMC, but with poorer statistics.
We therefore only show model comparisons with the LMC, but have checked that all of our conclusions also apply to the SMC, although with larger uncertainties.
\subsection{Comparison with Model~1: Sudden Mass-Dependent Disruption}
We first consider Model~1, which assumes that clusters are disrupted
suddenly with a power-law dependence on the initial mass, $\tau_d(M_0)=\tau_*(M_0/M_*)^k$, with $M_*=10^4\,M_{\odot}$. We adopt the specific values $k=0.6$ and $\tau_*=8\times10^9$~yr, which have been claimed to give good fits to clusters in the LMC (de Grijs \& Anders 2006) and in the SMC (Boutloukos \& Lamers 2003; Lamers et~al.\ 2005), based on the
indirect methodology developed by Boutloukos \& Lamers. A comparison between the predicted and observed $\bar{g}(M)$ distributions (Figure~12a)
shows that while the {\em shapes} match over the observed range, the {\em amplitudes} do not, because they are predicted to be constant with age but
are observed to decrease. The situation is even worse for the $\bar{g}(\tau)$ distributions shown in Figure~12b, because the predicted and observed shapes are quite different, with $\bar{g}(\tau)$ predicted to be flat initially,
with a feature determined by the characteristic disruption time, whereas the observations decline as a power law with no obvious feature. {\em We conclude that predictions for the specific combination $k=0.6$ and
$\tau_*=8\times10^9$~yr for Model~1, which has been claimed in several works to give a good description of cluster disruption in the Magellanic Clouds, does not actually match the observed mass and age distributions in either galaxy.}
Why did the Boutloukos \& Lamers approach fail for clusters in the LMC and SMC? The disruption predicted by the combination $k=0.6$, $\tau_*=8\times10^9$~yr is shown as the dashed diagonal line in Figures~3 and 5. According to Model~1, all clusters above this line survive and should be observed, while the region below the dashed line should be empty because the clusters have been destroyed, leading to features in $g(M,\tau)$, $\bar{g}(M)$, and $\bar{g}(\tau)$. However the Hunter et~al.\ data do not reach the masses and ages covered by the dashed line, and therefore these data cannot logically return the claimed values of $\tau_*$ and $k$. Indeed, de Grijs \& Anders (2006) found no evidence for mass-dependent cluster disruption in the LMC when they plotted the mass function directly (i.e., they
found no bends or breaks). Yet when they applied the method developed by Boutloukos \& Lamers (2003) to the same data, they derived specific values for $k$ and $\tau_*$ that imply access to a mass-age domain for the clusters well beyond the domain that is actually available. The Boutloukos \& Lamers technique is an {\em indirect} one, and assumes that any bend observed in the mass and age distribution is due to mass-dependent disruption. However, bends can appear for other reasons, for example due to artifacts related to systematic errors or to statistical noise, or most problematically when the mass function is constructed from a sample that is limited in luminosity but includes clusters with a wide range of age (as we have shown in FCW09). Thus, the Boutloukos \& Lamers procedure may be more sensitive to selection boundaries in the sample than to any physical relations among the clusters themselves.
\subsection{Comparison with Model~2: Gradual Mass-Dependent Disruption}
Model~2 assumes that clusters are disrupted more gradually than Model~1,
but still on a timescale that has a power-law dependence on the initial cluster mass. Figure~13 shows that $\bar{g}(M)$ from Model~2 with $\tau_*=8\times10^9$~yr and $k=0.62$ is very similar to $\bar{g}(M)$ from Model~1, and does not match the observations. The predicted $\bar{g}(M)$ distributions for $\tau_* < \mbox{few}\times10^9$~yr with $k=0.62$
flatten earlier (i.e., at a higher mass) than in Figure~12a, causing mismatches with the observations in both shape and amplitude. Higher values of $k$ lead to faster evolution (stronger curvature) in the predicted $\bar{g}(M)$ distributions, while lower values of $k$ lead to slower evolution (weaker curvature). Because the observed mass function for star clusters in the LMC has the same shape but different amplitude for different ages, and the observed age distribution has the same shape but different amplitudes for different masses, no combination of $\tau_*$ and $k>0$ can reproduce the observations. These statements are also true for the mass and age distributions of star clusters in the SMC, as seen from Figures~8 and 11.
A critical point here is that mass-dependent disruption, whether of the power-law form adopted in Models~1 and 2 or any other form, always imprints features in the $\bar{g}(M)$ distribution related to the characteristic disruption timescale. Unless such a feature is observed, there is no evidence that mass-dependent disruption affects a cluster system. {\em We conclude that the disruption of clusters in the Magellanic Clouds has little or no dependence on mass, at least over the $M$--$\tau$ domain studied here.} This conclusion is not affected by the formation history of the clusters, and holds whether clusters form at a constant or variable rate, as long as they form with the same initial mass function (see also FCW09, especially their footnote~7, for further discussion of this point).
Model~2, with $k=0$, is an example of mass-{\em in}dependent disruption.
This preserves the power-law shape of $\bar{g}(M)$, but not $\bar{g}(\tau)$.
Predictions for $\bar{g}(M)$ distributions in the case $k=0$ are shown for
two values of $\tau_*$: $\tau_*=1\times10^8$~yr (Figure~14a) and $\tau_*=2\times10^7$~yr (Figure~14b). While the predicted shape for $\bar{g}(M)$ with $k=0$ matches the observations, these figures show that no single
value of $\tau_*$ gives the correct amplitudes for clusters at different ages in the LMC, with the predictions dropping off faster at older ages than the observations. The reason for this mismatch in amplitudes can be understood from Figure~15, which shows the $\bar{g}(\tau)$ predictions for $k=0$ and several values of $\tau_*$, including the two shown in Figures~14a and 14b.
The predicted distributions start off flat, but change shape at a characteristic age, and then drop exponentially (i.e., faster than any power law). The observed $\bar{g}(\tau)$ distributions on the other hand, are essentially featureless power laws. Therefore, no combination of $k$ and $\tau_*$ can reproduce the observations, and Model~2, like Model~1, must be rejected for clusters in the Magellanic Clouds.
\subsection{Comparison with Model~3: Gradual Mass-Independent Disruption}
We now compare the observations with predictions from Model~3, which posits that the population of clusters is disrupted gradually at a rate that is independent of their mass (although as noted in Section~2, individual clusters may be disrupted gradually or rapidly). Model~3 has $\bar{g}(M) \propto M^{\beta}$ and $\bar{g}(\tau) \propto \tau^{\gamma}$, a direct consequence of the bivariate mass-age distribution $g(M,\tau) \propto M^{\beta} \tau^{\gamma}$. Recall that this model gives a good, first-order description of young clusters in the Antennae (FCW09). In Figure~16a, we compare the observed and predicted $\bar{g}(M)$ distributions in the LMC,
where we have assumed $\gamma=-0.8$ based on the results from
Section~3.2. There is a good match in both shape and in amplitude. The orthogonal projection $\bar{g}(\tau)$ in Figure~16b also shows excellent agreement between predictions and observations. {\em Model~3, with
$\beta=-1.8\pm0.2$ and $\gamma=-0.8\pm0.2$, provides a good, first-order description of the observed mass--age distribution for star clusters in the LMC. } We discuss the physical implications of this result in Section~5.1.
As we have just shown, the joint distribution of masses and ages $g(M,\tau)$ for clusters in the LMC is similar to that found for the Antennae. A nice feature of the present work is that the distributions in the Magellanic Clouds
extend to clusters with lower masses and to older ages than can be currently distinguished from stars in the Antennae. We show this graphically in Figures~17a and 17b, where we plot $\bar{g}(M)$ and $\bar{g}(\tau)$ distributions for the Antennae, LMC, and SMC, showing that these are all
essentially parallel, and hence that Model~3 also describes---at least approximately---the clusters in the SMC. While studies of young clusters in the solar neighborhood are not as extensive, the results compiled by Lada \& Lada (2003) support mass and age distributions of the form $\psi(M) \propto M^{-2}$ and $\chi(\tau) \propto \tau^{-1}$, for clusters that have masses lower by an order of magnitude than those studied here.\footnote{Figure~2 of
Lada \& Lada (2003) shows $MdN/d\log M\approx$\,const for embedded (i.e., very young) clusters with $50\,M_{\odot} \lea M \lea 1000\,M_{\odot}$. Figure~3 of Lada \& Lada (2003) shows $dN/d\log\tau \approx$\,const for embedded and non-embedded clusters with $10^6$yr$\lea \tau \lea 10^8$~yr. These results are equivalent to $\psi(M) \propto M^{-2}$ and $\chi(\tau) \propto \tau^{-1}$.} Taken together, {\em these results for clusters in the Antennae, solar neighborhood, LMC, and SMC support the view that there are similar patterns in the $g(M,\tau)$ distributions of young star cluster systems in different, well-studied galaxies.}
\section{INTERPRETATION}
\subsection{Formation and Early Evolution of Star Clusters}
We now seek to understand the results for the LMC, the SMC, and the Antennae in terms of the physical processes that operated during the formation and early evolution of the clusters. This section follows the more extensive discussion in FCW09. The LMC is particularly interesting because its populations of molecular clouds and star clusters have been studied extensively. The mass function of both kinds of objects are approximate power laws, with $\beta\approx-1.7$ for clouds (Rosolowsky 2005; Fukui et~al.\ 2008), and $\beta\approx-1.8$ for clusters (this paper). Thus, the complex processes of converting giant molecular clouds into young star clusters apparently happens in a manner that preserves the shape of the mass function. This suggests that the average efficiency of star formation in
protoclusters---the ratio of the final stellar mass to the initial interstellar mass---is approximately independent of the mass of the protoclusters.
Following their formation, several physical processes are responsible for disrupting star clusters. The dominant processes, which likely operate in the following, approximate sequence and timescales are:
(1)~removal of ISM by stellar feedback, $\tau \la 10^7$~yr;
(2)~continued stellar mass loss, $10^7~ {\rm yr} \la \tau \la 10^8~ {\rm yr}$;
(3)~tidal disturbances by passing molecular clouds, $\tau \ga 10^8$~yr; and
(4)~escape of stars driven by internal two-body relaxation (hereafter referred to as ``evaporation''), for clusters with $M \lea M_p$, as given by Equation~(7) below. We discuss the first three processes here, and return to the last process in the next subsection.
One of the earliest processes responsible for the disruption of star clusters is
the removal of left-over ISM by the feedback activity of massive stars (e.g., photoionization, radiation pressure, stellar winds, and supernovae). An important consequence of this rapid mass loss is that many protoclusters cannot remain gravitationally bound (e.g., Hills 1980). Afterwards, clusters continue to lose mass due to stellar evolution, and will be depleted by $\sim$40\% over $\sim\mbox{few}\times10^8$~yr, with most of this mass lost in the first $\mbox{few}\times10^7$~yr. This continued mass loss can unbind fragile clusters in a tidal field. Both of these disruption processes are internally driven, and will occur regardless of the specific properties of the host galaxy. Therefore, we would expect these early disruption processes to affect young clusters in different galaxies in a similar way. This expectation is supported by observations of clusters in the solar neighborhood, the merging Antennae galaxies, and the LMC and SMC, which all have a similar shape for $\bar{g}(\tau)$ (as shown in Figure~17 and discussed in Section~4).
It has been suggested, on theoretical grounds, that the removal of the ISM strongly affects the shape of the cluster mass function by disproportionately disrupting lower mass clusters. Some models predict that an initial power-law mass function with $\beta=-2$ will evolve quickly after removal of the ISM, on timescales $\tau \lea 10^7$~yr, resulting in a flatter mass function with $\beta\approx-1$ (see Figure~4 in Baumgardt et~al.\
2008),\footnote{Baumgardt et~al.\ (2008) state that gas expulsion will introduce a peak in the mass function near $M_p\approx10^5\,M_{\odot}$. However their Figure~4 shows a different result. This figure suggests that while the predicted mass function is significantly flatter after gas expulsion, a peak only occurs on much longer timescales, after other disruption processes, particularly relaxation-driven stellar evaporation, have had a chance
to further erode the cluster system.} or one that has a peak near $10^5\,M_{\odot}$ (e.g., Kroupa \& Boily 2002; Parmentier et~al.\ 2008). However, these predictions are ruled out by observations. We see from Figures~7 and 8 that the mass functions of clusters in the Magellanic Clouds with ages $\tau < 10^7$~yr and $\tau > 10^7$~yr have essentially the same shape,
indicating that early disruption processes do not depend strongly on the masses of the clusters. This result is explained nicely by models in which the stellar feedback is momentum-driven rather than energy-driven (Fall et~al.\ 2010).
Clusters are also subject to tidal disturbances by passing molecular clouds (Spitzer 1958; Binney \& Tremaine 2008). The timescale for this process depends on the number and density of molecular clouds in a galaxy. For example, Binney \& Tremaine (2008) estimate a disruption time of $\tau_d\approx3\times10^8$~yr for clusters in the solar neighborhood, due to interactions with molecular clouds. If conditions in the Magellanic Clouds are similar, this estimate of $\tau_d$ might also apply there. We argued in FCW09 that interactions with molecular clouds would have little or no effect
on the shape of the mass function of clusters. In contrast, Gieles et~al.\ (2006) claim that such interactions would preferentially disrupt low-mass clusters. We see no evidence for this effect, either in the Antennae or in the Magellanic Clouds over the ranges of masses and ages we have examined.
We discuss the physical reasons for and implications of this result in a separate paper (S.~M.\ Fall \& R.~Chandar, in prep).
It is likely that the rate of disruption is somewhat different for each of
the processes discussed above, and that with perfect data one might observe
features marking the transition between them. However, it is also likely that the combination of processes, plus the errors in the masses and ages caused by imperfect observations, have washed out such features, leaving a relatively smooth age distribution. For these reasons, our power-law model,
$\bar{g}(\tau) \propto \tau^{\gamma}$ is likely to be a simple approximation to a complex situation including several different physical processes which predominate at different times, rather than an exact description of a single process working in isolation.
For the Antennae, we argued that $\bar{g}(\tau)$ {\em primarily} reflects
the disruption rather than the formation of clusters, because it has the same sharp peak at $\tau \lea 10^7$~yr in several large regions inside the Antennae galaxies separated by distances of order 10~kpc. There are no hydrodynamical processes that could synchronize a burst of cluster formation this precisely over such large separations (FCW05; WCF07). In the present work, we found that star clusters in the LMC and SMC, two galaxies which are currently in a more quiescent phase than the Antennae, nevertheless have similar age distributions. A recent analysis of the star clusters in the (closest) collisional ring galaxy NGC~922, also shows a similar form for the age distribution (Pellerin et~al.\ 2009). It is far more likely that the clusters in all of these very different galaxies have similar disruption histories, than it is that they have similar formation histories and that we also happen to be observing them all at the same special time when the rate of formation is just now peaking.
An alternative explanation, that the declining shape of $\bar{g}(\tau)$ in the LMC over the last few Gyr is due to a large increase (by a factor $\approx100$) in the rate at which clusters formed from the past to the present, was suggested by Parmentier \& de Grijs (2008). In addition to the problem discussed above, another powerful argument against this suggestion is that the star formation history of the SMC and LMC determined from millions of individual stars is nearly constant over this timescale, with variations of only a factor of two or less (Harris \& Zaritsky 2004, 2009). Nevertheless, we expect variations in the formation rate of clusters to have some effect on their age distribution, as reflected in minor differences in the exponent $\gamma$ among different galaxies, likely at the level $\Delta\gamma \approx 0.2$. This is consistent with our finding $\gamma = -1.0 \pm 0.2$ for the Antennae and $\gamma = -0.8 \pm 0.2$ for the LMC and SMC.
\subsection{Late Evolution of Star Clusters}
The most important long-term disruptive process for star clusters is the
gradual escape of stars driven by internal two-body relaxation (e.g., Fall \& Zhang 2001; Prieto \& Gnedin 2008). This process depletes the mass of each cluster approximately linearly with time, $M(\tau) \approx M_0 - \mu_\mathrm{ev}\tau$, equivalent to $k=1$ in Model~2, at a rate $\mu_\mathrm{ev}$ that depends primarily on the mean internal density of the cluster. The evaporation rate $\mu_\mathrm{ev}$ has a slightly different
dependence on density for tidally limited clusters in the cases of standard and retarded evaporation, with $\mu_\mathrm{ev} \propto \rho_t^{1/2}$
(with $\rho_t=3M/4\pi r_t^3$) for standard evaporation, and $\mu_\mathrm{ev} \propto \Sigma_t^{3/4}$ (with $\Sigma_t = M/\pi r_t^2$) for retarded evaporation. See McLaughlin \& Fall (2008) for a detailed discussion of these
dependencies. Elson et~al.\ (1987) found that the density profiles of young clusters in the LMC (with $\tau \sim 10^8$~yr) extend out to and even beyond their tidal radii (sometimes also referred to as Jacobi radii) as determined from a detailed study of the tidal field of the LMC (see also Lupton
et~al.\ 1989). Thus, these clusters already fill their Roche lobes, and satisfy the main assumption underlying the simple evaporation model of McLaughlin \& Fall.
For a population of clusters with a single age, which is effectively true for ancient globular clusters (GC), a peak or turnover in the mass function at $M_p$ results from the higher fractional mass loss rates ($d\ln M/d\tau$) for low-mass clusters than for high-mass clusters. This peak should increase
with the internal density of the clusters, since stars evaporate more quickly from denser clusters, leading to their earlier disruption relative to lower density clusters. MF08 found this exact trend of $M_p$ with $\rho_h$ in the Milky Way globular cluster system, and presented a simple evaporation model with $\mu_\mathrm{ev} \propto \rho_h^{1/2}$, where $\rho_h$ is the internal half-mass density (a good approximation to both $\mu_\mathrm{ev} \propto \rho_t^{1/2}$ and $\mu_\mathrm{ev} \propto \Sigma_t^{3/4}$ in practice). Putting in the coefficients, the peak mass\footnote{We note that the mass function strictly only shows a peak when it is plotted as $dN/d\log M$. In this work, we have plotted $dN/dM$ throughout. In this form, the ``peak'' of the mass function $M_p$ corresponds to the point where the distribution flattens, or has a ``knee.''} is given by
\begin{equation}
M_p = 3.1 \times (\tau / 13\times10^9~\mathrm{yr}) (\rho_h / 10^3)^{1/2} 10^5\,M_{\odot} .
\end{equation}
Assuming the median half-mass density for GCs in the Milky Way of $\hat{\rho_h}=246\,M_{\odot}~\mbox{pc}^{-3}$ (MF08), at an age $\tau=13$~Gyr, Equation~(7) reproduces the observed peak mass $M_p=1.6\times10^{5}\,M_{\odot}$ for the Galactic GC system. The same equation, with a similar coefficient, also explains the observations of GCs in the Sombrero galaxy (Chandar, Fall, \& McLaughlin 2007).
Here, we apply the MF08 model to clusters in the LMC for the first time. The LMC has 15 known ancient GCs, which reside in a rotating disk with their younger counterparts (e.g., Freeman et~al.\ 1983). We estimate their masses from the total $V$-band luminosities determined by McLaughlin \& van der Marel (2005) from King (1966) model fits to their radial profiles,
and assume a typical $M/L_V=1.5$, the median value found for the Milky Way GC system. This gives a median mass of $\hat{M} \approx 2\times10^{5}\,M_{\odot}$ for the GC system in the LMC, very similar to that found for the GC systems in other galaxies such as the Milky Way and the Sombrero.
We compute the internal half-mass density of clusters from $\rho_h = (M/2)/(4\pi r_h^3/3)$, where $r_h$, the 3D half-mass radius is determined from the 2D half-mass radius $r_\mathrm{hp}$ and the standard 3D--2D conversion $r_h = (4/3) r_\mathrm{hp}$ (Spitzer 1987). For the sample of 15 globular clusters in the LMC, we find a median half-mass density $\hat{\rho_h}=295\,M_{\odot}~\mbox{pc}^{-3}$, similar to the median density of
$246\,M_{\odot}~\mbox{pc}^{-3}$ found by MF08 for the Galactic globular cluster system. Globular clusters in the LMC have ages indistinguishable from their Galactic counterparts (e.g., Brocato et~al.\ 1996), and for an assumed age $\tau=13$~Gyr and the observed value of $\hat{\rho_h}$, Equation~(7) gives $M_p \approx 1.7\times10^{5}\,M_{\odot}$, very similar to the observed value for $\hat{M}$.
Next, we consider the $10^8$--$10^9$~yr-old LMC clusters, for which we plotted $\bar{g}(M)$ in Figure~7. We find a median internal density of
$\rho_h\approx20\,M_{\odot}~\mbox{pc}^{-3}$ for the nine clusters in this age interval that have profile fits from McLaughlin \& van der Marel
(2005).\footnote{While Hunter et~al.\ 2003 provide FWHM measurements for the clusters in their sample, there is a poor correlation between their ground-based size estimates and those measured by McLaughlin \& van der Marel 2005 from high-resolution \emph{HST} images. Therefore we decided not to use the Hunter et~al.\ size estimates in this work.} This is approximately 14 times lower than that found for the ancient LMC clusters.
This difference in density is primarily due to the larger size of the $10^8$--$10^9$~yr-old clusters compared with that for the old GCs ($\hat{r_h}=8.2$~pc vs.\ $\hat{r_h}=3.4$~pc). Assuming that the median density of $\rho_h\approx20\,M_{\odot}~\mbox{pc}^{-3}$ for these nine clusters is representative of the entire $10^8$--$10^9$~yr old population, Equation~(6) predicts $M_p\approx10^3\,M_{\odot}$ for $\tau=3\times10^8$~yr, the median age of clusters in this bin. A flattening of $\bar{g}(M)$ near this mass would be near the edge and possibly below the plotted range in Figure~7. Therefore, the fact that we do not detect any flattening in the mass function for $10^8$--$10^9$~yr-old clusters in the LMC is consistent with the MF08 model predictions.
\section{SUMMARY AND CONCLUSIONS}
We have compared the mass-age distribution $g(M,\tau)$ of star clusters in the Magellanic Clouds with three idealized disruption models: (1)~sudden mass-dependent disruption, (2)~gradual mass-dependent disruption, and
(3)~gradual mass-{\em in}dependent disruption, to determine whether any of them describe the properties of star clusters in these relatively quiescent
galaxies. We compared the integrated \textit{UBVR} photometric measurements for 854 LMC and 239 SMC clusters cataloged by Hunter et~al.\ (2003), with predictions from population synthesis models in order to estimate $M$ and $\tau$ for each cluster. We constructed $\bar{g}(M)$ and $\bar{g}(\tau)$ (mass and age) distributions, and tested them for sensitivities with respect to stellar population models, dating techniques, and extinction corrections. These distributions in the LMC are fairly robust, particularly for $\tau \gea10^7$~yr, but in the SMC can flatten significantly given the most extreme (although unrealistic) assumptions. We find that the shape of the cluster mass and age distributions are in reasonably good agreement with the results from most previous studies in the LMC, and with several studies in the SMC (as explained in detail in the Appendix). While there are remaining uncertainties, such as not knowing the precise completeness of the cluster samples or the aperture corrections for individual objects, which affect the results presented here as well as those from previous works, we believe that these will not significantly alter our
conclusions. Our main findings are:
(1) Model~3, with mass-{\em in}dependent disruption provides a good match to the observed mass-age distribution of star clusters in the Magellanic Clouds, at least over the $M$-$\tau$ range studied here: roughly $\tau \lea 10^7(M/10^2\,M_{\odot})^{1.3}$~yr. The model can be expressed as $g(M,\tau) \propto M^{\beta} \tau^{\gamma}$, with best-fit exponents $\beta=-1.8\pm0.2$ and $\gamma=-0.8\pm0.2$.
(2) Models~1 and 2 give poor descriptions of the mass and age distributions of star clusters in the Magellanic Clouds, for all combinations of the adjustable parameters $k$ and $\tau_*$. This contradicts previously published
claims (Boutloukos \& Lamers 2003; Lamers et~al.\ 2005; de Grijs \& Anders 2006) that these models, with the specific parameters $k\approx0.6$ and $\tau_*=8\times10^9$~yr, provide a good description of these cluster systems.
(3) The luminosity function of star clusters in the Magellanic Clouds is well represented by a single power law, $dN/dL \propto L^{\alpha}$, with $\alpha=-1.8\pm0.2$. Within the uncertainties, the luminosity function has the same power-law exponent as the mass function. Fall (2006) has shown that this is a direct consequence of the fact that the cluster age and mass
distributions are independent of one another.
(4) The young clusters in the LMC, with $\tau \lea 10^9$~yr, have substantially lower internal (half-mass) densities than the old globular clusters
with $\tau \approx10^{10}$~yr. Thus, the young clusters have correspondingly lower rates of evaporation, and this explains the absence of a bend in their mass function, according to the McLaughlin \& Fall (2008) model.
Our main result is that the bivariate mass-age distribution for star clusters in
the LMC and likely in the SMC can be approximated by $g(M,\tau) \propto M^{\beta} \tau^{\gamma}$ with indices $\beta$ and $\gamma$ similar to those we found for star clusters in the Antennae galaxies (i.e., $\beta\approx-2$ and $\gamma\approx-1$; Zhang \& Fall 1999; FCW05; WCF07; FCW09).
The Magellanic Cloud clusters extend the domain of validity for this result to older ages and to lower masses than can currently be studied for clusters in the Antennae. The similarity found for $g(M,\tau)$ provides a strong test that
the processes which dominate the formation and early disruption of star clusters are similar from galaxy to galaxy, because the Antennae and the Magellanic Clouds represent different environments for star and cluster formation: two large, interacting galaxies versus two small, relatively quiescent galaxies. We hypothesize that this simple formula for $g(M,\tau)$ describes, at least approximately, the early evolution of all star clusters (open, globular, populous, proto-, and super-), in many if not most galaxies.
We discussed the following, approximate sequence for disruption processes
that shape $g(M,\tau)$ over the first $\approx10^9$~yr:
(1)~removal of the ISM by the activity of massive stars (e.g., winds, radiation, supernovae), $\tau \la 10^7$~yr,
(2)~continued mass loss due to stellar evolution, $10^7 {\rm yr} \la \tau \la 10^8~{\rm yr}$, and
(3)~tidal disturbances by passing molecular clouds, $\tau \ga 10^8$~yr.
These processes are expected to operate in a fashion that will not change the shape of the mass function with age, consistent with the observations.
If this picture turns out to be generally valid, it will mean that the main difference between populations of young clusters in different galaxies is simply in the normalization (amplitude) of $g(M, \tau)$, and hence in the number of clusters formed.
\acknowledgements{We are extremely grateful to Deidre Hunter and
Bruce Elmegreen for providing their Magellanic Cloud cluster catalogs and photometry, and for helpful discussions about them. We thank Dean McLaughlin, Francois Schweizer, and Bruce Elmegreen for comments on an earlier version of the manuscript, and the referee for helpful suggestions that improved our paper. SMF acknowledges support from the Ambrose Monell Foundation and from NASA grant AR-09539.1-A. }
|
\section{Introduction}
We consider finite, undirected graphs $G=(V,E)$ with vertex set $V=V(G)$
and edge set $E=E(G)$. The graphs might have multiple edges but no loops.
Let $G$ be a graph. The length of the shortest (odd) cycle of the
underlying simple graph of $G$ is called the {\em (odd) girth} of $G$, i.e.~the girth of $G$ is always at least three.
For $X\subseteq V(G)$ we denote by $\partial_{G}(X)$
the set of edges with precisely one end in $X$. The minimum and maximum degree
of $G$ is denoted by $\delta(G)$ and $\Delta(G)$, respectively. A partial proper $t$-edge-coloring of $G$ is an assignment
of colors $1,...,t$ to some edges of $G$ such that adjacent
edges receive different colors. Let $\theta$ be a partial proper $t$-edge-coloring
of $G$. The components of the subgraph which is induced by two colors $\alpha$ and $\beta$ are
called $(\alpha,\beta)$-Kempe-chains. That is, a Kempe-chain is either a path or an even cycle. In the case that it is a path $P$, we
also say, that $P$ is an $\alpha$-$\beta$-alternating path.
A partial proper $t$-edge-coloring of $G$ is called a proper
$t$-edge-coloring (or just $t$-edge-coloring) if all edges
are assigned some color. The smallest number $k$ for which $G$ has
a $k$-edge-coloring is called the {\em chromatic index} of $G$, and it is
denoted by $\chi'(G)$. A graph $G$ is {\em critical}, if
$\chi'(G)>\Delta(G)$ and $\chi'(G-e)<\chi'(G)$, for every edge $e \in E(G)$.
If $G$ is simple, we also say that $G$ is $\Delta(G)${\em-critical}.
Clearly, $\Delta(G) \leq \chi'(G)$, and by the classical theorems of Shannon and Vizing we have the following upper bounds
for the chromatic index of a graph.
\begin{theorem}\label{Shannon}(Shannon) Let $G$ be a graph, then $\chi'(G)\leq\lfloor\frac{3\Delta(G)}{2}\rfloor$.
\end{theorem}
\begin{theorem}\label{Vizing}(Vizing) Let $G$ be a graph, then
$\chi'(G)\leq\Delta(G)+\mu(G)$, where $\mu(G)$ is the maximum
multiplicity of an edge in $G$.
\end{theorem}
A graph $G$ with $\chi'(G) = \Delta(G) = \Delta$ is {\em class I},
otherwise it is {\em class II}.
There are long standing open conjectures on class II graphs, cf. \cite{Jensen_Toft}.
It is a notorious difficult open problem to characterize class II graphs or even to obtain
some insight into their structural properties. This paper focuses on the $\Delta$-edge-colorable part
of graphs. A subgraph $H$ of $G$ is called {\em maximum} $\Delta$-edge-colorable, if it is $\Delta$-edge-colorable
and contains as many edges as possible. The fraction $|E(H)|/|E(G)|$ is the subject of many papers,
e.g. lower bounds are proved for cubic, subcubic or 4-regular graphs, \cite{A_Haas,part1,Rizzi_2009}.
One aim of this paper is to prove a general best possible lower bound for all graphs.
Let $H$ be a maximum $\Delta$-edge-colorable
subgraph of $G$, which is properly colored with colors $1,...,\Delta$.
Usually, we will refer to edges of $E(G)\backslash E(H)$ as uncolored
edges. For a vertex $v$ of $G$ let $C(v)$ be the set of colors
that appear at $v$, and $\overline{C}(v)=\{1,...,\Delta\}\backslash C(v)$
be the set of colors which are missing at $v$.
Let $e=(v,u)\in E(G)\backslash E(H)$ be an uncolored edge, $\alpha\in\overline{C}(u)$,
$\beta\in\overline{C}(v)$. Since $H$ is a maximum $\Delta$-edge-colorable
subgraph of $G$, we have that $\alpha\in C(v)$ and $\beta\in C(u)$.
Consider the $\alpha$-$\beta$-alternating path $P$ starting from the
vertex $v$. Again, since $H$ is a maximum $\Delta$-edge-colorable
subgraph of $G$, the path $P$ ends in $u$. Thus $P$ is an even
path, which together with the edge $e$ forms an odd cycle. We will
denote this cycle by $C_{\alpha,\beta,H}^{e}$. If the subgraph $H$
is fixed, then we will shorten the notation to $C_{\alpha,\beta}^{e}$.
Kempe chains forming an odd cycle together with an uncolored edge $e$, $C_{\alpha,\beta}^e$, play a central role in the study of cubic graphs \cite{SteffenClassification,Steffen}. The second aim of this paper is to generalize
some of these results to arbitrary graphs, and to investigate the
maximum $\Delta$-edge-colorable subgraphs. We show that any set of
vertex-disjoint cycles of a graph $G$ with $\Delta(G)\geq3$ can be extended to a maximum
$\Delta$-edge-colorable subgraph of $G$. In particular, any $2$-factor
of a graph with maximum degree at least three can be extended to such a subgraph.
Let $\phi$ be a $\chi'(G)$-edge-coloring of a graph $G$ with
$\chi'(G)=\Delta(G)+k$ ($k\geq1$), and $r'_{\phi}(G)=\min\sum_{j=1}^{k}|\phi^{-1}(i_{j})|$.
We define $r'_{e}(G)=\min_{\phi}r'_{\phi}(G)$ as the minimum size
of the union of $k$ color-classes in a $\chi'(G)$-edge-coloring
of $G$. Let $r_{e}(G)$ denote the minimum
number of edges that should be removed from $G$ in order
to obtain a graph $H$ with $\chi'(H)=\Delta(G)$.
Clearly, $r_{e}(G)=|E(G)|-|E(H)|$, where $H$ is a maximum $\Delta$-edge-colorable
subgraph of $G$. In \cite{part2} it is shown that the complement of any maximum
$3$-edge-colorable subgraph of a cubic graph is a matching, and hence $r_{e}(G)=r'_{e}(G)$
for cubic graphs. This paper generalizes this result to simple graphs.
We further prove some bounds for the vertex degrees of a maximum $\Delta(G)$-edge-colorable
subgraph $H$.
\section{Maximum $\Delta$-edge-colorable subgraphs: Cycles}
The key property of cycles corresponding to uncolored edges in cubic graphs that is
used in \cite{SteffenClassification,Steffen} is their vertex-disjointness.
As the example from Figure \ref{FatTriangleExample} shows, they can
have even common edges in the general case. To see this, consider
the graph $G$ and its maximum $\Delta(G)$-edge-colorable subgraph
$H$ from Figure \ref{FatTriangleExample}. Note that $\overline{C}(a)=\{4\}$,
$\overline{C}(b)=\{3\}$, $\overline{C}(c)=\{1,2\}$,
and hence $E(C_{1,3}^{(b,c)})\cap E(C_{1,4}^{(a,c)})\neq\emptyset$.
\begin{center}
\begin{figure}[ht]
\begin{center}
\includegraphics[height=10pc, width=13pc]{FatTriangleExample.eps}\\
\caption{Cycles of uncolored edges may intersect.}\label{FatTriangleExample}
\end{center}
\end{figure}
\end{center}
Despite this, it turns out that, as Theorem \ref{assignment} demonstrates
below, the edge-disjointness of the cycles can be preserved. For the
proof we need the following Lemma, where we implicitly assume that the maximum $\Delta(G)$-edge-colorable
subgraph $H$ of a graph $G$ is colored with $\Delta=\Delta(G)$
colors. Before we proceed, let us note that if $C_{\alpha,\beta}^{e}$ and $C_{\gamma,\delta}^{e'}$ are two cycles corresponding to two different uncolored edges $e$ and $e'$ with respect to a maximum $\Delta(G)$-edge-colorable subgraph, then $E(C_{\alpha,\beta}^{e})\cap E(C_{\gamma,\delta}^{e'})\neq \emptyset$ implies $\{\alpha,\beta\}\cap \{\gamma,\delta\} \neq \emptyset$. The Lemma, that we are going to prove, describes the placement of the edges of $e$ and $e'$.
\begin{lemma}\label{CyclesIntersection} Let $H$ be any maximum
$\Delta(G)$-edge-colorable subgraph of a graph $G$, and $e$, $e'$
be two uncolored edges. If $E(C_{\alpha,\beta}^{e})\cap E(C_{\alpha,\gamma}^{e'})\neq\emptyset$,
then there is a vertex $v$ such that $e$ and $e'$ are
incident to $v$ and $\alpha \in \overline{C}(v)$. \end{lemma}
\begin{proof} Let $e$ and $e'$ be two uncolored edges with respect to $H$, and
$E(C_{\alpha,\beta}^{e})\cap E(C_{\alpha,\gamma}^{e'})\neq \emptyset$. If $\beta=\gamma$, then $e$ and $e'$ are parallel, therefore the statement of the Lemma is trivial. Thus, we can assume that $\beta\neq\gamma$, and therefore every edge $e''\in E(C_{\alpha,\beta}^{e})\cap E(C_{\alpha,\gamma}^{e'})$ is colored with color $\alpha$.
We first show that $e$ and $e'$ are adjacent. Assume to the contrary that
this is not the case. Let $P'$ be a subpath of $C_{\alpha,\gamma}^{e'}$ that connects $e'$ to $(v,w)=e''\in E(C_{\alpha,\beta}^{e})\cap E(C_{\alpha,\gamma}^{e'})$, where $e''$ is chosen so that $E(P')\cap E(C_{\alpha,\beta}^{e})\cap E(C_{\alpha,\gamma}^{e'})=\emptyset$.
Note, that this is always possible. Suppose that $P'$ connects $e'$ to the vertex $w$. Let $P$ be a subpath of $C_{\alpha,\beta}^{e}$ that connects $e$ to the vertex $v$, and does not pass through $w$.
There are edges $f\in E(P+e)$,
$f'\in E(P'+e')$ which are adjacent to $e''$. If $f\neq e$, then $f$ is colored with
color $\beta$, and similarly, if $f'\neq e'$, $f'$ is colored with color $\gamma$. Moreover, $f$ and $f'$ do not share a vertex, $f$ is
incident to $v$, and $f'$ to $w$.
Now recolor $E(P)+e$ by leaving
$e''$ uncolored and color the remaining edges with colors $\alpha$,
$\beta$ alternately, to obtain another maximum $\Delta$-edge-colorable
subgraph $H'$of $G$.
By the choice of $P$ and $P'$, no edge of the subpath $P'$ of $C_{\alpha,\gamma}^{e'}$
is involved in the recoloring process. Thus $P'+e'$ can be colored
with colors $\alpha$, $\gamma$ and we obtain a $\Delta$-edge-colorable
subgraph $H^{*}$ with $|E(H^{*})|>|E(H)|$, contradicting our choice
of $H$. Thus $e$ and $e'$ are adjacent.
If $e$ and $e'$ are parallel, then our statement follows easily. It remains to consider the case when $e$ and $e'$ share precisely
one vertex, say $v$. Assume to the contrary that $\alpha\not\in\overline{C}(v)$,
i.e. $\alpha\in C(v)$ . Then, interchanging colors $\alpha$ and $\beta$ in $C_{\alpha,\beta}^{e}$
allows us to color $e'$ with color $\alpha$, contradicting the maximality
of $H$. Hence it follows that $\alpha\in\overline{C}(v)$, and the statement
is proved.$\square$ \end{proof}
\begin{theorem}\label{assignment}Let $H$ be any maximum $\Delta(G)$-edge-colorable
subgraph of a graph $G$, and let $E(G)-E(H)=\{e_{i}=(u_{i},v_{i})|1\leq i\leq n\}$
be the set of uncolored edges. Assume that $H$ is properly edge-colored
with colors $1,\ldots,\Delta(G)$. Then there is an assignment of
colors $\alpha_{1}\in\overline{C}(u_{1}),\beta_{1}\in\overline{C}(v_{1}),\ldots,\alpha_{n}\in\overline{C}(u_{n}),\beta_{n}\in\overline{C}(v_{n})$
to the uncolored edges, such that $
E(C_{\alpha_{i},\beta_{i}}^{e_{i}})\cap E(C_{\alpha_{j},\beta_{j}}^{e_{j}})=\emptyset$,
for all $1\leq i<j\leq n$. \end{theorem}
\begin{proof} We prove the statement by induction on the number $n$
of uncolored edges.
If $n=1$, then the statement is trivial.
Let $n \geq 2$, and $G$ be a graph with $|E(G)-E(H)|=n$. We
need to consider two cases.
Case 1: There is a $k$ ($1\leq k\leq n$), such that $\Delta(G-e_{k})=\Delta(G)$.
Then, $H$ is a maximum $\Delta(G-e_{k})$-edge-colorable
subgraph of a graph $G-e_{k}$. Thus by the induction hypothesis, there
are $\alpha_{1}\in\overline{C}(u_{1}),\beta_{1}\in\overline{C}(v_{1}),\ldots,\alpha_{k-1}\in\overline{C}(u_{k-1}),\beta_{k-1}\in\overline{C}(v_{k-1}),\alpha_{k+1}\in\overline{C}(u_{k+1}),\beta_{k+1}\in\overline{C}(v_{k+1}),\ldots,\alpha_{n}\in\overline{C}(u_{n}),\beta_{n}\in\overline{C}(v_{n})$
such that \[
E(C_{\alpha_{i},\beta_{i}}^{e_{i}})\cap E(C_{\alpha_{j},\beta_{j}}^{e_{j}})=\emptyset,
\mbox{ for all } 1\leq i<j\leq n \hspace{.5cm} (i,j\neq k). \]
Define
$I_{k}=\{e_{i}|i\neq k\textrm{, and }e_{i} \textrm{ is incident to }u_{k}\}$, and $J_{k}=\{e_{i}|i\neq k\textrm{, and }e_{i}\textrm{ is incident to }v_{k}\}$.
Then
$|\overline{C}(u_{k})|\geq 1+|I_{k}|$, $|\overline{C}(v_{k})|\geq 1+|J_{k}|$,
and hence there are $\alpha_{k}\in\overline{C}(u_{k}),\beta_{k}\in\overline{C}(v_{k})$,
such that
$\alpha_{k}\neq\alpha_{i},e_{i}\in I_{k}$ and $\beta_{k}\neq\beta_{j},e_{j}\in J_{k}$.
Choose the colors $\alpha_{k},\beta_{k}$ to create $C_{\alpha_{k},\beta_{k}}^{e_{k}}$, then Lemma
\ref{CyclesIntersection} implies that
$ E(C_{\alpha_{i},\beta_{i}}^{e_{i}})\cap E(C_{\alpha_{j},\beta_{j}}^{e_{j}})=\emptyset$,
for all $1\leq i<j\leq n$.
Case 2: For each $k$ ($1\leq k\leq n$) we have $\Delta(G-e_{k})<\Delta(G)$.
Then, there is a vertex $v$ of maximum degree $\Delta(G)$, such that
$v$ is incident to all uncolored edges $e_{1},...,e_{n}$ (Figure
\ref{AllUncoloreds}).
Thus, without loss of generality, we can assume that $v=v_{1}=...=v_{n}$.
Let $u^{(1)},...,u^{(l)}$ ($l \leq n$) be all the neighbors of $v$ such that
$v$ and $u^{(i)}$ are connected by $k_{i}$ uncolored edges, $k_{i}\geq1$ , $i=1,...,l$
(Figure \ref{AllUncoloreds}). Clearly,
$k_{1}+...+k_{l}=n$.
We may assume that the edges $e_{1},...,e_{k_{1}}$ are incident
to $u^{(1)}$, the edges $e_{k_{1}+1},...,e_{k_{1}+k_{2}}$ are incident to
$u^{(2)}$,..., the edges $e_{k_{1}+...+k_{l-1}+1},...,e_{n}$ are incident to
$u^{(l)}$. Moreover, let $\overline{C}(v)=\{c_{1},...,c_{n}\}$ and
$\overline{C}(u^{(j)})=\{c_{1}^{(j)},...,c_{t_{j}}^{(j)}\},1\leq j\leq l$.
\begin{center}
\begin{figure}[ht]
\begin{center}
\includegraphics[height=13pc, width=20pc]{edge_colorable_subg_3.eps}\\
\caption{All uncolored edges are incident to the vertex $v$.}
\label{AllUncoloreds}
\end{center}
\end{figure}
\end{center}
Note that $t_{j}\geq k_{j}$, for $j=1,...,l$. Define
\begin{eqnarray*}
\beta_{1}=c_{1},...,\beta_{n}=c_{n},\\
\alpha_{1}=c_{1}^{(1)},...,\alpha_{k_{1}}=c_{k_{1}}^{(1)},\\
\alpha_{k_{1}+1}=c_{1}^{(2)},...,\alpha_{k_{1}+k_{2}}=c_{k_{2}}^{(2)},\\
\dots\\
\alpha_{k_{1}+...+k_{l-1}+1}=c_{1}^{(l)},...,\alpha_{n}=c_{k_{l}}^{(l)}.\end{eqnarray*}
Now, it is not hard to see that the choice of $\alpha_{1},\beta_{1},...,\alpha_{n},\beta_{n}$
and Lemma \ref{CyclesIntersection} imply that
$E(C_{\alpha_{i},\beta_{i}}^{e_{i}})\cap E(C_{\alpha_{j},\beta_{j}}^{e_{j}})=\emptyset$,
for all $1\leq i<j\leq n$. $\square$
\end{proof}
The next Theorem generalizes the result of \cite{part2} that any
$2$-factor of a cubic graph can be extended to a maximum $3$-edge-colorable
subgraph to arbitrary graphs.
\begin{theorem}\label{CycleSystemExtension}Let $\overline{F}$ be any
set of vertex-disjoint cycles of a graph $G$ with $\Delta=\Delta(G)\geq3$. Then there is a maximum
$\Delta$-edge-colorable subgraph $H$ of $G$, such that $E(\overline{F})\subseteq E(H)$.
\end{theorem}
\begin{proof}Let $\Delta=\Delta(G)\geq3$. For $\overline{F}$ consider a maximum
$\Delta$-edge-colorable subgraph $H$ of $G$ with
$|E(\overline{F})\cap E(H)|\textrm{ is maximum}$.
We will show that $E(\overline{F})\subseteq E(H)$. Assume to the contrary that there is an edge
$e=(u,v)\in E(\overline{F})$, such that $e$ does not belong to $E(H)$.
Let us assume that $H$ is properly colored with colors $1,2,...,\Delta$.
Case 1: There are $\alpha\in C(u)\backslash C(v)$ and $\beta\in C(v)\backslash C(u)$,
such that $C_{\alpha,\beta}^{e}$ contains an edge $f$ that does
not belong to a cycle of $\overline{F}$.
Consider a proper partial $\Delta$-edge-coloring of $G$ obtained
from the coloring of $H$ by shifting the colors on the cycle $C_{\alpha,\beta}^{e}$
and leaving the edge $f$ uncolored. The new partial $\Delta$-edge-coloring
corresponds to a maximum $\Delta$-edge-colorable subgraph $H'$ of
$G$ with
$|E(\overline{F})\cap E(H')|>|E(\overline{F})\cap E(H)|$,
contradicting the choice of $H$.
Case 2: For all $\alpha\in C(u)\backslash C(v)$ and $\beta\in C(v)\backslash C(u)$,
the edges of $C_{\alpha,\beta}^{e}$ lie on a cycle of $\overline{F}$.
Since the cycles of $\overline{F}$ are vertex-disjoint, it follows that
$|C(u)\backslash C(v)|=|C(v)\backslash C(u)|=1$ and there is only
one cycle $C_{\alpha,\beta}^{e}=C_{e}$, which, in its turn, is a cycle
of $\overline{F}$. This implies that $d_H(u)=d_H(v)=\Delta(G)-1\geq2$ and $d(u)=d(v)=\Delta(G)\geq3$.
Let us assume that $C_{e}-e$ is colored by the colors $\Delta-1$
and $\Delta$, alternately. For an edge $f$ of $C_e$, define $H_f$ obtained from $H$ as follows,
\begin{enumerate}
\item leave the edge $f$ uncolored,
\item color the edges of the even path $C_{e}-f$ with colors $\Delta-1$
and $\Delta$, alternately,
\item leave the colors of the rest of edges unchanged.
\end{enumerate}
Note that the subgraphs $H_e$ and $H$ are the same, though they may have received different colorings, and for each edge $f$ of $C_e$, $H_f$ is a maximum $\Delta$-edge-colorable subgraph of $G$ with
$|E(\overline{F})\cap E(H_f)|=|E(\overline{F})\cap E(H)|$. Similar to the consideration of the case 1 with respect to $H_f$, it can be shown that we can assume that if $f=(p,q)$ is an edge of $C_e$, then $d_{H_f}(p)=d_{H_f}(q)=\Delta(G)-1\geq2$ and $d(p)=d(q)=\Delta(G)\geq3$. This implies that the vertices of the cycle $C_e$ are of degree $\Delta$. Since $\Delta\geq3$, each vertex of $C_e$ is incident to $\Delta-2\geq 1$ edges that do not lie on $C_{e}$, and which are colored by the colors $1,...,\Delta-2$ in every $H_f$.
Let $\theta$ be a proper partial $\Delta$-edge-coloring
of $G$ which is obtained from the coloring of $H$ by deleting the
colors of all edges of $C_{e}$. Since $C_{e}$ is an odd cycle, it
follows that there is a $1$-$\Delta$-alternating path $P_{w}$ (with respect to the edge-coloring $\theta$)
that starts from a vertex $w\in V(C_{e})$
and does not end at a vertex of $C_{e}$. Choose an edge $g=(w,z)\in E(C_{e})$,
and let $h$ be the other edge of $C_{e}$ that is incident to $w$.
Consider a proper partial $\Delta$-edge-coloring of $G$ obtained
from $\theta$ as follows,
\begin{enumerate}
\item shift the colors on the path $P_{w}$, and clear the color of the
edge that is incident to $z$ and has a color $1$ in $\theta$,
\item color $g$ with color $1$, and color the edges of the even path $C_{e}-g$
with colors $\Delta-1$ and $\Delta$ alternately, starting from the edge
$h$.
\end{enumerate}
This new partial $\Delta$-edge-coloring
of $G$ corresponds to a maximum $\Delta$-edge-colorable subgraph
$H'$ of $G$, which satisfies $
|E(\overline{F})\cap E(H')|>|E(\overline{F})\cap E(H_g)|=|E(\overline{F})\cap E(H)|$,
contradicting the choice of $H$. $\square$ \end{proof}
\begin{corollary}\label{2factorExtension}Let $\overline{F}$ be any 2-factor
of a graph $G$ with $\Delta(G)\geq3$. Then, there is a maximum $\Delta(G)$-edge-colorable
subgraph $H$ of $G$, such that $E(\overline{F})\subseteq E(H)$. \end{corollary}
\section{Maximum $\Delta$-edge-colorable subgraphs: Cuts and size}
\begin{theorem}\label{cutCondition} Let $H$ be any maximum $\Delta(G)$-edge-colorable
subgraph of a graph $G$. Then
\begin{enumerate}
\item $|\partial_{H}(X)|\geq\lceil\frac{|\partial_{G}(X)|}{2}\rceil$ for each $X\subseteq V(G)$,
\item $d_{H}(x)\geq\lceil\frac{d_{G}(x)}{2}\rceil$ for each vertex $x$ of $G$, and
\item $\delta(H)\geq\lceil\frac{\delta(G)}{2}\rceil$.
\end{enumerate}
Furthermore, the bounds are best possible.
\end{theorem}
\begin{proof} 1. Let $X\subseteq V(G)$, and assume
that $\partial_{G}(X)$ contains $k$ uncolored edges.
By Theorem \ref{assignment}, there is
an assignment of colors to uncolored edges with respect to $H$ such
that the corresponding cycles do not intersect. Since every
cycle $C_{\alpha,\beta}^{e}$ of an uncolored edge $e$ of $\partial_{G}(X)$,
intersects $\partial_{G}(X)$ at least twice, there are at least
$k$ pairwise different edges of $H$ that belong to $\partial_{G}(X)$.
Statements 2. and 3. follow directly from the first. It remains to show that the bounds are best possible.
Let $H_r$ be the complete bipartite graph $K_{2r+1,2r+1}$ with a subdivided edge. It is well known, that $H_r$ is
$(2r+1)$-critical. Therefore it has a $(2r+1)$-edge-coloring, that leaves precisely one
edge, which is incident to a vertex of degree 2, uncolored. Take $r$ copies
of $H_r$ and identify the vertices of degree two to obtain the graph $H$.
Now take two copies of $H$
and connect the vertices of degree $2r$ with an edge to obtain the
graph $G$. It is not hard to observe that $G$ has a maximum $(2r+1)$-edge-colorable subgraph of minimum degree $r+1$.
This implies that the bounds are best possible (for statement 3 as well, since $G$ is regular). Note, that the graphs are simple.
For multi-graphs, take $k > 1$ copies of the graph with three vertices, one vertex of degree 2 and the other two
of degree $2k$. These graphs have a $2k$-edge-colorable subgraph that leaves precisely one
edge uncolored, which is incident to a bivalent vertex. Identify the bivalent vertices to obtain a
$2k$-regular graph with the desired properties.
$\square$
\end{proof}
Next we will construct some graphs to which we will refer in the following to show that the bounds of Theorem \ref{Multigraphgirth} are best possible. Recall that, by definition, the girth of a graph is always at least three.
\begin{lemma} \label{Example}
For each $k \geq 1$, there is a graph $G_k$ with girth $g = 2k+1$, such that $|E(H_k)| = \frac{2k}{2k+1}|E(G_k)|$,
for each maximum $\Delta(G_k)$-edge-colorable subgraph $H_k$.
Furthermore, there is a maximum $\Delta(G_k)$-edge-colorable subgraph $H_k^*$
with $\Delta(H_k^*) = \lceil \frac{2k}{2k+1} \Delta(G_k) \rceil$.
\end{lemma}
\begin{proof}
For $k \geq 1$, let $G_k = C^{2k}_{2k+1}$ be a cycle of length $2k+1$, where each edge has multiplicity $2k$.
Then $\Delta(C^{2k}_{2k+1})=4k=\Delta$, and $|E(C^{2k}_{2k+1})|=2k(2k+1)$. Every color class contains at most $k$ edges, hence
$|E(H_k)| \leq 4k^2$, for every maximum $\Delta$-edge-colorable subgraph.
Let $H_k^*$ be the subgraph of $G_k$ with
$4k^2$ edges, one edge of multiplicity $2k$ and the remaining $2k$ edges of multiplicity $2k-1$. It is easy to see that
$H_k^*$ is $4k$-edge-colorable, thus $|E(H_k)| = 4k^2$ for all maximum $\Delta$-edge-colorable subgraphs of $G_k$.
Furthermore $\Delta(H_k^*) = 4k-1 = \lceil 4k(1 - \frac{1}{2k+1}) \rceil$. $\square$
\end{proof}
\begin{theorem}\label{Multigraphgirth} If $G$ is a graph with girth $g \in \{2k, 2k+1\}$
($k\geq 1$), and $H$ a maximum $\Delta(G)$-edge-colorable
subgraph of $G$, then $|E(H)| \geq \frac{2k}{2k+1}|E(G)|$,
and the bound is best possible.
\end{theorem}
\begin{proof} Let $H$ be a maximum $\Delta(G)$-edge-colorable subgraph of $G$, and
$\{e_1, \dots, e_n\}$ be the set of uncolored edges. Let $C_{\alpha_i, \beta_i}^{e_i}$ be the pairwise edge-disjoint
cycles of Theorem \ref{assignment}, $G'=(V(G), \bigcup_{i=1}^n E(C_{\alpha_i, \beta_i}^{e_i}))$, and $H'$ the colored
subgraph of $G'$. Then it follows by Theorem \ref{assignment}, that $|E(H')| \geq \frac{2k}{2k+1}|E(G')|$. Furthermore,
$E(G)- E(G')=E(H)- E(H')$, and hence $|E(H)|= |E(H')|+|E(G)-E(G')| \geq \frac{2k}{2k+1}|E(G')| + |E(G)-E(G')| \geq \frac{2k}{2k+1}|E(G)|$.
The graphs $G_k$ of Lemma \ref{Example} show that the bound is best possible.
$\square$
\end{proof}
Theorem \ref{cutCondition} implies that $\Delta(H) \geq \lceil \frac{\Delta(G)}{2} \rceil$.
However for the maximum vertex degree of a maximum $\Delta(G)$-edge-colorable subgraph $H$ of $G$ much better
bounds can be proved. For this we need the well known fact, that
(*) $\chi'(G) \leq \frac{g}{g-1}\Delta(G) + \frac{g-3}{g-1}$ for graphs $G$ with $\chi'(G) > \Delta(G) +1$
and odd girth $g$. This result is easily deducible from the results of Kierstead \cite{Kierstead} (about acceptable paths),
and it implies that
$\Delta(G) = \chi'(H) \leq \frac{g}{g-1}\Delta(H) + \frac{g-3}{g-1}$. In \cite{girthEckhard} it is proved
that $\chi'(G) \leq \Delta(G) + \lceil \frac{\mu(G)}{\lfloor \frac{g}{2}\rfloor}\rceil$. We summarize the consequences of these
two results in the following corollary.
\begin{corollary} If $G$ is a graph with girth $g \in \{2k, 2k+1\}$
$(k\geq 1)$, and $H$ a maximum $\Delta(G)$-edge-colorable subgraph of $G$, then
\begin{enumerate}
\item $\chi'(H) \geq \max \{\chi'(G) - \lceil \frac{\mu(G)}{k}\rceil, \frac{2k}{2k+1}\chi'(G) - \frac{2k-2}{2k+1}\}$,
\item $\Delta(H) \geq \max \{ \Delta(G) - \lceil \frac{\mu(G)}{k}\rceil, \frac{2k}{2k+1}\Delta(G) - \frac{2k-2}{2k+1} \}$.
\end{enumerate}
\end{corollary}
\subsection{Simple graphs}
\begin{theorem}\label{MatchingComplement} Every simple graph
$G$ contains a maximum $\Delta$-edge-colorable subgraph, such
that the uncolored edges form a matching. \end{theorem}
\begin{proof}
Take a maximum $\Delta$-edge-colorable subgraph $H$ of $G$ which
minimizes the number of pairs of adjacent uncolored edges. The proof
will be completed if we show that this number is zero.
Assume, on the contrary, that there is a vertex $v$ which is incident
to two uncolored edges, one of which is $(u,v)$. Let $H$ be $\Delta$-edge-colored
and assume that $\alpha_{0}\in\overline{C}(v)$, $\beta\in\overline{C}(u)$. Since $H$
is a maximum $\Delta$-edge-colorable subgraph of $G$, there is an
edge $(u,v_{0})$ that is colored with color $\alpha_{0}$. Consider a maximal
fan beginning from the vertex $v_{0}$, that is a maximal sequence
$(v_{0},\alpha_{0},v_{1},\alpha_{1},...,v_{k},\alpha_{k})$ such that
\begin{enumerate}
\item $(u,v_{i})\in E(G)$, and for $i=0,...,k$, the edge $(u,v_{i})$ is colored with color $\alpha_{i}$,
\item $\alpha_{i}\in\overline{C}(v_{i-1})$, for $i=1,...,k$, and
\item $\alpha_0, \ldots ,\alpha_k$ are distinct colors.
\end{enumerate}
We show that $\overline{C}(v_{k})\neq\emptyset$. Suppose that $\overline{C}(v_{k})=\emptyset$.
This means that $v_{k}$ is a vertex of maximum degree $\Delta(G)$
and all edges incident to it are from $H$. Consider a maximum $\Delta$-edge-colorable
subgraph $H'$ of $G$ obtained from $H$ as follows: Color $(u,v)$
by $\alpha_{0}$, for $i=0,...,k-1$ color $(u,v_{i})$ with color $\alpha_{i+1}$, and leave the edge
$(u,v_{k})$ uncolored. Then there are less pairs of adjacent
uncolored edges in $H'$ than in $H$, which
contradicts the choice of $H$.
Thus, there exist $\alpha_{k+1}\in\overline{C}(v_{k})$. The maximality
of the sequence $(v_{0},\alpha_{0},...,v_{k},\alpha_{k})$ implies that $\alpha_{k+1}\in\{\alpha_{0},...,\alpha_{k-1}\}$,
say $\alpha_{k+1}=\alpha_{i}$, for an $0\leq i\leq k-1$.
Recall that $\beta \in\overline{C}(u)$. The maximality of $H$ implies that
$\beta \in C(v_{k})$. Consider the maximal $\beta$-$\alpha_{k+1}=\beta$-$\alpha_{i}$ alternating
path $P$ beginning from the vertex $v_{k}$. We will show that our assumptions imply
that $H$ is not a maximum $\Delta$-edge-colorable subgraph of $G$, contradicting the choice of $H$.
Case 1: $P$ does not reach $u,v_{i},v_{i-1}$.
The following recoloring yields the desired contradiction.
Color the edge $(u,v)$ by $\alpha_{0}$, $(u,v_{j})$ by $\alpha_{j+1}$, $j=0,...,k-1$,
exchange the colors on the path $P$ and color $(u,v_{k})$ by $\beta$.
Case 2: $P$ reaches $v_{i}$.
Since the edge $(u,v_{i})$ is colored by $\alpha_{i}$ it follows
that the path $P$ enters $v_{i}$ by an edge colored $\beta$. To get the contradiction, color
the edge $(u,v)$ by $\alpha_{0}$, $(u,v_{j})$ by $\alpha_{j+1}$, $j=0,...,i-1$,
exchange the colors on the path $P$ and color $(u,v_{i})$ by $\beta$.
Case 3: $P$ reaches $v_{i-1}$.
Since $\alpha_{i}\in\overline{C}(v_{i-1})$, it follows that the path
$P$ reaches $v_{i-1}$ by an edge colored $\beta$. Color the edge
$(u,v)$ by $\alpha_{0}$, $(u,v_{j})$ by $\alpha_{j+1}$, $j=0,...,i-2$,
exchange the colors on the path $P$ and color $(u,v_{i-1})$ by $\beta$ to get the desired
contradiction. $\square$ \end{proof}
Theorem \ref{MatchingComplement} is equivalent to
\begin{theorem} If $G$ is a simple graph, then $r_{e}(G)=r'_{e}(G)$.
\end{theorem}
Lemma \ref{Example} shows that a maximum $\Delta$-edge-colorable subgraph of a multigraph can be class II as well.
This cannot be the case for simple graphs as the following theorem shows.
\begin{theorem} \label{equality} If $H$ is a maximum $\Delta(G)$-edge-colorable
subgraph of a simple graph $G$, then $\Delta(H)=\Delta(G)$, i.e. $H$ is class I.
\end{theorem}
\begin{proof} Let $e = (v,w) \in E(G)-E(H)$ be an uncolored edge. Then $\chi'(H+e) = \Delta(G)+1$, and hence, since $H+e$ is simple, $\Delta(H+e)= \Delta(G)$. The graph $H+e$ contains a
$\Delta(G)$-critical subgraph $H'$, which clearly contains the edge $e$. By Vizing's Adjacency Lemma \cite{Vizing}, every vertex
of $H'$ is adjacent to at least two vertices of maximum degree $\Delta(G)$.
Thus there is a vertex $x \not= v,w$ of $H'$ with $d_{H'}(x)=\Delta(G)$, and hence
$\Delta(H) = \Delta(G)$. $\square$
\end{proof}
\section{Discussion and some Conjectures}
Theorem \ref{MatchingComplement} says, that every simple class II graph $G$ has a
maximum $\Delta$-edge-colorable subgraph $H$, such that $\chi'(G \backslash E(H))=1$.
We believe that this can be generalized to multigraphs.
\begin{conjecture} \label{k_matching} If $G$ is a graph with $\chi'(G)=\Delta(G)+k$
$(k \geq 0)$,
then there is a maximum $\Delta(G)$-edge-colorable subgraph $H$ of $G$,
such that $\chi'(G\backslash E(H))=k$.
\end{conjecture}
This conjecture is equivalent to the following statement.
\begin{conjecture} Let $G$ be a graph, then $r_{e}(G)=r'_{e}(G)$. \end{conjecture}
\begin{acknowledgement}
We would like to thank our reviewers for their useful comments that helped us to improve the presentation of the paper.
Our special thanks to the second reviewer, who has pointed out a mistake in the earlier version of the proof of Lemma \ref{CyclesIntersection}.
\end{acknowledgement}
|
\section{Introduction}
A Hamiltonian governing the total energy of many electrons hopping and interacting on a finite lattice is defined as a self-adjoint operator on the finite dimensional Hilbert space of all states of electrons. Correlation functions in the system at non-zero temperature are formulated as a quotient of trace operations over the Hilbert space. Despite the explicitness of their mathematical definitions, to rigorously analyze the behavior of the correlation functions still requires restrictive assumptions and remains to be solved in a general setting.
The method using Bogoliubov's inequality initiated by Hohenberg \cite{H}, Mermin and Wagner \cite{MW} has been applied to prove decay of various order parameters in the one- and two-dimensional Hubbard models. It has been shown that the magnetic order parameters (\cite{WR},\cite{G},\cite{U}) and the electron-pairing order parameters (\cite{SSZ},\cite{SS}) vanish in the thermodynamic limit as the amplitude of the corresponding external field goes to zero. In these theories the application of Bogoliubov's inequality demonstrates that the thermodynamic limit of the order parameter under consideration is bounded from above by a constant times the inverse of an integral of the form
$$
\prod_{j=1}^d\int_{-\pi}^{\pi}dk_j\frac{1}{\sum_{j=1}^dk_j^2+\lambda}
$$
with the space dimension $d$ and the amplitude $\lambda>0$ of the external field. If this integral diverges as the amplitude $\lambda$ is sent to zero, which is true for $d=1,2$, not true for $d\ge 3$, one concludes that the order parameter accordingly converges to zero. In order to reach this conclusion, thus, these theories require the space dimension to be less than or equal to $2$.
On the other hand, the method proposed by McBryan and Spencer \cite{MS} to prove decay properties of correlations in classical spin systems has been extended to bound the correlation functions in the one- and two-dimensional Hubbard models in \cite{KT}, \cite{MR}. These theories make use of $U(1)$-symmetry property of the system and deduce that the electron pairing-pairing correlation function for $2$ separate sites ${\mathbf{x}}$, ${\mathbf{y}}$ is bounded from above by
$$e^{-C_1(\theta_{{\mathbf{x}}}-\theta_{{\mathbf{y}}})+C_2\sum_{{\mathbf{u}},{\mathbf{v}}}|t_{{\mathbf{u}},{\mathbf{v}}}|(\cosh(\theta_{{\mathbf{u}}}-\theta_{{\mathbf{v}}})-1)},
$$
where $C_1,C_2>0$ are constants, $\{\theta_{{\mathbf{u}}}\}$ are arbitrary taken real parameters indexed by every site on the lattice, $(t_{{\mathbf{u}},{\mathbf{v}}})_{{\mathbf{u}},{\mathbf{v}}}$ is the hopping matrix and the sum with respect to ${\mathbf{u}},{\mathbf{v}}$ is taken over all the sites. A suitable choice of the parameters $\{\theta_{{\mathbf{u}}}\}$ yields a decaying upper bound on the correlation function. As in \cite{MS} Koma and Tasaki \cite{KT} took $\{\theta_{{\mathbf{u}}}\}$ to be the fundamental solution of a Laplace equation on the lattice and concluded that the pairing-pairing correlation function decays as $|{\mathbf{x}}-{\mathbf{y}}|\to +\infty$. Macris and Ruiz \cite{MR} referred to a list of the possible parameters $\{\theta_{{\mathbf{u}}}\}$ summarized in \cite{MMR} and extended the decay bounds obtained by Koma and Tasaki to be valid for the Hubbard models with long range hopping matrix as well. Since appropriate choices of $\{\theta_{{\mathbf{u}}}\}$ have been found in one and two dimensions, this approach verifies the decay of the correlation functions in these low dimensional cases at present, to the best of the author's knowledge.
Apart from these analysis to bound the correlation functions in low dimensions, Kubo and Kishi \cite{KK} proved that the susceptibilities for the Hubbard models at non-zero temperature are bounded from above by the inverse of modulus of the coupling constant in any space dimension under a few assumptions on the sign of the coupling constant and the lattice.
In this paper we consider the equal-time correlation functions for a class of the Hubbard models at non-zero temperature and show that the correlation functions decay exponentially in the distance between the center of positions of the electrons and the center of positions of the holes in any space dimension if a norm of the interaction term of the Hamiltonian is sufficiently small (see Theorem \ref{thm_exponential_decay} and Theorem \ref{thm_exponential_decay_hubbard_model} in Section \ref{sec_models_results} for the precise statements). As in \cite{KT}, \cite{MR} our approach essentially uses the $U(1)$-symmetry property of the model. We start from characterizing the correlation functions as a limit of Grassmann integrals over finite Grassmann variables. The Grassmann integral formulations called the Schwinger functions are mathematical objects defined on a rigorous base. The $U(1)$-symmetry property of the model is simply implemented in the Grassmann integral and enables us to convert the formulation into a form of multi-contour integral of the Schwinger function with respect to newly introduced complex variables contained in the covariance matrix. Our objective is, thus, set to find an upper bound on the Schwinger function inside the multi-contour integral. The evaluation of the Schwinger function is done perturbatively. We find a volume-independent upper bound on each term of the Taylor expansion of the Schwinger function with respect to the interaction around zero. We require the interaction to be small so that the perturbation series of the Schwinger function converges.
One advantage of this approach is that the space dimension of the system causes no technical difficulty and the resulting decay bounds on the correlation functions are valid in any dimension. However, as we need to go through the perturbation theory with respect to the interaction, the additional assumption is imposed on the magnitude of the interaction.
Rigorous frameworks have been developed to control the perturbation theory for many-Fermion on lattice independently of the volume factor. In \cite{PS} Pedra and Salmhofer extended the notion of Gram's inequality to be applicable to the determinant of the covariance matrices appearing in many-Fermion systems. Their abstract theorem \cite[\mbox{Theorem 1.3}]{PS} is general enough to cover the modified covariance coming into play in our construction and bounds its determinant independently of volume and temperature. The tree formula for partial derivatives of logarithm of the Grassmann Gaussian integral summarized in \cite{SW} coupled with the determinant bound of the covariance makes it feasible to find volume-independent upper bounds on the perturbative expansion of the Schwinger functions. Though in this paper we employ the tree formula to bound the Schwinger functions, the same goal can also be achieved via a concise representation of the Schwinger functions established by Feldman, Kn\"orrer and Trubowitz in \cite{FKTrep}. Indeed, their bound \cite[\mbox{Theorem I.9}]{FKTrep} proved in a quite general context needs only the determinant bound and the $L^1$-bound of the covariance, which are the same inputs to return the upper bounds on the Schwinger functions by means of the tree formula.
We often refer to the recent article \cite{K} for basic lemmas needed in our construction. In fact, this work should be regarded as a continuation of \cite{K}, which intended to explain mathematical tools to analyze the perturbation theory for many electrons in detail.
As in \cite{K} we directly treat the perturbation theory of the original model without introducing any multi-scale technique. Though the results can be presented explicitly in a simple manner through the single scale analysis, it costs the temperature dependency of the interaction. In our analysis the norm of the interaction is restricted to be less than a constant times $\beta^{-d-1}$ for the inverse temperature $\beta$ in $d$-dimensional case. As a renormalization group analysis in the theoretical front let us remark that for a wide class of the Fermionic lattice models various mathematical properties of the correlation functions have been intensively studied by Pedra \cite{P} in a generalized form. Pedra's multi-scale analysis remarkably concludes that the correlation functions can be qualitatively analyzed if a norm of the interaction is bounded by a constant times $(\log \beta)^{-1}$ in a $2$-dimensional case, or less than a constant times $\beta^n$ with some $n\in (-d-1,-d/2+1/2]$ in $d$-dimensional cases $(d\ge 2)$ as well.
This paper is organized as follows. In Section \ref{sec_models_results} we define the model Hamiltonians, prepare notations and state the main results of this paper. In Section \ref{sec_formulation} the correlation functions are formulated into a limit of the finite dimensional Grasssmann integrals. In Section \ref{sec_perturbative_bound} upper bounds on each term of the Taylor series of the Grassmann integral formulations are obtained. In Section \ref{sec_exponential_decay_correlation} we first prove that the covariance matrix fulfills the necessary requirements for the Grassmann integral formulation to be bounded perturbatively. By using the perturbative bounds on the Grassmann integral formulations, we then complete the proof of our main results on the exponential decay property of the correlation functions. Appendix \ref{app_anti_symmetricity} shows that the coefficient function defining the interaction term can be replaced by a unique anti-symmetric function. Appendix \ref{app_thermodynamic_limit} presents a proof of existence of the thermodynamic limit of the correlation functions.
\section{Model Hamiltonians and main results}\label{sec_models_results}
In this section we define Hamiltonian operators and correlation functions together with notations used in this paper and state the main results. We use the standard terminology concerning the Fermionic Fock space without providing the definitions. They are documented, e.g, in the book \cite{BR} or briefly in \cite[\mbox{Appendix A}]{K}.
\subsection{The Hamiltonian operator}
We are going to define our Hamiltonian operator on the Fermionic Fock space on the $d$-dimensional hyper-cubic lattice $\G=({\mathbb{Z}}}% \Z == \mathbb{Z/(L{\mathbb{Z}}}% \Z == \mathbb{Z))^d$ $(L,d\in{\mathbb{N}}}% \N == \mathbb{N)$ and the spin coordinate ${\{\uparrow,\downarrow\}}$. The lattice $\G^*$ of momentum is given by $\G^*:=((2\pi{\mathbb{Z}}}% \Z == \mathbb{Z/L)/(2\pi{\mathbb{Z}}}% \Z == \mathbb{Z))^d$. We admit conventions that Kronecker's delta $\delta_{{\mathbf{x}},{\mathbf{y}}}$ takes $1$ if the element ${\mathbf{x}}$ is identical with ${\mathbf{y}}$ in the set they belong to, $0$ otherwise and the function $1_P$ of a proposition $P$ returns $1$ if $P$ is true, $0$ otherwise. For any vectors ${\mathbf{a}}=(a_1,\cdots,a_n)$, ${\mathbf{b}}=(b_1,\cdots,b_n)$ of algebra, let $\<{\mathbf{a}},{\mathbf{b}}\>$ denote the sum $\sum_{j=1}^na_jb_j$. Let $\|\cdot\|_{{\mathbb{R}}}% \R == \mathbb{R^n}$ be the Euclidean norm in ${\mathbb{R}}}% \R == \mathbb{R^n$, $\<\cdot,\cdot\>_{{\mathbb{C}}}% \C == \mathbb{C^n}$ be the inner product of ${\mathbb{C}}}% \C == \mathbb{C^n$ and $\|\cdot\|_{{\mathbb{C}}}% \C == \mathbb{C^n}$ denote the norm of ${\mathbb{C}}}% \C == \mathbb{C^n$ induced by $\<\cdot,\cdot\>_{{\mathbb{C}}}% \C == \mathbb{C^n}$. For any finite set $S$ let $\sharp S$ denote the number of elements of $S$.
With the creation operator $\psi^*_{{\mathbf{x}}\xi}$ and the annihilation operator $\psi_{{\mathbf{x}}\xi}$ for any ${\mathbf{x}}\in\G$ and $\xi\in{\{\uparrow,\downarrow\}}$ the free part $H_0$ is defined by
$$H_0:= \sum_{{\mathbf{x}},{\mathbf{y}}\in \G}\sum_{\xi,\phi\in{\{\uparrow,\downarrow\}}}T({\mathbf{x}}\xi,{\mathbf{y}}\phi)
\psi_{{\mathbf{x}}\xi}^*\psi_{{\mathbf{y}}\phi},$$
where the short range hopping matrix $\{T({\mathbf{x}}\xi,{\mathbf{y}}\phi)\}_{({\mathbf{x}},\xi),({\mathbf{y}},\phi)\in \G\times{\{\uparrow,\downarrow\}}}$ is given by
\begin{equation*}
\begin{split}
T&({\mathbf{x}}\xi,{\mathbf{y}}\phi):=\delta_{\xi,\phi}\Bigg(-t\sum_{j=1}^d(\delta_{{\mathbf{x}},{\mathbf{y}}-{\mathbf{e}}_j}+\delta_{{\mathbf{x}},{\mathbf{y}}+{\mathbf{e}}_j})\\
&-t'\cdot 1_{d\ge 2}\sum_{j,k=1 \atop j<k}^d(\delta_{{\mathbf{x}},{\mathbf{y}}-{\mathbf{e}}_j-{\mathbf{e}}_k}+\delta_{{\mathbf{x}},{\mathbf{y}}-{\mathbf{e}}_j+{\mathbf{e}}_k}+\delta_{{\mathbf{x}},{\mathbf{y}}+{\mathbf{e}}_j-{\mathbf{e}}_k}+\delta_{{\mathbf{x}},{\mathbf{y}}+{\mathbf{e}}_j+{\mathbf{e}}_k})-\mu \delta_{{\mathbf{x}},{\mathbf{y}}}\Bigg).
\end{split}
\end{equation*}
The real parameters $t,t'$ and $\mu$ are called the nearest neighbor hopping amplitude, the next to nearest neighbor hopping amplitude, and the chemical potential, respectively. The vectors ${\mathbf{e}}_j\in\G$ $(j\in\{1,\cdots,d\})$ are defined by ${\mathbf{e}}_j(l):=\delta_{j,l}$ for all $j,l\in \{1,\cdots,d\}$.
Since our problem becomes trivial otherwise (see Remark \ref{rem_H0_trivialcase}), we assume that
\begin{equation}\label{eq_nonzero_hopping}
|t|+|t'|1_{d\ge 2}\neq 0
\end{equation}
throughout the paper.
To define the interacting part of the Hamiltonian we introduce functions $U_l:({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l\times {\{\uparrow,\downarrow\}}^{2l}\to {\mathbb{C}}}% \C == \mathbb{C$ $(l\in\{1,\cdots,{\tilde{n}}\})$ satisfying the equality
\begin{equation}\label{eq_condition_U}
\begin{split}
&\overline{U_l(({\mathbf{x}}_{1},\cdots,{\mathbf{x}}_{l}),(\xi_{1},\cdots,\xi_{l}),(\phi_{1},\cdots,\phi_{l}))}\\
&\quad =U_l(({\mathbf{x}}_{1},\cdots,{\mathbf{x}}_{l}),(\phi_{1},\cdots,\phi_{l}),(\xi_{1},\cdots,\xi_{l}))\ (\forall l\ge 1)
\end{split}
\end{equation}
and the translation invariance
\begin{equation}\label{eq_translation_invariance}
\begin{split}
&U_l(({\mathbf{x}}_1,{\mathbf{x}}_2,\cdots,{\mathbf{x}}_l),(\xi_1,\xi_2,\cdots,\xi_l),(\phi_1,\phi_2,\cdots,\phi_l))\\
&\quad = U_l(({\mathbf{x}}_1+{\mathbf{y}},{\mathbf{x}}_2+{\mathbf{y}},\cdots,{\mathbf{x}}_l+{\mathbf{y}}),(\xi_1,\xi_2,\cdots,\xi_l),(\phi_1,\phi_2,\cdots,\phi_l))\ (\forall l\ge 2)
\end{split}
\end{equation}
for all $({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l)\in ({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l$, $(\xi_{1},\cdots,\xi_{l}),(\phi_{1},\cdots,\phi_{l})\in {\{\uparrow,\downarrow\}}^l$ and ${\mathbf{y}}\in {\mathbb{Z}}}% \Z == \mathbb{Z^d$.
We define a restriction of $U_l$ by periodicity in the following way. Let $\lfloor L/2 \rfloor$ denote the largest integer not exceeding $L/2$. For any ${\mathbf{x}}\in{\mathbb{Z}}}% \Z == \mathbb{Z^d$ there uniquely exists ${\mathbf{x}}^L\in (\{-\lfloor L/2 \rfloor,-\lfloor L/2 \rfloor+1,\cdots,-\lfloor L/2 \rfloor+L-1\})^d$ such that ${\mathbf{x}}={\mathbf{x}}^L$ in $\G$. By using this identification we define $U_{L,l}:({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l\times {\{\uparrow,\downarrow\}}^{2l}\to {\mathbb{C}}}% \C == \mathbb{C$ $(l\in\{1,\cdots,{\tilde{n}}\})$ by
\begin{equation}\label{eq_reduction_periodicity}
\begin{split}
&U_{L,1}({\mathbf{x}},\xi,\phi):=U_{1}({\mathbf{x}}^L,\xi,\phi),\\
&U_{L,l}(({\mathbf{x}}_1,{\mathbf{x}}_2,\cdots,{\mathbf{x}}_{l-1},{\mathbf{x}}_l),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))\\
&\quad:=U_{l}((({\mathbf{x}}_1-{\mathbf{x}}_l)^L,({\mathbf{x}}_2-{\mathbf{x}}_l)^L,\cdots,({\mathbf{x}}_{l-1}-{\mathbf{x}}_l)^L,{\mathbf{0}}),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))
\end{split}
\end{equation}
for $l\in\{2,\cdots,{\tilde{n}}\}$. Note that $U_{L,l}$ is periodic with respect to the spacial variables and obeys \eqref{eq_condition_U}-\eqref{eq_translation_invariance} and
\begin{equation}\label{eq_limit_equivalence}
\begin{split}
&\lim_{L\to +\infty\atop L\in{\mathbb{N}}}% \N == \mathbb{N}U_{L,l}(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))\\
&\quad = U_{l}(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))\end{split}
\end{equation}
for all $({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l)\in {\mathbb{Z}}}% \Z == \mathbb{Z^l$, $(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l)\in\G^l$ $(\forall l\in\{1,\cdots,{\tilde{n}}\})$.
The interacting part $V$ is defined as follows.
\begin{equation}\label{eq_definition_V}
\begin{split}
V:=\sum_{l=1}^{{\tilde{n}}}\sum_{{\mathbf{x}}_{j}\in\G\atop\forall j\in\{1,\cdots,l\}}&\sum_{\xi_{j},\phi_{j}\in{\{\uparrow,\downarrow\}} \atop \forall j\in\{1,\cdots,l\}}U_{L,l}(({\mathbf{x}}_{1},\cdots,{\mathbf{x}}_{l}),(\xi_{1},\cdots,\xi_{l}),(\phi_{1},\cdots,\phi_{l}))\\
&\cdot\psi_{{\mathbf{x}}_{1}\xi_{1}}^*\psi_{{\mathbf{x}}_{2}\xi_{2}}^*\cdots\psi_{{\mathbf{x}}_{l}\xi_{l}}^*\psi_{{\mathbf{x}}_{l}\phi_{l}}\psi_{{\mathbf{x}}_{l-1}\phi_{l-1}}\cdots\psi_{{\mathbf{x}}_{1}\phi_{1}}.
\end{split}
\end{equation}
Note that the condition \eqref{eq_condition_U} makes $V$ self-adjoint. The following examples motivate us to generalize the interacting part of the Hamiltonian as defined above.
\begin{example}[the density-density interaction]
With real functions $U^{dd}_l:({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l\times{\{\uparrow,\downarrow\}}^l\to {\mathbb{R}}}% \R == \mathbb{R\ (l\in\{1,\cdots,n\})$ satisfying the translation invariance
$$U^{dd}_l(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l),(\xi_1,\cdots,\xi_l))=U^{dd}_l(({\mathbf{x}}_1+{\mathbf{y}},\cdots,{\mathbf{x}}_l+{\mathbf{y}}),(\xi_1,\cdots,\xi_l))$$
and $U^{dd}_l(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l),(\xi_1,\cdots,\xi_l))=0$ if $({\mathbf{x}}_j,\xi_j)= ({\mathbf{x}}_k,\xi_k)$ with $j\neq k$ $(\forall l\ge 2)$,
let us define the density-density interaction $V_{dd}$ by
\begin{equation*}
V_{dd}:=\sum_{l=1}^n\sum_{{\mathbf{x}}_{j}\in\G\atop \forall j\in\{1,\cdots,l\}}\sum_{\xi_{j}\in{\{\uparrow,\downarrow\}}\atop \forall j\in\{1,\cdots,l\}}U_{L,l}^{dd}(({\mathbf{x}}_{1},\cdots,{\mathbf{x}}_{l}),(\xi_{1},\cdots,\xi_{l}))\prod_{j=1}^l\psi_{{\mathbf{x}}_{j}\xi_{j}}^*\psi_{{\mathbf{x}}_{j}\xi_{j}},
\end{equation*}
where the function $U_{L,l}^{dd}$ is derived from $U_{l}^{dd}$ in the same way as in \eqref{eq_reduction_periodicity}. The operator $V_{dd}$ can be rewritten as
\begin{equation*}
\begin{split}
V_{dd}=&\sum_{l=1}^n\sum_{{\mathbf{x}}_{j}\in\G\atop \forall j\in\{1,\cdots,l\}}\sum_{\xi_{j},\phi_{j}\in{\{\uparrow,\downarrow\}}\atop \forall j\in\{1,\cdots,l\}}U_{L,l}^{dd}(({\mathbf{x}}_{1},\cdots,{\mathbf{x}}_{l}),(\xi_{1},\cdots,\xi_{l}))\prod_{j=1}^l\delta_{\xi_{j},\phi_{j}}\\
&\cdot\psi_{{\mathbf{x}}_{1}\xi_{1}}^*\psi_{{\mathbf{x}}_{2}\xi_{2}}^*\cdots\psi_{{\mathbf{x}}_{l}\xi_{l}}^*\psi_{{\mathbf{x}}_{l}\phi_{l}}\psi_{{\mathbf{x}}_{l-1}\phi_{l-1}}\cdots\psi_{{\mathbf{x}}_{1}\phi_{1}},
\end{split}
\end{equation*}
which shows that $V_{dd}$ has the form \eqref{eq_definition_V}.
\end{example}
\begin{example}[the spin operator coupled with a local magnetic field]\label{ex_spin_field}
Introduce a local magnetic field ${\mathbf{B}}_{{\mathbf{x}}}=(B_{{\mathbf{x}}}^{(1)},B_{{\mathbf{x}}}^{(2)},B_{{\mathbf{x}}}^{(3)}):{\mathbb{Z}}}% \Z == \mathbb{Z^d\to {\mathbb{R}}}% \R == \mathbb{R^3$ and let ${\mathbf{B}}_{L,{\mathbf{x}}}=$\\
$(B_{L,{\mathbf{x}}}^{(1)},B_{L,{\mathbf{x}}}^{(2)},B_{L,{\mathbf{x}}}^{(3)}):\G\to {\mathbb{R}}}% \R == \mathbb{R^3$ be the restriction of ${\mathbf{B}}_{{\mathbf{x}}}$ on $\G$ by periodicity as defined in \eqref{eq_reduction_periodicity}. With the spin operator ${\mathbf{S}}_{{\mathbf{x}}}=(S_{{\mathbf{x}}}^{(1)},S_{{\mathbf{x}}}^{(2)},S_{{\mathbf{x}}}^{(3)})$ $({\mathbf{x}}\in\G)$ given by
$$S_{{\mathbf{x}}}^{(l)}:=\frac{1}{2}\sum_{\xi,\phi\in\{{\uparrow},{\downarrow}\}}P_{\xi,\phi}^{(l)}\psi_{{\mathbf{x}}\xi}^*\psi_{{\mathbf{x}}\phi}\ (l\in\{1,2,3\})$$
with the Pauli matrices
$$P^{(1)}=\left(\begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array}\right),\ P^{(2)}=\left(\begin{array}{cc} 0 & -i \\ i & 0 \end{array}\right),\ P^{(3)}=\left(\begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array}\right),$$
we define a self-adjoint operator $V_s$ by
$$V_s:=\sum_{{\mathbf{x}}\in\G}\<{\mathbf{B}}_{L,{\mathbf{x}}},{\mathbf{S}}_{{\mathbf{x}}}\>.$$
Since
$$V_s=\sum_{{\mathbf{x}}\in\G}\sum_{\xi,\phi\in{\{\uparrow,\downarrow\}}}\left(\frac{1}{2}\sum_{l=1}^3B_{L,{\mathbf{x}}}^{(l)}P_{\xi,\phi}^{(l)}\right)\psi_{{\mathbf{x}}\xi}^*\psi_{{\mathbf{x}}\phi}$$
and $\overline{P_{\xi,\phi}^{(l)}}=P_{\phi,\xi}^{(l)}$ $(\forall l\in\{1,2,3\},\forall \xi,\phi\in{\{\uparrow,\downarrow\}})$, the operator $V_s$ provides one example of the interactions of the form \eqref{eq_definition_V}.
\end{example}
\begin{example}[the spin-spin interaction]
With the spin operator ${\mathbf{S}}_{{\mathbf{x}}}$ introduced in Example \ref{ex_spin_field} and a real function $w({\mathbf{x}}):{\mathbb{Z}}}% \Z == \mathbb{Z^d\to{\mathbb{R}}}% \R == \mathbb{R$, we define the spin-spin interaction $V_{ss}$ by
$$V_{ss}:=\sum_{{\mathbf{x}},{\mathbf{y}}\in\G}w_L({\mathbf{x}}-{\mathbf{y}})\<{\mathbf{S}}_{{\mathbf{x}}},{\mathbf{S}}_{{\mathbf{y}}}\>,$$
where the coefficient $w_L$ is the restriction of $w$ on $\G$ by periodicity.
A calculation shows that
\begin{equation*}
\begin{split}
&V_{ss}=\sum_{{\mathbf{x}}\in\G}\sum_{\xi,\phi\in{\{\uparrow,\downarrow\}}}\left(\frac{w_L({\mathbf{0}})}{4}\sum_{l=1}^3\sum_{\tau\in{\{\uparrow,\downarrow\}}}P_{\xi,\tau}^{(l)}P_{\tau,\phi}^{(l)}\right)\psi_{{\mathbf{x}}\xi}^*\psi_{{\mathbf{x}}\phi}\\
&\quad+\sum_{{\mathbf{x}}_1,{\mathbf{x}}_2\in\G}\sum_{\xi_1,\xi_2,\phi_1,\phi_2\in{\{\uparrow,\downarrow\}}}\left(\frac{w_L({\mathbf{x}}_1-{\mathbf{x}}_2)}{4}\sum_{l=1}^3P_{\xi_1,\phi_1}^{(l)}P_{\xi_2,\phi_2}^{(l)}\right)\psi_{{\mathbf{x}}_1\xi_1}^*\psi_{{\mathbf{x}}_2\xi_2}^*\psi_{{\mathbf{x}}_2\phi_2}\psi_{{\mathbf{x}}_1\phi_1}.\end{split}
\end{equation*}
Hence $V_{ss}$ can be written in the form \eqref{eq_definition_V}. Consequently, our $V$ covers the interaction of the form $V_{dd}+V_s+V_{ss}$.
\end{example}
In this paper we treat the Hamiltonian operator $H=H_0+V$.
\subsection{Main results}
We employ norms $\|\cdot\|_{L,l}$ and $\|\cdot\|_l$ to measure the magnitude of the interaction.
\begin{equation*}
\begin{split}
&\|U_{L,1}\|_{L,1}:=\max_{{\mathbf{x}}\in\G}\max_{\xi\in{\{\uparrow,\downarrow\}}}\sum_{\phi\in{\{\uparrow,\downarrow\}}}|U_{L,1}({\mathbf{x}},\xi,\phi)|,\\
&\|U_{L,l}\|_{L,l}:=\max_{j\in\{1,\cdots,l\}}\max_{\xi_j\in{\{\uparrow,\downarrow\}}}\sum_{{\mathbf{x}}_k\in\G\atop \forall k\in\{1,\cdots,l-1\}}\sum_{\xi_k\in{\{\uparrow,\downarrow\}}\atop \forall k\in\{1,\cdots,l\}\backslash \{j\}}\sum_{\phi_k\in{\{\uparrow,\downarrow\}}\atop \forall k\in\{1,\cdots,l\}}\\
&\qquad\qquad\qquad\cdot|U_{L,l}(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_{l-1},{\mathbf{0}}),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))|\ (\forall l\ge 2),\\
&\|U_{1}\|_1:=\sup_{{\mathbf{x}}\in{\mathbb{Z}}}% \Z == \mathbb{Z^d}\max_{\xi\in{\{\uparrow,\downarrow\}}}\sum_{\phi\in{\{\uparrow,\downarrow\}}}|U_1({\mathbf{x}},\xi,\phi)|,\\
&\|U_{l}\|_l:=\max_{j\in\{1,\cdots,l\}}\max_{\xi_j\in{\{\uparrow,\downarrow\}}}\sum_{{\mathbf{x}}_k\in{\mathbb{Z}}}% \Z == \mathbb{Z^d\atop \forall k\in\{1,\cdots,l-1\}}\sum_{\xi_k\in{\{\uparrow,\downarrow\}}\atop \forall k\in\{1,\cdots,l\}\backslash \{j\}}\sum_{\phi_k\in{\{\uparrow,\downarrow\}}\atop \forall k\in\{1,\cdots,l\}}\\
&\quad\qquad\qquad\cdot|U_{l}(({\mathbf{x}}_1,\cdots,{\mathbf{x}}_{l-1},{\mathbf{0}}),(\xi_1,\cdots,\xi_l),(\phi_1,\cdots,\phi_l))|\ (\forall l\ge 2).
\end{split}
\end{equation*}
For any operator ${\mathcal{O}}$ we define the thermal average $\<{\mathcal{O}}\>_L$ by
$$\<{\mathcal{O}}\>_L:=\frac{\mathop{\mathrm{Tr}}(e^{-\beta H}{\mathcal{O}})}{\mathop{\mathrm{Tr}} e^{-\beta H}},$$
where the trace operation is taken over the Fermionic Fock space on $\G\times {\{\uparrow,\downarrow\}}$ and the positive constant $\beta$ is proportional to the inverse of temperature.
Define the function $F_{t,t',d}:{\mathbb{R}}}% \R == \mathbb{R\to{\mathbb{R}}}% \R == \mathbb{R_{>0}$ by
\begin{equation}\label{eq_definition_keyfunction}
F_{t,t',d}(r):=\frac{r}{2(|t|+2(d-1)|t'|)}+\sqrt{\frac{r^2}{4(|t|+2(d-1)|t'|)^2}+1}.
\end{equation}
For ${\mathbf{x}}\in{\mathbb{Z}}}% \Z == \mathbb{Z^d$ let us define $\psi_{{\mathbf{x}}\xi}^{(*)}$ by considering ${\mathbf{x}}$ as a site of $\G$ by periodicity.
The main results of this paper are stated as follows.
\begin{theorem}\label{thm_exponential_decay}
Assume that there exists $R\in (0,1)$ such that
\begin{equation}\label{eq_assumption_U_convergence}
\sum_{l=1}^{{\tilde{n}}}l 16^{l}\|U_l\|_{l}<\beta^{-1} \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{-d}R.
\end{equation}
For any ${\hat{m}}\in{\mathbb{N}}}% \N == \mathbb{N$ and any ${\hat{\mathbf{x}}}_j,{\hat{\mathbf{y}}}_j\in{\mathbb{Z}}}% \Z == \mathbb{Z^d$, $\hat{\xi}_j,{\hat{\phi}}_j\in{\{\uparrow,\downarrow\}}$ $(\forall j\in\{1,\cdots,{\hat{m}}\})$ the thermodynamic limit
\begin{equation}\label{eq_thermodynamic_limit}
\begin{split}
\lim_{L\to +\infty\atop L\in{\mathbb{N}}}% \N == \mathbb{N}&\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\psi_{{\hat{\mathbf{x}}}_2\hat{\xi}_2}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}-1}{\hat{\phi}}_{{\hat{m}}-1}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}\\
&\quad+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\psi_{{\hat{\mathbf{y}}}_{2}{\hat{\phi}}_{2}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}-1}\hat{\xi}_{{\hat{m}}-1}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L
\end{split}
\end{equation}
exists and satisfies the following inequality.
\begin{equation}\label{eq_main_exponential_decay}
\begin{split}
\lim_{L\to +\infty\atop L\in{\mathbb{N}}}% \N == \mathbb{N}&\Big|\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\psi_{{\hat{\mathbf{x}}}_2\hat{\xi}_2}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}-1}{\hat{\phi}}_{{\hat{m}}-1}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}\\
&\quad+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\psi_{{\hat{\mathbf{y}}}_{2}{\hat{\phi}}_{2}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}-1}\hat{\xi}_{{\hat{m}}-1}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L\Big|\\
&\le (4^{{\hat{m}}+1}-{\hat{m}} 4^{2{\hat{m}}+1}\log(1-R))\cdot F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\|\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j\|_{{\mathbb{R}}}% \R == \mathbb{R^d}}.
\end{split}
\end{equation}
\end{theorem}
In the case that the interaction $V$ is the on-site interaction \\
$U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}$ $(U\in{\mathbb{R}}}% \R == \mathbb{R)$,
the 4 point correlation functions can be bounded more sharply as follows.
\begin{theorem}\label{thm_exponential_decay_hubbard_model}
Assume that the Hamiltonian operator $H$ is given by
$$H=H_0+U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}$$
with the coupling constant $U\in{\mathbb{R}}}% \R == \mathbb{R$ satisfying
\begin{equation}\label{eq_assumption_U_convergence_hubbard_model}
|U|\le (108\beta)^{-1} \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{-d}.
\end{equation}
For any ${\hat{\mathbf{x}}}_1,{\hat{\mathbf{x}}}_2,{\hat{\mathbf{y}}}_1,{\hat{\mathbf{y}}}_2\in{\mathbb{Z}}}% \Z == \mathbb{Z^d$ the thermodynamic limit
\begin{equation}\label{eq_thermodynamic_limit_hubbard}
\lim_{L\to +\infty\atop L\in{\mathbb{N}}}% \N == \mathbb{N}\<\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}^*\psi_{{\hat{\mathbf{x}}}_2{\downarrow}}^*\psi_{{\hat{\mathbf{y}}}_2{\downarrow}}\psi_{{\hat{\mathbf{y}}}_1{\uparrow}}+\psi_{{\hat{\mathbf{y}}}_1{\uparrow}}^*\psi_{{\hat{\mathbf{y}}}_2{\downarrow}}^*\psi_{{\hat{\mathbf{x}}}_2{\downarrow}}\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}\>_L
\end{equation}
exists and satisfies
\begin{equation*}
\begin{split}
\lim_{L\to +\infty\atop L\in{\mathbb{N}}}% \N == \mathbb{N}&|\<\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}^*\psi_{{\hat{\mathbf{x}}}_2{\downarrow}}^*\psi_{{\hat{\mathbf{y}}}_2{\downarrow}}\psi_{{\hat{\mathbf{y}}}_1{\uparrow}}+\psi_{{\hat{\mathbf{y}}}_1{\uparrow}}^*\psi_{{\hat{\mathbf{y}}}_2{\downarrow}}^*\psi_{{\hat{\mathbf{x}}}_2{\downarrow}}\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}\>_L|\\
&\quad \le 324\cdot F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\|{\hat{\mathbf{x}}}_1+{\hat{\mathbf{x}}}_2-{\hat{\mathbf{y}}}_1-{\hat{\mathbf{y}}}_2\|_{{\mathbb{R}}}% \R == \mathbb{R^d}}.
\end{split}
\end{equation*}
\end{theorem}
Theorem \ref{thm_exponential_decay} and Theorem \ref{thm_exponential_decay_hubbard_model} will be proved in Section \ref{sec_exponential_decay_correlation}.
\begin{remark}
Theorem \ref{thm_exponential_decay_hubbard_model} does not follow Theorem \ref{thm_exponential_decay}. To see this, we write
\begin{equation*}
\begin{split}
&U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}\\
&\quad=\sum_{{\mathbf{x}}_j,{\mathbf{y}}_j\in\G\atop \forall j\in\{1,2\}}\sum_{\xi_j,\phi_j\in{\{\uparrow,\downarrow\}}\atop \forall j\in\{1,2\}}f_{c}(({\mathbf{x}}_1,\xi_1),({\mathbf{x}}_2,\xi_2),({\mathbf{y}}_1,\phi_1),({\mathbf{y}}_2,\phi_2))\psi_{{\mathbf{x}}_1\xi_1}^*\psi_{{\mathbf{x}}_2\xi_2}^*\psi_{{\mathbf{y}}_1\phi_1}\psi_{{\mathbf{y}}_2\phi_2}
\end{split}
\end{equation*}
with
\begin{equation*}
\begin{split}
&f_c(({\mathbf{x}}_1,\xi_1),({\mathbf{x}}_2,\xi_2),({\mathbf{y}}_1,\phi_1),({\mathbf{y}}_2,\phi_2))\\
&\quad:=-\frac{U}{4}\delta_{{\mathbf{x}}_1,{\mathbf{x}}_2}\delta_{{\mathbf{y}}_1,{\mathbf{y}}_2}\delta_{{\mathbf{x}}_1,{\mathbf{y}}_1}(\delta_{\xi_1,{\uparrow}}\delta_{\xi_2,{\downarrow}}-\delta_{\xi_1,{\downarrow}}\delta_{\xi_2,{\uparrow}})(\delta_{\phi_1,{\uparrow}}\delta_{\phi_2,{\downarrow}}-\delta_{\phi_1,{\downarrow}}\delta_{\phi_2,{\uparrow}}).
\end{split}
\end{equation*}
The function $f_c $ satisfies the anti-symmetricity \eqref{eq_anti_symmetricity} and
\begin{equation*}
\begin{split}
\max\Bigg\{&\max_{({\mathbf{x}}_1,\xi_1)\in\G\times{\{\uparrow,\downarrow\}}}\sum_{({\mathbf{x}}_2,\xi_2),({\mathbf{y}}_j,\phi_j)\in\G\times{\{\uparrow,\downarrow\}}\atop \forall j\in\{1,2\}}|f_{c}(({\mathbf{x}}_1,\xi_1),({\mathbf{x}}_2,\xi_2),({\mathbf{y}}_1,\phi_1),({\mathbf{y}}_2,\phi_2))|,\\
&\max_{({\mathbf{y}}_1,\phi_1)\in\G\times{\{\uparrow,\downarrow\}}}\sum_{({\mathbf{y}}_2,\phi_2),({\mathbf{x}}_j,\xi_j)\in\G\times{\{\uparrow,\downarrow\}}\atop \forall j\in\{1,2\}}|f_{c}(({\mathbf{x}}_1,\xi_1),({\mathbf{x}}_2,\xi_2),({\mathbf{y}}_1,\phi_1),({\mathbf{y}}_2,\phi_2))|\Bigg\}\\
&=\frac{|U|}{2}.
\end{split}
\end{equation*}
Thus, Lemma \ref{lem_anti_symmetrization} proved in Appendix \ref{app_anti_symmetricity} ensures that if a function $U_c:({\mathbb{Z}}}% \Z == \mathbb{Z^d)^2\times {\{\uparrow,\downarrow\}}^4\to {\mathbb{C}}}% \C == \mathbb{C$ satisfies \eqref{eq_condition_U}-\eqref{eq_translation_invariance} and its restriction $U_{c,L}:\G^2\times {\{\uparrow,\downarrow\}}^4\to {\mathbb{C}}}% \C == \mathbb{C$ by periodicity obeys
\begin{equation}\label{eq_interaction_hubbard_rewritten}
\begin{split}
&U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}\\
&\quad=\sum_{{\mathbf{x}}_1,{\mathbf{x}}_2\in\G}\sum_{\xi_1,\xi_2,\phi_1,\phi_2\in{\{\uparrow,\downarrow\}}}U_{c,L}(({\mathbf{x}}_1,{\mathbf{x}}_2),(\xi_1,\xi_2),(\phi_1,\phi_2))\psi_{{\mathbf{x}}_1\xi_1}^*\psi_{{\mathbf{x}}_2\xi_2}^*\psi_{{\mathbf{x}}_2\phi_2}\psi_{{\mathbf{x}}_1\phi_1},
\end{split}
\end{equation}
the inequalities
\begin{equation}\label{eq_norm_inequality}
\frac{|U|}{2}\le\|U_{c,L}\|_{L,2}\le \|U_{c}\|_{2}
\end{equation}
must hold.
Let us apply Theorem \ref{thm_exponential_decay} to the interaction \eqref{eq_interaction_hubbard_rewritten}. The inequalities \eqref{eq_assumption_U_convergence} and \eqref{eq_norm_inequality} imply
$$
|U|<(256\beta)^{-1} \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{-d}R,
$$
which is a stricter constraint than \eqref{eq_assumption_U_convergence_hubbard_model}. Moreover, for $R$ close to $1$, the coefficient $4^3-2\cdot4^5\log(1-R)$ is larger than $324$.
\end{remark}
\begin{remark}\label{rem_beta_dependency}One can prove that for any $b>0$ there exist constants $C_{t,t',d,b}>0$ depending only on $t,t',d, b$ and $C_{t,t',d}'>0$ depending only on $t,t',d$ such that
$$C_{t,t',d,b}\beta^{-d-1}\le \beta^{-1} \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{-d}\le C_{t,t',d}'\beta^{-d-1}$$
for any $\beta \ge b$. Hence, the interaction needs to be small to claim the decay bounds in low temperatures.
\end{remark}
\begin{remark}
To generalize the results to many-Fermion systems with finite coordinates of colors is straightforward. We present the results only for the spins ${\{\uparrow,\downarrow\}}$ in order to refer to proved materials for many-electron systems summarized in \cite{K}.
\end{remark}
\begin{remark}
We use the translation invariance \eqref{eq_translation_invariance} to prove the existence of the thermodynamic limit \eqref{eq_thermodynamic_limit} in Lemma \ref{lem_thermodynamic_limit}. Without assuming \eqref{eq_translation_invariance} we can also prove the inequality of the form \eqref{eq_main_exponential_decay} with $\limsup_{L\to +\infty, L\in{\mathbb{N}}}% \N == \mathbb{N}$ in the left side instead of $\lim_{L\to +\infty,L\in{\mathbb{N}}}% \N == \mathbb{N}$ under an appropriate modification of the norm of $U_l$.
\end{remark}
\begin{remark}\label{rem_H0_trivialcase}
If $|t|+|t'|1_{d\ge 2}=0$, the correlation functions decay trivially. To prove this, let us take any real parameters $\{\theta_{{\mathbf{x}}}\}_{{\mathbf{x}}\in\G}$ and define the unitary operator $A_{\theta}$ by
$$A_{\theta}:=\prod_{({\mathbf{x}},\xi)\in\G\times{\{\uparrow,\downarrow\}}}e^{i\theta_{{\mathbf{x}}}\psi_{{\mathbf{x}}\xi}^*\psi_{{\mathbf{x}}\xi}}.$$
In this case, $A_{\theta}HA_{\theta}^*=H$ and thus
\begin{equation}\label{eq_remark_unitary}
\begin{split}
&\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_{1}\hat{\xi}_{1}}\>_L\\
&=\<A_{\theta}\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}A_{\theta}^*\>_L\\
&\qquad\qquad+\<A_{\theta}^*\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_{1}\hat{\xi}_{1}}A_{\theta}\>_L\\
&= e^{i\sum_{j=1}^{{\hat{m}}}(\theta_{{\hat{\mathbf{x}}}_j}-\theta_{{\hat{\mathbf{y}}}_j})}\\
&\qquad\cdot\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_{1}\hat{\xi}_{1}}\>_L.
\end{split}
\end{equation}
If $\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j\neq\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j$ in $\G$, we can choose $\{\theta_{{\mathbf{x}}}\}_{{\mathbf{x}}\in\G}$ to satisfy $\sum_{j=1}^{{\hat{m}}}\theta_{{\hat{\mathbf{x}}}_j}-\sum_{j=1}^{{\hat{m}}}\theta_{{\hat{\mathbf{y}}}_j}\neq 0$ in ${\mathbb{R}}}% \R == \mathbb{R/(2\pi{\mathbb{Z}}}% \Z == \mathbb{Z)$ and the equality \eqref{eq_remark_unitary} implies that
\begin{equation*}
\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_{1}\hat{\xi}_{1}}\>_L=0.
\end{equation*}
\end{remark}
\section{Grassmann integral formulation of the correlation functions}\label{sec_formulation}
In this section we formulate the correlation functions as a limit of finite dimensional Grassmann integrals. To attain this goal, we follow steps. As a preliminary let us fix the way to abbreviate the notations.
\subsection{The correlation functions}\label{subsec_correlation_functions}
To simplify presentations, from now we write $X^l=({\mathbf{x}}_1,\cdots,{\mathbf{x}}_l),Y^l=({\mathbf{y}}_1,\cdots,{\mathbf{y}}_l)\in({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l$, $\Xi^l=(\xi_1,\cdots,\xi_l),\Phi^l=(\phi_1,\cdots,\phi_l)\in{\{\uparrow,\downarrow\}}^l$ $(\forall l\in{\mathbb{N}}}% \N == \mathbb{N)$. To indicate the sites and the spins on which the correlation functions are defined, we use the notation $\hat{\cdot}$ and write
\begin{equation*}
\begin{split}
&{\hat{X}}^{{\hat{m}}}=({\hat{\mathbf{x}}}_1,\cdots,{\hat{\mathbf{x}}}_{{\hat{m}}}),\ {\hat{Y}}^{{\hat{m}}}=({\hat{\mathbf{y}}}_1,\cdots,{\hat{\mathbf{y}}}_{{\hat{m}}})\in({\mathbb{Z}}}% \Z == \mathbb{Z^d)^{{\hat{m}}},\\
&\hat{\Xi}^{{\hat{m}}}=(\hat{\xi}_1,\cdots,\hat{\xi}_{{\hat{m}}}),\ {\hat{\Phi}}^{{\hat{m}}}=({\hat{\phi}}_1,\cdots,{\hat{\phi}}_{{\hat{m}}})\in{\{\uparrow,\downarrow\}}^{{\hat{m}}}.
\end{split}
\end{equation*}
We identify $X^l\in({\mathbb{Z}}}% \Z == \mathbb{Z^d)^l$ as an element of $\G^l$ by periodicity without remarking when we are considering a problem defined on $\G^l$.
To derive the correlation functions systematically, we introduce real parameters \\
$\{\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\}_{X^{{\hat{m}}},Y^{{\hat{m}}}\in\G^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}}\in{\{\uparrow,\downarrow\}}^{{\hat{m}}}}$ and define the coefficient \\
$U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)$ by
\begin{equation*}
\begin{split}
U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l):=&1_{l\le {\tilde{n}}}U_{L,l}(X^l, \Xi^l,\Phi^l)\delta_{X_l,Y_l} \\
&+ 1_{l={\hat{m}}}(\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})+\lambda(Y^{{\hat{m}}},X^{{\hat{m}}},\Phi^{{\hat{m}}},\Xi^{{\hat{m}}}))
\end{split}
\end{equation*}
for all $X^{l},Y^{l}\in\G^{l},\ \Xi^{l},\Phi^{l}\in{\{\uparrow,\downarrow\}}^{l}$, $l\in \{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}$. We see that
\begin{equation}\label{eq_condition_U_lambda}
\overline{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)}=U_{\lambda,l}(Y^l,X^l,\Phi^l,\Xi^l).
\end{equation}
Let us modify the interaction $V$ to contain the coefficients $\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\}$ and define
\begin{equation*}
\begin{split}
V_{\lambda}:=&\sum_{l=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\\
&\cdot\psi_{{\mathbf{x}}_{1}\xi_{1}}^*\cdots\psi_{{\mathbf{x}}_{l}\xi_{l}}^*\psi_{{\mathbf{y}}_{l}\phi_{l}}\cdots\psi_{{\mathbf{y}}_{1}\phi_{1}}.
\end{split}
\end{equation*}
We set $H_{\lambda}:=H_0+V_{\lambda}$, which is self-adjoint by the equality \eqref{eq_condition_U_lambda}. Also note that
\begin{equation*}
H_{\lambda}\Big|_{\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})=0\atop \forall (X^{{\hat{m}}},Y^{{\hat{m}}}, \Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\in\G^{2{\hat{m}}}\times{\{\uparrow,\downarrow\}}^{2{\hat{m}}}}=H.
\end{equation*}
\begin{lemma}\label{lem_derivative_correlation}
\begin{equation}\label{eq_derivative_correlation}
\begin{split}
&\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L\\
&\qquad=-\frac{1}{\beta}\frac{\partial}{\partial \lambda({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\log\left(\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}\right)\Bigg|_{\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})=0\atop \forall (X^{{\hat{m}}},Y^{{\hat{m}}}, \Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\in\G^{2{\hat{m}}}\times{\{\uparrow,\downarrow\}}^{2{\hat{m}}}}.
\end{split}
\end{equation}
\end{lemma}
\begin{proof} The proof is parallel to that of \cite[\mbox{Lemma 2.1}]{K}, based on \cite[\mbox{Lemma 2.3}]{K}.
\end{proof}
\subsection{The perturbation series}
In order to characterize the correlation functions as a limit of finite dimensional Grassmann integrals, we proceed in the following steps. Firstly we expand the partition function $\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0}$ into a perturbation series of the variables \\
$\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\}$. Secondly we replace the integrals over $[0,\beta)$ contained in the perturbation series by the Riemann sums to derive a fully discrete analog of the perturbation series. We then show that the discretized perturbation series converges to the original one by passing the parameter defining the Riemann sums to infinity. The discretized perturbation series is formulated into the Grassmann Gaussian integral involving only finite Grassmann variables. Combining the Grassmann integral formulation of the discretized partition function with the equality \eqref{eq_derivative_correlation} completes the characterization.
The first step results in the following proposition.
\begin{proposition}\label{prop_perturbation_series}
\begin{equation}\label{eq_perturbation_series}
\begin{split}
&\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}=1+\sum_{m=1}^{\infty}\frac{(-1)^m}{m!}\\
&\cdot\prod_{k=1}^m\left(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\int_0^{\beta}ds_k\right)\\
&\cdot\det(C((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k},
\end{split}
\end{equation}
where $X_k^{l_k}:=({\mathbf{x}}_{k,1},{\mathbf{x}}_{k,2},\cdots,{\mathbf{x}}_{k,l_k})$, $Y_k^{l_k}:=({\mathbf{y}}_{k,1},{\mathbf{y}}_{k,2},\cdots,{\mathbf{y}}_{k,l_k})$,\\
$\Xi_k^{l_k}:=(\xi_{k,1},\xi_{k,2},\cdots,\xi_{k,l_k})$, $\Phi_k^{l_k}:=(\phi_{k,1},\phi_{k,2},\cdots,\phi_{k,l_k})$, and
\begin{equation}\label{eq_covariance_argument}
(\widetilde{{\mathbf{x}}\xi s})_p:=({\mathbf{x}}_{v+1,u},\xi_{v+1,u},s_{v+1}),\ (\widetilde{{\mathbf{y}}\phi s})_p:=({\mathbf{y}}_{v+1,u},\phi_{v+1,u},s_{v+1})
\end{equation}
for $p=\sum_{k=1}^vl_k+u$, $u\in\{1,\cdots,l_{v+1}\}$, $v\in\{0,\cdots,m-1\}$.
The covariance $C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)$ is given by
\begin{equation}\label{eq_covariance}
C({\mathbf{x}}\xi x,{\mathbf{y}}\phi y):=\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}e^{-(y-x)E_{{\mathbf{k}}}}\left(\frac{1_{y-x\le
0}}{1+e^{\beta E_{{\mathbf{k}}}}}-\frac{1_{y-x>
0}}{1+e^{-\beta E_{{\mathbf{k}}}}}\right)
\end{equation}
$(\forall ({\mathbf{x}},\xi, x),({\mathbf{y}}, \phi, y)\in\G\times{\{\uparrow,\downarrow\}}\times [0,\beta))$ with the dispersion relation
\begin{equation}\label{eq_dispersion_relation}
E_{{\mathbf{k}}}:=-2t\sum_{j=1}^d\cos(\<{\mathbf{k}},{\mathbf{e}}_j\>) - 4t'\cdot 1_{d\ge 2}\sum_{j,k = 1\atop j<k}^d\cos(\<{\mathbf{k}},{\mathbf{e}}_j\>)\cos(\<{\mathbf{k}},{\mathbf{e}}_k\>)-\mu.
\end{equation}
\end{proposition}
\begin{proof}
For any operator ${\mathcal{O}}$ defined on the Fermionic Fock space, let ${\mathcal{O}}(s)$ denote $e^{sH_0}{\mathcal{O}} e^{-sH_0}$ for $s\in{\mathbb{R}}}% \R == \mathbb{R$ and $\<{\mathcal{O}}\>_{0,L}$ denote $\mathop{\mathrm{Tr}} (e^{-\beta H_0}{\mathcal{O}})/\mathop{\mathrm{Tr}} e^{-\beta H_0}$. We can apply \cite[\mbox{Lemma B.3}]{K} to derive the equality
$$e^{-\beta H_{\lambda}}=e^{-\beta H_0}+e^{-\beta H_0}\sum_{m=1}^{\infty}(-1)^m\int_{[0,\beta)^m}ds_1\cdots ds_m1_{s_1>\cdots> s_m}V_{\lambda}(s_1)\cdots V_{\lambda}(s_m),$$
which leads to
\begin{equation}\label{eq_partition_expansion}
\begin{split}
&\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}=1+\sum_{m=1}^{\infty}(-1)^m\int_{[0,\beta)^m}ds_1\cdots ds_m1_{s_1>\cdots>s_m}\\
&\qquad\cdot\prod_{k=1}^m\left(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\right)\\
&\qquad\cdot\<\psi_{{\mathbf{x}}_{1,1}\xi_{1,1}}^*(s_1)\cdots\psi_{{\mathbf{x}}_{1,l_1}\xi_{1,l_1}}^*(s_1)\psi_{{\mathbf{y}}_{1,l_1}\phi_{1,l_1}}(s_1)\cdots \psi_{{\mathbf{y}}_{1,1}\phi_{1,1}}(s_1)\\
&\qquad\quad \cdots \psi_{{\mathbf{x}}_{m,1}\xi_{m,1}}^*(s_m)\cdots\psi_{{\mathbf{x}}_{m,l_m}\xi_{m,l_m}}^*(s_m)\psi_{{\mathbf{y}}_{m,l_m}\phi_{m,l_m}}(s_m)\cdots \psi_{{\mathbf{y}}_{m,1}\phi_{m,1}}(s_m)\>_{0,L}.
\end{split}
\end{equation}
By recalling the definition \cite[\mbox{Definition B.2}]{K} of the temperature-ordering operator and its properties \cite[\mbox{Lemma B.7, Lemma B.9}]{K}, we can deduce \eqref{eq_perturbation_series} from \eqref{eq_partition_expansion}. The covariance $C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)$ and the dispersion relation $E_{{\mathbf{k}}}$ have been proved to have the forms \eqref{eq_covariance}-\eqref{eq_dispersion_relation} in \cite[\mbox{Lemma B.10}]{K}.
\end{proof}
As the second step toward the Grassmann integral formulation, we introduce the discrete analog of the expansion \eqref{eq_perturbation_series}. By taking a parameter $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$, we define the discrete sets $[0,\beta)_h$ and $[-\beta,\beta)_h$ by
$$[0,\beta)_h:=\left\{0,\frac{1}{h},\frac{2}{h},\cdots,\beta - \frac{1}{h}\right\},\ [-\beta,\beta)_h:=\left\{-\beta,-\beta+\frac{1}{h},\cdots,\beta - \frac{1}{h}\right\}.
$$
Note that $\sharp[0,\beta)_h=\beta h$, $\sharp[-\beta,\beta)_h=2\beta h$. If the temperature variables $x,y$ are confined in $[0,\beta)_h$, we write
$$C_h({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)=C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y).$$
We then define $(\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$ by
\begin{equation}\label{eq_discrete_perturbation_series}
\begin{split}
&\left(\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}\right)_h:=1+\sum_{m=1}^{2L^d\beta h}\frac{(-1)^m}{m!}\\
&\cdot\prod_{k=1}^m\left(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\frac{1}{h}\sum_{s_k\in [0,\beta)_h}\right)\\
&\qquad \cdot\det(C_h((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k},
\end{split}
\end{equation}
where the variables $(\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_p\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h$ are defined by the same rule as \eqref{eq_covariance_argument}. Since any determinant made up of the elements $C_h({\mathbf{x}}\xi x, {\mathbf{y}}\phi y)$ vanishes if the size of the matrix exceeds $2L^d\beta h (=\sharp \G\times{\{\uparrow,\downarrow\}}\times [0,\beta)_h)$, the sum with respect to $m$ is taken only up to $m=2L^d\beta h$ in \eqref{eq_discrete_perturbation_series}.
\begin{remark}
The diagonalization of the covariance matrix \\
$(C_h({\mathbf{x}} \xi x,{\mathbf{y}} \phi y))_{({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}} \times [0,\beta)_h}$ was presented in \cite[\mbox{Appendix C}]{K} for any $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$. To refer to this result we take $h$ from $2{\mathbb{N}}}% \N == \mathbb{N/\beta$.
\end{remark}
The following lemma states that the partition function $\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0}$ in Lemma \ref{lem_derivative_correlation} can be replaced by $(\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$.
\begin{lemma}\label{lem_correlation_h_limit}
For any $r>0$ there exists $N_0\in{\mathbb{N}}}% \N == \mathbb{N$ such that $\mathop{\mathrm{Re}} (\mathop{\mathrm{Tr}}
e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$\\
$>0$ for all $h\in
2{\mathbb{N}}}% \N == \mathbb{N/\beta$ with $h\ge 2N_0/\beta$ and all $\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\in {\mathbb{R}}}% \R == \mathbb{R$ with \\
$|\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})|\le r$. Moreover,
\begin{equation}\label{eq_correlation_h_limit}
\begin{split}
&\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L\\
&=-\frac{1}{\beta}\lim_{h\to +\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}\frac{\partial}{\partial \lambda({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\log\left(\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}\right)_h\Bigg|_{\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})=0\atop\forall (X^{{\hat{m}}},Y^{{\hat{m}}}, \Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\in\G^{2{\hat{m}}}\times{\{\uparrow,\downarrow\}}^{2{\hat{m}}}},
\end{split}
\end{equation}
where for $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $\mathop{\mathrm{Re}} z>0$, $\log z$ is defined by taking the principal value;
$$\log z:= \log|z| +i\mathop{\mathrm{Arg}} z,\ \mathop{\mathrm{Arg}} z\in (-\pi/2,\pi/2).$$
\end{lemma}
\begin{proof}
By considering
$\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\}_{l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}, (X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}$ as mutually independent, complex multi-variables, we define the functions \\
$P(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})$ and
$P_h(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})$ by the right hand side of
\eqref{eq_perturbation_series} and
\eqref{eq_discrete_perturbation_series}, respectively. We show that
$P_h$ converges to $P$ as $h\to +\infty$ locally uniformly with respect
to the variables.
Pedra-Salmhofer's determinant bound \cite[\mbox{Theorem 2.4}]{PS} (see also Proposition \ref{prop_extended_determinant_bound} below for an extended statement) implies that
\begin{equation}\label{eq_determinant_bound_appl}
|\det(C({\mathbf{x}}_p\xi_ps_p,{\mathbf{y}}_q\phi_qt_q))_{1\le p,q\le n}|\le 4^{n}
\end{equation}
for any $({\mathbf{x}}_p,\xi_p,s_p),({\mathbf{y}}_p,\phi_p,t_p)\in \G\times{\{\uparrow,\downarrow\}} \times
[0,\beta)$ $(\forall p\in\{1,\cdots,n\})$. Thus, if all the variables satisfy the inequality $|U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)|\le r$ for an $r>0$, we have\begin{equation}\label{eq_perturbation_bound}
\begin{split}
\Bigg|&\frac{(-1)^m}{m!}\prod_{k=1}^m\Bigg(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\Bigg)\\
&\quad\cdot\Bigg(\prod_{k=1}^m\int_0^{\beta}dt_k\det(C((\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q))_{1\le p,q\le \sum_{k=1}^ml_k}\\
&\quad\qquad-\prod_{k=1}^m\frac{1}{h}\sum_{s_k\in [0,\beta)_h}\det(C_h((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k}\Bigg)\Bigg|\\
&\le\frac{2}{m!}\prod_{k=1}^{m}\left(\beta r\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}(2L^d)^{2l_k}\right) 4^{\sum_{k=1}^ml_k},
\end{split}
\end{equation}
where $(\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q,(\widetilde{{\mathbf{x}}\xi
s})_p,(\widetilde{{\mathbf{y}}\phi s})_q$ are defined in the same way as in
\eqref{eq_covariance_argument}. The right hand side of \eqref{eq_perturbation_bound} is summable over $m\in{\mathbb{N}}}% \N == \mathbb{N$.
Fix any $(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}$ $(\forall k\in\{1,\cdots,m\})$. We define a function $g_h:[0,\beta)^m\to {\mathbb{C}}}% \C == \mathbb{C$ by
$$g_h(t_1,\cdots,t_m):=\det(C_h((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k},$$
where $s_j\in [0,\beta)_h$ satisfies $t_j\in [s_j,s_j+1/h)$ $(\forall j\in\{1,\cdots,m\})$ and
$(\widetilde{{\mathbf{x}}\xi s})_p$, $(\widetilde{{\mathbf{y}}\phi s})_q$ \\
$(p,q\in \{1,\cdots,\sum_{k=1}^ml_k\})$ are defined by the same rule as \eqref{eq_covariance_argument}.
Since $C({\mathbf{x}}\xi s,{\mathbf{y}} \phi t)$ is continuous with respect to the variables
$(s,t)\in[0,\beta)^2$ almost everywhere, so is $\det(C((\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q))_{1\le p,q\le \sum_{k=1}^ml_k}$ with respect to \\
$(t_1,\cdots,t_m)\in[0,\beta)^m$, and thus
$$\lim_{h\to+\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}g_h(t_1,\cdots,t_m)=\det(C((\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q))_{1\le p,q\le \sum_{k=1}^ml_k}$$
for a.e. $(t_1,\cdots,t_m)\in [0,\beta)^m$. Hence, by \eqref{eq_determinant_bound_appl} and the dominated convergence theorem for $L^1([0,\beta)^m)$, we have
\begin{equation}\label{eq_convergence_determinant}
\begin{split}
&\lim_{h\to +\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}\Bigg|\prod_{k=1}^m\int_0^{\beta}dt_k\det(C((\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q))_{1\le p,q\le \sum_{k=1}^ml_k}\\
&\qquad\qquad\qquad-\prod_{k=1}^m\frac{1}{h}\sum_{s_k\in [0,\beta)_h}\det(C_h((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k}\Bigg|\\
&=\lim_{h\to +\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}\Bigg|\prod_{k=1}^m\int_0^{\beta}dt_k\left(\det(C((\widetilde{{\mathbf{x}}\xi t})_p,(\widetilde{{\mathbf{y}}\phi t})_q))_{1\le p,q\le \sum_{k=1}^ml_k}-g_h(t_1,\cdots,t_m)\right)\Bigg|\\
&=0.
\end{split}
\end{equation}
By \eqref{eq_perturbation_bound} and \eqref{eq_convergence_determinant}
we can again apply the dominated convergence theorem for $l^1({\mathbb{N}}}% \N == \mathbb{N)$ to show that\begin{equation}\label{eq_convergence_partition_function}
\begin{split}
\lim_{h\to +\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}\sup_{\forall
U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\in{\mathbb{C}}}% \C == \mathbb{C\atop \text{with }|U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)|\le r}\Bigg|&P\left(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\}\right)\\
&-P_h\left(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\}\right)\Bigg|=0
\end{split}
\end{equation}
for all $r>0$.
Since the multi-variable function $P-P_h$ is entirely
analytic, the convergence property
\eqref{eq_convergence_partition_function} and Cauchy's integral formula
prove that
\begin{equation}\label{eq_convergence_derivative}
\begin{split}
\lim_{h\to +\infty\atop h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta}&\sup_{\forall
U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\in{\mathbb{C}}}% \C == \mathbb{C\atop\text{with } |U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)|\le r}\\
&\cdot\Bigg|\frac{\partial}{\partial U_{\lambda,k}(X^k,Y^k, \Xi^k,\Phi^k)}\Bigg( P\left(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\}\right)\\
&\quad-P_h\left(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\}\right)\Bigg)\Bigg|=0
\end{split}
\end{equation}
for all $r>0$ and any $k\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}$, $(X^{k},Y^{k}, \Xi^{k},\Phi^{k})\in\G^{2k}\times{\{\uparrow,\downarrow\}}^{2k}$.
Since $\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0}>0$, the uniform
convergence property \eqref{eq_convergence_partition_function} verifies the first statement of the lemma on $\mathop{\mathrm{Re}} (\mathop{\mathrm{Tr}}
e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$.
Note the equality that
\begin{equation}\label{eq_equality_differential_operator}
\begin{split}
&\frac{\partial}{\partial
\lambda({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\log\left(\frac{\mathop{\mathrm{Tr}}
e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}\right)_h\\
&\quad=\frac{\left(\frac{\partial}{\partial
U_{\lambda,{\hat{m}}}({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}+\frac{\partial}{\partial
U_{\lambda,{\hat{m}}}({\hat{Y}}^{{\hat{m}}},{\hat{X}}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}})}\right)P_h(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\})}{P_h(\{U_{\lambda,l}(X^l,Y^l, \Xi^l,\Phi^l)\})}.
\end{split}
\end{equation}
Combining the convergence properties \eqref{eq_convergence_partition_function}-\eqref{eq_convergence_derivative} and the equality \eqref{eq_equality_differential_operator} with \eqref{eq_derivative_correlation} yields the equality \eqref{eq_correlation_h_limit}.
\end{proof}
\subsection{The Grassmann integral formulation}
As the third step we formulate $(\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$ into a Grassmann Gaussian integral on the finite dimensional Grassmann algebra \\
$\{{\overline{\psi}}_{{\mathbf{x}}\xi x},\psi_{{\mathbf{x}}\xi x}\}_{({\mathbf{x}},\xi, x)\in \G\times{\{\uparrow,\downarrow\}} \times[0,\beta)_h}$. Basic properties of the finite dimensional Grassmann integral have been summarized in the books \cite{FKT}, \cite{S}. We assume that each element $({\mathbf{x}},\xi,x)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h$ is numbered so that we can write
$$\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h=\{({\mathbf{x}}\xi x)_j\ |\ j\in\{1,\cdots,N\}\}$$
with $N:=2L^d\beta h$.
The Grassmann integral $\int\cdot d\psi_{({\mathbf{x}}\xi x)_N}\cdots
d\psi_{({\mathbf{x}}\xi x)_1}d{\overline{\psi}}_{({\mathbf{x}}\xi x)_N}\cdots d{\overline{\psi}}_{({\mathbf{x}}\xi x)_1}$
is a linear functional on the complex linear space ${\mathbb{C}}}% \C == \mathbb{C[{\overline{\psi}}_{({\mathbf{x}}\xi
x)_j},\psi_{({\mathbf{x}}\xi x)_j}\ |\ j\in\{1,\cdots, N\}]$ of Grassmann monomials and satisfies that
\begin{equation}\label{eq_definition_grassmann_integral}
\begin{split}
&\int{\overline{\psi}}_{({\mathbf{x}}\xi x)_1}\cdots{\overline{\psi}}_{({\mathbf{x}}\xi x)_N}\psi_{({\mathbf{x}}\xi
x)_1}\cdots\psi_{({\mathbf{x}}\xi x)_N}\\
&\qquad \cdot d\psi_{({\mathbf{x}}\xi x)_N}\cdots d\psi_{({\mathbf{x}}\xi x)_1}d{\overline{\psi}}_{({\mathbf{x}}\xi x)_N}\cdots d{\overline{\psi}}_{({\mathbf{x}}\xi x)_1}=1,\\
&\int{\overline{\psi}}_{({\mathbf{x}}\xi x)_{j_1}}\cdots{\overline{\psi}}_{({\mathbf{x}}\xi x)_{j_k}}\psi_{({\mathbf{x}}\xi
x)_{p_1}}\cdots\psi_{({\mathbf{x}}\xi x)_{p_q}}\\
&\qquad \cdot d\psi_{({\mathbf{x}}\xi x)_N}\cdots d\psi_{({\mathbf{x}}\xi x)_1}d{\overline{\psi}}_{({\mathbf{x}}\xi
x)_N}\cdots d{\overline{\psi}}_{({\mathbf{x}}\xi x)_1}=0,
\end{split}
\end{equation}
if $k<N$ or $q< N$.
For simplicity let ${\overline{\psi}}_X$, $\psi_X$ denote the vectors of the Grassmann variables \\
$({\overline{\psi}}_{({\mathbf{x}}\xi x)_1},\cdots,{\overline{\psi}}_{({\mathbf{x}}\xi x)_N})$, $(\psi_{({\mathbf{x}}\xi x)_1},\cdots,\psi_{({\mathbf{x}}\xi x)_N})$, respectively. We also write $d{\overline{\psi}}_X=d{\overline{\psi}}_{({\mathbf{x}}\xi x)_N}\cdots d{\overline{\psi}}_{({\mathbf{x}}\xi x)_1}$ and $d\psi_X=d\psi_{({\mathbf{x}}\xi x)_N}\cdots d\psi_{({\mathbf{x}}\xi x)_1}$.
For any $f({\overline{\psi}}_X,\psi_X)\in {\mathbb{C}}}% \C == \mathbb{C[{\overline{\psi}}_{({\mathbf{x}}\xi x)_j},\psi_{({\mathbf{x}}\xi
x)_j}\ |\ j\in\{1,\cdots, N\}]$ the value of the Grassmann integral
$\int f({\overline{\psi}}_X,\psi_X)d\psi_Xd{\overline{\psi}}_X$ can be computed by linearity
and the anti-commutation relations of the Grassmann variables.
For any $N\times N$ matrix $G_h$ with $\det G_h\neq 0$, let ${\mathbf{G}}_h$ denote the $2N\times2N$ skew symmetric matrix
\begin{equation*}
\left(\begin{array}{cc} 0 & G_h \\ -G_h^t & 0 \end{array}\right).
\end{equation*}
From the definition and the assumption that $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ we see that
\begin{equation*}
\begin{split}
\int e^{-\frac{1}{2}\<({\overline{\psi}}_X,\psi_X)^t,{\mathbf{G}}_h^{-1}({\overline{\psi}}_X,\psi_X)^t\>}d\psi_Xd{\overline{\psi}}_X&=\int e^{-\<\psi_X^t,G_h^{-1}{\overline{\psi}}_X^t\>}d\psi_Xd{\overline{\psi}}_X\\
&=(-1)^{N(N-1)/2}(\det G_h)^{-1}=(\det G_h)^{-1}.
\end{split}
\end{equation*}
\begin{definition}\label{def_grassmann_gaussian_integral}
For any $N\times N$ matrix $G_h$ with $\det G_h\neq 0$ the Grassmann Gaussian integral $\int\cdot d\mu_{G_h}({\overline{\psi}}_X,\psi_X):{\mathbb{C}}}% \C == \mathbb{C[{\overline{\psi}}_{({\mathbf{x}}\xi x)_j},\psi_{({\mathbf{x}}\xi x)_j}\ |\ j\in\{1,\cdots,N\}]\to{\mathbb{C}}}% \C == \mathbb{C$ is defined by
\begin{equation*}
\int f({\overline{\psi}}_X,\psi_X)d\mu_{G_h}({\overline{\psi}}_X,\psi_X):=\frac{\int f({\overline{\psi}}_X,\psi_X)e^{-\frac{1}{2}\<({\overline{\psi}}_X,\psi_X)^t,{\mathbf{G}}_h^{-1}({\overline{\psi}}_X,\psi_X)^t\>}d\psi_Xd{\overline{\psi}}_X}{\int e^{-\frac{1}{2}\<({\overline{\psi}}_X,\psi_X)^t,{\mathbf{G}}_h^{-1}({\overline{\psi}}_X,\psi_X)^t\>}d\psi_Xd{\overline{\psi}}_X}.
\end{equation*}
\end{definition}
Transforming $(\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}/\mathop{\mathrm{Tr}} e^{-\beta H_0})_h$ into a
Grassmann Gaussian integral essentially relies on the following equality.
\begin{lemma}\label{lem_grassmann_gaussian_equality}
\begin{equation*}
\begin{split}
&\int{\overline{\psi}}_{({\mathbf{x}}\xi x)_{j_k}}\cdots {\overline{\psi}}_{({\mathbf{x}}\xi x)_{j_1}}
\psi_{({\mathbf{x}}\xi x)_{p_1}}\cdots\psi_{({\mathbf{x}}\xi
x)_{p_k}}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)\\
&\quad =\det(G_h(({\mathbf{x}}\xi x)_{j_u},({\mathbf{x}}\xi x)_{p_v}))_{1\le u,v\le k}.
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
This equality follows \cite[\mbox{Problem I.13}]{FKT} by replacing
the notations $\psi_X$, ${\overline{\psi}}_X$ by ${\overline{\psi}}_X$, $\psi_X$, respectively.
\end{proof}
\begin{remark}
Our definition of the Grassmann Gaussian integral Definition
\ref{def_grassmann_gaussian_integral} differs from that summarized in
\cite[\mbox{Problem I.13}]{FKT} or its follower \cite[\mbox{Section
3.2}]{K} and corresponds to the statement derived by changing $\psi_X$ for ${\overline{\psi}}_X$ and ${\overline{\psi}}_X$ for $\psi_X$ respectively in \cite[\mbox{Problem I.13}]{FKT} and \cite[\mbox{Section 3.2}]{K}. In \cite[\mbox{Proposition 3.7}]{K} the discretized partition function of the Hubbard model was formulated in a Grassmann Gaussian integral. The formulation \cite[\mbox{Proposition 3.7}]{K} required the symmetry assumption on the coefficients $U_{{\mathbf{x}},{\mathbf{y}},{\mathbf{z}},{\mathbf{w}}}$ $({\mathbf{x}},{\mathbf{y}},{\mathbf{z}},{\mathbf{w}}\in\G)$, namely $U_{{\mathbf{x}},{\mathbf{y}},{\mathbf{z}},{\mathbf{w}}}=U_{{\mathbf{z}},{\mathbf{w}},{\mathbf{x}},{\mathbf{y}}}$. Under Definition \ref{def_grassmann_gaussian_integral}, however, we do not need any additional assumption on the coefficients $U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)$ to complete the desired formulation below.
\end{remark}
\begin{lemma}\label{lem_grassmann_gaussian_partition}
\begin{equation}\label{eq_grassmann_gaussian_partition}
\begin{split}
&\left(\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta H_0}}\right)_h=\\
&\qquad\int
e^{\sum_{l=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)V_{h,X^l,Y^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}\\
&\qquad\qquad\cdot d\mu_{C_h}({\overline{\psi}}_X,\psi_X),
\end{split}
\end{equation}
where
\begin{equation*}
\begin{split}
&V_{h,X^l,Y^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)\\
&\quad
:=-\frac{1}{h}\sum_{x\in[0,\beta)_h}{\overline{\psi}}_{{\mathbf{x}}_{1}\xi_{1}x}{\overline{\psi}}_{{\mathbf{x}}_{2}\xi_{2}x}\cdots {\overline{\psi}}_{{\mathbf{x}}_{l}\xi_{l}x}\psi_{{\mathbf{y}}_{l}\phi_{l}x}\psi_{{\mathbf{y}}_{l-1}\phi_{l-1}x}\cdots\psi_{{\mathbf{y}}_{1}\phi_{1}x}.
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
First note that $\det C_h\neq 0$ for any $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ by
\cite[\mbox{Proposition C.7}]{K}.
By applying Lemma \ref{lem_grassmann_gaussian_equality} to $\det(C_h((\widetilde{{\mathbf{x}}\xi s})_p,(\widetilde{{\mathbf{y}}\phi s})_q))_{1\le p,q\le \sum_{k=1}^ml_k}$ in \eqref{eq_discrete_perturbation_series} and using the fact that $\int1d\mu_{C_h}({\overline{\psi}}_X,\psi_X)=1$, we have
\begin{equation*}
\begin{split}
&\left(\frac{\mathop{\mathrm{Tr}} e^{-\beta H_{\lambda}}}{\mathop{\mathrm{Tr}} e^{-\beta
H_0}}\right)_h=1+\sum_{m=1}^{2L^d\beta h}\frac{(-1)^m}{m!}\\
&\cdot\prod_{k=1}^m\left(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\frac{1}{h}\sum_{s_k\in [0,\beta)_h}\right)\\
&\qquad \cdot\int{\overline{\psi}}_{{\mathbf{x}}_{m,l_m}\xi_{m,l_m}s_m}\cdots{\overline{\psi}}_{{\mathbf{x}}_{m,1}\xi_{m,1}s_m}\cdots{\overline{\psi}}_{{\mathbf{x}}_{1,l_1}\xi_{1,l_1}s_1}\cdots{\overline{\psi}}_{{\mathbf{x}}_{1,1}\xi_{1,1}s_1}\\
&\qquad\quad\cdot\psi_{{\mathbf{y}}_{1,1}\phi_{1,1}s_1}\cdots\psi_{{\mathbf{y}}_{1,l_1}\phi_{1,l_1}s_1}\cdots\psi_{{\mathbf{y}}_{m,1}\phi_{m,1}s_m}\cdots\psi_{{\mathbf{y}}_{m,l_m}\phi_{m,l_m}s_m}d\mu_{C_h}({\overline{\psi}}_X,\psi_X)
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
&=\int\Bigg(1+\sum_{m=1}^{2L^d\beta h}\frac{1}{m!}\\
&\cdot \prod_{k=1}^m\Bigg(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{\lambda,l_k}(X_k^{l_k},Y_k^{l_k}, \Xi_k^{l_k},\Phi_k^{l_k})\frac{-1}{h}\sum_{s_k\in [0,\beta)_h}\\
&\qquad\quad\cdot{\overline{\psi}}_{{\mathbf{x}}_{k,1}\xi_{k,1}s_k}\cdots{\overline{\psi}}_{{\mathbf{x}}_{k,l_k}\xi_{k,l_k}s_k}\psi_{{\mathbf{y}}_{k,l_k}\phi_{k,l_k}s_k}\cdots\psi_{{\mathbf{y}}_{k,1}\phi_{k,1}s_k}\Bigg)\Bigg)d\mu_{C_h}({\overline{\psi}}_X,\psi_X),
\end{split}
\end{equation*}
which is \eqref{eq_grassmann_gaussian_partition}.
\end{proof}
Here we complete our Grassmann integral formulation of the correlation function.\begin{proposition}\label{prop_correlation_grassmann_integral}
\begin{equation}\label{eq_correlation_grassmann_integral}
\begin{split}
&\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L\\
&=-\frac{1}{\beta}\lim_{h\to +\infty\atop h\in
2{\mathbb{N}}}% \N == \mathbb{N/\beta}\int \left(V_{h,{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}^{{\hat{m}}}({\overline{\psi}}_X,\psi_X)+V_{h,{\hat{Y}}^{{\hat{m}}},{\hat{X}}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}, \hat{\Xi}^{{\hat{m}}}}^{{\hat{m}}}({\overline{\psi}}_X,\psi_X)\right)\\
&\quad\cdot e^{\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{C_h}({\overline{\psi}}_X,\psi_X)\\
&\quad\cdot\Bigg/\int e^{\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{C_h}({\overline{\psi}}_X,\psi_X).
\end{split}
\end{equation}
\end{proposition}
\begin{proof}
The equality \eqref{eq_correlation_grassmann_integral} is obtained by substituting
\eqref{eq_grassmann_gaussian_partition} into the right hand side of
\eqref{eq_correlation_h_limit} and differentiating the Grassmann polynomial
$$e^{\sum_{l=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)V_{h,X^l,Y^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}$$
by the variable $\lambda({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})$
inside the Grassmann integral (see \\
\cite[\mbox{Problem I.3}]{FKT} for differentiation of Grassmann polynomials).
\end{proof}
\section{Perturbative bounds on the Grassmann integral formulation}\label{sec_perturbative_bound}
In this section we find upper bounds on each term of a perturbative expansion of the Grassmann integrals of the form
\eqref{eq_correlation_grassmann_integral} for fixed $h\in
2{\mathbb{N}}}% \N == \mathbb{N/\beta$. Here we generalize the problem. The covariance is
assumed to be any matrix $G_h:(\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h)^2\to {\mathbb{C}}}% \C == \mathbb{C$ with $\det G_h\neq 0$.
Set an integral of $G_h$ by
\begin{equation*}
\begin{split}
{\mathcal{D}}:=\max\Bigg\{&\max_{({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}\frac{1}{h}\sum_{({\mathbf{x}},\xi,x)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}|G_h({\mathbf{x}}\xi x,{\mathbf{y}}\phi y)|,\\
&\max_{({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}\frac{1}{h}\sum_{({\mathbf{x}},\xi,x)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}|G_h({\mathbf{y}}\phi y,{\mathbf{x}}\xi x)|\Bigg\}.
\end{split}
\end{equation*}
We define the Schwinger function by
\begin{equation}\label{eq_schwinger_functional}
\begin{split}
&S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(G_h,\eta)\\
&:=-\frac{1}{\beta}\int V_{h,{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}^{{\hat{m}}}({\overline{\psi}}_X,\psi_X)\\
&\quad\cdot e^{\eta\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}} U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)\\
&\quad\cdot\Bigg/\int e^{\eta\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)
\end{split}
\end{equation}
for any ${\hat{m}}\in{\mathbb{N}}}% \N == \mathbb{N$ and $({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})\in
\G^{2{\hat{m}}}\times{\{\uparrow,\downarrow\}}^{2{\hat{m}}}$ and any $\eta\in{\mathbb{C}}}% \C == \mathbb{C$ for which the denominator of \eqref{eq_schwinger_functional} does not vanish. In fact, since
\begin{equation*}
\begin{split}
\lim_{\eta\to 0}&\int e^{\eta\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}\\
&\cdot d\mu_{G_h}({\overline{\psi}}_X,\psi_X)=1,
\end{split}
\end{equation*}
$S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(G_h,\eta)$ is analytic with respect to $\eta$ in a neighborhood of origin.
The purpose of this section is to prove the following proposition.
\begin{proposition}\label{prop_schwinger_functional_bound}
Assume that there exists a positive constant ${\mathcal{B}}>0$ such that for any $p\in{\mathbb{N}}}% \N == \mathbb{N$, $({\mathbf{x}}_{j_l},\xi_{j_l},x_{j_l}),({\mathbf{y}}_{k_l},\phi_{k_l},y_{k_l})\in \G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h$ $(\forall l\in\{1,\cdots,p\})$ and $n\in{\mathbb{N}}}% \N == \mathbb{N$,
\begin{equation}\label{eq_assumption_determinant_bound}
\sup_{{\mathbf{u}}_l,{\mathbf{v}}_l\in{\mathbb{C}}}% \C == \mathbb{C^n\atop \|{\mathbf{u}}_l\|_{{\mathbb{C}}}% \C == \mathbb{C^n},\|{\mathbf{v}}_l\|_{{\mathbb{C}}}% \C == \mathbb{C^n}\le 1\ \forall l\in\{1,\cdots,p\}}|\det\left(\<{\mathbf{u}}_l,{\mathbf{v}}_m\>_{{\mathbb{C}}}% \C == \mathbb{C^n}G_h({\mathbf{x}}_{j_l}\xi_{j_l}x_{j_l},{\mathbf{y}}_{k_m}\phi_{k_m}y_{k_m})\right)_{1\le l,m\le p}|\le {\mathcal{B}}^p.
\end{equation}
Let $b_m$ denote the coefficient of $\eta^m$ in the Taylor series of the function $\eta\mapsto $\\
$S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(G_h,\eta)$ around $0\in{\mathbb{C}}}% \C == \mathbb{C$. The following bounds hold.
\begin{equation*}
\begin{split}
&|b_0|\le {\mathcal{B}}^{{\hat{m}}},\\
&|b_m|\le \frac{{\hat{m}} 4^{{\hat{m}}}{\mathcal{B}}^{{\hat{m}}}}{m}\left(\sum_{l=1}^{{\tilde{n}}}l4^l{\mathcal{B}}^{l-1}\|U_{L,l}\|_{L,l}{\mathcal{D}}\right)^m\ (\forall m\in{\mathbb{N}}}% \N == \mathbb{N).
\end{split}
\end{equation*}
\end{proposition}
\begin{proof} Let us consider $\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\}_{l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}, (X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}$ as mutually independent, complex multi-variables. In a neighborhood of origin we can define a multi-variable analytic function $W(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})$ by
\begin{equation*}
\begin{split}
&W(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\\
&:=\log\Bigg(\int
e^{\sum_{l=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X^l,Y^l,\Xi^l,\Phi^l)\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)V_{h,X^l,Y^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}\\
&\qquad\qquad \cdot d\mu_{G_h}({\overline{\psi}}_X,\psi_X)\Bigg).
\end{split}
\end{equation*}
Its Taylor expansion becomes
\begin{equation*}
\begin{split}
&W(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})=\sum_{m=0}^{\infty}\frac{1}{m!}\prod_{k=1}^m\left(\sum_{l_k=1}^{\max\{{\hat{m}},{\tilde{n}}\}}\sum_{(X^{l_k}_k,Y^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)\in\G^{2l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}\right)\\
&\cdot \prod_{k=1}^m\frac{\partial}{\partial U_{\lambda,l_k}(X^{l_k}_k,Y^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)}W(\{U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\Bigg|_{\substack{U_{\lambda,l}(X^{l},Y^{l},\Xi^{l},\Phi^{l})=0\\ \forall l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}\\ \forall (X^{l},Y^{l},\Xi^{l},\Phi^{l})\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}}\\
&\cdot \prod_{k=1}^mU_{\lambda,l_k}(X^{l_k}_k,Y^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k).
\end{split}
\end{equation*}
We can take a small $r>0$ so that for all $\eta\in{\mathbb{C}}}% \C == \mathbb{C\backslash \{0\}$ with $|\eta|<r$
\begin{equation}\label{eq_schwinger_functional_expansion}
\begin{split}
&S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(G_h,\eta)\\
&\quad =-\frac{1}{\beta\eta}\frac{\partial}{\partial
U_{\lambda,{\hat{m}}}({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\\
&\qquad \cdot W(\{\eta U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\Bigg|_{\lambda(X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})=0\atop\forall (X^{{\hat{m}}},Y^{{\hat{m}}},\Xi^{{\hat{m}}},\Phi^{{\hat{m}}})\in\G^{2{\hat{m}}}\times{\{\uparrow,\downarrow\}}^{2{\hat{m}}}}\\
&\quad =
-\frac{1}{\beta}\sum_{m=1}^{\infty}\frac{1}{(m-1)!}\prod_{k=1}^{m-1}\left(\sum_{l_k=1}^{{\tilde{n}}}\sum_{(X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)\in\G^{l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}\right)\\
&\qquad \cdot\frac{\partial}{\partial
U_{\lambda,{\hat{m}}}({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\prod_{k=1}^{m-1}\frac{\partial}{\partial U_{\lambda,l_k}(X^{l_k}_k,X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)}\\
&\qquad \cdot W(\{ U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\Bigg|_{\substack{U_{\lambda,l}(X^{l},Y^{l},\Xi^{l},\Phi^{l})=0\\ \forall l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}\\ \forall (X^{l},Y^{l},\Xi^{l},\Phi^{l})\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}}\\
&\qquad\cdot\prod_{k=1}^{m-1}U_{L,l_k}(X_k^{l_k},\Xi_k^{l_k},\Phi_k^{l_k})\cdot\eta^{m-1}.
\end{split}
\end{equation}
By comparing with the definition \eqref{eq_schwinger_functional} we see that the equality \eqref{eq_schwinger_functional_expansion} also
holds true for $\eta =0$.
We will bound the coefficient of $\eta^{m}$ in the right hand side of
\eqref{eq_schwinger_functional_expansion}. To organize the argument we
divide the rest of the proof by 4 steps.
(Step 1) First remark that by Lemma \ref{lem_grassmann_gaussian_equality}.
\begin{equation}\label{eq_bound_for_m_1}
\begin{split}
b_0&=\\
&\ -\frac{1}{\beta}\frac{\partial}{\partial U_{\lambda,{\hat{m}}}({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}W(\{ U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\Bigg|_{\substack{U_{\lambda,l}(X^{l},Y^{l},\Xi^{l},\Phi^{l})=0\\ \forall l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}\\ \forall (X^{l},Y^{l},\Xi^{l},\Phi^{l})\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}}\\
&=\frac{1}{\beta h}\sum_{s\in[0,\beta)_h}\int{\overline{\psi}}_{{\hat{\mathbf{x}}}_1\hat{\xi}_1s}\cdots{\overline{\psi}}_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}s}\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}s}\cdots \psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}s}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)\\
&=\det(G_h({\hat{\mathbf{x}}}_j\hat{\xi}_{j}0,{\hat{\mathbf{y}}}_k{\hat{\phi}}_k0))_{1\le j,k\le {\hat{m}}}.
\end{split}
\end{equation}
Thus, by the assumption \eqref{eq_assumption_determinant_bound} $|b_0|\le {\mathcal{B}}^{{\hat{m}}}$.
(Step 2) Consider the coefficient of $\eta^{m}$ in
\eqref{eq_schwinger_functional_expansion} for $m\ge 1$. Here we apply
the tree formula \cite[\mbox{Theorem 3}]{SW} to characterize the partial derivatives of $W(\{ U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})$ evaluated at origin. We then obtain
\begin{equation}\label{eq_tree_expansion}
\begin{split}
&b_{m}\\
&=-\frac{1}{\beta}\frac{1}{m!}\prod_{k=1}^{m}\left(\sum_{l_k=1}^{{\tilde{n}}}\sum_{(X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)\in\G^{l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{L,l_k}(X_k^{l_k},\Xi_k^{l_k},\Phi_k^{l_k})\right)\\
&\quad\cdot\frac{\partial}{\partial
U_{\lambda,{\hat{m}}}({\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}})}\prod_{k=1}^{m}\frac{\partial}{\partial U_{\lambda,l_k}(X^{l_k}_k,X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)}\\
&\quad\cdot W(\{ U_{\lambda,l}(X^l,Y^l,\Xi^l,\Phi^l)\})\Bigg|_{\substack{U_{\lambda,l}(X^{l},Y^{l},\Xi^{l},\Phi^{l})=0\\ \forall l\in\{1,\cdots,\max\{{\hat{m}},{\tilde{n}}\}\}\\ \forall (X^{l},Y^{l},\Xi^{l},\Phi^{l})\in\G^{2l}\times{\{\uparrow,\downarrow\}}^{2l}}}\\
&= -\frac{1}{\beta}\frac{1}{m!}\prod_{k=1}^{m}\left(\sum_{l_k=1}^{{\tilde{n}}}\sum_{(X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)\in\G^{l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}U_{L,l_k}(X_k^{l_k},\Xi_k^{l_k},\Phi_k^{l_k})\right)\\
&\cdot \sum_{T\in{\mathbb{T}}}% \T == \mathbb{T(\{0,1,\cdots,m\})}\prod_{\{p,q\}\in
T}(\D_{p,q}+\D_{q,p})\int_{[0,1]^{m}}d{\mathbf{s}} \sum_{\pi\in
{\mathbb{S}}}% \S == \mathbb{S_{m+1}(T)}\chi(T,\pi,{\mathbf{s}})\cdot e^{\D(M(T,\pi,{\mathbf{s}}))}\\
&\cdot V_{h,{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}^{{\hat{m}}}({\overline{\psi}}_X^0,\psi_X^0)\prod_{k=1}^{m}V_{h,X_k^{l_k},X_k^{l_k},\Xi_k^{l_k},\Phi_k^{l_k}}^{l_k}({\overline{\psi}}_X^k,\psi_X^k)\Bigg|_{{\overline{\psi}}_X^q=\psi_X^q={\mathbf{0}}\atop \forall q\in\{0,1,\cdots,m\}}.
\end{split}
\end{equation}
Definitions of the newly introduced notations are in order; ${\mathbb{T}}}% \T == \mathbb{T(\{0,1,\cdots,m\})$ is the set of all trees on the vertices $\{0,1,\cdots,m\}$, ${\mathbb{S}}}% \S == \mathbb{S_{m+1}(T)$ is a $T$-dependent set of permutations over $m+1$ elements, $\chi(T,\pi,{\mathbf{s}})$ is a $(T,\pi,{\mathbf{s}})$-dependent non-negative real function obeying
\begin{equation}\label{eq_phi_equality}
\int_{[0,1]^{m}}d{\mathbf{s}}\sum_{\pi\in {\mathbb{S}}}% \S == \mathbb{S_{m+1}(T)}\chi(T,\pi,{\mathbf{s}})=1,
\end{equation}
$M(T,\pi,{\mathbf{s}})$ is a $(T,\pi,{\mathbf{s}})$-dependent non-negative real symmetric
$(m+1)\times (m+1)$ matrix satisfying $M(T,\pi,{\mathbf{s}})_{p,p}=1$ for all $p \in \{0,1,\cdots,m\}$ and the Laplacian operators $\D_{p,q}$ and $\D(M(T,\pi,{\mathbf{s}}))$ are defined by
$$\D_{p,q}:=-\<\left(\frac{\partial}{\partial {\overline{\psi}}_X^p}\right)^t,G_h\left(\frac{\partial}{\partial \psi_X^{q}}\right)^t\>,\ \D(M(T,\pi,{\mathbf{s}})):=\sum_{p,q=0}^{m}M(T,\pi,{\mathbf{s}})_{p,q}\D_{p,q},$$
where
$$\frac{\partial}{\partial {\overline{\psi}}_X^q}:=\left(\frac{\partial}{\partial {\overline{\psi}}_{({\mathbf{x}}\xi x)_1}^q},\cdots,\frac{\partial}{\partial {\overline{\psi}}_{({\mathbf{x}}\xi x)_N}^q}\right),\ \frac{\partial}{\partial \psi_X^q}:=\left(\frac{\partial}{\partial \psi_{({\mathbf{x}}\xi x)_1}^q},\cdots,\frac{\partial}{\partial \psi_{({\mathbf{x}}\xi x)_N}^q}\right)$$
are vectors of the Grassmann left derivatives corresponding to the labeled Grassmann variables
$\{{\overline{\psi}}_{({\mathbf{x}}\xi x)_j}^q,\psi_{({\mathbf{x}}\xi x)_j}^q\ |\
j\in\{1,\cdots,N\}\}$ $(\forall q\in\{0,1,\cdots,m\})$.
(Step 3) We will bound the right hand side of \eqref{eq_tree_expansion} by replacing the sum over trees by the sum over incidence numbers in the next step. As the preparation let us summarize some facts. Let
$d_0,d_1,\cdots,d_{m}$ denote the incidence numbers of a tree $T$
corresponding to the vertices $0,1,\cdots,m$, respectively. Every term of the expansion of $\prod_{\{p,q\}\in T}(\D_{p,q}+\D_{q,p})$ is a product of $m$ Laplacians as the tree $T$ has $m$ lines. Each product
of the $m$ Laplacians contains $d_q$ derivatives with respect to the
Grassmann variables with label $q$ for all
$q\in\{0,1,\cdots,m\}$. Moreover, the $d_q$ derivatives consist of
$\alpha_q$ derivatives with respect to the variables $\{\psi_{{\mathbf{x}}\xi
x}^q\}_{({\mathbf{x}},\xi, x)\in \G\times {\{\uparrow,\downarrow\}}\times [0,\beta)_h}$ and
${\tilde{\alpha}}_q$ derivatives with respect to the variables $\{{\overline{\psi}}_{{\mathbf{x}}\xi x}^q\}_{({\mathbf{x}},\xi, x)\in \G\times {\{\uparrow,\downarrow\}}\times [0,\beta)_h}$ for some non-negative integers $\alpha_q,{\tilde{\alpha}}_q$ satisfying $d_q=\alpha_q+{\tilde{\alpha}}_q$.
This product of the $m$ Laplacians produces
$$1_{\alpha_0\le {\hat{m}}}1_{{\tilde{\alpha}}_0\le {\hat{m}}}\left(\begin{array}{c} {\hat{m}} \\ \alpha_0\end{array}\right)\alpha_0!\left(\begin{array}{c} {\hat{m}} \\ {\tilde{\alpha}}_0\end{array}\right){\tilde{\alpha}}_0!\prod_{k=1}^{m}1_{\alpha_k\le l_k}1_{{\tilde{\alpha}}_k\le l_k}\left(\begin{array}{c} l_k \\ \alpha_k\end{array}\right)\alpha_k!\left(\begin{array}{c} \l_k \\ {\tilde{\alpha}}_k\end{array}\right){\tilde{\alpha}}_k!$$
monomials when the term acts on the monomial
\begin{equation}\label{eq_grassmann_monomial}
\begin{split}
&{\overline{\psi}}_{{\hat{\mathbf{x}}}_1\hat{\xi}_1 s_0}{\overline{\psi}}_{{\hat{\mathbf{x}}}_2\hat{\xi}_2 s_0}\cdots
{\overline{\psi}}_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}} s_0}\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}
s_0}\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}-1}{\hat{\phi}}_{{\hat{m}}-1} s_0}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}
s_0}\\
&\cdot\prod_{k=1}^{m}\Big({\overline{\psi}}_{{\mathbf{x}}_{k,1}\xi_{k,1}s_k}^k{\overline{\psi}}_{{\mathbf{x}}_{k,2}\xi_{k,2}s_k}^k\cdots{\overline{\psi}}_{{\mathbf{x}}_{k,l_k}\xi_{k,l_k}s_k}^k\\
&\qquad\quad\cdot\psi_{{\mathbf{y}}_{k,l_k}\phi_{k,l_k}s_k}^k\psi_{{\mathbf{y}}_{k,l_k-1}\phi_{k,l_k-1}s_k}^k\cdots\psi_{{\mathbf{y}}_{k,1}\phi_{k,1}s_k}^k\Big).
\end{split}
\end{equation}
Moreover, every remaining monomial left after the product of the $m$ Laplacians acting on \eqref{eq_grassmann_monomial} consists of ${\hat{m}}-\alpha_0+\sum_{k=1}^{m}(l_k-\alpha_k)$ products of the Grassmann variables of the form $\psi_{{\mathbf{x}}\xi s}^q$ and ${\hat{m}}-{\tilde{\alpha}}_0+\sum_{k=1}^{m}(l_k-{\tilde{\alpha}}_k)$ products of the Grassmann variables of the form ${\overline{\psi}}_{{\mathbf{x}}\xi s}^q$ and is acted by the operator $e^{\D(M(T,\pi,{\mathbf{s}}))}$. The determinant bound \eqref{eq_assumption_determinant_bound} and the non-negative symmetric property of $M(T,\pi,{\mathbf{s}})$ together with the fact that all the diagonal elements of $M(T,\pi,{\mathbf{s}})$ are $1$ validate the inequality
\begin{equation}\label{eq_determinant_bound_application}
\begin{split}
&\Bigg|e^{\D(M(T,\pi,{\mathbf{s}}))}{\overline{\psi}}_{{\mathbf{x}}_{0,1}\xi_{0,1}s_0}^0\cdots {\overline{\psi}}_{{\mathbf{x}}_{0,{\hat{m}}-{\tilde{\alpha}}_0}\xi_{0,{\hat{m}}-{\tilde{\alpha}}_0}s_0}^0\psi_{{\mathbf{y}}_{0,1}\phi_{0,1}s_0}^0\cdots \psi_{{\mathbf{y}}_{0,{\hat{m}}-\alpha_0}\phi_{0,{\hat{m}}-\alpha_0}s_0}^0\\
&\quad\cdot\prod_{k=1}^{m}{\overline{\psi}}_{{\mathbf{x}}_{k,1}\xi_{k,1}s_k}^k\cdots
{\overline{\psi}}_{{\mathbf{x}}_{k,l_k-{\tilde{\alpha}}_k}\xi_{k,l_k-{\tilde{\alpha}}_k}s_k}^k\\
&\qquad\quad\cdot\psi_{{\mathbf{y}}_{k,1}\phi_{k,1}s_k}^k\cdots \psi_{{\mathbf{y}}_{k,l_k-\alpha_k}\phi_{k,l_k-\alpha_k}s_k}^k\Big|_{{\overline{\psi}}_X^q=\psi_X^q={\mathbf{0}}\atop \forall q\in\{0,1,\cdots,m\}}\Bigg|\\
&\quad\le 1_{\sum_{q=0}^{m}\alpha_q=\sum_{q=0}^{m}{\tilde{\alpha}}_q}{\mathcal{B}}^{{\hat{m}}-\alpha_0+\sum_{k=1}^{m}(l_k-\alpha_k)}
\end{split}
\end{equation}
(see \cite[\mbox{Lemma 4.5}]{K} for a detailed proof of the same bound in the case ${\mathcal{B}}=4$).
The coefficient of each remaining monomial left after the product of the $m$ Laplacians acting on the monomial \eqref{eq_grassmann_monomial} is a product of $m$ elements of the covariance
$(G_h({\mathbf{x}}\xi x,{\mathbf{y}}\phi y))_{({\mathbf{x}},\xi, x),({\mathbf{y}},\phi, y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}$ and going to be multiplied by \\
$\prod_{k=1}^{m}\frac{1}{h}U_{L,l_k}(X_k^{l_k},\Xi_{k}^{l_k},\Phi_k^{l_k})$ and summed over the variables $(X_k^{l_k},\Xi_{k}^{l_k},\Phi_k^{l_k},s_k)\in$\\
$\G^{l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}\times[0,\beta)_h$ for all $k\in\{1,\cdots,m\}$. By considering that the tree $T$ starts from the vertex $0$ and using the properties \eqref{eq_condition_U}-\eqref{eq_translation_invariance}, a recursive argument through all the lines of $T$ running from the later generation to the earlier generation shows that this sum is bounded by
\begin{equation}\label{eq_recurcive_bound}
\prod_{k=1}^{m}\|U_{L,l_k}\|_{L,l_k}{\mathcal{D}}^{m}.
\end{equation}
Note that to derive the upper bound \eqref{eq_recurcive_bound} we repeatedly use the inequality of the form
$$\frac{1}{h}\sum_{s\in[0,\beta)_h}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^l\times {\{\uparrow,\downarrow\}}^{2l}}|U_{L,l}(X^l,\Xi^l,\Phi^l)||G_h({\mathbf{y}}\s y,{\mathbf{x}}_j\xi_j s)|\le\|U_{L,l}\|_{L,l}{\mathcal{D}},$$
where $({\mathbf{y}},\s,y)\in \G\times{\{\uparrow,\downarrow\}}\times [0,\beta)_h$ is any fixed element and ${\mathbf{x}}_j$ and $\xi_j$ are $j$-th component of $X^l$ and $\Xi^l$ respectively for some $j\in\{1,\cdots,l\}$.
Lastly remark that the number of terms containing $\alpha_q$ derivatives with respect to the variables $\{\psi_{{\mathbf{x}}\xi x}^q\}_{({\mathbf{x}},\xi,x)\in\G\times {\{\uparrow,\downarrow\}}\times [0,\beta)_h
}$ $(\forall q\in\{0,1,\cdots,m\})$ in the expansion $\prod_{\{p,q\}\in
T}(\Delta_{p,q}+\Delta_{q,p})$ is at most
$$\prod_{q=0}^{m}\left(\begin{array}{c} d_q \\ \alpha_q \end{array}\right).$$
(Step 4) By summing up all these considerations, replacing the sum over trees by the sum over possible incidence numbers $d_0,\cdots,d_{m}$ satisfying the condition $\sum_{q=0}^{m}d_q=2m$, and using Cayley's theorem on the number of trees with fixed incidence numbers (see, e.g, \cite[\mbox{Corollary 2.2.4}]{W}) and the equality \eqref{eq_phi_equality}, we have
\begin{equation*}
\begin{split}
&|b_{m}|\\
&\le \frac{1}{m!}\frac{1}{\beta
h}\sum_{s_0\in[0,\beta)_h}\prod_{k=1}^{m}\\
&\quad\cdot\left(\sum_{l_k=1}^{{\tilde{n}}}\sum_{(X^{l_k}_k,\Xi^{l_k}_k,\Phi^{l_k}_k)\in\G^{l_k}\times{\{\uparrow,\downarrow\}}^{2l_k}}|U_{L,l_k}(X_k^{l_k},\Xi_k^{l_k},\Phi_k^{l_k})|\frac{1}{h}\sum_{s_k\in[0,\beta)_h}\right)\\
&\quad\cdot \sum_{T\in{\mathbb{T}}}% \T == \mathbb{T(\{0,1,\cdots,m\})}\sup_{{\mathbf{s}}\in[0,1]^{m}\atop \pi
\in {\mathbb{S}}}% \S == \mathbb{S_{m+1}(T)}\\
&\quad\cdot\Bigg|e^{\D(M(T,\pi,{\mathbf{s}}))}\prod_{\{p,q\}\in
T}(\D_{p,q}+\D_{q,p})\cdot(\text{the monomial
}\eqref{eq_grassmann_monomial})\Big|_{{\overline{\psi}}_X^q=\psi_X^q={\mathbf{0}}\atop
\forall q\in\{0,1,\cdots,m\}}\Bigg|\\
&\le \frac{1}{m!}\frac{1}{\beta
h}\sum_{s_0\in[0,\beta)_h}\prod_{k=1}^{m}\left(\sum_{l_k=1}^{{\tilde{n}}}\|U_{L,l_k}\|_{L,l_k}{\mathcal{D}}\right)\sum_{\substack{d_0\in\{1,\cdots,2{\hat{m}}\}\\ d_k\in\{1,\cdots,2l_k\}\\ \forall k\in\{1,\cdots,m\}}}\frac{(m-1)!}{\prod_{q=0}^{m}(d_q-1)!}1_{\sum_{q=0}^{m}d_q=2m}\\
&\quad\cdot\sum_{\alpha_q,{\tilde{\alpha}}_q\in\{0,\cdots,d_q\}\atop \forall q\in\{0,1,\cdots,m\}}\prod_{q=0}^{m}1_{\alpha_q+{\tilde{\alpha}}_q=d_q}1_{\sum_{q=0}^{m}\alpha_q=\sum_{q=0}^{m}{\tilde{\alpha}}_q}\prod_{q=0}^{m}\left(\begin{array}{c} d_q \\ \alpha_q \end{array}\right){\mathcal{B}}^{{\hat{m}}-\alpha_0+\sum_{k=1}^{m}(l_k-\alpha_k)}\\
&\quad\cdot1_{\alpha_0\le{\hat{m}}}1_{{\tilde{\alpha}}_0\le{\hat{m}}}\left(\begin{array}{c} {\hat{m}} \\
\alpha_0\end{array}\right)\alpha_0!\left(\begin{array}{c} {\hat{m}} \\ {\tilde{\alpha}}_0\end{array}\right){\tilde{\alpha}}_0!\prod_{k=1}^{m}1_{\alpha_k\le l_k}1_{{\tilde{\alpha}}_k\le l_k}\left(\begin{array}{c} l_k \\ \alpha_k\end{array}\right)\alpha_k!\left(\begin{array}{c} \l_k \\ {\tilde{\alpha}}_k\end{array}\right){\tilde{\alpha}}_k!\\
&=\frac{1}{m}\prod_{k=1}^{m}\left(\sum_{l_k=1}^{{\tilde{n}}}\|U_{L,l_k}\|_{L,l_k}{\mathcal{D}}\right){\mathcal{B}}^{{\hat{m}}+\sum_{k=1}^{m}l_k-m}\sum_{\substack{\alpha_0,{\tilde{\alpha}}_0\in\{0,\cdots,{\hat{m}}\}\\ \alpha_k,{\tilde{\alpha}}_k\in\{0,\cdots,l_k\}\\ \forall k\in\{1,\cdots,m\}}}1_{\sum_{q=0}^m\alpha_q=\sum_{q=0}^m{\tilde{\alpha}}_q=m}\\
&\quad\cdot\prod_{q=0}^{m}(\alpha_q+{\tilde{\alpha}}_q)\left(\begin{array}{c} {\hat{m}} \\ \alpha_0\end{array}\right)\left(\begin{array}{c} {\hat{m}} \\ {\tilde{\alpha}}_0\end{array}\right)\prod_{k=1}^{m}\left(\begin{array}{c} l_k \\ \alpha_k\end{array}\right)\left(\begin{array}{c} \l_k \\ {\tilde{\alpha}}_k\end{array}\right).
\end{split}
\end{equation*}
Then, by dropping the constraint $1_{\sum_{q=0}^m\alpha_q=\sum_{q=0}^m{\tilde{\alpha}}_q=m}$ and using the equality
$$\sum_{\alpha=0}^l\sum_{{\tilde{\alpha}}=0}^l(\alpha+{\tilde{\alpha}})\left(\begin{array}{c} l \\ \alpha\end{array}\right)\left(\begin{array}{c} l \\
{\tilde{\alpha}}\end{array}\right)=l4^l,
$$
we obtain
\begin{equation*}
\begin{split}
|b_{m}|&\le \frac{1}{m}\prod_{k=1}^{m}\left(\sum_{l_k=1}^{{\tilde{n}}}\|U_{L,l_k}\|_{L,l_k}{\mathcal{D}}\right){\mathcal{B}}^{{\hat{m}}+\sum_{k=1}^{m}l_k-m}\hm4^{{\hat{m}}}\prod_{k=1}^{m}l_k4^{l_k}\\
&= \frac{{\hat{m}} 4^{{\hat{m}}}{\mathcal{B}}^{{\hat{m}}}}{m}\left(\sum_{l=1}^{{\tilde{n}}}l4^l{\mathcal{B}}^{l-1}\|U_{L,l}\|_{L,l}{\mathcal{D}}\right)^{m}.
\end{split}
\end{equation*}
\end{proof}
The evaluation of the tree expansion presented above overcounts the
combinatorial factor, which is defined as the number of monomials appearing in the expansion $\prod_{\{p,q\}\in T}(\D_{p,q}+\D_{q,p})\cdot(\text{the monomial
}\eqref{eq_grassmann_monomial})$.
If we consider the $4$ point correlation functions for the on-site interaction $V=U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}$ $(U\in{\mathbb{R}}}% \R == \mathbb{R)$, the
exact calculation of the combinatorial factor without overcounting is possible as proved in \cite[\mbox{Lemma 4.7}]{K}. Consequently, the perturbative bound is improved in this case.
\begin{proposition}\label{prop_schwinger_functional_hubbard_model_bound}
Assume that a non-singular matrix $G_h$ satisfies \eqref{eq_assumption_determinant_bound}. For any $m\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$ let $c_m$ denote the coefficient of $\eta^m$ in the Taylor series of the function
\begin{equation*}
\begin{split}
&\eta\longmapsto \\
&\frac{1}{\beta h}\sum_{x\in[0,\beta)_h}\int{\overline{\psi}}_{{\hat{\mathbf{x}}}_1{\uparrow} x}{\overline{\psi}}_{{\hat{\mathbf{x}}}_2{\downarrow} x}\psi_{{\hat{\mathbf{y}}}_2{\downarrow} x}\psi_{{\hat{\mathbf{y}}}_1{\uparrow} x}e^{-\eta\frac{U}{h}\sum_{s\in[0,\beta)_h}{\overline{\psi}}_{{\mathbf{x}}{\uparrow} s}{\overline{\psi}}_{{\mathbf{x}}{\downarrow} s}\psi_{{\mathbf{x}}{\downarrow} s}\psi_{{\mathbf{x}}{\uparrow} s}}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)\\
&\cdot \Big/\int e^{-\eta\frac{U}{h}\sum_{s\in[0,\beta)_h}{\overline{\psi}}_{{\mathbf{x}}{\uparrow} s}{\overline{\psi}}_{{\mathbf{x}}{\downarrow} s}\psi_{{\mathbf{x}}{\downarrow} s}\psi_{{\mathbf{x}}{\uparrow} s}}d\mu_{G_h}({\overline{\psi}}_X,\psi_X)
\end{split}
\end{equation*}
around $0\in{\mathbb{C}}}% \C == \mathbb{C$. The following bound holds. For any $m\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$
$$|c_m|\le\frac{4{\mathcal{B}}^2}{3m+4}\left(\begin{array}{c} 3m+4 \\ m \end{array}\right)({\mathcal{D}}{\mathcal{B}}|U|)^m.$$
\end{proposition}
\begin{proof} The argument in \cite[\mbox{Section 4}]{K} leading to the bound \cite[\mbox{Lemma 4.8}]{K} straightforwardly applies to deduce the claimed bound.
\end{proof}
\section{Exponential decay of the correlation functions}\label{sec_exponential_decay_correlation}
In this section we prove the exponential decay property of the
correlation functions. The Grassmann integral formulation
\eqref{eq_correlation_grassmann_integral} serves the base of our analysis.
In the subsection
\ref{subsec_exponential_decay_correlation} we will see that the $U(1)$-invariance property of the Grassmann integral leads to a
replacement of the covariance $C_h$ by a modified covariance in the
Grassmann integral formulation. In the following subsections
\ref{subsec_determinant_bound}-\ref{subsec_exponential_decay_covariance}
we study properties of such perturbed covariance matrices as
preliminaries for proving our main theorems.
\subsection{Determinant bound on modified covariance matrices}\label{subsec_determinant_bound}
Here we prove the determinant bound on a generalized covariance matrix of the form
\begin{equation}\label{eq_extended_covariance}
C_{{\mathcal{E}}}({\mathbf{x}}\xi x,{\mathbf{y}}\phi y):=\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}e^{-(y-x){{\mathcal{E}}}_{{\mathbf{k}}}}\left(\frac{1_{y-x\le
0}}{1+e^{\beta {{\mathcal{E}}}_{{\mathbf{k}}}}}-\frac{1_{y-x>
0}}{1+e^{-\beta {{\mathcal{E}}}_{{\mathbf{k}}}}}\right)
\end{equation}
$(\forall ({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta))$. In
\eqref{eq_extended_covariance} ${\mathcal{E}}_{{\mathbf{k}}}:\G^*\to{\mathbb{C}}}% \C == \mathbb{C$ is any function satisfying
\begin{equation}\label{eq_regular_condition}
|\mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}}}|<\frac{\pi}{\beta}\ (\forall{\mathbf{k}}\in\G^*).
\end{equation}
Note that the inequality \eqref{eq_regular_condition} implies $|1+e^{\beta{\mathcal{E}}_{{\mathbf{k}}}}|,\ |1+e^{-\beta{\mathcal{E}}_{{\mathbf{k}}}}|>0$ $(\forall{\mathbf{k}}\in\G^*)$, and thus $C_{{\mathcal{E}}}$ is well-defined. The covariance \eqref{eq_extended_covariance} is a generalization of the covariance \eqref{eq_covariance} in which the dispersion relation $E_{{\mathbf{k}}}$ is real-valued.
The volume- and temperature-independent determinant bound on the covariance \eqref{eq_covariance} was established by Pedra and Salmhofer in \cite{PS} by extending the notion of the Gram bound on determinant. However, the Gram representation of the covariance presented in \cite[\mbox{Section 4.1}]{PS} does not apply to the case that the dispersion relation is complex-valued. We need to construct a Gram representation of the covariance matrix \eqref{eq_extended_covariance} to bound its determinant.
We use the following abstract framework proved by Pedra and Salmhofer as a generalization of Gram's inequality.
\begin{lemma}{\cite[\mbox{Theorem 1.3}]{PS}}\label{lem_abstract_determinant_bound}
Assume that $K,k\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$ satisfy $k+K\ge 1$. Let ${\mathbb{X}}$ be any
set. Assume that $J$ is a set totally ordered under `$\preceq$' and
`$\prec$' is a strict total order in $J$. Let $\zeta_l$, $\zeta_l'$ be
maps from ${\mathbb{X}}$ to $J$ $(\forall l\in\{1,\cdots,k+K\})$. Let ${\mathcal{H}}$ be a
Hilbert space with the inner product $\<\cdot,\cdot\>_{{\mathcal{H}}}$ and
$\|\cdot\|_{{\mathcal{H}}}$ denote the norm induced by $\<\cdot,\cdot\>_{{\mathcal{H}}}$. Define a matrix $M_{x,y}$ $(x,y\in{\mathbb{X}})$ by
\begin{equation}\label{eq_gram_representation}
M_{x,y}:=\<f_x^0,g_y^0\>_{{\mathcal{H}}}+\sum_{l=1}^k\<f_x^l,g_y^l\>_{{\mathcal{H}}}1_{\zeta_l'(x)\succ\zeta_l(y)}+\sum_{l=k+1}^{K+k}\<f_x^l,g_y^l\>_{{\mathcal{H}}}1_{\zeta_l'(x)\succeq\zeta_l(y)},
\end{equation}
where $f_x^l$, $g_y^l\in{\mathcal{H}}$. If
$$\sup_{x\in{\mathbb{X}}}\max\{\|f_x^l\|_{{\mathcal{H}}},\|g_x^l\|_{{\mathcal{H}}}\}\le\g_l\ (\forall l\in\{0,\cdots,K+k\}),$$
the following bound holds. For any $m, n\in{\mathbb{N}}}% \N == \mathbb{N$ and $x_1,\cdots,x_n,y_1,\cdots,y_n\in{\mathbb{X}}$
\begin{equation}\label{eq_PS_bound}
\sup_{\substack{{\mathbf{u}}_j,{\mathbf{v}}_j\in{\mathbb{C}}}% \C == \mathbb{C^m\\
\|{\mathbf{u}}_j\|_{{\mathbb{C}}}% \C == \mathbb{C^m},\|{\mathbf{v}}_j\|_{{\mathbb{C}}}% \C == \mathbb{C^m}\le 1\\ \forall j\in
\{1,\cdots,n\}}}|\det(\<{\mathbf{u}}_j,{\mathbf{v}}_k\>_{{\mathbb{C}}}% \C == \mathbb{C^m}M_{x_j,y_k})_{1\le j,k\le
n}|\le \left(\sum_{l=0}^{k+K}\g_l\right)^{2n}.
\end{equation}
\end{lemma}
\begin{remark}
The determinant bound \cite[\mbox{Theorem 1.3}]{PS} was originally claimed
for the case that $m=n$ in \eqref{eq_PS_bound}. The bound for general $m$ immediately follows from \eqref{eq_PS_bound} without
the coefficient $\<{\mathbf{u}}_j,{\mathbf{v}}_k\>_{{\mathbb{C}}}% \C == \mathbb{C^m}$. In fact it is sufficient to redefine ${\mathcal{H}}$ by ${\mathbb{C}}}% \C == \mathbb{C^m \otimes {\mathcal{H}}$ and $M_{x,y}$ by
\begin{equation*}
\begin{split}
M_{x,y}:=&\<{\mathbf{u}}_x\otimes f_x^0,{\mathbf{v}}_y\otimes g_y^0\>_{{\mathbb{C}}}% \C == \mathbb{C^m \otimes{\mathcal{H}}}+\sum_{l=1}^k\<{\mathbf{u}}_x\otimes f_x^l,{\mathbf{v}}_y\otimes g_y^l\>_{{\mathbb{C}}}% \C == \mathbb{C^m \otimes{\mathcal{H}}}1_{\zeta_l'(x)\succ\zeta_l(y)}\\
&+\sum_{l=k+1}^{K+k}\<{\mathbf{u}}_x\otimes f_x^l, {\mathbf{v}}_y\otimes g_y^l\>_{{\mathbb{C}}}% \C == \mathbb{C^m \otimes{\mathcal{H}}}1_{\zeta_l'(x)\succeq\zeta_l(y)}
\end{split}
\end{equation*}
for any vectors ${\mathbf{u}}_x,{\mathbf{v}}_x\in{\mathbb{C}}}% \C == \mathbb{C^m$ with $\|{\mathbf{u}}_x\|_{{\mathbb{C}}}% \C == \mathbb{C^m},\|{\mathbf{v}}_x\|_{{\mathbb{C}}}% \C == \mathbb{C^m}\le 1$.
\end{remark}
As an application of Lemma \ref{lem_abstract_determinant_bound} we prove the following statement.
\begin{proposition}\label{prop_extended_determinant_bound}
If the inequality
\begin{equation}\label{eq_imaginary_condition}
|\mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}}}|\le\frac{\pi}{2\beta}\ (\forall {\mathbf{k}}\in\G^*)
\end{equation}
holds, then for any $m,n\in{\mathbb{N}}}% \N == \mathbb{N$ and $({\mathbf{x}}_j,\xi_j,x_j),({\mathbf{y}}_j,\phi_j,y_j)\in\G\times{\{\uparrow,\downarrow\}}\times [0,\beta)$ $(\forall j\in\{1,\cdots,n\})$
\begin{equation}\label{eq_extended_determinant_bound}
\sup_{\substack{{\mathbf{u}}_j,{\mathbf{v}}_j\in{\mathbb{C}}}% \C == \mathbb{C^m\\ \|{\mathbf{u}}_j\|_{{\mathbb{C}}}% \C == \mathbb{C^m},\|{\mathbf{v}}_j\|_{{\mathbb{C}}}% \C == \mathbb{C^m}\le 1\\ \forall j\in \{1,\cdots,n\}}}|\det(\<{\mathbf{u}}_j,{\mathbf{v}}_k\>_{{\mathbb{C}}}% \C == \mathbb{C^m}C_{{\mathcal{E}}}({\mathbf{x}}_j\xi_jx_j,{\mathbf{y}}_k\phi_ky_k))_{1\le j,k\le n}|\le 4^n.
\end{equation}
\end{proposition}
\begin{remark}The upper bound $4^n$ claimed above is same as that obtained in \cite[\mbox{Theorem 2.4}]{PS} for the covariance matrices with real-valued dispersion relations.
\end{remark}
\begin{proof}[Proof of Proposition \ref{prop_extended_determinant_bound}]
Our argument here closely follows \cite[\mbox{Section 4.1}]{PS}.
Since $\G^*$ is a finite set, for a sufficiently small $\eps>0$ we may assume that $\mathop{\mathrm{Re}}({\mathcal{E}}_{{\mathbf{k}}}+\eps)\neq 0$ for all ${\mathbf{k}}\in\G^*$. Set ${\mathcal{E}}_{{\mathbf{k}},\eps}:={\mathcal{E}}_{{\mathbf{k}}}+\eps$ $(\forall {\mathbf{k}}\in\G^*)$ and define the parameterized covariance $C_{{\mathcal{E}}}^{\eps}$ by
$$C_{{\mathcal{E}}}^{\eps}({\mathbf{x}}\xi x,{\mathbf{y}} \phi y):=\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}e^{-(y-x){{\mathcal{E}}}_{{\mathbf{k}},\eps}}\left(\frac{1_{y-x\le
0}}{1+e^{\beta {{\mathcal{E}}}_{{\mathbf{k}},\eps}}}-\frac{1_{y-x>
0}}{1+e^{-\beta {{\mathcal{E}}}_{{\mathbf{k}},\eps}}}\right).
$$
We construct a Gram representation of $C_{{\mathcal{E}}}^{\eps}$ fitting in the form \eqref{eq_gram_representation} to apply Lemma \ref{lem_abstract_determinant_bound}. Observe the following decomposition.
\begin{equation}\label{eq_covariance_decomposition}
\begin{split}
& C_{{\mathcal{E}}}^{\eps}({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)=\\
&1_{y-x\le 0}\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}\left(1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}>0}\frac{e^{-(\beta+y-x){{\mathcal{E}}_{{\mathbf{k}},\eps}}}}{1+e^{-\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}}+1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}<0}\frac{e^{-(x-y){(-{\mathcal{E}}_{{\mathbf{k}},\eps})}}}{1+e^{\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}}\right)\\
&-1_{y-x> 0}\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}\left(1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}>0}\frac{e^{-(y-x){{\mathcal{E}}_{{\mathbf{k}},\eps}}}}{1+e^{-\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}}+1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}<0}\frac{e^{-(\beta+x-y){(-{\mathcal{E}}_{{\mathbf{k}},\eps})}}}{1+e^{\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}}\right).\end{split}
\end{equation}
Note that for $t\ge 0$ and ${\mathcal{E}}\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} {\mathcal{E}}|<\pi/\beta$ and $\mathop{\mathrm{Re}} {\mathcal{E}}>0$
\begin{equation}\label{eq_fermi_formula}
\begin{split}
\frac{e^{-t{\mathcal{E}}}}{1+e^{-\beta{{\mathcal{E}}}}}&=\frac{e^{-it\mathop{\mathrm{Im}} {\mathcal{E}}}(1+e^{-\beta{\overline{{\mathcal{E}}}}})}{|1+e^{-\beta{{\mathcal{E}}}}|^2}e^{-t\mathop{\mathrm{Re}} {\mathcal{E}}}\\
&= \frac{e^{-it\mathop{\mathrm{Im}} {\mathcal{E}}}(1+e^{-\beta{\overline{{\mathcal{E}}}}})}{|1+e^{-\beta{{\mathcal{E}}}}|^2}\frac{\mathop{\mathrm{Re}} {\mathcal{E}}}{\pi}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv\frac{e^{itv}}{v^2+(\mathop{\mathrm{Re}} {\mathcal{E}})^2},
\end{split}
\end{equation}
where $\overline{{\mathcal{E}}}$ denotes the complex conjugate of ${\mathcal{E}}$.
By substituting the formula \eqref{eq_fermi_formula} into \eqref{eq_covariance_decomposition} we obtain
\begin{equation}\label{eq_covariance_reformation}
\begin{split}
& C_{{\mathcal{E}}}^{\eps}({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)=\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}\\
&\cdot\Bigg(1_{y-x\le 0}\Bigg(1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}>0}\frac{e^{-i(\beta +y -x)\mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}},\eps}}(1+e^{-\beta{\overline{{\mathcal{E}}_{{\mathbf{k}},\eps}}}})}{|1+e^{-\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}|^2}\frac{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}}{\pi}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv\frac{e^{i(\beta+y-x)v}}{v^2+(\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps})^2}\\
&+1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}<0}\frac{e^{-i(x -y)\mathop{\mathrm{Im}}(- {\mathcal{E}}_{{\mathbf{k}},\eps})}(1+e^{\beta\overline{{\mathcal{E}}_{{\mathbf{k}},\eps}}})}{|1+e^{\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}|^2}\frac{\mathop{\mathrm{Re}}(- {\mathcal{E}}_{{\mathbf{k}},\eps})}{\pi}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv\frac{e^{i(x-y)v}}{v^2+(\mathop{\mathrm{Re}} (-{\mathcal{E}}_{{\mathbf{k}},\eps}))^2}\Bigg)\\
&-1_{y-x> 0}\Bigg(1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}>0}\frac{e^{-i(y -x)\mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}},\eps}}(1+e^{-\beta{\overline{{\mathcal{E}}_{{\mathbf{k}},\eps}}}})}{|1+e^{-\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}|^2}\frac{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}}{\pi}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv\frac{e^{i(y-x)v}}{v^2+(\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps})^2}\\
&+1_{\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}<0}\frac{e^{-i(\beta+ x -y)\mathop{\mathrm{Im}}(- {\mathcal{E}}_{{\mathbf{k}},\eps})}(1+e^{\beta\overline{{\mathcal{E}}_{{\mathbf{k}},\eps}}})}{|1+e^{\beta{{\mathcal{E}}_{{\mathbf{k}},\eps}}}|^2}\frac{\mathop{\mathrm{Re}}(- {\mathcal{E}}_{{\mathbf{k}},\eps})}{\pi}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv\frac{e^{i(\beta+x-y)v}}{v^2+(\mathop{\mathrm{Re}} (-{\mathcal{E}}_{{\mathbf{k}},\eps}))^2}\Bigg)\Bigg).
\end{split}
\end{equation}
Let us define the Hilbert space $L^2(\G^*\times {\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)$ equipped with the inner product $\<\cdot,\cdot\>_{L^2(\G^*\times {\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}$ given by
$$\<f,g\>_{L^2(\G^*\times {\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}:=\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}\sum_{\xi\in{\{\uparrow,\downarrow\}}}\int_{{\mathbb{R}}}% \R == \mathbb{R}dv \overline{f({\mathbf{k}},\xi,v)}g({\mathbf{k}},\xi,v).
$$
For $({\mathbf{x}},\xi,x)\in\G\times{\{\uparrow,\downarrow\}}\times[-\beta,\beta]$ and $a\in\{-1,1\}$ define vectors $f_{{\mathbf{x}},\xi, x}^{a}$, $g_{{\mathbf{x}},\xi, x}^{a}\in L^2(\G^*\times {\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)$ by
\begin{equation*}
\begin{split}
&f_{{\mathbf{x}},\xi, x}^{a}({\mathbf{k}},\phi,v)\\
&\ :=1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\delta_{\xi,\phi}e^{i\<{\mathbf{k}},{\mathbf{x}}\>}e^{-ix(\mathop{\mathrm{Im}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})-v)}\frac{1+e^{-\beta a {\mathcal{E}}_{{\mathbf{k}},\eps}} }{|1+e^{-\beta a{\mathcal{E}}_{{\mathbf{k}},\eps}}|}\sqrt{\frac{|\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}|}{\pi}}\frac{1}{iv +\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})},\\
&g_{{\mathbf{x}},\xi, x}^{a}({\mathbf{k}},\phi,v)\\
&\ :=1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\delta_{\xi,\phi}e^{i\<{\mathbf{k}},{\mathbf{x}}\>}e^{-ix(\mathop{\mathrm{Im}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})-v)}\frac{1}{|1+e^{-\beta a{\mathcal{E}}_{{\mathbf{k}},\eps}}|}\sqrt{\frac{|\mathop{\mathrm{Re}} {\mathcal{E}}_{{\mathbf{k}},\eps}|}{\pi}}\frac{1}{iv +\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}.
\end{split}
\end{equation*}
Note that
\begin{equation}\label{eq_vector_orthogonal}
\begin{split}
\<f_{{\mathbf{x}},\xi ,x}^{a},f_{{\mathbf{y}}, \phi, y}^{-a}\>_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}&=\<f_{{\mathbf{x}},\xi, x}^{a},g_{{\mathbf{y}}, \phi, y}^{-a}\>_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}\\
&=\<g_{{\mathbf{x}},\xi, x}^{a},g_{{\mathbf{y}}, \phi, y}^{-a}\>_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}=0,
\end{split}
\end{equation}
and
\begin{equation}\label{eq_vector_norm}
\begin{split}
&\|f_{{\mathbf{x}},\xi, x}^{a}\|^2_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}=\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0},\\
& \|g_{{\mathbf{x}},\xi, x}^{a}\|^2_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}=\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},,\eps})>0}\frac{1}{|1+e^{-\beta a{\mathcal{E}}_{{\mathbf{k}},\eps}}|^2}.
\end{split}
\end{equation}
The equalities \eqref{eq_vector_orthogonal}-\eqref{eq_vector_norm} imply that
\begin{equation}\label{eq_gram_norm_estimate_f}
\|f_{{\mathbf{x}},\xi, x}^{a}+f_{{\mathbf{y}},\phi, y}^{-a}\|^2_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}=\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}\left(1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}+1_{\mathop{\mathrm{Re}}(-a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\right)= 1
\end{equation}
and under the assumption that $|\mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}}}|\le \pi/(2\beta)$
\begin{equation}\label{eq_gram_norm_estimate_g}
\begin{split}
&\|g_{{\mathbf{x}},\xi, x}^{a}+g_{{\mathbf{y}},\phi, y}^{-a}\|^2_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}\\
&\le\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}\Bigg(1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\frac{1}{e^{-2\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+2\cos(\beta \mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}},\eps})e^{-\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+1}\\
&\qquad\qquad\qquad +1_{\mathop{\mathrm{Re}}(-a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\frac{1}{e^{2\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+2\cos(\beta \mathop{\mathrm{Im}} {\mathcal{E}}_{{\mathbf{k}},\eps})e^{\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+1}\Bigg)\\
&\le\frac{1}{L^d}\sum_{{\mathbf{k}}\in\G^*}\left(1_{\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\frac{1}{e^{-2\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+1}+1_{\mathop{\mathrm{Re}}(-a{\mathcal{E}}_{{\mathbf{k}},\eps})>0}\frac{1}{e^{2\beta\mathop{\mathrm{Re}}(a{\mathcal{E}}_{{\mathbf{k}},\eps})}+1}\right)\le 1
\end{split}
\end{equation}
for all $({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[-\beta,\beta]$ and $a\in\{-1,1\}$.
The equality \eqref{eq_covariance_reformation} can be written as
\begin{equation}\label{eq_covariance_gram}
\begin{split}
C_{{\mathcal{E}}}^{\eps}({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)
&=1_{y-x\le 0}\<f_{{\mathbf{x}},\xi, x-\beta}^{1}+f_{{\mathbf{x}},\xi, -x}^{-1},g_{{\mathbf{y}},\phi, y}^{1}+g_{{\mathbf{y}},\phi, -y}^{-1}\>_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}\\
&\quad +1_{y-x> 0}\<-f_{{\mathbf{x}},\xi, x}^{1}-f_{{\mathbf{x}},\xi, -x}^{-1},g_{{\mathbf{y}},\phi, y}^{1}+g_{{\mathbf{y}},\phi, \beta -y}^{-1}\>_{L^2(\G^*\times{\{\uparrow,\downarrow\}}\times{\mathbb{R}}}% \R == \mathbb{R)}.
\end{split}
\end{equation}
The representation \eqref{eq_covariance_gram} and the bounds \eqref{eq_gram_norm_estimate_f}-\eqref{eq_gram_norm_estimate_g} enable us to apply Lemma \ref{lem_abstract_determinant_bound} to obtain the bound \eqref{eq_extended_determinant_bound} for $C_{{\mathcal{E}}}^{\eps}$. Then, by sending $\eps\searrow 0$ and the continuity of $C_{{\mathcal{E}}}^{\eps}$ with respect to $\eps$ we complete the proof.
\end{proof}
\subsection{Exponential decay of the covariance}\label{subsec_exponential_decay_covariance}
In this subsection we show the exponential decay property of the covariance and find a volume-independent upper bound on its integral as the corollary.
Later in our analysis we will need to deal with a modified covariance whose dispersion relation is given by $E_{{\mathbf{k}}+z{\mathbf{e}}_p}$ with $z\in{\mathbb{C}}}% \C == \mathbb{C$, $p\in\{1,\cdots,d\}$. It is convenient to prove the $L^1$-bound on the modified covariances which include the normal covariance \eqref{eq_covariance} with the real-valued dispersion relation \eqref{eq_dispersion_relation} as a special case at this stage.
Furthermore, to bound the covariance with the dispersion relation $E_{{\mathbf{k}}+z{\mathbf{e}}_p}$ needs to control the covariance with the perturbed dispersion relation $E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}$ defined by
\begin{equation*}
\begin{split}
&E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}:=-2t\sum_{j=1}^d\cos(\<{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q,{\mathbf{e}}_j\>)\\
&\qquad - 4t'\cdot 1_{d\ge 2}\sum_{j,k = 1\atop j<k}^d\cos(\<{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q,{\mathbf{e}}_j\>)\cos(\<{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q,{\mathbf{e}}_k\>)-\mu
\end{split}
\end{equation*}
for $z,w\in{\mathbb{C}}}% \C == \mathbb{C$ and $p,q\in\{1,\cdots,d\}$. Let us study properties of $E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}$.
First we clarify a sufficient condition for the inequalities
$$|\mathop{\mathrm{Im}} (E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q})|< \frac{\pi}{\beta},\ |\mathop{\mathrm{Im}} (E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q})|\le \frac{\pi}{2\beta},$$
which will ensure the analyticity of the modified covariance with respect to the variables $z,w$ and make the determinant bound Proposition \ref{prop_extended_determinant_bound} applicable.
\begin{lemma}\label{lem_dispersion_relation_condition}
Let $z,w\in{\mathbb{C}}}% \C == \mathbb{C$ and $r>0$. If $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|< \frac{1}{2}\log F_{t,t',d}(r)$ with the function $F_{t,t',d}(\cdot)$ defined in \eqref{eq_definition_keyfunction}, then $|\mathop{\mathrm{Im}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}|<r$ holds for all ${\mathbf{k}}\in\G^*$ and $p,q\in\{1,\cdots,d\}$. The parallel statement with `$\le$'s in place of `$<$'s holds as well.
\end{lemma}
\begin{proof}
Let us write $z=a+b i$, $w=u+v i$ $(a,b,u,v\in{\mathbb{R}}}% \R == \mathbb{R)$. By using the formula
$$\cos(x+iy)=\cos(x)\cosh(y)-i\sin(x)\sinh(y)\ (x,y\in{\mathbb{R}}}% \R == \mathbb{R)$$
and writing ${\mathbf{k}}=(k_1,\cdots,k_d)$,
we find that
\begin{equation*}
\begin{split}
\mathop{\mathrm{Im}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}
&= 2t\sum_{j=1}^d\sin(k_j+a\delta_{p,j}+u\delta_{q,j})\sinh(b\delta_{p,j}+v\delta_{q,j})\\
&\quad+4t'1_{d\ge 2}\sum_{j,l=1\atop j\neq l}^d\cos(k_j+a\delta_{p,j}+u\delta_{q,j})\cosh(b\delta_{p,j}+v\delta_{q,j})\\
&\qquad\qquad\qquad\quad\cdot\sin(k_l+a\delta_{p,l}+u\delta_{q,l})\sinh(b\delta_{p,l}+v\delta_{q,l}).
\end{split}
\end{equation*}
Assume that $|b|,|v|\le s$ for $s>0$.
\begin{equation*}
\begin{split}
&|\mathop{\mathrm{Im}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}|\\
&\le 2|t|\sum_{j=1}^d\sinh(s(\delta_{p,j}+\delta_{q,j}))+4|t'|1_{d\ge 2}\sum_{j,l=1\atop j\neq l}^d\cosh(s(\delta_{p,j}+\delta_{q,j}))\sinh(s(\delta_{p,l}+\delta_{q,l}))\\
&\le 1_{p=q}(2|t|\sinh(2s)+4|t'|(d-1)\sinh(2s))\\
&\quad + 1_{p\neq q}(4|t|\sinh(s)+8|t'|(d-1)\cosh(s)\sinh(s))\\
&\le 2\sinh(2s)(|t|+2(d-1)|t'|),
\end{split}
\end{equation*}
where we used the facts that for $s\ge0$
$$1\le \cosh(s),\ 2\sinh(s)\cosh(s)=\sinh(2s),\ 2\sinh(s)\le \sinh(2s).$$
Thus, the inequality
\begin{equation}\label{eq_sinh_inequality}
2\sinh(2s)(|t|+2(d-1)|t'|)<r
\end{equation}
implies that $|\mathop{\mathrm{Im}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}|<r$. Since
$$\sinh^{-1}(X)=\log\left(X+\sqrt{X^2+1}\right) (\forall X\in{\mathbb{R}}}% \R == \mathbb{R),$$
the inequality \eqref{eq_sinh_inequality} is equivalent to the inequality $s<\frac{1}{2}\log F_{t,t',d}(r)$.
\end{proof}
We define the perturbed covariance $C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ as follows.
\begin{equation*}
\begin{split}
C({\mathbf{x}}\xi x,{\mathbf{y}}\phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q):=&\frac{\delta_{\xi,\phi}}{L^d}\sum_{{\mathbf{k}}\in\G^*}e^{i\<{\mathbf{k}},{\mathbf{y}}-{\mathbf{x}}\>}e^{-(y-x){E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}\\
&\cdot\left(\frac{1_{y-x\le
0}}{1+e^{\beta {E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}}-\frac{1_{y-x>
0}}{1+e^{-\beta {E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}}\right)
\end{split}
\end{equation*}
for any $({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)$, $p,q\in\{1,\cdots,d\}$ and $z,w\in {\mathbb{C}}}% \C == \mathbb{C$. Moreover, if $x,y\in[0,\beta)_h$ we write
$$C_h({\mathbf{x}}\xi x,{\mathbf{y}}\phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q)=C({\mathbf{x}}\xi x,{\mathbf{y}}\phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q).$$The properties of $C(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ are summarized as follows.
\begin{lemma}\label{lem_properties_covariance}
\begin{enumerate}[(i)]
\item\label{item_analyticity_covariance} For any $({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)$, $p,q\in\{1,\cdots,d\}$, the function $(z,w)\mapsto C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ is analytic in the domain
$$\left\{(z,w)\in{\mathbb{C}}}% \C == \mathbb{C^2\ \Big|\ |\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|<\frac{1}{2}\log F_{t,t',d}\left(\frac{\pi}{\beta}\right)\right\}.$$
\item\label{item_bound_covariance} The inequality $|C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q)|\le 1$ holds for any $({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)$, $p,q\in\{1,\cdots,d\}$ and $z,w\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|\le \frac{1}{2}\log F_{t,t',d}(\pi/(2\beta))$.
\item\label{item_determinantbound_covariance} For any $z,w\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|\le\frac{1}{2}\log F_{t,t',d}(\pi/(2\beta))$ and $p,q\in$\\
$\{1,\cdots,d\}$, the covariance $C(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ satisfies the determinant bound \eqref{eq_extended_determinant_bound}.
\item\label{item_determinant_covariance} For any $p,q\in\{1,\cdots,d\}$, $z,w\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|< \frac{1}{2}\log F_{t,t',d}(\pi/\beta)$ and $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$,
$$\det(C_h({\mathbf{x}}\xi x,{\mathbf{y}} \phi y))_{({\mathbf{x}},\xi, x),({\mathbf{y}},\phi, y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}=\prod_{{\mathbf{k}}\in\G^*}\frac{1}{(1+e^{\beta {E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}})^2}\neq 0.$$
\end{enumerate}
\end{lemma}
\begin{remark}
In \eqref{item_determinant_covariance} and in its proof below let us think that each element of $\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h$ has been given a number from $1$ to $N(=2\beta h L^d)$ and according to this numbering the $N\times N$ matrix $C_h(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ is defined. However, the claim \eqref{item_determinant_covariance} implies that the value of the determinant is independent of how to number the elements of $\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h$.
\end{remark}
\begin{proof}[Proof of Lemma \ref{lem_properties_covariance}]
If $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|<\frac{1}{2}\log F_{t,t',d}(\pi/\beta)$, Lemma \ref{lem_dispersion_relation_condition} ensures that $|\mathop{\mathrm{Im}} {E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}|<\pi/\beta$. Then, for all ${\mathbf{k}}\in\G^*$ and $a\in\{1,-1\}$
$$|1+e^{\beta a E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}|^2>\left(e^{\beta a\mathop{\mathrm{Re}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}-1\right)^2\ge 0.$$
Thus, the denominators in $C({\mathbf{x}}\xi x, {\mathbf{y}}\phi y)(z{\mathbf{e}}_p+w{\mathbf{e}}_q)$ do not vanish, which proves \eqref{item_analyticity_covariance}.
If $|\mathop{\mathrm{Im}} z|,|\mathop{\mathrm{Im}} w|\le\frac{1}{2}\log F_{t,t',d}(\pi/(2\beta))$, by Lemma \ref{lem_dispersion_relation_condition} $|\mathop{\mathrm{Im}} {E}_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}|\le$\\
$\pi/(2\beta)$. Thus, the claim \eqref{item_determinantbound_covariance} holds true by Proposition \ref{prop_extended_determinant_bound}. Moreover, for any $x\in [0,\beta)$ and ${\mathbf{k}}\in\G^*$
$$\left|\frac{e^{xE_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}}{1+e^{\beta E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}}\right|\le \frac{e^{x\mathop{\mathrm{Re}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}}}{(1+e^{2\beta \mathop{\mathrm{Re}} E_{{\mathbf{k}}+z{\mathbf{e}}_p+w{\mathbf{e}}_q}})^{1/2}}\le 1,$$
from which the claim \eqref{item_bound_covariance} follows.
To show \eqref{item_determinant_covariance}, define an $h$-dependent finite set $M_h$ of Matsubara frequencies by
$$M_h:=\left\{{\omega} \in \frac{\pi (2{\mathbb{Z}}}% \Z == \mathbb{Z+1)}{\beta}\ \big|\ -\pi h< {\omega} < \pi h\right\}.$$
Define the unitary matrix $Y=$\\
$(Y({\mathbf{k}}\phi{\omega},{\mathbf{x}}\xi x))_{({\mathbf{k}},\phi,{\omega})\in\G^*\times{\{\uparrow,\downarrow\}}\times M_h, ({\mathbf{x}},\xi, x)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)_h}$ by
$$Y({\mathbf{k}}\phi{\omega},{\mathbf{x}}\xi x):=\frac{\delta_{\phi,\xi}}{\sqrt{\beta hL^d}}e^{i\<{\mathbf{k}},{\mathbf{x}}\>}e^{-i{\omega} x}.$$
Using the assumption that $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$, the argument parallel to \cite[\mbox{Appendix C}]{K} demonstrates that
\begin{equation}\label{eq_diagonalization_covariance}
(YC_hY^*)({\mathbf{k}}\phi {\omega},\hat{{\mathbf{k}}}\hat{\phi}\hat{{\omega}})=\delta_{{\mathbf{k}},\hat{{\mathbf{k}}}}\delta_{\phi,\hat{\phi}}\delta_{{\omega},\hat{{\omega}}}\frac{1}{1-e^{-i{\omega}/h+E_{{\mathbf{k}}+z{\mathbf{e}}_p + w{\mathbf{e}}_q}/h}}.
\end{equation}
As in \cite[\mbox{Proposition C.7}]{K}, the diagonalization \eqref{eq_diagonalization_covariance} deduces \eqref{item_determinant_covariance}.
\end{proof}
Here we present a calculation, which shows the essence of our analysis to bound the correlation functions. Take any $n\in{\mathbb{N}}}% \N == \mathbb{N$ and set $r_n:=\frac{1}{2n}\log F_{t,t',d}(\beta/(2\pi))>0$.
By periodicity we observe that for any $p,q\in\{1,\cdots,d\}$ and $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|\le \frac{1}{2}\log F_{t,t',d}(\beta/(2\pi))$
\begin{equation}\label{eq_reform_covariance}
e^{i\frac{2\pi}{L}\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>}C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)=C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)\left(z{\mathbf{e}}_p+\frac{2\pi}{L}{\mathbf{e}}_q\right).
\end{equation}
Moreover, by Cauchy's integral formula we have
\begin{equation*}
\begin{split}
\frac{e^{i\frac{2\pi}{L}\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>}-1}{2\pi/L}&C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)=\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta\frac{d}{d\theta}C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p+\theta {\mathbf{e}}_q)\\
&=\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta\frac{1}{2\pi i}\oint_{|w-\theta|=r_n}dw\frac{C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p+w {\mathbf{e}}_q)}{(w-\theta)^2},
\end{split}
\end{equation*}
where $\oint_{|w-\theta|=r_n}dw$ stands for the contour integral along the contour $\{w\in{\mathbb{C}}}% \C == \mathbb{C\ |\ |w-\theta|=r_n\}$ oriented counter clock-wise. With this choice of $r_n$, Lemma \ref{lem_properties_covariance} \eqref{item_analyticity_covariance} allows us to repeat this operation $n$ times to obtain
\begin{lemma}\label{lem_covariance_integral_formula}
For any $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|\le \frac{1}{2}\log F_{t,t',d}(\beta/(2\pi))$ and $p,q\in \{1,\cdots,d\}$
\begin{equation}\label{eq_covariance_integral_formula}
\begin{split}
&\left(\frac{e^{i\frac{2\pi}{L}\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>}-1}{2\pi/L}\right)^nC({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)\\
&\quad=\prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|w_j-\theta_j|=r_n}dw_j\frac{1}{(w_j-\theta_j)^2}\right)\\
&\quad\qquad\cdot C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)\left(z{\mathbf{e}}_p+\sum_{j=1}^nw_j{\mathbf{e}}_q\right),
\end{split}
\end{equation}
where $r_n=\frac{1}{2n}\log F_{t,t',d}(\beta/(2\pi))$.
\end{lemma}
Evaluating both sides of \eqref{eq_covariance_integral_formula} leads to the following bounds.
\begin{proposition}\label{prop_covariance_exponential_decay}
For any $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|\le \frac{1}{2}\log F_{t,t',d}(\beta/(2\pi))$, $p\in\{1,\cdots,d\}$, the following inequalities hold.
\begin{equation}\label{eq_covariance_exponential_decay}
|C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)|\le 2 F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\sum_{q=1}^d\left|\frac{e^{i2\pi\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>/L}-1}{2\pi/L}\right|}
\end{equation}
for all $({\mathbf{x}},\xi,x),({\mathbf{y}},\phi,y)\in\G\times{\{\uparrow,\downarrow\}}\times[0,\beta)$. Especially,
\begin{equation}\label{eq_covariance_exponential_decay_limit}
|C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)|\le2 F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{2e \pi d}\sum_{q=1}^d|\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>|}
\end{equation}
if ${\mathbf{x}}-{\mathbf{y}}\in (\{-\lfloor L/2 \rfloor,-\lfloor L/2\rfloor + 1,\cdots,-\lfloor L/2\rfloor+L-1\})^d$.
\end{proposition}
\begin{proof}
By using Lemma \ref{lem_properties_covariance} \eqref{item_bound_covariance} and the inequality
\begin{equation}\label{eq_evaluation_multi_integral}
\left|\prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|w_j-\theta_j|=r_n}dw_j\frac{1}{(w_j-\theta_j)^2}\right)\right|\le\left(\frac{2n}{\log F_{t,t',d}(\pi/(2\beta))}\right)^n,\end{equation}
we have
\begin{equation}\label{eq_covariance_difference_evaluation}
|(\text{The right hand side of \eqref{eq_covariance_integral_formula}})|\le\left(\frac{2n}{\log F_{t,t',d}(\pi/(2\beta))}\right)^n.
\end{equation}
Stirling's formula states that for any $n\in {\mathbb{N}}}% \N == \mathbb{N$ there exists $\eps(n)\in (0,1)$ such that
\begin{equation}\label{eq_stirling_formula}
n^n=n!(2\pi n)^{-1/2}e^{n-\eps(n)/(12n)}
\end{equation}
(see, e.g, \cite[\mbox{Eq.6.1.38}]{AS}).
On the convention that $0^0=0!=1$, the formula \eqref{eq_stirling_formula} gives the inequality
\begin{equation}\label{eq_stirling_inequality}
n^n\le n!e^n \ (\forall n\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}).
\end{equation}
By combining \eqref{eq_stirling_inequality} with \eqref{eq_covariance_difference_evaluation} we obtain
\begin{equation}\label{eq_covariance_weight_evaluation}
\frac{1}{n!}\left(\frac{\log F_{t,t',d}(\pi/(2\beta))}{2e}\right)^n\left|\frac{e^{i\frac{2\pi}{L}\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>}-1}{2\pi/L}\right|^n|C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)|\le 1.
\end{equation}
By multiplying both sides of \eqref{eq_covariance_weight_evaluation} by $2^{-n}$ and summing over ${\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$, we have
$$
|C({\mathbf{x}}\xi x,{\mathbf{y}} \phi y)(z{\mathbf{e}}_p)|\le 2 F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4e}\left|\frac{e^{i2\pi\<{\mathbf{x}}-{\mathbf{y}},{\mathbf{e}}_q\>/L}-1}{2\pi/L}\right|}
$$
for any $q\in\{1,\cdots,d\}$, which yields \eqref{eq_covariance_exponential_decay}. By applying the inequality that $|e^{i\theta}-1|\ge 2|\theta|/\pi$ $(\forall \theta\in [-\pi,\pi])$ to \eqref{eq_covariance_exponential_decay}, we can derive \eqref{eq_covariance_exponential_decay_limit}.
\end{proof}
\begin{corollary}\label{cor_covariance_L1_bound}
For any $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|\le \frac{1}{2}\log F_{t,t',d}(\beta/(2\pi))$ and $p\in\{1,\cdots,d\}$,
\begin{equation}\label{eq_covariance_L1_bound}
\frac{1}{h}\sum_{x\in[-\beta,\beta)_h}\sum_{{\mathbf{x}}\in\G}|C_h({\mathbf{x}}\xi x,{\mathbf{0}}\xi 0)(z{\mathbf{e}}_p)| \le 4\beta \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{d}.
\end{equation}
\end{corollary}
\begin{proof}
By using \eqref{eq_covariance_exponential_decay_limit} and periodicity we have that
\begin{equation*}
\begin{split}
&\sum_{{\mathbf{x}}\in\G}|C({\mathbf{x}}\xi x,{\mathbf{0}} \xi 0)(z{\mathbf{e}}_p)|=\sum_{x_j=-\lfloor L/2\rfloor\atop \forall j\in\{1,\cdots,d\}}^{-\lfloor L/2\rfloor+L-1}|C({\mathbf{x}}\xi x,{\mathbf{0}} \xi 0)(z{\mathbf{e}}_p)|\\
&\quad\le 2\left(1+ 2\sum_{l=1}^{\infty}F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{l}{2e\pi d}}\right)^d= 2\left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d,
\end{split}
\end{equation*}
which gives \eqref{eq_covariance_L1_bound}.
\end{proof}
\begin{remark}
Remark \ref{rem_beta_dependency} and the inequality \eqref{eq_covariance_L1_bound} imply that
$$\frac{1}{h}\sum_{x\in [-\beta,\beta)_h}\sum_{{\mathbf{x}}\in\G}|C_h({\mathbf{x}}\xi x,{\mathbf{0}}\xi 0)(z{\mathbf{e}}_p)| =O(\beta^{d+1})$$
as $\beta \to +\infty$.
\end{remark}
\begin{remark}
It is the same procedure as the proof of Proposition \ref{prop_covariance_exponential_decay} to derive the decay bound on the determinant of the covariance. By using the equality that
\begin{equation*}
\begin{split}
&e^{i\frac{2\pi}{L}\<\sum_{j=1}^{n}{\mathbf{x}}_j-\sum_{j=1}^{n}{\mathbf{y}}_j,{\mathbf{e}}_q\>}\det(C({\mathbf{x}}_j\xi_jx_j,{\mathbf{y}}_k\phi_ky_k))_{1\le j,k\le n}\\
&\qquad=
\det\left(C({\mathbf{x}}_j\xi_jx_j,{\mathbf{y}}_k\phi_ky_k)\left(\frac{2\pi}{L}{\mathbf{e}}_q\right)\right)_{1\le j,k\le n}
\end{split}
\end{equation*}
and the determinant bound Proposition \ref{prop_extended_determinant_bound} one can prove that for any $({\mathbf{x}}_j,\xi_j,x_j)$,\\
$({\mathbf{y}}_j,\phi_j,y_j)\in\G\times{\{\uparrow,\downarrow\}}\times [0,\beta)$ $(\forall j\in\{1,\cdots,n\})$
\begin{equation*}
\begin{split}
&\left|\det(C({\mathbf{x}}_j\xi_jx_j,{\mathbf{y}}_k\phi_ky_k))_{1\le j,k\le n}\right|\\
&\quad\le 2\cdot 4^{n}F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\sum_{q=1}^d\left|\frac{e^{i2\pi\<\sum_{j=1}^n{\mathbf{x}}_j-\sum_{j=1}^n{\mathbf{y}}_j,{\mathbf{e}}_q\>/L}-1}{2\pi/L}\right|}.
\end{split}
\end{equation*}
\end{remark}
\subsection{Proof of the main theorems}\label{subsec_exponential_decay_correlation}
We prepare some lemmas and give the proofs of Theorem \ref{thm_exponential_decay} and Theorem \ref{thm_exponential_decay_hubbard_model}.
The following lemma fundamentally supports the validity of our argument in this subsection.
\begin{lemma}\label{lem_schwinger_functional_analytic}
For any $R_r>0$, $R_i\in (0,\frac{1}{2}\log F_{t,t',d}(\pi/\beta))$ and $p\in\{1,\cdots,d\}$ there exists $Q>0$ such that the function
$$(z,\eta)\longmapsto S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h(z{\mathbf{e}}_p),\eta)$$
is analytic in the domain
$$\{ (z,\eta)\in{\mathbb{C}}}% \C == \mathbb{C^2\ |\ |\mathop{\mathrm{Re}} z|<R_r, \ |\mathop{\mathrm{Im}} z|<R_i,\ |\eta|<Q\}.$$
\end{lemma}
\begin{proof}
Note that by Lemma \ref{lem_properties_covariance} \eqref{item_analyticity_covariance} for any $R_r>0$ and $R_i\in (0,\frac{1}{2}\log F_{t,t',d}(\pi/\beta))$ the function $z\mapsto C_h({\mathbf{x}}\xi x,{\mathbf{y}}\phi y)(z{\mathbf{e}}_p)$ is bounded in the compact set $\{z\in{\mathbb{C}}}% \C == \mathbb{C\ |\ |\mathop{\mathrm{Re}} z|\le R_r,\ |\mathop{\mathrm{Im}} z|\le R_i\}$.
Thus, by Lemma \ref{lem_grassmann_gaussian_equality}
\begin{equation*}
\begin{split}
&\lim_{\eta\to 0}\sup_{z\in{\mathbb{C}}}% \C == \mathbb{C\atop |\mathop{\mathrm{Re}} z|\le R_r,\ |\mathop{\mathrm{Im}} z|\le R_i}\\
&\cdot\Bigg|\int e^{\eta\sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{C_h(z{\mathbf{e}}_p)}({\overline{\psi}}_X,\psi_X)\\
&\qquad-1\Bigg|=0.
\end{split}
\end{equation*}
Hence, we can take a small $Q>0$ such that the denominator of
$S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}$\\
$(C_h(z{\mathbf{e}}_p),\eta)$ does not vanish for any $\eta\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\eta|<Q$ and $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Re}} z|< R_r$, $|\mathop{\mathrm{Im}} z|< R_i$. This proves the analyticity of $S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h(z{\mathbf{e}}_p),\eta)$ in the same domain.
\end{proof}
For $p\in\{1,\cdots,d\}$ we define the matrix ${\mathcal{U}}_p=({\mathcal{U}}_p(({\mathbf{x}}\xi x)_j,({\mathbf{x}}\xi x)_k))_{1\le j,k\le N}$ by
$${\mathcal{U}}_p({\mathbf{x}}\xi x,{\mathbf{y}} \phi y):=\delta_{{\mathbf{x}},{\mathbf{y}}}\delta_{\xi,\phi}\delta_{x,y}e^{i2\pi\<{\mathbf{x}},{\mathbf{e}}_p\>/L}.$$
The following equality is based on the $U(1)$-invariance property of the Grassmann integral $\int\cdot d\psi_Xd{\overline{\psi}}_X$.
\begin{lemma}\label{lem_U_1_symmetry}
For any $p\in\{1,\cdots,d\}$, $f({\overline{\psi}}_X,\psi_X)\in{\mathbb{C}}}% \C == \mathbb{C[{\overline{\psi}}_{({\mathbf{x}}\xi x)_j},\psi_{({\mathbf{x}}\xi x)_j}\ |\ j\in\{1,\cdots,N\}]$ and $z\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\mathop{\mathrm{Im}} z|\le\frac{1}{2}\log F_{t,t',d}(\pi/(2\beta))$, the following equality holds.
\begin{equation*}
\begin{split}
\int& f((\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t)^t,({\mathcal{U}}_p\psi_X^t)^t)d \mu_{C_h((z+2\pi/L){\mathbf{e}}_p)}({\overline{\psi}}_X,\psi_X)\\
&=\int f({\overline{\psi}}_X,\psi_X)d\mu_{C_h(z{\mathbf{e}}_p)}({\overline{\psi}}_X,\psi_X).
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
Observe the invariance that
$$\int f((\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t)^t,({\mathcal{U}}_p\psi_X^t)^t)d\psi_Xd{\overline{\psi}}_X=\int f({\overline{\psi}}_X,\psi_X)d\psi_Xd{\overline{\psi}}_X,$$
which implies that
\begin{equation}\label{eq_grassmann_reform}
\begin{split}
\int& f({\overline{\psi}}_X,\psi_X)d\mu_{C_h(z{\mathbf{e}}_p)}({\overline{\psi}}_X,\psi_X)\\
&=\int f((\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t)^t,({\mathcal{U}}_p\psi_X^t)^t)e^{-\<{{\mathcal{U}}_p}\psi_X^t,C_h(z{\mathbf{e}}_p)^{-1}\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t\>}d\psi_Xd{\overline{\psi}}_X\\
&\quad\cdot\Bigg/\int e^{-\<{{\mathcal{U}}_p}\psi_X^t,C_h(z{\mathbf{e}}_p)^{-1}\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t\>}d\psi_Xd{\overline{\psi}}_X.
\end{split}
\end{equation}
Moreover, by using the equality \eqref{eq_reform_covariance} we have
\begin{equation}\label{eq_covariance_grassmann_reform}
\begin{split}
\<{{\mathcal{U}}_p}\psi_X^t,C_h(z{\mathbf{e}}_p)^{-1}\overline{{\mathcal{U}}_p}{\overline{\psi}}_X^t\>&=\<\psi_X^t,({{\mathcal{U}}_p}C_h(z{\mathbf{e}}_p)\overline{{\mathcal{U}}_p})^{-1}{\overline{\psi}}_X^t\>\\
&=\<\psi_X^t,C_h((z+2\pi/L){\mathbf{e}}_p)^{-1}{\overline{\psi}}_X^t\>.
\end{split}
\end{equation}
By substituting \eqref{eq_covariance_grassmann_reform} into \eqref{eq_grassmann_reform}, we obtain the desired equality.
\end{proof}
Combining Lemma \ref{lem_schwinger_functional_analytic} with Lemma \ref{lem_U_1_symmetry} shows
\begin{lemma}\label{lem_schwinger_functional_formula}
For any $n\in{\mathbb{N}}}% \N == \mathbb{N$, set $r_n:=\frac{1}{2n}\log F_{t,t',d}(\pi/(2\beta))$. Let $p\in \{1,\cdots,d\}$. There exists $Q_n>0$ such that for any $\eta\in{\mathbb{C}}}% \C == \mathbb{C$ with $|\eta|<Q_n$,
\begin{equation}\label{eq_schwinger_functional_formula}
\begin{split}
&\left(\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,\eta)\\
&\quad = \prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|z_j-\theta_j|=r_n}dz_j\frac{1}{(z_j-\theta_j)^2}\right)\\
&\qquad\qquad\cdot S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h\left(\sum_{j=1}^nz_j{\mathbf{e}}_p\right),\eta\right),
\end{split}
\end{equation}
where $\oint_{|z_j-\theta_j|=r_n}dz_j$ stands for the contour integral along the contour $\{z_j\in{\mathbb{C}}}% \C == \mathbb{C\ |\ |z_j-\theta_j|=r_n\}$ oriented counter clock-wise.
\end{lemma}
\begin{proof}
Take any $R_i\in (\frac{1}{2}\log F_{t,t',d}(\pi/(2\beta)),\frac{1}{2}\log F_{t,t',d}(\pi/\beta))$ and $R_r>nr_n+\frac{2\pi n}{L}$. By Lemma \ref{lem_schwinger_functional_analytic} we can take a small $Q_n>0$ so that $(z,\eta)\mapsto S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}$\\$(C_h(z{\mathbf{e}}_p),\eta)$ is analytic in the domain $\{(z,\eta)\in{\mathbb{C}}}% \C == \mathbb{C^2\ |\ |\mathop{\mathrm{Re}} z|<R_r,\ |\mathop{\mathrm{Im}} z|<R_i,\ \ |\eta|<Q_n\}$.
Fix any $\eta \in{\mathbb{C}}}% \C == \mathbb{C$ with $|\eta|<Q_n$. By noting this domain of analyticity, Lemma \ref{lem_U_1_symmetry}, the fundamental theorem of calculus and Cauchy's integral theorem verify the following equalities.
\begin{equation}\label{eq_schwinger_integral_1}
\begin{split}
&\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,\eta)\\
&\quad= \frac{L}{2\pi}\left(S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h\left(\frac{2\pi}{L}{\mathbf{e}}_p\right),\eta\right)-S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h,\eta\right)\right)\\
&\quad= \frac{L}{2\pi}\int^{2\pi/L}_0d\theta_1\frac{d}{d\theta_1}S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h\left(\theta_1{\mathbf{e}}_p\right),\eta\right)\\
&\quad= \frac{L}{2\pi}\int^{2\pi/L}_0d\theta_1\frac{1}{2\pi i}\oint_{|z_1-\theta_1|=r_n}dz_1 \frac{1}{(z_1-\theta_1)^2}S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h\left(z_1{\mathbf{e}}_p\right),\eta\right).
\end{split}
\end{equation}
Then, we multiply both sides of \eqref{eq_schwinger_integral_1} by $(e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1)/(2\pi/L)$ and do the same calculation for the integrand $S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}\left(C_h\left(z_1{\mathbf{e}}_p\right),\eta\right)$. Repeating this procedure $n$ times results in \eqref{eq_schwinger_functional_formula}.
\end{proof}
\begin{lemma}\label{lem_schwinger_functional_integral_bound}
For any $m,n\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$ let $b_{n,m}$ denote the coefficient of $\eta^m$ in the Taylor series of the function
$$\eta\longmapsto\left(\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,\eta)$$
around $0\in{\mathbb{C}}}% \C == \mathbb{C$. The following bounds hold. For any $n\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$, $m\in{\mathbb{N}}}% \N == \mathbb{N$,
\begin{equation*}
\begin{split}
&|b_{n,0}|\le 2^{n}n^n\left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}4^{{\hat{m}}},\\
&|b_{n,m}|\le 2^{n}n^n\left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}{\hat{m}} 16^{{\hat{m}}}\\
&\qquad\qquad\cdot\frac{1}{m}\left(\beta\left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d\sum_{l=1}^{{\tilde{n}}}l 16^{l}\|U_{L,l}\|_{L,l}\right)^m,
\end{split}
\end{equation*}
where $0^0=1$.
\end{lemma}
\begin{proof}
We apply Proposition \ref{prop_schwinger_functional_bound} to bound the integrand\\ $S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h(\sum_{j=1}^nz_j{\mathbf{e}}_p),\eta)$ in \eqref{eq_schwinger_functional_formula}. Since $|\mathop{\mathrm{Im}} \sum_{j=1}^nz_j|\le$\\
$\frac{1}{2}\log F_{t,t',d}(\pi/(2\beta))$, Lemma \ref{lem_properties_covariance} \eqref{item_determinantbound_covariance}, \eqref{item_determinant_covariance} imply that the covariance $C_h(\sum_{j=1}^nz_j{\mathbf{e}}_p)$ satisfies the assumptions of Proposition \ref{prop_schwinger_functional_bound} with ${\mathcal{B}}=4$. Moreover, Corollary \ref{cor_covariance_L1_bound}, the periodicity and the translation invariance of $C_h$ ensure that
$${\mathcal{D}}\le 4\beta\left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d.$$
Therefore, if $\tilde{b}_m$ denotes the coefficient of $\eta^m$ $(\forall m\in{\mathbb{N}}}% \N == \mathbb{N\cup \{0\})$ in the Taylor series of the function
$\eta\mapsto S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h(\sum_{j=1}^nz_j{\mathbf{e}}_p),\eta)$, we obtain the inequalities that $|\tilde{b}_0|\le 4^{{\hat{m}}}$,
\begin{equation}\label{eq_each_order_bound}
|\tilde{b}_m|\le \frac{{\hat{m}} 16^{{\hat{m}}}}{m}\left(\beta \left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d\sum_{l=1}^{{\tilde{n}}}l16^{l}\|U_{L,l}\|_{L,l}\right)^m\ (\forall m\in{\mathbb{N}}}% \N == \mathbb{N).
\end{equation}
In the right hand side of \eqref{eq_schwinger_functional_formula} we expand $S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h(\sum_{j=1}^nz_j{\mathbf{e}}_p),\eta)$ into the power series of $\eta$. Here note that by Fubini-Tonelli's theorem we may exchange the integral $\prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|z_j-\theta_j|=r_n}dz_j\frac{1}{(z_j-\theta_j)^2}\right)\sum_{m=0}^{\infty}$ by\\
$\sum_{m=0}^{\infty}\prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|z_j-\theta_j|=r_n}dz_j\frac{1}{(z_j-\theta_j)^2}\right)$ if $|\eta|$ is sufficiently small, and thus the equality
$$b_{n,m}=\prod_{j=1}^n\left(\frac{L}{2\pi}\int_{0}^{2\pi/L}d\theta_j\frac{1}{2\pi i}\oint_{|z_j-\theta_j|=r_n}dz_j\frac{1}{(z_j-\theta_j)^2}\right)\tilde{b}_m\ $$
holds for all $m\in{\mathbb{N}}}% \N == \mathbb{N\cup \{0\}$. Combining the inequality \eqref{eq_evaluation_multi_integral} with the bounds \eqref{eq_each_order_bound} yields the desired bound on $|{b}_{n,m}|$.
\end{proof}
\begin{corollary}\label{cor_analytic_continuation}
Assume \eqref{eq_assumption_U_convergence} with $R\in (0,1)$ and use the same notation as in Lemma \ref{lem_schwinger_functional_integral_bound}. For any $n\in {\mathbb{N}}}% \N == \mathbb{N\cup \{0\}$ and $p\in \{1,\cdots,d\}$ there exists $N_0\in{\mathbb{N}}}% \N == \mathbb{N$ such that
\begin{equation}\label{eq_analytic_continuation}
\left(\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,1)=\sum_{m=0}^{\infty}{b}_{n,m},
\end{equation}
and
\begin{equation}\label{eq_analytic_continuation_bound}
\begin{split}
& \left|\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right|^n\left|S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,1)\right|\\
&\quad\le 2^{n}n^n \left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}\left(4^{{\hat{m}}}-\hm16^{{\hat{m}}}\log(1-R)\right),
\end{split}
\end{equation}
for all $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ with $h\ge 2N_0/\beta$.
\end{corollary}
\begin{proof}
Since $\mathop{\mathrm{Tr}} e^{-\beta(H_0+\eta V)}/\mathop{\mathrm{Tr}} e^{-\beta H_0}>0$ for all $\eta\in{\mathbb{R}}}% \R == \mathbb{R$, the uniform convergence property \eqref{eq_convergence_partition_function} ensures that there exists $N_0\in {\mathbb{N}}}% \N == \mathbb{N$ such that
\begin{equation}\label{eq_regular_inequality}
\left|\int e^{\eta \sum_{l=1}^{{\tilde{n}}}\sum_{(X^l,\Xi^l,\Phi^l)\in\G^{l}\times{\{\uparrow,\downarrow\}}^{2l}}U_{L,l}(X^l,\Xi^l,\Phi^l)V_{h,X^l,X^l,\Xi^l,\Phi^l}^l({\overline{\psi}}_X,\psi_X)}d\mu_{C_h}({\overline{\psi}}_X,\psi_X)\right|>0
\end{equation}
for all $\eta\in{\mathbb{R}}}% \R == \mathbb{R$ with $|\eta|\le 1$ and $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ with $h\ge 2N_0/\beta$. Let us fix such a large $h$. Since the Grassmann Gaussian integral in \eqref{eq_regular_inequality} is a polynomial of $\eta$, we can take a domain $O_h\subset {\mathbb{C}}}% \C == \mathbb{C$ such that $[-1,1]\subset O_h$ and the inequality \eqref{eq_regular_inequality} holds for all $\eta\in O_h$. This proves that the function
\begin{equation*}
\eta\longmapsto \left(\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,\eta)
\end{equation*}
is analytic in the domain $O_h$.
On the other hand, Lemma \ref{lem_schwinger_functional_integral_bound} and the inequality $\|U_{L,l}\|_{L,l}\le \|U_l\|_l$ imply that if $\eta\in{\mathbb{C}}}% \C == \mathbb{C$ satisfies
\begin{equation}\label{eq_condition_eta}
|\eta|<\left(\beta\left(\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d\sum_{l=1}^{{\tilde{n}}}l16^{l}\|U_l\|_l\right)^{-1}R,
\end{equation}
\begin{equation}\label{eq_bound_sum_eta}
\begin{split}
\sum_{m=0}^{\infty}|{b}_{n,m}||\eta^m|&\le 2^{n}n^n
\left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}4^{{\hat{m}}}\\
&\quad+2^{n}n^n \left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}{\hat{m}} 16^{{\hat{m}}}\sum_{m=1}^{\infty}\frac{1}{m}R^m\\
&= 2^{n}n^n \left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}\left(4^{{\hat{m}}}-\hm16^{{\hat{m}}}\log\left(1-R\right)\right)<\infty.
\end{split}
\end{equation}
Since by the assumption \eqref{eq_assumption_U_convergence} the right hand side of \eqref{eq_condition_eta} is larger than $1$, the identity theorem for analytic functions proves the equality
\begin{equation*}
\left(\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,\eta)=\sum_{m=0}^{\infty}{b}_{n,m}\eta^m
\end{equation*}
for all $\eta\in [-1,1]$. The bound \eqref{eq_bound_sum_eta} for $\eta=1$ gives \eqref{eq_analytic_continuation_bound}.
\end{proof}
As the last lemma let us claim
\begin{lemma}\label{lem_thermodynamic_limit}
Assume \eqref{eq_assumption_U_convergence}. The thermodynamic limit \eqref{eq_thermodynamic_limit} exists and takes a finite value.
\end{lemma}
For continuity of our argument in this subsection we present the proof of Lemma \ref{lem_thermodynamic_limit} in Appendix \ref{app_thermodynamic_limit}. We now proceed to
\begin{proof}[Proof of Theorem \ref{thm_exponential_decay}]
By substituting \eqref{eq_stirling_inequality} into \eqref{eq_analytic_continuation_bound} and dividing both sides by $4^ne^nn!\cdot\left(\log F_{t,t',d}\left(\pi/(2\beta)\right)\right)^{-n}$ we have
\begin{equation}\label{eq_identity_bound}
\begin{split}
& \frac{1}{n!}\left(\frac{\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)}{4e}\left|\frac{e^{i\frac{2\pi}{L}\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>}-1}{2\pi/L}\right|\right)^n\left|S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,1)\right|\\
&\quad\le 2^{-n}\left(4^{{\hat{m}}}-\hm16^{{\hat{m}}}\log(1-R)\right)
\end{split}
\end{equation}
for all $n\in {\mathbb{N}}}% \N == \mathbb{N\cup \{0\}$ and $p\in\{1,\cdots,d\}$. Summing \eqref{eq_identity_bound} over $n\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$ yields
\begin{equation}\label{eq_schwinger_pre_bound}
\begin{split}
\left|S_{{\hat{X}}^{{\hat{m}}},{\hat{Y}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}}}(C_h,1)\right|\le& 2\left(4^{{\hat{m}}}-{\hat{m}} 16^{{\hat{m}}}\log(1-R)\right)\\
&\cdot F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\sum_{p=1}^d\left|\frac{e^{i2\pi\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>/L}-1}{2\pi/L}\right|}
\end{split}
\end{equation}
for all $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ with $h\ge 2N_0/\beta$.
To derive the same bound on $S_{{\hat{Y}}^{{\hat{m}}},{\hat{X}}^{{\hat{m}}},{\hat{\Phi}}^{{\hat{m}}},\hat{\Xi}^{{\hat{m}}}}$\\
$(C_h,1)$ from \eqref{eq_schwinger_pre_bound} is immediate. Then, Proposition \ref{prop_correlation_grassmann_integral} ensures that
\begin{equation*}
\begin{split}
&\left|\<\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}^*\cdots\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}+\psi_{{\hat{\mathbf{y}}}_{1}{\hat{\phi}}_{1}}^*\cdots\psi_{{\hat{\mathbf{y}}}_{{\hat{m}}}{\hat{\phi}}_{{\hat{m}}}}^*\psi_{{\hat{\mathbf{x}}}_{{\hat{m}}}\hat{\xi}_{{\hat{m}}}}\cdots\psi_{{\hat{\mathbf{x}}}_1\hat{\xi}_1}\>_L\right|\\
&\quad\le (4^{{\hat{m}}+1}-{\hat{m}} 4^{2{\hat{m}}+1}\log(1-R))\\
&\qquad\cdot F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\sum_{p=1}^d\left|\frac{e^{i2\pi\<\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{x}}}_j-\sum_{j=1}^{{\hat{m}}}{\hat{\mathbf{y}}}_j,{\mathbf{e}}_p\>/L}-1}{2\pi/L}\right|}.
\end{split}
\end{equation*}
Finally, by Lemma \ref{lem_thermodynamic_limit} we can take the limit $L\to +\infty$ in the both sides to obtain the claimed inequality.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm_exponential_decay_hubbard_model}]
Since the construction of Theorem \ref{thm_exponential_decay_hubbard_model} is parallel to that of Theorem \ref{thm_exponential_decay}, we only outline the proof. The main difference is that we use Proposition \ref{prop_schwinger_functional_hubbard_model_bound} in place of Proposition \ref{prop_schwinger_functional_bound}.
(Step 1) For any $m,n\in{\mathbb{N}}}% \N == \mathbb{N\cup\{0\}$ let $c_{n,m}$ denote the coefficient of $\eta^m$ in the Taylor expansion of
\begin{equation*}
\eta\longmapsto \left(\frac{e^{i\frac{2\pi}{L}\<{\hat{\mathbf{x}}}_1+{\hat{\mathbf{x}}}_2-{\hat{\mathbf{y}}}_1-{\hat{\mathbf{y}}}_2,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{({\hat{\mathbf{x}}}_1,{\hat{\mathbf{x}}}_2), ({\hat{\mathbf{y}}}_1,{\hat{\mathbf{y}}}_2),({\uparrow},{\downarrow}), ({\uparrow},{\downarrow})}(C_h,\eta)
\end{equation*}
around $0\in{\mathbb{C}}}% \C == \mathbb{C$. By repeating the same argument as in Lemma \ref{lem_schwinger_functional_integral_bound} using Proposition \ref{prop_schwinger_functional_hubbard_model_bound} in stead of Proposition \ref{prop_schwinger_functional_bound} we obtain
\begin{equation*}
\begin{split}
|c_{n,m}|\le &2^n n^n \left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}\\
&\cdot\frac{64}{3m+4}\left(\begin{array}{c} 3m+4 \\ m\end{array}\right)\left(16\beta \left(
\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^d|U|\right)^m.
\end{split}
\end{equation*}
(Step 2) Recall the fact that the radius of convergence of the power series
$$\sum_{m=0}^{\infty}\frac{4}{3m+4}\left(\begin{array}{c} 3m+4 \\ m\end{array}\right)x^m$$
is $4/27$ and
\begin{equation*}
\sum_{m=0}^{\infty}\frac{4}{3m+4}\left(\begin{array}{c} 3m+4 \\ m\end{array}\right)\left(\frac{4}{27}\right)^m=\frac{81}{16}
\end{equation*}
(see \cite[\mbox{Lemma 4.9}]{K}).
Therefore, the argument involving the identity theorem parallel to the proof of Corollary \ref{cor_analytic_continuation} shows that for all $U\in{\mathbb{R}}}% \R == \mathbb{R$ satisfying
\begin{equation*}
|U|<(108\beta)^{-1}\left(
\frac{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}+1}{F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{1/(2e\pi d)}-1}\right)^{-d}
\end{equation*}
and for sufficiently large $h\in 2{\mathbb{N}}}% \N == \mathbb{N/\beta$ the equality
\begin{equation*}
\left(\frac{e^{i\frac{2\pi}{L}\<{\hat{\mathbf{x}}}_1+{\hat{\mathbf{x}}}_2-{\hat{\mathbf{y}}}_1-{\hat{\mathbf{y}}}_2,{\mathbf{e}}_p\>}-1}{2\pi/L}\right)^nS_{({\hat{\mathbf{x}}}_1,{\hat{\mathbf{x}}}_2),({\hat{\mathbf{y}}}_1,{\hat{\mathbf{y}}}_2),({\uparrow},{\downarrow}),({\uparrow},{\downarrow})}(C_h,1)=\sum_{m=0}^{\infty}{c}_{n,m}
\end{equation*}
and the inequality
\begin{equation}\label{eq_pre_bound_Hubbard}
\begin{split}
& \left|\frac{e^{i\frac{2\pi}{L}\<{\hat{\mathbf{x}}}_1+{\hat{\mathbf{x}}}_2-{\hat{\mathbf{y}}}_1-{\hat{\mathbf{y}}}_2,{\mathbf{e}}_p\>}-1}{2\pi/L}\right|^n\left|S_{({\hat{\mathbf{x}}}_1,{\hat{\mathbf{x}}}_2),({\hat{\mathbf{y}}}_1,{\hat{\mathbf{y}}}_2),({\uparrow},{\downarrow}),({\uparrow},{\downarrow})}\left(C_h,1\right)\right|\\
&\quad\le 2^{n}n^n \left(\log F_{t,t',d}\left(\frac{\pi}{2\beta}\right)\right)^{-n}\cdot 81
\end{split}
\end{equation}
hold. By using the fact that $S_{({\hat{\mathbf{x}}}_1,{\hat{\mathbf{x}}}_2),({\hat{\mathbf{y}}}_1,{\hat{\mathbf{y}}}_2),({\uparrow},{\downarrow}),({\uparrow},{\downarrow})}\left(C_h,1\right)$ is a rational function of $U$, we can prove \eqref{eq_pre_bound_Hubbard} for $U\in{\mathbb{R}}}% \R == \mathbb{R$ with \eqref{eq_assumption_U_convergence_hubbard_model} as well.
(Step 3) By repeating the same calculation as in the proof of Theorem \ref{thm_exponential_decay} or Proposition \ref{prop_covariance_exponential_decay} we reach the bound
\begin{equation}\label{eq_decay_bound_almost}
\begin{split}
&\left|\<\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}^*\psi_{{\hat{\mathbf{x}}}_{2}{\downarrow}}^*\psi_{{\hat{\mathbf{y}}}_{2}{\downarrow}}\psi_{{\hat{\mathbf{y}}}_{1}{\uparrow}}+\psi_{{\hat{\mathbf{y}}}_{1}{\uparrow}}^*\psi_{{\hat{\mathbf{y}}}_{2}{\downarrow}}^*\psi_{{\hat{\mathbf{x}}}_{2}{\downarrow}}\psi_{{\hat{\mathbf{x}}}_1{\uparrow}}\>_L\right|\\
&\quad\le 324\cdot F_{t,t',d}\left(\frac{\pi}{2\beta}\right)^{-\frac{1}{4ed}\sum_{p=1}^d\left|\frac{e^{i2\pi\<{\hat{\mathbf{x}}}_1+{\hat{\mathbf{x}}}_2-{\hat{\mathbf{y}}}_1-{\hat{\mathbf{y}}}_2,{\mathbf{e}}_p\>/L}-1}{2\pi/L}\right|}
\end{split}
\end{equation}
for $U\in{\mathbb{R}}}% \R == \mathbb{R$ satisfying \eqref{eq_assumption_U_convergence_hubbard_model} from \eqref{eq_pre_bound_Hubbard}.
(Step 4) By writing
\begin{equation*}
\begin{split}
&U\sum_{{\mathbf{x}}\in\G}\psi_{{\mathbf{x}}{\uparrow}}^*\psi_{{\mathbf{x}}{\downarrow}}^*\psi_{{\mathbf{x}}{\downarrow}}\psi_{{\mathbf{x}}{\uparrow}}\\
&\quad =\sum_{{\mathbf{x}}_1,{\mathbf{x}}_2\in\G}\sum_{\xi_1,\xi_2,\phi_1,\phi_2\in{\{\uparrow,\downarrow\}}}U_o(({\mathbf{x}}_1,{\mathbf{x}}_2),(\xi_1,\xi_2),(\phi_1,\phi_2))\psi_{{\mathbf{x}}_1\xi_1}^*\psi_{{\mathbf{x}}_2\xi_2}^*\psi_{{\mathbf{x}}_2\phi_2}\psi_{{\mathbf{x}}_1\phi_1}
\end{split}
\end{equation*}
with $U_o(({\mathbf{x}}_1,{\mathbf{x}}_2),(\xi_1,\xi_2),(\phi_1,\phi_2)):=U\delta_{{\mathbf{x}}_1,{\mathbf{x}}_2}\delta_{\xi_1,{\uparrow}}\delta_{\xi_2,{\downarrow}}\delta_{\xi_1,\phi_1}\delta_{\xi_2,\phi_2}$,
one can translate the proof of Lemma \ref{lem_thermodynamic_limit} to confirm the existence of the thermodynamic limit \eqref{eq_thermodynamic_limit_hubbard}. Then, by sending $L\to +\infty$ in \eqref{eq_decay_bound_almost} we obtain the claimed bound for $U$ satisfying \eqref{eq_assumption_U_convergence_hubbard_model}.
\end{proof}
|
\section{\label{sec:intro}Introduction}
The ternary nitrides or carbides with the general formula AXM$_3$ (
A: divalent or trivalent element; X: carbon or nitrogen; and M:
transition metal) are already known for several decades
~\cite{Goodenough70,Chern92,Jager93}. These compounds crystalize in
a cubic anti-perovskite structure (A: cube-corner position; X:
body-center position; M: face-center position) and exhibit a wide
range of interesting physical properties~\cite{Goodenough70}, such
as giant magneto-resistance~\cite{Kim2001} and nearly zero
temperature coefficient resistivity~\cite{Chi2001}. They have
renewedly attracted considerable attention due to the discovery of
superconductivity at $\sim$8 K in intermetallic compound
MgCNi$_3$~\cite{He01}.
Considering the Ni-rich composition, it is expected that the
ferromagnetism could exist in MgCNi$_3$. However, the absence of
ferromagnetism was observed in experiment~\cite{He01} for MgCNi$_3$. From the
electronic structures obtained by the first-principles
calculations~\cite{Singh01,Shim01,Rosner02,Wu2009251}, the
non-ferromagnetic ground state of MgCNi$_3$ is ascribed to a reduced
Stoner factor that results from a strong hybridization between the
Ni-3$d$ and C-2$p$ electrons. For other Ni-based ternary carbides
ACNi$_3$ (e.g., A = Al, Ca, In, Zn, and Cd), the first-principles
calculations~\cite{Wu2009251,Okoye05,Sieberer07,Zhong07,Wu20084232}
also show that the ground states of these compounds are
non-magnetic and the C-Ni bonding exhibits nearly same
characteristics as the one in MgCNi$_3$. Therefore, this indicates
that the change of composition A could not induce the
ferromagnetism in ACNi$_3$. On the other side, it raises a question
whether the change of composition X or M can lead to the appearance
of ferromagnetism in AXM$_3$ or not.
Very recently, the antiperovskite-type compounds InN$_y$Co$_3$ and
InN$_y$Ni$_3$ ($y\sim$ 1.0 and 0.8, respectively) have been
synthesized by solid-gas reactions of metal powders with NH$_3$ and
they have been reported to have spin-glass-like properties based on
the measurements of temperature dependence
magnetization.~\cite{Cao093353} The recent first-principles
calculations~\cite{Sieberer07} showed that the non-stoichiometry
could affect the magnetic properties of ACNi$_3$ (e.g., AlCNi$_3$
and GaCNi$_3$) and suggested that the tendencies toward magnetism
found in experiments~\cite{Dong05,Tong06Al,Tong06Ga,Tong07} for
these compounds should be explained by the deviation of the Ni/C
atomic ratio from the ideal stoichiometry. To shed more light on the
understanding on the magnetic properties of InNCo$_3$ and InNNi$_3$
reported in experiment~\cite{Cao093353}, it is of great importance
to theoretically study the electronic structures of these two compounds
as well as the nature of the N-Co and N-Ni bondings.
In order to completely understand the electronic structures and
magnetic ground states of InNCo$_3$ and
InNNi$_3$ with cubic anti-perovskite structure, we carried out the first-principles calculations on these
compounds using the pseudopotentials method with plane-wave basis
set within the local density approximation and the generalized
gradient approximation. Since the elastic properties of a solid are highly associated with various fundamental solid-state properties such as phonon spectra, specific heat, Debye temperature, and so on, we have calculated the independent elastic constant and the elastic moduli of InNCo$_3$ and InNNi$_3$.
\section{\label{sec:method}Computational details}
All calculations on antiperovskite-type InNCo$_3$ and InNNi$_3$ were
performed using the Quantum ESPRESSO code~\cite{Baroni}, which is
based on the density functional theory (DFT)~\cite{Kohn65}. The
electronic exchange-correlation potential was calculated within the
local density approximation (LDA)~\cite{Ceperley80,Perdew81} and the
generalized gradient approximation using the scheme of
Perdew-Burke-Ernzerhof (PBE)~\cite{Perdew96}. The spin polarization
was also considered in the calculation in order to assess the
magnetic properties of these compounds. Electron-ion interaction was
represented by the norm-conserving optimized~\cite{Rappe90} designed
nonlocal pseudopotentials. The 4\emph{d} electrons are explicitly
included in the valence of In. The electronic wavefunctions were
expanded by the plane waves up to a kinetic energy cutoff of 55 Ry.
The \textbf{k}-point sampling in Brillouin zone (BZ) of simple cubic
lattice was treated with the Monkhorst-Pack scheme
\cite{Monkhorst76} and a 20$\times$20$\times$20 \textbf{k}-point
mesh (i.e., 286 irreducible points in the first BZ) was used. The
chosen plane-wave cutoff and number of \textbf{k} points were
carefully checked to ensure that the total energy was converged to
be better than 1 mRy/cell. The total energies are obtained as a
function of volume and they are fitted with the Birch-Murnaghan
3rd-order equation of states (EoS)~\cite{Birch47} to give the
equilibrium lattice constant and other ground state properties.
During the calculation of density of states (DOS), a dense
\textbf{k}-point mesh of 30$\times$30$\times$30 is used, the total
DOS is computed by the tetrahedron method~\cite{Lehmann77}, and the
atomic-projected DOS is calculated by the L\"{o}wdin
populations~\cite{Portal95}.
For a cubic crystal, its independent elastic constants are $c_{11}$,
$c_{12}$, and $c_{44}$. To determine the elastic constants of
InNCo$_3$ and InNNi$_3$ by means of the curvature of the internal
energy versus the strain curves~\cite{Mehl93,Wu07425}, three strain
modes~\cite{Hou081651} are adopted and their nonzero strains are as
follows: (1) $\epsilon_{11}=\epsilon_{22}=\delta$,
$\epsilon_{33}=(1+\delta)^{-2}-1$; (2)
$\epsilon_{11}=\epsilon_{22}=\epsilon_{33}=\delta$; and (3)
$\epsilon_{12}=\epsilon_{21}=\delta/2$,
$\epsilon_{33}=\delta^2/(4-\delta^2)$. The deformation magnitudes
$\delta$ from -0.012 to 0.012 in the step of 0.03 are applied in the
first and second strain modes, and $\delta$ from -0.04 to 0.04 in
the step 0.01 are adopted in the third strain mode. Once the
independent elastic constants for single crystal properties are
obtained through the above procedure, the elastic moduli (e.g., the
shear modulus and the bulk modulus) of polycrystalline aggregates
can be estimated according to the Voigt-Reuss-Hill
approximation~\cite{Voigt28,Reuss29,Hill52}. In the Voigt
average~\cite{Voigt28}, the shear modulus and the bulk modulus of
cubic lattice are given by
\begin{equation}\label{eq:1}
G_V=\frac{1}{5}\left[(c_{11}-c_{12})+3c_{44}\right]
\end{equation}
and
\begin{equation}\label{eq:2}
B_V= \frac{1}{3}(c_{11}+2c_{12}),
\end{equation}
while in the Reuss average~\cite{Reuss29} they are given by
\begin{equation}\label{eq:3}
G_R=\frac{5}{4(s_{11}-s_{12})+3s_{44}}
\end{equation}
and
\begin{equation}\label{eq:4}
B_R= \frac{1}{3s_{11}+6s_{12}}
\end{equation}
with the relations $c_{44}=s_{44}^{-1}$,
$c_{11}-c_{12}=(s_{11}-s_{12})^{-1}$, and
$c_{11}+2c_{12}=(s_{11}+2s_{12})^{-1}$ in the cubic lattice, where
$s_{ij}$ are the elastic compliance constants. Therefore, $G_R$ and
$B_R$ in the cubic lattice can be rewritten as
\begin{equation}\label{eq:5}
G_R=\left[\frac{4}{5}(c_{11}-c_{12})^{-1}+\frac{3}{5}c_{44}^{-1}\right]^{-1},
\end{equation}
and
\begin{eqnarray}\label{eq:6}
\nonumber
B_R&=&\frac{1}{3}\left[(c_{11}+2c_{12})\right]\\
&=&B_V.
\end{eqnarray}
In the Hill empirical average~\cite{Hill52}, the shear modulus and
the bulk modulus are taken as $G=\frac{1}{2}(G_V+G_R)$ and
$B=\frac{1}{2}(B_V+B_R)$, respectively. Knowing $G$ and $B$, the
Young's modulus $E$ and Poisson's ratio $\nu$, which are frequently
measured for polycrystalline materials when investigating their
hardness, can be calculated from the isotropic relations:
\begin{equation}\label{eq:7}
E=\frac{9BG}{3B+G}
\end{equation}
and
\begin{equation}\label{eq:8}
\nu=\frac{3B-2G}{2(3B+G)} .
\end{equation}
\section{Results and Discussions}
\subsection{\label{sec:struct}Structural properties}
In experiment with the powder X-ray diffraction patterns, Cao et
al~\cite{Cao093353} have reported that InN$_y$Co$_3$ and
InN$_y$Ni$_3$ ($y\sim$ 1.0 and 0.8, respectively) have the cubic
anti-perovskite structure with the space group 221($Pm\bar{3}m$) and
the corresponding lattice parameters were 3.854 \AA~and 3.844 \AA,
respectively. Starting from the experimental data, we have
calculated the total energies of unit cell at a series of volumes
for each compound in the paramagnetic (PM) and ferromagnetic (FM)
states. The results are presented in Fig.~\ref{fig:1}. It is found
that the energy difference between the FM and PM states is -0.0397
eV (-0.226 eV) in the LDA (GGA) calculations for InNCo$_3$ and 0.0
eV for InNNi$_3$. The total magnetic moment of InNCo$_3$ is about
2.14 $\mu_B$ (2.91 $\mu_B$) and the local magnetic moment of each In ion is about 0.69 $\mu_B$ (0.94 $\mu_B$) in LDA (GGA) calculations. The total magnetic moment of InNNi$_3$ and the local magnetic moment of each Ni atom are zero. These indicate that
the ferromagnetic state is energetically favorable to InNCo$_3$ and the ground state of InNNi$_3$ is paramagnetic state
(non-magnetic). The obtained equilibrium lattice constant
($a_0$), bulk modulus ($B$), and first pressure derivative of bulk
modulus ($B^{\prime}$) of InNCo$_3$ and InNNi$_3$ are listed in
Table~\ref{tab:1}. In our calculations, the predicted lattice constant of InNCo$_3$ is slightly larger than that of InNNi$_3$, which is oppsite to the trend reported in experiment~\cite{Cao093353}. This may be due to the deviation of the Ni/N atomic ratio from the ideal stoichiometry in experiment for InN$_y$Ni$_3$. In addition, it can be seen that the deviations of the LDA (GGA)
lattice constants of both the InNCo$_3$ and InNNi$_3$ with respect
to the experimental values are less than 2.6\% (1.0\%). That is to
say, the calculated equilibrium lattice constants of InNCo$_3$ and
InNNi$_3$ are in excellent agreement with the experimental
data~\cite{Cao093353}.
\subsection{\label{sec:poly}Elastic properties}
The calculated independent elastic constants for single crystal of
InNCo$_3$ and InNNi$_3$ are listed in Table~\ref{tab:2}. Based on
the Voigt-Reuss-Hill approximation~\cite{Voigt28,Reuss29,Hill52},
the elastic moduli of InNCo$_3$ and InNNi$_3$ are estimated and the
results are listed in Table~\ref{tab:2}. For the bulk moduli of
InNCo$_3$ and InNNi$_3$, the estimations based on the independent
elastic constants agree well those obtained by the fit of the
Birch-Murnaghan 3rd-order EoS. To the best of our knowledge, no
experimental data or theoretical results for the elasticity of
InNCo$_3$ and InNNi$_3$ compounds have been reported up to now.
Considering that the elastic properties of ZnNNi$_3$, InNSc$_3$, and InCNi$_3$ with cubic
anti-perovskite structure have been studied recently and the
theoretical results are available in
literature~\cite{Maurizio09,Wu20084232,lichong09,Shein10}, it will
be meaningful to compare them with those of InNCo$_3$ and InNNi$_3$
compounds. The order of bulk moduli of these five compounds from low to high is:
\textit{B}(InNSc$_3$) $<$ \textit{B}(InNNi$_3$) $<$
\textit{B}(InCNi$_3$) $<$ \textit{B}(InNCo$_3$) $<$
\textit{B}(ZnNNi$_3$). This could be understood from the trend in
lattice constants ($a$) of these compounds (i.e., $a$(InNSc$_3$) $>$
\textit{a}(InNNi$_3$) $>$ \textit{a}(InCNi$_3$) $>$
\textit{a}(InNCo$_3$) $>$ \textit{a}(ZnNNi$_3$)) as well as the
relationship between bulk modulus and equilibrium volume (i.e.,
$B\sim V^{-1}$)~\cite{Cohen85}. The GGA lattice constants of
InNSc$_3$, ZnNNi$_3$, and InCNi$_3$ are 4.411 \AA~\cite{Maurizio09},
3.77 \AA~\cite{lichong09}, and 3.880 \AA~\cite{Wu20084232},
respectively. The order of shear modulus from low to high is:
\textit{G}(ZnNNi$_3$ and InNNi$_3$) $<$ \textit{G}(InCNi$_3$) $<$
\textit{G}(InNCo$_3$). Pugh~\cite{Pugh54} has proposed that a high
$B/G$ ratio may be associated with better ductility whereas a low
value would correspond to a more brittleness, and the critical value
separating ductile and brittle materials is around 1.75. From the
results of $B/G$ ratio for InNCo$_3$, InNNi$_3$, InNSc$_3$,
ZnNNi$_3$, and InCNi$_3$, it is found that only InNSc$_3$ can be
classified as brittle materials and others may be ductile materials.
Furthermore, InNNi$_3$ seems to be more ductile than InNCo$_3$.
For a cubic crystal, its mechanical stability requires that its
three independent elastic constants should satisfy the following
relations~\cite{Wallace72}:
\begin{equation}\label{eq:9}
(c_{11}-c_{12})>0, c_{11}>0, c_{44}>0, (c_{11}+2c_{12})>0.
\end{equation}
These conditions also lead to a restriction on the magnitude of $B$:
\begin{equation}\label{eq:10}
c_{12}<B<c_{11}.
\end{equation}
The predicted $c_{ij}$ values (see Table~\ref{tab:2}) for InNCo$_3$ and InNNi$_3$ satisfy these conditions,
indicating that cubic antiperovskite-type compounds InNCo$_3$ and
InNNi$_3$ are mechanically stable.
\subsection{\label{sec:estruct}Electronic structures}
In order to understand the different magnetic ground states for InNCo$_3$
and InNNi$_3$, we examined the electronic structures of these two compounds. For the simplicity in discussion, only the GGA results are presented below. The calculated
electronic band structures along the high symmetry directions in the
Brillouin zone are shown in Fig.~\ref{fig:2}. As discussed above,
for InNCo$_3$ the ferromagnetic state is energetically preferable to
the paramagnetic state, and hence it is clearly seen that the
spin-splitting occurs in the bands around the $E_F$. For InNCo$_3$ the profile of
majority spin bands looks roughly similar to the one of
minority spin bands, however only one
band in the majority spin bands crosses the Fermi level and the corresponding band with same dispersion in the minority spin bands is
unoccupied. In addition, two bands in the minority spin bands across the
Fermi level of InNCo$_3$ (see Fig.~\ref{fig:2}(a)). For InNNi$_3$ the
ferromagnetic state is not energetically preferable to the
paramagnetic state, consequently, the majority and minority spin bands are
degenerated, that is to say, no spin-splitting occurs in the
band structure (see Fig.~\ref{fig:2}(b)). It is interesting to note
that the whole feature of majority spin bands of InNNi$_3$ is very
similar to the one of InNCo$_3$. In order to reveal the detailed
character of band structure, the total density of states (DOS) and
the angular-momentum-projected DOS of each atom in InNCo$_3$ and
InNNi$_3$ are presented in Fig.~\ref{fig:3}. For both InNCo$_3$ and
InNNi$_3$, the four bands from -9 eV to -5 eV come mainly from the N
2\emph{p} states and In 5\emph{s} states, and the five bands roughly
from -5 eV to -2.5 eV have significant contribution from
3\emph{d}-$t_{2g}$ states of transition metal atoms (Co/Ni) and the
5\emph{p} state of In atom. For the bands from -2.5 eV to 0 eV (i.e., the Fermi level), they are dominated by the 3\emph{d} states of
transition metal atoms (Co/Ni) and have small contribution from the
2\emph{p} states of N atom. Because the number of 3\emph{d}
electrons in Co is one less than that of Ni, two minority spin
bands composed of the 3\emph{d}-$t_{2g}$ states around the Fermi level
are unoccupied in InNCo$_3$, while the counterpart in the InNNi$_3$ are occupied. This also results in different behavior of
spin-splitting in InNCo$_3$ and InNNi$_3$. Due to the significant spin-plitting
around the Fermi level in InNCo$_3$, the hybridization between the
Co 3\textit{d} and N 2\textit{p} states in InNCo$_3$ are slightly
weaker than the one between Ni 3\textit{d} and N 2\textit{p} states
in InNNi$_3$(see Fig.~\ref{fig:3}).
The contributions of each kind of atoms to the DOS at the $E_F$ of
InNCo$_3$ and InNNi$_3$ are listed in Table~\ref{tab:3}. The total
DOS at the $E_F$ of InNCo$_3$ is about 2.761 states/eV
per formula unit (f.u.) in the GGA calculations, which is larger than that of InNNi$_3$, and its main contribution comes from Co 3$d$ states which accounts for 87\%. For InNNi$_3$ in the GGA calculations, the contribution of Ni 3\textit{d} states to the
the total DOS at the $E_F$ (i.e., 1.803 states/eV.f.u.) accounts for 72\%. These indicate the 3\textit{d} states of transition metal atoms in InNCo$_3$ and InNNi$_3$ play dominant roles in the total dnesity of states of these compounds.
In order to understand the bonding nature among the ions in
InNCo$_3$ and InNNi$_3$, we analyzed the charge density contours of
InNCo$_3$ and InNNi$_3$ in the(110) plane, as shown in
Fig.~\ref{fig:4}. From Fig.~\ref{fig:4}, it is found that a certain
amount of charges are accumulated in the intermediate region between
N and Co atoms in InNCo$_3$, and slightly more charges are
accumulated intermediate region between N and Ni atoms in InNNi$_3$.
This gives an evidence for the strong hybridization between N and
transition metal (Co/Ni) atoms, indicating that the N-Co and N-Ni
bondings exhibit strong covalent characteristics and the latter is slightly stronger than the former. The similar
bonding characteristics for Ni-N atoms or Ni-C atoms were also found in
other Ni-based ternary nitrides or carbides
AXNi$_3$~\cite{Wu2009251,Wu20084232,lichong09,Shein10}. Therefore, our results suggest that the magnetic properties of InNNi$_3$ reported in experiment~\cite{Cao093353} are very likely due to the non-stoichiometry effect, which was also found in the cases of AlCNi$_3$ and GaCNi$_3$~\cite{Dong05,Tong06Al,Tong06Ga,Tong07,Sieberer07}.
\section{Conclusions}
In summary, we performed the first-principles calculations to study the
elastic and electronic properties of cubic antiperovskites InNCo$_3$
and InNNi$_3$. Based on the Voigt, Reuss and Hill bounds, the shear,
Young's moduli and Poisson's ratio have also been estimated for the
InNCo$_3$ and InNNi$_3$ polycrystals. The theoretically predicted equilibrium lattice
parameters are in good agreement with the available
experimental data. Our calculations
show that the 3\textit{d} states of transition metal atoms in
InNCo$_3$ and InNNi$_3$ play dominant roles near the Fermi levels.
InNCo$_3$ energetically prefers to the ferromagnetic state. The
magnetic ground state of InNNi$_3$, which is same to other Ni-based
ternary nitrides or carbides with a cubic anti-perovskite structure,
is a stable paramagnetic (non-magnetic) state. This could be understood from
that the hybridization between Ni-3$d$ and N-2$p$ states in
InNNi$_3$ is slightly stronger than the one between Co-3$d$ and
N-2$p$ states in InNCo$_3$ because of the more 3\textit{d} electrons
in Ni.
\section*{Acknowledgments}
The author acknowledges support from National Natural Science
Foundation of China under Grant No. 10674028.
\bibliographystyle{elsarticle-num}
|
\section{Evolutionary and Photoionization Synthesis Models}
We present the PopStar evolutionary synthesis model in
\cite{mol09}(Paper I). Selected isochrones are a new Padova set,
specifically computed for this piece of work with a broad age and
metallicity coverage, and a detailed treatment of mass-loss for both,
young (O, B, WR) and old ages (post-AGB until planetary nebula). The
spectral energy distributions (SEDs) are calculated for each Single
Stellar Population (SSP) by including the nebular contribution.
Colors are calculated for these SEDs for different photometric
systems. They result redder for the youngest stellar populations,
mainly for low Zmet, than those obtained by \cite{bc03} but they are
similar to those obtained by \cite{stb99} for ages younger than 1
Gyr. The old stellar populations show the same colors than
\cite{bc03}. Both results imply that our models are equally well tuned
for young as for old stellar populations.
We have then computed photoionization models (CLOUDY) with the
previously described SSP-SEDs, obtaining the emission line spectra due
to the youngest stellar populations \cite{PaperII}(Paper II). Some
intense emission lines fall in the broad band filters and therefore
their contribution must be included into the magnitudes. Colors of
stellar clusters changes appreciably, such as we may see in Fig.~3
from \cite[][Paper III]{gv09}, in particular when the emission lines
proceed from very young stellar clusters but also when they are as old
as 10 Myr. In the color-color diagrams, when a burst of star formation
takes places, points go out of the stellar population region, falling
in a region impossible to reach in any other way.
\section{Spiral Evolution Models}
The same basic chemical evolution numerical model was applied to a
wide grid of theoretical galaxies with different total masses and
variable star formation efficiencies in \cite{mol05}. This grid
reproduces well the observational data for local spiral
galaxies. By using the resulting star formation, $\Psi(t)$, and metal
enrichment, Z(t), histories, we calculate the SED, $F_{\lambda}(t)$,
for each galaxy from the equation:
\begin{equation}
F_{\lambda}(t)=\int_{0}^{t} S_{\lambda}(\tau,Z(t'))\Psi(t')dt',
\label{Flujo}
\end{equation}
where $\tau=t-t'$ and $S_{\lambda}(\tau,Z)$ is the SED of each stellar
generation or SSP of a given age $\tau$ and a metallicity $Z$, taken
from Paper I spectra. Once these SEDs calculated, we compute
magnitudes and colors in the Johnson and SDSS systems. They reproduce
well the data of our local universe, demonstrating that the grid
of models is well calibrated.
\section{Results and Conclusions}
\begin{figure}
\begin{center}
\includegraphics[width=2.5in,angle=0]{molla_fig.eps}
\caption{Color-Color diagrams with and without the emission lines
contribution}
\label{fig}
\end{center}
\end{figure}
We check if the observed colors of galaxies change when emission line
intensities are included in the calculations. We compute the observed
and rest-frame colors evolution along the redshift for a MWG-type
galaxy with continuous star formation. The rest-frame absolute colors
U-B and B-V evolution are shown in Fig.~\ref{fig}. The emission
lines produced by the last stellar generations change the magnitudes
in ~0.2 mag in any band.
In summary, the contribution of the emission lines when a burst takes
place changes the broad band magnitudes, therefore it is essential to
take it into account when star-forming galaxies are studied. This is
important specially at redshifts in which the star formation history
reaches its maximum. Therefore to interpreting high redshift colors by
using evolutionary synthesis models without taking into account the
star formation history of galaxies and the subsequent emission lines
may yield erroneous conclusions.
|
\section{\large Introduction}
Let $(G,X)$ and $(H,Y)$ be permutation groups (all actions in the paper are faithful). The
permutational wreath product $H\wr G$ is the semidirect product $H^{|X|}\rtimes G$ with a natural
action on $X\times Y$, where $G$ acts on $H^{|X|}$ by permuting the copies of~$H$. Given a sequence
of finite permutation groups $(G_n,X_n)$ the inverse limit of iterated wreath products
\[
W=\lim_{\longleftarrow} \ (G_n\wr\ldots \wr G_2\wr G_1)
\]
is a profinite group called the infinitely iterated wreath product $\ldots\wr G_2\wr G_1$. The goal
of this note is to prove the following theorem.
\begin{thm}\label{thm_main}
Let $(G_n,X_n)$ be a sequence of finite transitive permutation groups with uniformly bounded number
of generators. Then the profinite group $\ldots\wr G_2\wr G_1$ is finitely generated if and only if
the profinite abelian group $\prod_{n\geq 1} G_n/G'_n$ is finitely generated.
\end{thm}
Since the group $\prod_{n\geq 1} G_n/G'_n$ is an epimorphic image of the group $\ldots\wr G_2\wr
G_1$, in one direction the statement is obvious. For the converse we construct a finitely generated
dense subgroup. The construction is based on the notions of directed automorphisms and branch
groups introduced by R.~Grigorchuk \cite{branch:gri}, and it is basically the same construction
used by P.~Neumann \cite{neumann86}, L.~Bartholdi \cite{uniform:bart}, D.~Segal \cite{fin_im:segal}
and others to construct certain groups with interesting properties.
M.~Bhattacharjee \cite{bhattach} showed that the infinitely iterated wreath products of alternating
groups of degree $\geq 5$ can be generated by two elements even with positive probability. This was
generalized to wreath products of non-abelian simple groups with transitive actions by M.~Quick
\cite{quick1,quick2}. D.~Segal \cite{fin_im:segal} showed that the infinitely iterated wreath
products of perfect groups with certain conditions on the actions are finitely generated. The
iterated wreath products of cyclic groups of pairwise coprime orders are two-generated by result of
A.~Woryna \cite{woryna}.
The profinite group $W$ is finitely generated if and only if the sequence $d(G_n\wr \ldots \wr
G_2\wr G_1)$ is bounded, where $d(G)$ is the minimal number of generators of the group $G$. The
behavior of generating sequence $d(G^n)$ for the direct product $G^n$ of a finite group $G$ was
described in a series of papers by J.~Wiegold (see \cite{growth_seqII,growth_fgg} and references
therein): the sequence $d(G^n)$ is roughly logarithmic if $G$ is perfect, and is linear otherwise.
The generating sequence $d_n^{wr}(G)=d(G\wr\ldots\wr G\wr G)$ for the wreath power of a finite
transitive group $G$ happens to be even more simple. As a corollary of Theorem~\ref{thm_main} we
get that $d_n^{wr}(G)$ is bounded if $G$ is perfect, and grows linearly otherwise.
The wreath powers of finite groups are related to self-similar groups of finite type introduced by
R.~Grigorchuk \cite[Section~7]{solved:gri} to describe profinite completion of certain branch
groups. Every such group is given by a finite pattern, and there is an open question
\cite[Problem~7.3]{solved:gri}: which patterns define finitely generated groups. Every self-similar
group of finite type given by pattern of size $1$ is isomorphic to the infinite wreath power
$\ldots\wr G\wr G$. Hence it will be finitely generated if and only if the group $G$ is perfect.
Moreover, in this case the finitely generated dense subgroup constructed in the proof of
Theorem~\ref{thm_main} is a branch group, which has maximal subgroups of infinite index. This
answers Question~14 in \cite[p.~1107]{branch_groups}.\\[-0.3cm]
\noindent\textbf{Acknowledgments.} I would like to thank Rostyslav Kravchenko and Dmytro Savchuk
for fruitful discussions.
\section{\large Automorphisms of spherically homogeneous rooted tree}
Let $\tree$ be the spherically homogeneous rooted tree given by the alphabets $X_i$, where the
$n$-th level is $\tree_n=X_1\times\ldots\times X_n$ (here $\tree_0$ consists of the root
$\emptyset$), and each vertex $v\in \tree_n$ is connected with $vx\in \tree_{n+1}$ for every $x\in
X_{n+1}$. Denote by $\tree^{[n]}$ the finite truncated rooted tree consisting of levels from $0$ to
$n$. For a vertex $v\in\tree$ denote by $\tree_v$ the rooted tree hanging ``below'' the vertex $v$,
which consists of vertices $u$ such that the geodesic connecting $u$ with the root $\emptyset$
passes through~$v$. The set $vX_n$ is the first level of tree $\tree_v$ for $v\in\tree_{n-1}$.
Every automorphism $g\in\Aut\tree$ induces a map $vX_n\rightarrow g(v)X_n$ for every vertex
$v\in\tree_{n-1}$. Forgetting the prefixes $v$ and $g(v)$ we get a permutation on the set $X_n$,
which is called the \textit{vertex permutation of $g$ at $v$} and denoted $g \at v\in\Sym(X_{n})$.
Every automorphism can be given by its vertex permutations at each vertex.
The automorphism group of the tree $\tree^{[n]}$ is the iterated permutational wreath product
\[
\Aut\tree^{[n]}=\Sym(X_{n})\wr \ldots\wr \Sym(X_1),
\]
and the group $\Aut\tree$ is the inverse limit of these groups, which is also the infinite
permutational wreath product $\ldots\wr\Sym(X_2)\wr\Sym(X_1)$. The group $\ldots\wr G_2\wr G_1$ can
be naturally considered as a subgroup of $\Aut\tree$. We also identify the group $G_1$ with the
subgroup of $\Aut\tree$, which consists of automorphisms whose vertex permutations at the root form
the group $G_1$ and are trivial at the other vertices.
Let $G$ be a subgroup of $\Aut\tree$. The \textit{vertex stabilizer} $\st_G(v)$ of a vertex
$v\in\tree$ is the subgroup of all $g\in G$ such that $g(v)=v$. The \textit{$n$-th level
stabilizer} $\st_G(n)$ is the subgroup of all $g\in G$ such that $g(v)=v$ for every
$v\in\tree_{n}$. The \textit{rigid vertex stabilizer} $\rs_G(v)$ of a vertex $v\in\tree$ is the
subgroup of all $g\in G$ such that the vertex permutations of $g$ at vertices outside the subtree
$\tree_v$ are trivial. The \textit{rigid level stabilizer} $\rs_G(n)$ is the subgroup generated by
$\rs_G(v)$ for $v\in\tree_n$. The group $G$ is called \textit{branch}, if it acts transitively on
the levels $\tree_n$ and every rigid level stabilizer $\rs_G(n)$ is of finite index in $G$.
We say that the rigid stabilizer $\rs_G(v)$ for $v\in\tree_{n-1}$ contains a subgroup
$H\in\Sym(X_n)$ as a \textit{rooted subgroup}, if $\rs_G(v)$ contains a subgroup whose vertex
permutations at $v$ form the group $H$ and all the other vertex permutations are trivial.
\section{\large Proof of Theorem~\ref{thm_main} and Corollaries}
\begin{proof}(of Theorem~\ref{thm_main})
The trivial groups $G_n$ can be omitted and we can assume $|X_n|\geq 2$ for all $n$. Also we can
assume that every group $G_n$ is non-abelian, otherwise we can pass to a new sequence given by a
different arrangement of brackets $\ldots\wr(G_6\wr G_5)\wr (G_4\wr G_3)\wr (G_2\wr G_1)$ for which
it is true (the permutational wreath product is associative). By the same reason (passing to the
same arrangement of brackets with already non-abelian groups) we can assume that for every action
$(G'_n,X_n)$ there are points in the same orbit with different stabilizers. Fix letters $x_n,y_n\in
X_n$ and permutations $\tau_n,\pi_n\in G'_n$ such that $\tau_n(x_n)=y_n$, $\pi_n(x_n)=x_n$, and
$\pi_n(y_n)\neq y_n$.
Consider the decomposition of the groups $G_n/G'_n$ is the direct sum of cyclic groups of
prime-power order. Since the group $\prod_{n\geq 1} G_n/G'_n$ is finitely generated, there is an
absolute bound on the number of cyclic $p$-groups in these decomposition for any particular prime
number $p$. Hence there is a finite generating set $a_1=(a_1^{(n)})_n,\ldots, a_e=(a_e^{(n)})_n$
such that for every fixed $i$ two elements $a_i^{(j)}$ and $a_i^{(j')}$ have coprime orders for
every $j,j'$. Let $g_i^{(n)}\in G_n$ be a preimage of $a_i^{(n)}$ under the canonical projection
$G_n\rightarrow G_n/G'_n$. Then $G_n=\langle g_1^{(n)},\ldots, g_e^{(n)},G'_n\rangle$ for every
$n$. Since there is a uniform bound on the number of generators of $G_n$, we can complete the
elements $g_1^{(n)},\ldots, g_e^{(n)}$ to a generating set of $G_n$ by some elements
$g_{e+1}^{(n)},\ldots, g_m^{(n)}$, i.e. $G_n=\langle g_1^{(n)},\ldots, g_m^{(n)}\rangle$ for every
$n$ with fixed $m$. Define the automorphisms $g_1,\ldots, g_m$ of the tree $\tree$ by their vertex
permutations
\begin{equation}\label{eq_thm_defi_gi}
g_i\at v=\left\{
\begin{array}{ll}
g_i^{(k+1)}, & \hbox{if $v=y_1y_2\ldots y_{k-1}x_k$;} \\
1, & \hbox{otherwise.}
\end{array}
\right.
\end{equation}
Consider the group $G=\langle g_1^{(1)},\ldots,g_m^{(1)},g_1,\ldots,g_m\rangle$ and let us prove
that $G$ is dense in the group $W$. Fix $n$ and let us show that the restriction of $G$ on the tree
$\tree^{[n]}$ coincides with the group $G_{n}\wr\ldots\wr G_2\wr G_1$.
Since the group $G$ acts transitively on the set $X_1$ and the vertex stabilizer $\st_G(y_1\ldots
y_{n-2}x_{n-1})$ acts transitively on $y_1\ldots y_{n-2}x_{n-1}X_n$ for every $n$, inductively we
get that the group $G$ acts transitively on the levels of $\tree$.
Let us prove that the rigid vertex stabilizers $\rs_G(y_1\ldots y_{k-1}y_k)$ and $\rs_G(y_1\ldots
y_{k-1}x_k)$ contain the group $G'_{k+1}$ as a rooted subgroup. Since elements $g_1^{(1)},\ldots,
g_m^{(1)}$ generate the group $G_1$ as a rooted subgroup at the root of the tree, the statement
holds at zero level. Assume that we have proved it for all levels $<k$. By inductive hypothesis
there exist $\pi,\tau\in G$ such that $\pi\at {y_1\ldots y_{k-1}}=\pi_k$, $\tau \at {y_1\ldots
y_{k-1}}=\tau_k$, and all the other vertex permutations are trivial. Then
\begin{eqnarray*}
\pi^{-1}g_i\pi \at {y_1\ldots y_{k-1}x_k}=g_i^{(k+1)},\quad \pi^{-1}g_i\pi \at {y_1\ldots
y_{s-1}x_s}=1\quad\mbox{ for } \ s>k \\ \Rightarrow\qquad [{\pi^{-1}g_i\pi,g_j}] \at {y_1\ldots
y_{l-1}x_l}=[g_i^{(l+1)},g_j^{(l+1)}]\quad\mbox{ for } \ l\leq k
\end{eqnarray*}
and all the other vertex permutations of $[\pi^{-1}g_i\pi,g_j]$ are trivial for all $g_i,g_j$. By
inductive\linebreak hypothesis, we can multiply $[\pi^{-1}g_i\pi,g_j]$ on the appropriate elements
from $\rs_G(y_1\ldots y_{l-1}x_l)$ for $l<k$ to remove all commutators at vertices $y_1\ldots
y_{l-1}x_l$, and get elements from $\rs_G(y_1\ldots y_{k-1}x_k)$. Conjugating by generators
$g_1,\ldots,g_m$ we get that $\rs_G(y_1\ldots y_{k-1}x_k)$ contains $G'_{k+1}$ as a rooted
subgroup. Conjugating by $\tau$ we get that $\rs_G(y_1\ldots y_{k-1}y_k)$ contains $G'_{k+1}$ as a
rooted subgroup.
The elements $a_i^{(1)},\ldots,a_i^{(n)}$ have pairwise coprime orders (here $i\leq e$), and for a
fixed $k\leq n$ we can choose a power $\alpha$ such that $(g_i)^{\alpha} \at {y_1\ldots
y_{k-2}x_{k-1}}=g_i^{(k)}h_i$ for some $h_i\in G'_{k}$, $(g_i)^{\alpha} \at {y_1\ldots
y_{l-2}x_{l-1}}\in G'_l$ for all $l\leq n$, $l\neq k$, and all the other vertex permutations at the
vertices of $\tree^{[n]}$ are trivial. Multiplying $(g_i)^{\alpha}$ on the corresponding elements
from $\rs_G(y_1\ldots y_{l-2}x_{l-1})$ we can remove the elements from commutants, and get
element\linebreak $f_i\in G$ such that $f_i \at {y_1\ldots y_{k-2}x_{k-1}}=g_i^{(k)}$ with all the
other vertex permutations at the vertices of $\tree^{[n]}$ being trivial. It follows that the group
$G$ contains a subgroup $H$ such that $H \at y_1\ldots y_{k-2}x_{k-1}=G_k$ and the vertex
permutations of every element of $H$ at the vertices of $\tree^{[n]}$ and outside $\tree_{y_1\ldots
y_{k-2}x_{k-1}}$ are trivial. Since the action is transitive this holds for every vertex
$v\in\tree_{k-1}$. The result follows.
\end{proof}
\begin{rem}
We got the bound on the number of generators. In some cases one can reduce the number of generators
as it was done in \cite{fin_im:segal} using Lemma~2 there.
\end{rem}
The next corollary is a generalization of result in \cite{woryna}.
\begin{cor}\label{cor_abel}
For a sequence $A_n$ of finite abelian transitive groups with pairwise coprime orders and uniformly
bounded number of generators the iterated wreath product $\ldots\wr A_2\wr A_1$ is finitely
generated.
\end{cor}
The last corollary also follows from Theorem~2 in \cite{gen_wreath_aug}, which says that if $(G,X)$
is a finite transitive permutation group and $H$ is a finite solvable group, then
\begin{equation}\label{eq_solvable1}
d(H\wr G)=\max\left(d(H/H'\wr G), \left[\frac{d(H)-2}{|X|}\right]+2\right).
\end{equation}
Moreover, by Corollary~6 in \cite{gen_wreath_aug}, if $H$ is abelian of coprime order with $|G|$,
then
\begin{equation}\label{eq_solvable2}
d(H\wr G)=\max\left(d(G), d(H)+1\right).
\end{equation}
Hence $d(A_n\wr\ldots\wr A_2\wr A_1)=\max_{2\leq i\leq n} (d(A_1), d(A_i)+1)$ under conditions of
Corollary~\ref{cor_abel}.
The next example shows that we cannot remove assumption on the number of generators in
Theorem~\ref{thm_main}.
\begin{ex}
Take a sequence of finite solvable transitive groups $(G_n,X_n)$ such that the groups $G_n/G'_n$
are cyclic with pairwise coprime orders and $d(G_n)$ is greater than $n|X_1||X_2|\cdots|X_{n-1}|$.
Then the sequence $d(G_n\wr\ldots\wr G_2\wr G_1)$ is not bounded by (\ref{eq_solvable1}), the group
$W$ is not finitely generated, while the group $\prod_{n\geq 1} G_n/G'_n$ is procyclic.
\end{ex}
The next examples show that the uniform bound on the number of generators in not necessary for the
group $W$ to be finitely generated.
\begin{ex}
Take a sequence of finite solvable transitive groups $(G_n,X_n)$ such that every group $G_n/G'_n$
is cyclic of order coprime with $G_{n-1}\wr\ldots\wr G_2\wr G_1$, and
$|X_1||X_2|\cdots|X_{n-1}|>d(G_n)\rightarrow\infty$. Then the sequence $d(G_n\wr\ldots\wr G_2\wr
G_1)$ is bounded by (\ref{eq_solvable1}) and (\ref{eq_solvable2}), hence the group $W$ is finitely
generated.
\end{ex}
\begin{ex}
Using result in \cite{gen_wreath_two} one can construct a sequence of finite perfect transitive
groups $(G_n,X_n)$ for $n\geq 2$ with the following properties: there are generating sets
$G_n=\langle g_1^{(n)},\ldots,g_{m_n}^{(n)}\rangle$ such that the elements $g_i^{(j)}$ have
pairwise coprime orders for all $i,j$, and $|X_{n-1}|>d(G_n)=m_n\rightarrow\infty$ when
$n\rightarrow\infty$. Take the cyclic group $G_1=\langle a\rangle$ of order $d(G_2)+1$ with the
regular action $(G_1,X_1)$. Fix different letters $x_n^{(1)},\ldots,x_{n}^{(m_n)},y_n\in X_n$ for
every $n$ and define the automorphism $b$ of the tree $\tree$ by its vertex permutations
\[
b \at v=\left\{
\begin{array}{ll}
g_i^{(n+1)}, & \hbox{if $v=y_1y_2\ldots y_{n-1}x_n^{(i)}$;} \\
1, & \hbox{otherwise.}
\end{array}
\right.
\]
\end{ex}
Since the elements $g_i^{(j)}$ have coprime orders we can take a power of $b$ to get any
$g_i^{(j)}$ rooted at vertex $y_1\ldots y_{j-2}x_{j-1}^{(i)}$ with all the other vertex
permutations at vertices of $\tree^{[n]}$ being trivial for every fixed $n$. Since the group
$\langle a,b\rangle$ acts transitively on the levels of $\tree$, conjugating we get the group $G_j$
as a rooted subgroup at every vertex of $(j-1)$-th level with all the other vertex permutations at
vertices of $\tree^{[n]}$ being trivial. Hence the group $\langle a,b\rangle$ is dense in $W$,
while the abelianization of $W$ is the finite cyclic group $G_1$ and $d(G_n)\rightarrow\infty$.
\begin{cor}
Let $H$ be a finite transitive group. The infinite wreath power $\ldots\wr H\wr H$ is finitely
generated if and only if the group $H$ is perfect.
\end{cor}
Consider the previous corollary in more details. Let $H=\langle h_1,\ldots,h_m\rangle$ (we assume
the group $H$ satisfies the first paragraph of the proof of Theorem~\ref{thm_main},\linebreak
otherwise take $H\wr H$), then the dense group $G$ constructed in the proof of\linebreak
Theorem~\ref{thm_main} is generated by $h_1,\ldots, h_m$ and $g_1,\ldots,g_m$, where every
automorphism $g_i$ is defined recursively by
\[
g_i(v)=\left\{
\begin{array}{ll}
xg_i(u), & \hbox{$v=xu$;} \\
yh_i(u), & \hbox{$v=yu$;} \\
v, & \hbox{otherwise.}
\end{array}
\right.
\]
($x=x_1$ and $y=y_1$ are from the theorem). Here $\langle g_1,\ldots,g_m\rangle\simeq H$, and hence
the group $G$ is perfect, because it is generated by perfect subgroups. In the same way as in the
proof of Theorem~\ref{thm_main} we get that $\st_G(n)=\rs_G(n)\simeq G\times\ldots\times G$. In
particular, the group $G$ is branch (actually regular branch over itself, see definition in
\cite{branch:gri}). It is also just-infinite by \cite[Theorem~3]{branch:gri}, and satisfies the
congruence subgroup property by \cite[Proposition~2]{branch:gri}, i.e. every subgroup of finite
index contains the stabilizer of some level. In particular, the group $W=\ldots \wr H\wr H$ is not
only the closure of the group $G$ in $\Aut\tree$, but also the profinite completion of~$G$.
Consider the subgroup $F<G$, which consists of all automorphisms from $G$ that have only finitely
many non-trivial vertex permutations (finitary automorphisms). Since $\rs_G(v)$ contains $H$ as a
rooted subgroup for every vertex $v$, the group $F$ consists of all automorphisms of the tree
$\tree$, which have only finitely many non-trivial vertex permutations and all vertex permutations
are from the group $H$. In particular, the subgroup $F$ is dense in the groups $W$ and $G$. Also
$F$ is a proper subgroup of $G$ (here $g_1,\ldots,g_m$ are not in $F$). Since the group $G$ is
finitely generated, the subgroup $F$ is contained in some maximal subgroup $M$, which is also dense
in $G$ and hence of infinite index. Indeed, if $M$ had finite index then it would contain some
level stabilizer $\st_G(n)$; but $M/\st_G(n)=G/\st_G(n)$ and we get $M=G$, which contradicts
maximality of $M$. This answers Question~14 in\linebreak \cite[p.~1107]{branch_groups} (see
discussion in \cite[Section~6]{solved:gri}).
\begin{cor}
Let $H$ be a finite transitive group and put $d_n^{wr}(H)=d(H\wr \ldots \wr H\wr H)$. The sequence
$d_n^{wr}(H)$ is bounded if $H$ is perfect, and grows linearly otherwise.
\end{cor}
\begin{proof}
As shown above, if $H$ is perfect then
\[
d(H)\leq d_n^{wr}(H)\leq 2d(H\wr H).
\]
Let us show that if $H$ is non-perfect, then
\[
nd(H/H')\leq d^{wr}_n(H)\leq 2d(H\wr H\wr H\wr H)+nd(H/H')
\]
(here $H\wr H\wr H\wr H$ is to make sure that the action of commutant satisfies the first paragraph
of the proof of Theorem~\ref{thm_main}). The lower bound is clear. For the upper bound we proceed
as in the proof of Theorem~\ref{thm_main}. If $H=\langle h_1,\ldots,h_m\rangle$ then we define the
automorphisms $g_1,\ldots,g_m$ of the tree $\tree^{[n]}$ by (\ref{eq_thm_defi_gi}). For every
generator $a$ of $H/H'$ take its preimage $b$ in $H$ and construct $n-1$ automorphisms
$b^{(1)},\ldots,b^{(n-1)}$, where $b^{(i)}\at v=b$ if $v=y\ldots yx$ (here the length $=i$) and
trivial otherwise. The group generated by all constructed elements is isomorphic to $H\wr\ldots\wr
H\wr H$.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
\label{intro}
\footnotetext[1]{This publication is based on observations with the
MegaPrime/MegaCam, a joint project of the CFHT and CEA/DAPNIA, at the
Canada-France-HAwaii Telescope (CFHT), which is operated by the National
Research Council of Canada, the institut National des Sciences de
l'Univers of the Centre National de la Recherche Scientifique (CNRS),
and the University of Hawaii.}
In recent years, it has been increasingly recognised that the outskirts
of galaxies hold fundamental clues about their formation history.
It is into these regions that new material continues to arrive as part
of their assembly, by accretion of minor satellites, predominantly at
early epochs when large disk galaxies were assembling, as predicted by
the currently favored hierarchical formation models.
It was also in the outer regions of galaxies that material was deposited
during the violent interations in the galaxy's past. Most present-day
disk galaxies are suspected to have experienced mergers during the
last few billions of years \citep[e.g.][and references therein]{hammer07}.
Based on the systematic deviation of the Milky Way from a number of
galaxy scaling relations, \citet{hammer07} have agrued that our galaxy
had most likely escaped any significant major merger event over the
last $\sim 10$ Gyr. These authors suspect that the observed differences
between the Milky Way and its neighbour M31 are likely due to the
quiescent formation in the former case and to the merger-dominated
history for the latter. The observed properties of the stellar content
in the outskirts of M31 can be accounted for by either a succession of
minor mergers or a major merger, with this material most likely accreting
in the most recent half of the age of the Universe \citep{ibata05}.
By analysing the characteristics of spiral galaxy stellar halos
formed within a large grid of numerical chemo-dynamical simulations,
\citet{renda05} have shown that at any given total galactic mass,
the metallicities of simulated stellar halos span a range in excess
of $\sim$~1~dex. The underlying driver of this metallicity spread can
be traced back to the diversity of galactic mass assembly histories.
Galaxies with a more extended merging history possess halos which have
younger and more metal rich stellar populations than the stellar halos
associated with galaxies with a more abbreviated assembly. For a given
total mass, galaxies with more extended assembly histories also possess
more massive stellar halos.
The studies of the Galaxy, and to lesser extent the other large spiral in
the Local Group, M31, have been delivering the bulk of the observational
constraints on the properties of the stellar content of the outer regions
of galaxies. Evidence indicates that the Galaxy might be unrepresentative
of a typical spiral galaxy, and it may not even follow the standard
scenario of disk formation. The Milky Way halo seems to be populated by
old, metal-poor stars, while a few fields in the halo of M31 show a large
population of intermediate age stars with a much higher overall metallicity
\citep{brown06}. The fields in M31 in which the above results were obtained
have been found to be significantly contaminated by various accretion
events \citep{ibata07} casting doubt on the conclusion that the M31 halo
is globally younger and more metal-rich than that of the Milky Way.
The current observational evidence therefore demonstrates that halos are
complex structures. To establish comprehensively properties of stars in
the outskirts of galaxies, and to fully understand their nature and origin,
we need to undertake panoramic studies of the outer regions of spiral
galaxies beyond the Local Group. To do so, we have obtained deep and
wide-field optical imaging data of the nearby early-type spiral M81
(NGC~3031), resolving stars well below the tip of the red giant branch
of metal-poor stellar populations.
The M81 group of galaxies is one of the nearest groups to our own.
It contains one large spiral, two peculiar galaxies (M82 and NGC~3077),
two small spirals galaxies (NGC~2976 and IC~2574), as well as a large number
of dwarf galaxies \citep[e.g.][]{karachentsev85, karachentsev01, chiboucas08}.
The core galaxies of the group are strongly interacting. Atomic hydrogen
observations have revealed the presence of a large number of tidal streams
with large, dynamically complex atomic hydrogen clouds embedding M81,
M82, NGC~3077, and NGC~2976 \citep{vdh79,appleton81,yun94,boyce01}.
Close interactions between galaxies are capable of leaving tidal debris
that could be converted into new stellar systems \citep[e.g.][]{TT72}.
Compared to the Local Group, an interesting feature of the M81 group is
the presence of a population of stellar systems dominated by young stars
\citep[e.g.][]{durrell04,demello08a,davidge08}, which are suspected to be of
tidal origin \citep[e.g][]{makarova02}, and which have no counterparts in the
Local Group.
These young stellar systems, e.g. Holmberg IX, BK 3N, and Garland, are
embedded in H{\sc i} clouds \citep{boyce01}. Here, we take advantage of
our deep and wide field survey to study the spatial distribution of young
stellar populations in the outer regions of M81. We report the discovery
of new stellar systems in the tidal debris.
Analysis of the spatial distribution of old stellar populations, the
bi-dimensional distribution, the search for substructures, the metallicity
distribution functions, and globular cluster properties over the surveyed
area will be reported in forthcoming papers. The stellar populations of
immediate interest to the present paper are revealed by the upper main
sequence and the red supergiant stars. The layout of this paper is as
follows: in Section \ref{data} briefly represents the data set, while
section \ref{results} studies the young stellar content around M~81.
\section{Data}
\label{data}
\begin{figure}
\includegraphics[clip=,width=0.5\textwidth]{megacam_footprint_M81.ps}
\caption{CFHT/MegaCam footprint of the observations overlayed on the
Digitized Sky Survey image of M~81. The MegaCam field-of-view covers a
$0.96^{\circ} \times 0.94^{\circ}$ field. The three new young stellar clumps
reported here are identified as Clump I, II, III (see text for more details).
Also shown are the three tidal debris objects (TDO1, TDO2, TDO3) identified
by \citet{davidge08}. The locations of three control fields are also shown.}
\label{footprint}
\end{figure}
The MegaCam wide-field imager at the Canada-France-Hawaii Telescope (CFHT)
was used to map M81 over the area displayed in Figure \ref{footprint}.
MegaCam consists of a mosaic of thirty six $2048\times 4612$ EEV chips,
covering a $0.96^{\circ} \times 0.94^{\circ}$ field, with each pixel subtending
0.187 arcseconds on a side. The photometric depth and field of view achievable
with this instrument make it particularly powerful in regions of extremely low
stellar surface density. To observe as much of the galaxy halo and the extension
of the galaxy disk as possible, we positioned the centre of M81 in one corner
of the mosaic. The images used in this study are of a single pointing centred
at ${\rm R.A.} = 09^{\rm h}58^{\rm m}44.0^{\rm s}$,
${\rm Dec.} = +68\degr 51\arcmin 46.0\arcsec$ (J2000). The survey gives an
uninterrupted coverage out to approximately 50 kpc along the major axis, and
the inner halo out to $\sim 45$ kpc.
The observations were taken with the $g$ and $i$ filters, with total exposure
times of 14\,000 and 20\,000 seconds, respectively, in each of these two bands,
to reach $g\simeq 27.3$ and $i\simeq 25.9$. At the distance of M81, taken to
be 3.55 Mpc \citep{freedman94} throughout this paper, we detect approximately
the top 1.5 magnitudes of the red giant branch of metal-poor stellar
populations. The data were obtained in dark skies, with typical seeing of
0.9 and 0.8 arcseconds in the g- and i-bands, respectively (the relatively
poor seeing is due to the low elevation of the target as observed from Hawaii).
The individual exposures were recorded with a square-shaped dither pattern to
assist with the identification of bad pixels and the suppression of cosmic
rays.
\begin{figure*}
\includegraphics[clip=,width=0.45\textwidth]{cmd_fg_contours.ps}
\includegraphics[clip=,width=0.45\textwidth]{cmd_obs_contours.ps}
\caption{Left: The $i_{\circ}$ versus $(g-i)_{\circ}$ CMD of foreground stars
as predicted by The Besan\c{c}on Galactic population model in the direction
of M~81 over the same field-of-view as covered with a single MegaCam pointing.
Right: The $i_{\circ}$ versus $(g-i)_{\circ}$ CMD of objects detected in our
field and classified as stars. Regions of density less than ten stars per
bin of $0.15 \times 0.15$ magnitudes are plotted as points. Contours are
spaced by factors of two.}
\label{cmds}
\end{figure*}
The images were pre-processed by the CFHT Elixir pipeline for corrections
for bias, flat-fielding, and the fringing pattern. Photometric standards
observed over the season are used to determine the photometric zero point
in each passband. The images were then processed by the Cambridge Astronomical
Survey Unit (CASU) photometry pipeline \citep{irwin01}, in an identical manner
to that described in \citet{segall07}. The interested reader is referred to
this paper for more details. The software then proceeds to detect sources
and measures their photometry, the image profile, and shape. Based on the
information contained in the curve of growth, the algorithm classifies the
objects into noise detections, galaxies, and probable stars. Throughout the
paper we select as stars objects that have classification of -1 and -2 in both
bands. This corresponds to stars up to $2\,\sigma$ from the stellar locus.
To correct for the foreground extinction, we used the \citet{schlegel98} dust
map value of $E_{(B=V)}=0.08$, corresponding to $A_g = 0.303$ and $A_i = 0.167$
respectively.
\section{Results}
\label{results}
\begin{figure}
\includegraphics[clip=,width=0.45\textwidth]{cmd_m81_fg_corrected.ps}
\caption{Foreground-corrected CMD of stars in the MegaCam pointing.
The solid boxes mark the locations of the colour-magnitude diagram where
blue main sequence and red supergiant star candidates lie respectively.}
\label{cmd_fg_corrected}
\end{figure}
\begin{figure*}
\includegraphics[clip=,width=0.45\textwidth]{MS_HIcontours.ps}
\includegraphics[clip=,width=0.45\textwidth]{MS_RSG_HIcontours.ps}
\caption{(Left) Spatial distribution of main sequence star candidates,
shown as open circles, over the MegaCam pointing. The contours and
grayscale image show the H{\sc i} radio map from Yun et al. (1994).
(Right) similar to the left panel, but for both main sequence and red
supergiant star candidates. The H{\sc i} image adopted from Yun et al.'s
Figure 1 by permission from Macmillan Publishers Ltd: Nature, copyright
(1994). }
\label{young_stars_map}
\end{figure*}
\begin{figure}
\includegraphics[clip=,width=0.45\textwidth]{cmd_control_field_fg_corrected.ps}
\caption{Foreground-corrected CMD of stars in the three control fields.
The solid boxes mark the locations of the colour-magnitude diagram where
blue main sequence and red supergiant star candidates lie respectively.}
\label{cmd_control_field}
\end{figure}
\begin{figure}
\includegraphics[clip=,width=0.5\textwidth]{cmd_spiral_fg_corrected.ps}
\caption{The CMDs of stellar populations tracing the continuations of
both the nourthen and the southern H{\sc i} spiral arms as identified
in Fig.\,\ref{young_stars_map}. Superimposed on each CMD are the Padova
isochrones with a metallicity Z=0.008 and the indicated ages.
For a comparison purpose, the CMD of Holmberg IX is shown in the upper
left panel. }
\label{cmd_sp}
\end{figure}
As well as encompassing a large fraction of M81 and its outskirts, the
survey also intersects a non-negligible volume of the foreground Galaxy.
To subtract off the foreground counts, we used the Besan\c{c}on Galactic
population model \citep{robin03} to predict the foreground contamination.
This Galaxy model has shown its capacity to predict reliably the observed
foreground counts over large areas (see, e.g. \citealt{ibata07}).
To substract the background counts, we tessellated the survey area with
$0.25^{\circ}\times0.25^{\circ}$ bins and generated simulated catalogues
using the Besan\c{c}on model.
All stellar populations in the models with $i$-band magnitudes between
$17<i_{\circ}<26$ were accepted. To reduce noise in the randomly generated
catalogues, at each spatial bin we simulated a 30 times larger solid angle
and later corrected the density maps for this factor. Finally, the artificial
photometry was convolved with the observed magnitude-dependent error function.
The left panel of Fig.~\ref{cmds} shows the predicted
$i_{\circ}$ vs. $(g-i)_{\circ}$ colour-magnitude diagram (hereafter CMD) of
the Galaxy foreground stars in the direction of M81. Regions of the CMDs
with stellar densities higher than ten stars in a $0.15 \times 0.15$
magnitude bin are shown as contours with the contour levels spaced uniformly
by a factor of two. Stars are plotted as points when the densities are lower.
The foreground contamination is dominated by a prominent blue sequence,
at $0 \la (g-i)_{\circ} \la 0.8$, of halo stars at or close to the main
sequence turn-off at increasing distance through the Galactic halo.
On the red side of the CMD, a prominent nearly vertical sequence at
$2\la {g-i}_{\circ}\la 3$ and $19\la i_{\circ}\la 24$ is visible; the sequence
is the result of Galactic disk dwarf stars accumulating over a large
range of distances along the line of sight.
The right panel of figure \ref{cmds} shows the combined CMD of all objects
detected in our images and classified as stars in the deep MegaCam field.
The CMDs are shown as density contours to reveal features in otherwise
crowded regions. In addition to the blue and red sequences of foreground
contaminants, a prominent vertical sequence of stars bluer than stars
populating the blue foreground sequence, i.e., $(g-i)_{\circ} \le 0$, and
covering a wide range of magnitudes, i.e., $19 \la g_{\circ} \la 26$, is
present. A second sequence of bright, i.e., $21 \la i_{\circ} \la 22$, and
red, i.e., $1.5 \la (g-i)_{\circ} \la 2.5$, stars is present. A third CMD
feature is revealed by red giant branch stars.
Figure \ref{cmd_fg_corrected} shows the foreground-subtracted CMD of the
stellar population distributed over the MegaCam field. The CMD is typical
of stellar systems with extended star formation histories, with a prominent
young stellar component. We see the upper main sequence and probable
helium-burning blue stars, the red supergiant branch, and probably some
young and intermediate age asymptotic giant stars. At $i_{\circ}\sim 25$,
the contamination is dominated by background compact galaxies, unresolved
in ground-based images and prone to misclassification as stellar objects.
Characterising and correcting for the background contamination at faint
magnitudes is beyond the scope of the present contribution, and will be
addressed in a forthcoming paper. To reduce the contamination from
background objects, we have selected stars brighter than $g_{\circ}=25$.
A second potential source of contamination when tracing the spatial
distribution of young stars in the debris field is the population of
intermediate and old stars, either genuine members of the tidal debris
or belonging to M81 and/or NGC~3077. These stars start to dominate at
$i_{\circ}\sim 24$ and $(g-i)_{\circ}\sim 1$. We have excluded stars fainter
than the tip of the red giant branch magnitude at the distance of M81,
and with colours redder than the reddest colours predicted for red
supergiant stars. The criteria used to identify young star candidates
were defined by inspecting the foreground-corrected CMD and the predicted
locations of young stellar populations at the distance of M81.
The selection boxes of young stars are shown in Fig.\,\ref{cmd_fg_corrected}.
Objects within the blue selection box are defined as main sequence star
candidates, and those in the red selection box as red supergiant star
candidates. The blue selection box is affected by a modest contamination
from foreground Galactic stars and background galaxies, which tend to have
red colours, than the red selection box. Despite the selection criteria
imposed to select samples of young stars free of background contamination,
the red side of the CMD is expected to suffer still from contamination
(see \citet{davidge08} for more details). We have therefore adopted red
supergiant star candidates to be only secondary tracers of the debris field.
A potential source of errors when identifying bright stars in distant objects is
blending. As discussed in \citet{davidge08}, who had also used the
MegaPrime/MegaCam imager to investigate the bright stellar content of the
M81 group, the ages of the young stellar components estimated for
Holmberg IX, BK 3N, and Garland from MegaCam CMDs are consistent with
those obtained from CMDs obtained with the Hubble Space Telescope
\citep{makarova02}, suggesting that blending is not severe, if any.
The distribution of the young stellar clumps detected in the intergalactic
environment along H{\sc i} tidal arms rather than been randomly distributed
within the field supports this conclusion.
By comparing MegaCam images with those obtained with the Advanced Camera
for Surveys on board the Hubble Space Telescope, \citet{davidge08} had shown
that the bright blue sources in the MegaCam images are consistent with being
single main sequence stars. In addition, for the blue objects detected in our images
to be the results of blending, the blend stars have to be bright in $g$-band yet
faint in $i$-band, which is inconsistent with the fact that the bulk of stars in these
remote regions of galaxies are intrinsically faint and red.
\begin{figure*}
\includegraphics[clip=,width=0.85\textwidth]{cmd_southern_debris_fg_corrected.ps}
\caption{The CMDs of the stellar clumps identified along the southern H{\sc i}
tidal arm between M81 and NGC~3077, along with BK 3N and Garland.
Superimposed on each CMD are the Padova isochrones with a metallicity
Z=0.008 and the indicated ages. }
\label{cmd_td}
\end{figure*}
\begin{figure}
\includegraphics[clip=,width=0.5\textwidth]{cmd_Davidge_TDOs_fg_corrected.ps}
\caption{The CMDs of the young stellar grouping identified by \citet{davidge08}
in the M81 tidal debris. Superimposed on each CMD are the Padova isochrones
with a metallicity Z=0.008 and the indicated ages. }
\label{cmd_tdo_davidge08}
\end{figure}
Figure \ref{young_stars_map} shows the spatial distribution of main sequence
star candidates (left panel), and both main sequence and red supergiant
candidates (right panel) over the surveyed area. The contours and the
grayscale image show the H{\sc i} radio map from \citet{yun94}. The crowding
is such that individual stars are not well resolved in the central dense
region of M81 and NGC~3077, and their photometry is affected by large errors.
As expected, the bulk of young stellar populations are distributed over the
star-forming disk of M81. Beyond the young disk of M81, a number of stellar
concentrations distributed over a much larger scale are visible, of which a
few have been already reported in the literature, i.e., Holmberg IX, BK 3N,
Garland, M81-West, Arp's Loop (A0952+69), and the tidal debris objects
recently reported by \citet{davidge08}. In addition to the structure combining
M81-West and the first object in the list of \citet{davidge08}, tracing the
continuation of the H{\sc i} spiral structure of M81, a second arm of young
stars (M81-West2) tracing the continuation of the second H{\sc i} spiral
structure is visible. Similar structures of young stars tracing the
continuation of the H{\sc i} spiral structures (M81-SE) are observed to the
south along the major axis. Wide field GALEX images of M81 \citep{gildepaz07}
show that these structures of young star candidates are sites of Ultra-Violet
emission.
More strikingly, a number of previously unknown overdensities of young star
candidates can be seen along the tidal bridge of gas stretching from M81 to
its neighbour NGC~3077. Interestingly, BK3N and Garland appear to lie at
the two extremities of this chain of concentrations of young stars, close to M81
and NGC~3077 respectively. The southern H{\sc i} arm has been resolved
into a series of clouds with sizes of approximatively 4 kpc, which are embedded
in a more tenuous H{\sc i} distribution \citep{vdh79}. The new young stellar
systems to the south of M81 appear to trace the overdense clouds along this
H{\sc i} arm, with no clear bridge of young stars connecting these systems.
The foreground-corrected CMDs of three control fields sampling similar
areas on the sky as the new stellar clumps, and that are located away from
the M81-NGC~3077 and M81-M82 debris fields, are shown in
Fig.\,\ref{cmd_control_field}. The location of the control fields on the sky
are shown in Fig.\,\ref{footprint}. The selection boxes of both main sequence
and red supergiant star candidates are overplotted on the CMDs of the control
fields. The stellar contents of the control fields are all similarly dominated by
objects fainter than $g_{\circ}\ga 25$, i.e., red giant branch stars and
probable background compact galaxies, with a noticeable absence of both
stars with $(g-i)_{\circ} \la 0$, i.e., young main sequence star candidates,
and stars with properties similar to those expected for red supergiant stars
at the distance of M81, i.e., $(g-i)_{\circ} \sim 1.5$ and $i_{\circ} \sim 22$.
Fig.\,\ref{cmd_sp} shows the CMDs of the stellar populations that trace the
continuations of the H{\sc i} both to the north and to the south of the galaxy.
Superimposed on the CMD of each stellar system as indicated in each panel
are the Padova isochrones with a metallicity Z=0.008 and the indicated ages.
The locations of both main sequence stars and red supergiant stars in the
CMDs are best matched by Z=0.008 and Z=0.004 isochrones.
The plumes of bright main sequence stars are also reasonably matched
by isochrones of both higher and lower metallicities, however the predicted
locations of red supergiant stars are either bluer or redder than observed.
Fig.\,\ref{cmd_td} shows the CMDs of the stellar clumps identified along the
southern H{\sc i} tidal tail, and the CMDs of BK 3N and Garland, both located
at the two extremities of the same gaseous arm. Overplotted are the Padova
isochrones with a metallicity Z=0.008 and the indicated ages. All the stellar
systems were taken to be situated at the same distance as M81.
Compared to the control fields, the CMDs of the stellar clumps along the
southern H{\sc i} tidal tail show clear over-densities of objects with luminosities
and colours expected for main sequence and red supergiant star candidates.
Similar to the stellar populations tracing the continuations of the H{\sc i} spiral
structures, the CMDs of stellar systems tracing the southern H{\sc i} tail are best
matched by isochrones with metallicities ${\rm Z \sim 0.008}$. A similar conclusion
has been derived for the young stellar systems in the tidal bridge connecting M81
and M82 \citep{demello08a}, and is consistent with the gas-phase metallicity of
an H{\sc ii} region in Holmberg IX as measured by \citet{makarova02}.
The CMDs show that the three newly reported stellar systems, similar to BK 3N
and Garland, contain stars with properties consistent with a range of ages,
i.e., they are not simple stellar populations. \citet{davidge08} derived a similar
conclusion for three other young stellar systems of comparable extents and
stellar densities located respectively close to the area identified as M81-West
by \citet{sun05} and Holmberg IX. The ages of the youngest stars in
BK 3N (Garland) are consistent with the absence (the presence) of associated
H$\alpha$ emission \citep{karachentsev07}. Stars with photometry that are
best matched with isochrones of ages older than a few hundred Myr, or even
a few billion years, are present in abundance. We cannot however be sure
that these stars are genuine members of the stellar systems, whether they
belong to the stellar halo of M81, or whether they were ejected into the
intergalactic medium during the interaction.
The CMDs of the stellar structures tracing the continuation of the H{\sc i} spiral
structures of M81 (M81-West2, M81-SE) and Holmberg IX show the presence
of stars younger than these in the stellar over-densities distributed along the
southern H{\sc i} arm. The CMDs of the new stellar clumps show the presence
of stars with photometric properties consistent with their youngest stars formed
around 40\,Myr ago. BK 3N, located at the western end of the H{\sc i} tidal arm,
appears to be dominated by stars of similar ages to the other three clumps
along the tidal arm. The Garland structure, at the easter end of the H{\sc i} tidal
arm, however contains stars as young as those observed in Holmberg IX and
the spiral arms. This suggests that the star formation activity was likely truncated
earlier in the stellar systems within the tidal field that are away from large
galaxies than in those in their close vicinity.
The $i_{\circ}$ vs. $(g-i)_{\circ}$ CMDs of the recently reported stellar groupings
in the debris field of M81 by \citet{davidge08} are shown in
Fig.\,\ref{cmd_tdo_davidge08}. The locations of these stellar groupings on the
sky are indicated in Fig.\,\ref{footprint}. Superimposed on the CMD of each stellar
system as indicated in each panel are the Padova isochrones with a metallicity
Z=0.008 and the indicated ages. Similar to other young stellar systems within the
debris field of M81, the CMDs of those stellar systems are best matched by
isochrones with metallicities ${\rm Z \sim 0.008}$. The photometric properties of
turn-off stars in those groupings are consistent with their youngest stars formed
$\sim 40$\,Myr ago, consistent with the age estimates of \citet{davidge08}.
These stellar systems appear to be dominated by stellar populations similar to
those of the stellar clumps distributed along the southern H{\sc i} arm.
\section{Discussion}
\label{discussion}
Numerous multi-wavelength data sets have unambiguously identified localised
regions with signs of current and/or recent star formation activity distributed along
H{\sc i} tidal tails in interacting systems \citep[e.g.][]{weilbacher03,hibbard05,demello08b},
with a number of these regions suspected to be bounded, the so-called tidal dwarf
galaxies \citep[e.g.][]{duc98,braine01,hancock09}. The star formation activity along
the tidal tails appears to be distributed with a similar morphology to H{\sc i}
\citep[e.g.][]{hibbard05,neff05,hancock07}, in agreement with our finding of the new
stellar clumps reported here tracing the dense regions along the southern H{\sc i}
tidal tail.
\citet{hibbard05} have found that UV colours of localised regions of star formation
along the tidal tails of the archetypal merging system NGC 4038/39 are consistent
with continuing star formation. \citet{neff05} have argued however, for the case of
NGC 7769/71, NGC 5713/19, and the NGC 520 system, that most of young stars in
the tails have most likely formed in single bursts. The CMDs of stellar clumps in the
debris field of M81, e.g. Arp's loop region, Holmberg IX, contain stars with photometric
properties consistent with a wide range of ages, i.e., from $\sim10\,$Myr to
$\sim1\,$Gyr \citep[e.g.][]{demello08b,sabbi08}, suggesting extended star formation
histories. Deep imaging data show however a lack of any concentration of old stars
associated with the blue stars, suggesting that the ``old'' stellar component seen in
the CMDs of those stellar clumps were formed in the stellar disks of M81 and ejected
into the intergalactic medium during tidal passages, whereas the young stars have
formed in the tidal debris \citep{demello08b, weisz08}. The spatial distribution of
red giant branch stars over the region connecting M81 and NGC~3077 does not
show any noticeable concentrations of old stars associated to the young stellar
clumps identified along the southern H{\sc i} tidal tail. As for, e.g. Holmborg IX and
Arp's loop region, this suggests that the old stars seen in the CMDs of the newly
reported stellar clumps should have come from one of the interacting systems while,
since the stellar clumps along the southern tidal tail are considerably younger than
its dynamical age, e.g. $\sim 250\,$Myr \citep{yun99}, young stars formed on site.
Fitting single stellar population synthesis models to UV/optical colours of UV-bright
stellar substructures within the tidal tails of four ongoing galaxy mergers,
\citet{neff05} found that the star formation appears to be older near the parent
galaxies and younger at increasing distances. They have suggested that this could
be because the star formation occurs progressively along the tails, or because the
star formation has been inhibited near the galaxy/tail interface.
\citet{hibbard05} had reported negative UV and optical colour gradients along
the tidal tails of the ``Antennae'' system, indicative of negative age gradients when
moving outward along the tails \citep[see also][]{hibbard01}. Note that the observed
colour gradients could be accounted for alternatively by the presence of composite
stellar populations. The star formation within the tidal tails in M81 debris field
appears to be different. The CMDs of the bulk of stellar clumps in the debris field
of M81 indicate that they have ceased forming stars at similar epochs in the past,
i.e., $\sim40$\,Myr ago. This suggests that the star formation throughout M81 tidal
debris field could have been triggered by common events
\citep[see][for a similar conclusion]{davidge08}, and that the physical conditions
within the dense regions along the southern H{\sc i} tidal tail are comparable.
The systems within the debris field with younger stars, i.e., Holmberg IX and Garland,
are both in the close vicinity of M81 and NGC~3077 respectively, in contrast with the
findings of \citet{neff05}. The diversity of these star formation histories could be most
likely related to different gas contents and conditions within the tidal tails of those
interacting systems.
The exact nature of the previously known young stellar systems within the tidal
field of the M81 group, i.e., Holmberg IX, Gerland, and BK 3N, is not entirely
clear yet. Detailed modelling of the dynamics of the M81 group suggests that
the three largest galaxies in the system had an interaction $\sim 250$ Myr ago
for M81 and NGC~3077, and $\sim 200$ Myr ago for M82 and M81 \citep{yun99}.
The modelling of optical CMDs of Holmberg IX, BK 3N, and Arp's loop, of
comparable depth to ones presented here, suggested that these galaxies have
experienced star formation between about 20 and 200 Myr ago \citep{makarova02}.
It has been argued then that these galaxies are tidal dwarf condidates that
formed from dust and gas that was blown away from M81 and/or other galaxies in
the group \citep{boyce01,makarova02}. BK3N may be alternatively a pre-existing
dwarf irregular galaxy undergoing an interaction with M81 \citep{boyce01}.
The old stellar population associated spatially to Holmberg IX is suspected to
belong quite likely to the outer regions of M81, suggesting that this stellar system
is of a tidal origin \citep{sabbi08}.
The ages and metallicities of the isochrones that best reproduce the CMDs of
the stellar systems along the southern H{\sc i} arm are similar to these needed
to account for the properties of the stellar contents of previously known tidal
dwarf candidates in the tidal debris field. This suggests that they both could
share similar star formation histories and might be associated to similar events.
These isochrone metallicities are significantly higher than the typical
metallicities of dwarf galaxies of comparable luminosities \citep[e.g.][]{rm95}.
This suggests that these stellar systems have been assembled from pre-enriched
material. This is consistent with the conclusions of \citet{boone05} who found
that the abundances and physical conditions of the molecular complex situated
near the line of sight toward Holmberg IX are similar to those found in the disks
of spiral galaxies. The distribution of the stellar clumps along an H{\sc i} tail,
tracing the densest clouds within the gaseous arm, indicates that these systems
may have been assembled out of gas pulled from one of the large interacting
galaxies in the group. A primary criterion for the determination of the nature of
these objects is to measure their mass-to-light ratio, which tend to be low for
tidal dwarf galaxies due to the absence of dark matter \citep[e.g.][]{BH92, duc00}.
Unfortunately this cannot be measured from the dataset in hand. Without this
measurement, we can only conclude that the newly reported stellar clumps are
likely (among the nearest) tidal dwarf galaxies.
\begin{table}
\caption{Coordinates of the new young stellar clumps along the
southern H{\sc i} arm. }
\label{gal_prop_obs1}
\begin{tabular}{lll}
\hline
ID & R.A. (J2000) & Dec. (J2000) \\
\hline
Clump I & 09:57:21.2 & 68:42:55 \\
Clump II & 09:59:40.4 & 68:39:19\\
Clump III & 10:00:40.4 & 68:39:37 \\
\end{tabular}
\end{table}
\section*{Acknowledgments}
We would like to warmly thank Mike Irwin for (various) helpful discussions.
|
\section{Introduction}
A \emph{$2$-colouring} of a graph is an assignment of the black or white colour to the nodes so that each black node is adjacent only to white nodes, and vice versa. A \emph{weak $2$-colouring} assigns the colours so that each non-isolated black node is adjacent to at least one white node, and vice versa.
A graph can be $2$-coloured if and only if it is bipartite; a weak $2$-colouring always exists. Given a global view of the graph, it is easy to find a $2$-colouring of a bipartite graph and a weak $2$-colouring of any graph.
In a distributed setting, it is not possible to $2$-colour a bipartite graph without essentially global information of the whole graph. However, Naor and Stockmeyer~\cite{naor95what} showed in 1995 that one can find a weak $2$-colouring with a \emph{constant-time} synchronous distributed algorithm, assuming that the degree of each node is \emph{odd} and bounded by a constant.
Constant-time distributed algorithms are known as \emph{local algorithms} \cite{naor95what,suomela09survey} -- in a local algorithm, the output of each node depends only on its local neighbourhood, and the radius of the neighbourhood does not depend on the number of nodes in the network.
\subsection{Contributions}
We present local approximation algorithms for both $2$-coloured and weakly $2$-coloured graphs. We assume that a colouring is given in the input, i.e., that every node knows its colour. We study exactly how much this additional information helps from the perspective of local approximation algorithms.
We focus on two classical problems -- minimum dominating set and maximum matching. We consider bounded-degree graphs; we assume that there is a known constant $\Delta$ such that the degree of any node is at most~$\Delta$. The results are summarised in Table~\ref{tab:results}. All results are tight: there are matching upper and lower bounds.
\begin{table}
\centering
\newcommand{\hspace{1.5em}}{\hspace{1.5em}}
\begin{tabular}{l@{\hspace{1.5em}}l@{\hspace{1.5em}}l@{\hspace{1.5em}}l}
\toprule
Problem & approx. factor & upper bound & lower bound \\
\midrule
dominating set \\
-- no colouring, even $\Delta$ & $\Delta+1$ & trivial & \cite{czygrinow08fast,lenzen08leveraging} \\
-- no colouring, odd $\Delta$ & $\Delta$ & Theorem~\ref{thm:pos-ds-odd} & \cite{czygrinow08fast,lenzen08leveraging} \\
-- weak $2$-colouring & $(\Delta+1)/2$ & Theorem~\ref{thm:pos-ds-weak} & Theorem~\ref{thm:neg-ds-strong} \\
-- $2$-colouring & $(\Delta+1)/2$ & Theorem~\ref{thm:pos-ds-weak} & Theorem~\ref{thm:neg-ds-strong} \\
\midrule
matching \\
-- no colouring & none & --- & \cite{czygrinow08fast} \\
-- weak $2$-colouring & $(\Delta+1)/2$ & Theorem~\ref{thm:pos-m-weak} & Theorem~\ref{thm:neg-m-weak} \\
-- $2$-colouring & $1+\varepsilon$ & Theorem~\ref{thm:pos-m-strong} & --- \\
\bottomrule
\end{tabular}
\caption{The best possible approximation factors achievable by a deterministic local algorithm.}\label{tab:results}
\end{table}
In particular, we show that a weak $2$-colouring is as good as a $2$-colouring from the perspective of the local approximability of the dominating set problem. Furthermore, a weak $2$-colouring provides enough symmetry-breaking information so that an approximation of a maximum matching can be found locally. Finally, with a $2$-colouring, the maximum matching can be approximated to within an arbitrary constant.
We also look at a third problem, maximum independent set. There is a trivial local $\Delta$-approximation algorithm for independent set in $2$-coloured graphs: take all white nodes and all isolated black nodes. However, we show that in weakly $2$-coloured graphs, the problem does not admit any local constant-factor approximation algorithm (Theorem~\ref{thm:neg-is-weak}).
\subsection{Model of distributed computing}
Let $\mathcal{G} = (V, E)$ be a graph identified with a distributed system. Each node $v \in V$ is a device; there is an edge $\{u,v\} \in E$ if $u$ and $v$ can communicate with each other. To avoid trivialities, we assume that there are no isolated nodes in~$\mathcal{G}$.
Each node runs the same deterministic algorithm $\mathcal{A}$. Communication is synchronous: on every time step, all nodes first receive messages from their neighbours, then all nodes perform local computation, and finally all nodes send messages to their neighbours. The algorithm running in the node $v \in V$ knows the degree of $v$. Furthermore, if the node $v$ has a label (such as a colour or a unique identifier), then $\mathcal{A}$ has access to the label. If the edges are oriented, then $\mathcal{A}$ knows which edges are outgoing and which edges are incoming. Finally, every node knows the maximum node degree $\Delta$.
The distributed algorithm $\mathcal{A}$ is a local algorithm if there is a constant $T$ such that the algorithm completes in $T$ synchronous communication rounds, regardless of the input graph $\mathcal{G}$. The algorithm $\mathcal{A}$ and the constant $T$ may depend on the degree bound $\Delta$; however, the time $T$ cannot depend on the number of nodes in $\mathcal{G}$.
The results that we present are essentially oblivious to any other details of the model of distributed computing. All lower bounds (impossibility results) hold even if we use Linial's~\cite{linial92locality} model. We can assume that each node knows the total number of nodes $\mysize{V}$, each node is assigned a unique identifier from the set $\{1, 2, \dotsc, \mysize{V}\}$, local computation is free, and the size of a message is unbounded.
Our upper bounds (algorithms) do not need to exploit any of these assumptions. The nodes do not need to know $\mysize{V}$. Local computations are simple and messages are small; in particular, the size of a message does not depend on $\mysize{V}$. Furthermore, the algorithms do not require unique identifiers. With the exception of Theorem~\ref{thm:pos-ds-odd}, the algorithms only assume that there is a \emph{port numbering}~\cite{angluin80local}: each node imposes an ordering on incident edges.
\section{Prior work}\label{sec:prior}
Randomised local algorithms exist for dominating set~\cite{kuhn05constant-time,kuhn05price,kuhn06price}, matching~\cite{wattenhofer04distributed,hoepman06efficient,nguyen08constant-time}, and independent set~\cite{czygrinow08fast}. However, deterministic local approximation algorithms are scarce. The set of all nodes is a trivial ${(\Delta+1)}$-approximation of a minimum dominating set, and there are local constant-factor approximation algorithms for dominating set in planar graphs~\cite{czygrinow08fast,lenzen08what}. Some positive results are known for matchings in bounded-degree $2$-coloured graphs: local algorithms exist for finding a maximal matching~\cite{hanckowiak98distributed} and a constant-factor approximation of a maximum-weight matching~\cite{floreen09almost-stable}.
To present the earlier negative results on which we build our lower bounds, we need the following definition: a \emph{numbered directed $n$-cycle} $\mathcal{C}$ is a directed $n$-cycle where each node is assigned a unique identifier from the set $\{1, 2, \dotsc, n\}$. Each node has one incoming and one outgoing edge.
Linial's~\cite{linial92locality} seminal work shows that there is no local algorithm for finding a maximal independent set in $\mathcal{C}$. Recently, Czygrinow et al.\mbox{}{}~\cite{czygrinow08fast} and Lenzen and Wattenhofer~\cite{lenzen08leveraging} have extended this result to the approximability of the maximum independent set problem. We include here a proof of an adaptation of the inapproximability result, since the proofs of our lower bounds build directly upon it. We follow Czygrinow et al.\mbox{}{}'s~\cite{czygrinow08fast} techniques. In the proof, a \emph{$k$-set} is a set with $k$ elements.
\begin{theorem}[\mythmcite{czygrinow08fast,lenzen08leveraging}]\label{thm:neg-is-cycle}
For any $\alpha \ge 1$ and any local algorithm $\mathcal{A}$, there exists an integer $n_0$ such that for every $n\ge n_0$ there is a numbered directed $n$-cycle $\mathcal{C}$ where $\mathcal{A}$ does not produce an $\alpha$-approximation for maximum independent set.
\end{theorem}
\begin{proof}
Denote by $T$ the number of synchronous communication rounds that $\mathcal{A}$ takes. Let $m=\lceil 16 T \alpha\rceil$. By Ramsey's theorem~\cite{ramsey30problem}, there is a finite $N$ with the following property: Let $S$ be a set with at least $N$ elements, and assign an arbitrary label $f(X) \in \{0,1\}$ to each ${(2T+1)}$-set $X \subset S$. Then there is an $m$-set $A \subset S$ and a label $\ell \in \{0,1\}$ such that $f(X) = \ell$ for every ${(2T+1)}$-set $X \subset A$. We say that $A$ is an $\ell$-coloured $m$-subset of $S$.
Let $n_0=\lceil 8 N \alpha\rceil$, let $n \ge n_0$ and $S = \{1, 2, \dotsc, n\}$. We first assign a label $f(X) \in \{0,1\}$ to each ${(2T+1)}$-set $X \subset S$. Let $X = \{x_1, x_2, \dotsc, x_{2T+1}\}$ with $x_1 < x_2 < \dotso < x_{2T+1}$. Consider a fragment of a numbered directed $n$-cycle with the unique identifiers $x_1, x_2, \dotsc, x_{2T+1}$, in this order; let $f(X) \in \{0,1\}$ be the output of $\mathcal{A}$ for the node $x_{T+1}$, with $1$ denoting that the node joins the independent set. Observe that the output only depends on the set~$X$.
Let us next construct a numbered directed $n$-cycle $\mathcal{C}$ as follows. By the choice of $n_0$, we can find an $\ell_1$-coloured $m$-subset $A_1$ of $S$ for an $\ell_1 \in \{0,1\}$. As $\mysize{S \setminus A_1} \ge N$, we can then find an $\ell_2$-coloured $m$-subset $A_2$ of $S \setminus A_1$ for an $\ell_2 \in \{0,1\}$, etc. Overall we find $p=\lceil{(n-N)}/m\rceil$ disjoint sets $A_1, A_2, \dotsc, A_p$ such that $A_i$ is an $\ell_i$-coloured $m$-subset of $S$. Let $A_i = \{a_i^1, a_i^2, \dotsc, a_i^m\}$ with $a_i^1 < a_i^2 < \dotso < a_i^m$ for each $i$. Let $S \setminus (\bigcup_i A_i) = \{s_1, s_2, \dotsc, s_k\}$. Assign the unique identifiers in $\mathcal{C}$ in the order
\[
a_1^1, a_1^2, \dotsc, a_1^m, a_2^1, a_2^2, \dotsc, a_p^m, s_1, s_2, \dotsc, s_k.
\]
An optimal independent set of $\mathcal{C}$ contains at least $n/3$ nodes; to prove the theorem, it suffices to show that the algorithm $\mathcal{A}$ outputs $1$ for at most $n/(4\alpha)$ nodes. To see this, observe that for each $i$ the output of the nodes $a_i^{T+1}, a_i^{T+2}, \dotsc, a_i^{m-T}$ is $\ell_i$. Since they cannot all be in the independent set, we have $\ell_i = 0$. Hence there are at most $2Tp + k$ nodes that output~$1$. By construction, $2Tp \le 2Tn/m \le n/(8\alpha)$ and $k = n - mp \le N \le n/(8\alpha)$.
\end{proof}
This immediately gives a negative result for the approximability of a maximum matching as well: given a matching $M$ in a numbered directed $n$-cycle, we can construct an independent set $I = \{ u : (u,v) \in M \}$ with $\mysize{I} = \mysize{M}$.
\begin{corollary}[\mythmcite{czygrinow08fast}]\label{cor:neg-m-general}
There is no local constant-factor approximation algorithm for the maximum matching problem.
\end{corollary}
\section{Lower bounds and local reductions}\label{sec:lower}
In this section, we present local reductions that establish lower bounds for local approximation algorithms. All reductions are from the maximum independent set problem in numbered directed cycles (Theorem~\ref{thm:neg-is-cycle}). The reductions yield the strongest possible negative results, as there is a matching positive result for each of them. As an introduction to the local reductions, we begin with a known result for general graphs; Theorem~\ref{thm:neg-ds-general} is a restatement of the negative results for planar graphs~\cite{czygrinow08fast} and unit-disk graphs~\cite{lenzen08leveraging}.
\begin{theorem}[\mythmcite{czygrinow08fast,lenzen08leveraging}]\label{thm:neg-ds-general}
For any even $\Delta \ge 2$ and $\varepsilon > 0$, there is no local algorithm with approximation factor ${(\Delta+1-\varepsilon)}$ for the minimum dominating set problem.
\end{theorem}
\begin{proof}
Suppose that such an algorithm $\mathcal{A}$ exists for some $\Delta=2k$. Let $\alpha=\Delta(\Delta+1)/\varepsilon$. We will use $\mathcal{A}$ to find an independent set with at least $n/\alpha$ nodes in numbered directed $n$-cycles for any $n$ divisible by $\Delta+1$. This is a contradiction with Theorem~\ref{thm:neg-is-cycle}.
Given an $n$-cycle $\mathcal{C}$ (Figure~\ref{fig:reductions}a), we construct the $2k$-regular graph $\mathcal{G} = \mathcal{C}^k$ (Figure~\ref{fig:reductions}b illustrates the case $k=2$); the node identifiers are inherited from the cycle $\mathcal{C}$. We simulate the algorithm $\mathcal{A}$ in the graph $\mathcal{G}$. There is a dominating set of $\mathcal{G}$ with $n/(\Delta+1)$ nodes; hence $\mathcal{A}$ must output a dominating set $D$ with at most
\[
\Bigl(1 - \frac{\varepsilon}{\Delta+1}\Bigr) n
\]
nodes. Thus $\mysize{V \setminus D} \ge \varepsilon n/(\Delta+1)$. The subgraph of $\mathcal{C}$ induced by $V \setminus D$ consists of paths with at most $\Delta$ nodes each; hence there are at least
\[
\frac{\varepsilon n}{\Delta(\Delta+1)} = \frac{n}{\alpha}
\]
such paths. Construct an independent set $I$ with $\mysize{I} \ge n/\alpha$ by taking the first node of each such path.
\end{proof}
\begin{figure}
\centering
\input{fig1.pdf_t}
\caption{The local reductions for the lower bounds.}\label{fig:reductions}
\end{figure}
\begin{theorem}\label{thm:neg-ds-strong}
For any $\Delta \ge 2$ and $\varepsilon > 0$, there is no local algorithm with approximation factor ${(\Delta+1)/2-\varepsilon}$ for dominating sets in $2$-coloured graphs.
\end{theorem}
\begin{proof}
Suppose that such an algorithm $\mathcal{A}$ exists. Let $\alpha = (\Delta^2-1)/(2\varepsilon)$. We will use $\mathcal{A}$ to find an independent set with at least $n/\alpha$ nodes in numbered directed $n$-cycles for any $n$ divisible by $\Delta+1$. This is a contradiction with Theorem~\ref{thm:neg-is-cycle}.
Given an $n$-cycle $\mathcal{C}$ (Figure~\ref{fig:reductions}a), we construct a $\Delta$-regular $2$-coloured graph $\mathcal{G}$ as follows (Figure~\ref{fig:reductions}c shows the case $\Delta=3$). For each node $v$ in $\mathcal{C}$, there is a white node $v_1$ and a black node $v_2$ in $\mathcal{G}$. If the directed path from $u$ to $v$ in $\mathcal{C}$ has at most $\Delta-1$ edges, then there is an edge $\{u_1, v_2\}$ in $\mathcal{G}$. The node identifiers are inherited from the cycle~$\mathcal{C}$: for example, let $v_1 = 2v-1$ and $v_2 = 2v$.
We simulate the algorithm $\mathcal{A}$ in the graph $\mathcal{G}$. There is a dominating set of $\mathcal{G}$ with $2n/(\Delta+1)$ nodes; hence $\mathcal{A}$ must output a dominating set $D$ with at most
\[
\Bigl(1-\frac{2\varepsilon}{\Delta+1}\Bigr) n
\]
nodes. Let $B = \{ v \in V : v_1 \notin D\textrm{ and } v_2 \notin D \}$; we have
\[
\mysize{B} \ge \mysize{V} - \mysize{D} \ge \frac{2\varepsilon n}{\Delta+1}.
\]
The subgraph of $\mathcal{C}$ induced by $B$ consists of paths with at most $\Delta-1$ nodes each; hence there are at least
\[
\frac{2\varepsilon n}{(\Delta+1)(\Delta-1)} = \frac{n}{\alpha}
\]
such paths. Construct an independent set $I$ with $\mysize{I} \ge n/\alpha$ by taking the first node of each such path.
\end{proof}
\begin{theorem}\label{thm:neg-m-weak}
For any $\Delta \ge 3$ and $\varepsilon > 0$, there is no local algorithm with approximation factor ${(\Delta+1)/2-\varepsilon}$ for maximum matching in weakly $2$-coloured graphs.
\end{theorem}
\begin{proof}
Assume that such an algorithm $\mathcal{A}$ exists. Let
\[
\varepsilon' = \frac{2\varepsilon}{\Delta+1-2\varepsilon}, \qquad
\alpha = \frac{2\Delta-1}{\varepsilon'}.
\]
We will use $\mathcal{A}$ to find an independent set with at least $n/\alpha$ nodes in numbered directed $n$-cycles for any even $n$. This is a contradiction with Theorem~\ref{thm:neg-is-cycle}.
Given an $n$-cycle $\mathcal{C} = (V_{\mathcal{C}}, E_{\mathcal{C}})$, we construct a weakly $2$-coloured graph $\mathcal{G}$ as follows (Figure~\ref{fig:reductions}d shows the case $\Delta=3$). For each node $v$ in $\mathcal{C}$, there are $\Delta+1$ nodes in $\mathcal{G}$: white nodes $v_1, v_2, \dotsc, v_\Delta$ and a black node $v_0$. Each black node $v_0$ has degree $\Delta$: it is adjacent to all white nodes $v_1, v_2, \dotsc, v_\Delta$. Each white node has degree $3$: for each edge $(u,v)$ in $\mathcal{C}$, there are edges $\{u_1, v_1\},\allowbreak \{u_2, v_2\}, \dotsc, \allowbreak \{u_\Delta, v_\Delta\}$ in $\mathcal{G}$.
There is a matching with $(\Delta+1)n/2$ edges in $\mathcal{G}$. To see this, let $X$ be a perfect matching in $\mathcal{C}$, with $\mysize{X} = n/2$. Construct a perfect matching in $\mathcal{G}$ as follows: for each edge $(u,v)$ in $X$, choose the edges $\{v_0, v_\Delta\}$, $\{u_0, u_\Delta\}$, and $\{u_i, v_i\}$ for each $i \in \{1,2,\dotsc,\Delta-1\}$.
We simulate the algorithm $\mathcal{A}$ in the graph $\mathcal{G}$. The algorithm must output a matching $M$ with at least ${(1+\varepsilon')}n$ edges. Since there are $n$ black nodes in $\mathcal{G}$, there are at least $\varepsilon'n$ edges in $M$ that connect a pair of white nodes. For each $i = 1, 2, \dotsc, \Delta$, let
\[
I_i = \bigl\{ u \in V_{\mathcal{C}} : (u,v) \in E_{\mathcal{C}},\, \{ u_i, v_i \} \in M \bigr\}.
\]
Now each $I_i$ is an independent set in $\mathcal{C}$ and $\sum_i \mysize{I_i} \ge \varepsilon' n$.
We will now use the sets $I_i$ to construct an independent set $I$ in $\mathcal{C}$ with $\mysize{I} \ge n/\alpha$. At least one of the sets $I_i$ satisfies this condition, but a local algorithm cannot find the right index $i$; hence we proceed as follows. We begin with $I = \emptyset$. At each iteration $i = 1, 2, \dotsc, \Delta$, for each node $v \in I_i$ in parallel, we (i)~add $v$ to $I$, and (ii)~remove the copy of $v$ and its neighbours from $I_i, I_{i+1}, \dotsc, I_\Delta$.
In the end, $I$ is an independent set and each $I_i$ is empty. Furthermore, for each node added to $I$ there are at most $2\Delta-1$ nodes that we removed from $I_1, I_2, \dotsc, I_\Delta$; the worst case is that $I_1$ contains a node $v$ and each of $I_2, I_3, \dotsc, I_\Delta$ contains the two neighbours of $v$. Hence $\mysize{I} \ge \varepsilon' n / {(2\Delta-1)} = n/\alpha$.
\end{proof}
\begin{theorem}\label{thm:neg-is-weak}
There is no local constant-factor approximation algorithm for independent set in weakly $2$-coloured graphs.
\end{theorem}
\begin{proof}
The reduction is from an $n$-cycle $\mathcal{C}$ to the $3$-regular weakly $2$-coloured graph $\mathcal{G}$ illustrated in Figure~\ref{fig:reductions}e. If we can find an independent set with at least $k$ nodes in $\mathcal{G}$, then techniques similar to those in the proof of Theorem~\ref{thm:neg-m-weak} can be used to construct an independent set with at least $k/3$ nodes in $\mathcal{C}$. By Theorem~\ref{thm:neg-is-cycle}, we must have $k = o(n)$.
\end{proof}
\section{Algorithms for weakly coloured graphs}\label{sec:weak-algo}
In this section we give a local algorithm to find a spanning forest of stars in a weakly $2$-coloured graph. Once the stars are formed, it is simple to find a dominating set (roots of the stars) and a matching (one edge for each star).
To find a small dominating set, we would prefer large (high-degree) stars, and to find a large matching, we would prefer small (low-degree) stars. Nevertheless, the same approach -- find \emph{any} set of stars -- yields the \emph{same} approximation factor ${(\Delta+1)}/2$ for both problems. Moreover, this is the best possible (Theorems \ref{thm:neg-ds-strong} and \ref{thm:neg-m-weak}).
To build the stars, we can use an algorithm that is similar to the \emph{Balanced\_DOM} subroutine in Kutten and Peleg~\cite{kutten98fast}. We first construct a forest $F$ of rooted trees (Figure~\ref{fig:weak-algo}); the directed edges in $F$ point towards the trees' roots. The construction is simple:
\begin{enumerate}
\item Each black node $b$ chooses a white neighbour $w$; add the edge $(b,w)$ to~$F$ (Figure~\ref{fig:weak-algo}b).
\item Each white node $w$ which does not have any children in $F$ chooses a black neighbour $b$; add the edge $(w,b)$ to~$F$ (Figure~\ref{fig:weak-algo}c).
\end{enumerate}
\begin{figure}
\centering
\input{fig2.pdf_t}
\caption{Finding stars in a weakly $2$-coloured graph.}\label{fig:weak-algo}
\end{figure}
At this point, every node belongs to a tree; the depth of each tree is $1$ or $2$. Next we make local modifications within each tree, depending on its structure. Let $r$ be the root of the tree.
\begin{enumerate}
\item If all leaves are at depth $1$, do nothing.
\item If there are leaf nodes both at depth $1$ and at depth $2$, remove all edges $(c,r)$ where $c$ is a non-leaf child.
\item Otherwise, choose arbitrarily a child $x$ of the root. Remove all edges $(c,r)$ where $c$ is a child of the root, and $c \neq x$. Reverse the edge $(x,r)$.
\end{enumerate}
Now $F$ consists of stars, i.e., rooted trees of depth $1$ (Figure~\ref{fig:weak-algo}d). Each node is either a root node with at least one child, or a leaf node. The algorithm can be implemented by using only a port numbering; unique node identifiers are not needed. The port numbers are used both for representing the forest $F$ (e.g., a child does not know the identity of the parent node, but it knows the port number of the edge that leads to the parent node) and for breaking ties (e.g., when a black node has to choose one of its white neighbours).
Next we present the applications of the stars.
\begin{theorem}\label{thm:pos-ds-weak}
For any $\Delta \ge 1$, there is a local algorithm with approximation factor ${(\Delta+1)/2}$ for dominating set in weakly $2$-coloured graphs.
\end{theorem}
\begin{proof}
Let $D$ be the set of the roots of the stars. The set $D$ is a dominating set with at most $\mysize{V}/2$ nodes. Let $D^{*}$ be a minimum dominating set. Since a node cannot dominate more than $\Delta$ neighbours, $\mysize{D^{*}} \ge \mysize{V}/{(\Delta+1)}$. Therefore $D$ is a ${(\Delta+1)/2}$-approximation of a minimum dominating set.
\end{proof}
\begin{theorem}\label{thm:pos-m-weak}
For any $\Delta \ge 1$, there is a local algorithm with approximation factor ${(\Delta+1)/2}$ for maximum matching in weakly $2$-coloured graphs.
\end{theorem}
\begin{proof}
Let $M$ be the set of edges with one edge chosen arbitrarily from each star. The set $M$ is a matching. Each star contains at most $\Delta+1$ nodes; hence $\mysize{M} \ge \mysize{V}/{(\Delta+1)}$. For an optimal matching $M^{*}$, we have $\mysize{M^{*}} \le \mysize{V}/2$. Hence $M$ is a ${(\Delta+1)/2}$-approximation of a maximum matching.
\end{proof}
\section{Approximating dominating set if \texorpdfstring{$\Delta$}{Delta} is odd}
Now we are ready to present an application of the Naor--Stockmeyer algorithm for weak $2$-colouring~\cite{naor95what} and the techniques that we developed in Section~\ref{sec:weak-algo}. In this section we assume that the graph is not only port-numbered, but there is also an orientation: for each edge $\{u,v\} \in E$, exactly one direction $(u,v)$ or $(v,u)$ has been chosen.
The orientation can be used to break the symmetry in some cases. Specifically, if each node of $\mathcal{G}$ has an odd degree, then we can use Naor and Stockmeyer's algorithm to find a weak $2$-colouring; the algorithm does not require unique identifiers~\cite{mayer95local}. Theorem~\ref{thm:pos-ds-weak} then provides a factor ${(\Delta+1)/2}$ approximation for dominating set.
However, in this section we study the case where the degree bound $\Delta$ is odd, but nothing else is known about the degrees of the graph; that is, the case for which we have the lower bound $\Delta-\varepsilon$ from Theorem~\ref{thm:neg-ds-general}. A combination of weak colouring and Theorem~\ref{thm:pos-ds-weak} provides a matching upper bound.
\begin{theorem}\label{thm:pos-ds-odd}
For any odd $\Delta \ge 1$, there is a local algorithm with approximation factor $\Delta$ for dominating set in graphs with maximum degree~$\Delta$, assuming that there is a port numbering and an orientation.
\end{theorem}
\begin{proof}
Partition $V$ into $V = A \cup B \cup C$ such that $A$ consists of the odd-degree nodes, $B$ consists of the even-degree nodes adjacent to at least one node in $A$, and $C$ is the rest; in particular, the degree of each node in $B$ or $C$ is at most $\Delta-1$.
Consider the subgraph $\mathcal{H}$ induced by $A\cup B$. In the subgraph $\mathcal{H}$, the degree of each node in $A$ is odd, but some of the nodes in $B$ may have an even degree. Construct a new graph $\mathcal{H}_2$ by adding a new dummy node of degree $1$ as a neighbour of each even-degree node in the subgraph $\mathcal{H}$. Now every node in $\mathcal{H}_2$ has an odd degree, and we can use the Naor--Stockmeyer algorithm~\cite{mayer95local} to weakly $2$-colour it.
At this point, each node in $A$ is adjacent to a node of the opposite colour, but this does not necessarily hold for the nodes in $B$. However, we can easily find valid colours for each node in $B$ in parallel: if $b \in B$ and each $a \in A$ adjacent to $b$ has the same colour as $b$, then we reverse the colour of $b$. Now each node in $B$ has a neighbour with the opposite colour in $A$; furthermore, no node in $A$ lost a neighbour of the opposite colour.
Thus $\mathcal{H}$ is weakly $2$-coloured, and we can apply the algorithm of Theorem~\ref{thm:pos-ds-weak} to find a dominating set $D_\mathcal{H}$ with $\mysize{D_\mathcal{H}} \le \mysize{A \cup B}/2$ in the subgraph $\mathcal{H}$. The set $D = D_\mathcal{H} \cup C$ is now a dominating set of the original graph $\mathcal{G}$.
Let $D^*$ be a minimum dominating set of $V$. Let $D_1^* = D^* \cap A$ and $D_2^* = D^* \cap (B \cup C)$. Since a node with a degree $d$ can dominate at most $d+1$ nodes, and the nodes in $D_1^*$ are not adjacent to the nodes in $C$, the set $D^*$ must satisfy
\begin{align*}
(\Delta+1)\mysize{D_1^*}+\Delta\mysize{D_2^*} &\ge \mysize{A}+\mysize{B}+\mysize{C}, \\
\Delta \mysize{D_2^*} &\ge \mysize{C},
\intertext{which implies}
\mysize{D^*} = \mysize{D_1^*}+\mysize{D_2^*} &\ge \frac{\mysize{A}+\mysize{B}}{\Delta+1}+\frac{\mysize{C}}{\Delta}.
\end{align*}
Since $\mysize{D} \le (\mysize{A}+\mysize{B})/2+\mysize{C}$, we have $\mysize{D}/\mysize{D^*} \le \Delta$.
\end{proof}
\begin{remark}
It is necessary to assume that the graph is oriented in Theorem~\ref{thm:pos-ds-odd}. If the nodes are anonymous and there is a port numbering but no orientation, a deterministic distributed algorithm cannot have a better approximation factor than $\Delta+1$. To see this, consider the complete graph $K_{\Delta+1}$ on $\Delta+1$ nodes. Find an edge colouring of $K_{\Delta+1}$ with $\Delta$ colours -- this is possible, since we assumed that $\Delta$ is odd. Use the edge colouring to assign the port numbers: an edge with colour $k$ has the port number $k$ in both ends. Now from the perspective of distributed algorithms, the nodes are indistinguishable. Any deterministic algorithm has to produce the same output for each node; in particular, it has to output a dominating set with $\Delta+1$ nodes, while $1$ node would suffice.
\end{remark}
\section{Matching in two-coloured graphs}\label{sec:mm-scheme}
In Sections~\ref{sec:lower} and~\ref{sec:weak-algo} we proved that in weakly $2$-coloured graphs the maximum matching problem can be approximated to within a factor of $(\Delta + 1) / 2$, but not better. In this section we show that in $2$-coloured graphs the problem has a local approximation scheme.
Given a matching $M$, an augmenting path (w.r.t.\ $M$) is a path that starts and ends at an unmatched node and whose every other edge belongs to $M$. An augmenting tree is a tree whose every root--leaf path is an augmenting path. In Figures~\ref{fig:mm-scheme}a--c, a matching, two augmenting trees (rooted at black nodes), and two augmenting paths are shown.
\begin{figure}
\centering
\input{fig3.pdf_t}
\caption{Finding length-$3$ augmenting paths in $2$-coloured graphs with a local algorithm. (a)~The graph~$\mathcal{G}$. A matching~$M$ is highlighted with double lines. The matching is maximal, i.e., there is no length-$1$ augmenting path. However, there are several length-$3$ augmenting paths. (b)~Augmenting trees. The set of root nodes is a \emph{subset} of the black endpoints of the length-$3$ augmenting paths, while the set of leaf nodes is \emph{equal} to the set of white endpoints of the length-$3$ augmenting paths. (c)~Augmenting paths, one per tree. (d)~The new matching.}\label{fig:mm-scheme}
\end{figure}
The symmetric difference of $M$ and an augmenting path is a new matching whose size is larger than the size of $M$ by~1. If every augmenting path is longer than $2k - 1$, then the size of $M$ is at least $k/{(k+1)}$ times the size of the maximum matching (folklore). Hence we have the following $(1 + 1/k)$-approximation algorithm for maximum matching: Starting from an empty matching, for each $i = 1, 2, \dotsc, k$, find repeatedly augmenting paths of length $2i - 1$ and augment along the paths, until no such path exists. The iteration $i=1$ is equal to finding a maximal matching, which could be done locally by Ha\'{n}\'{c}kowiak et al.\mbox{}{}'s~\cite{hanckowiak98distributed} algorithm. Below we give a local algorithm that implements the iteration $i$ for a general $i \ge 1$. The algorithm uses techniques presented by Balas et al.\mbox{}{}~\cite{balas91parallel} in the context of parallel algorithms.
Our algorithm repeatedly invokes a subroutine that removes \emph{some} augmenting paths of length $h=2i - 1$, assuming that there is no shorter augmenting path. The subroutine consists of three phases (refer to Figure~\ref{fig:mm-scheme}).
\begin{enumerate}
\item In the \emph{flooding phase}, we construct a forest $F$ of disjoint augmenting trees, rooted at black nodes (Figure~\ref{fig:mm-scheme}b). Each root--leaf path has length $h$. Furthermore, if there is a length-$h$ augmenting path in the original graph between a black node $b$ and a white node $w$, then $w$ is a leaf node in a tree of $F$. (However, $b$ may or may not be a root node in a tree of $F$.)
\item In the \emph{proposal phase}, we choose one augmenting path in each tree (Figure~\ref{fig:mm-scheme}c).
\item Finally, in the \emph{augmenting phase}, we augment along the paths in parallel to find a new matching (Figure~\ref{fig:mm-scheme}d).
\end{enumerate}
To implement the flooding phase, each unmatched black node considers itself as a potential root of an augmenting tree. Every root node sends a message to each of its neighbours. When a white node receives messages, it chooses one of the senders as its parent, and forwards the message to its neighbour along an edge in the matching. When a black node receives a message, it chooses the sender as its parent, and forwards the message to its neighbours along each edge that is not in the matching. Messages are propagated for $h$ hops; messages that reach a matched white node or a dead end are simply discarded.
Since there is no augmenting path shorter than $h$, every unmatched white node that receives a message is an endpoint of a length-$h$ augmenting path. Conversely, all white endpoints of length-$h$ augmenting paths are reached by the messages. These unmatched white nodes become the leaves of the forest $F$. The edges of $F$ are defined by the links that point towards the parent nodes.
We now show that the trees of the forest $F$ are disjoint. To reach a contradiction, assume that $T_1$ and $T_2$ are two trees in $F$ and they share a node $v$. Let $b_j$ be the root node of the tree $T_j$; by assumption, $b_1 \ne b_2$. Let $P_j$ be an augmenting path in $T_j$ that begins from $b_j$, passes through $v$, and ends at a leaf node; let $\ell_j$ be the distance between $b_j$ and $v$ along $P_j$. If we had $\ell_1 = \ell_2$, the message initiated by the root $b_1$ would have reached the node $v$ on the same time step as the message initiated by the root $b_2$, and in our algorithm $v$ (or one of its ancestors) would have discarded one of the messages and joined only one of the trees. Hence we must have $\ell_1 \ne \ell_2$; but then it is possible to find an augmenting path (in the union of $P_1$ and $P_2$) that is strictly shorter than $h$, which contradicts our assumption.
Hence a local algorithm can find the forest $F$ with the above-mentioned properties. The other steps of the algorithm are straightforward. In the proposal phase, messages are initiated by the leaf nodes and propagated towards the root nodes; whenever several messages meet, all but one of them are discarded. Eventually, we have chosen exactly one augmenting path in each tree. Finally, in the augmenting phase, we augment along each of these paths in parallel.
To analyse how many invocations of the subroutine are needed, note that a white node can be an endpoint of at most $t_i = \Delta(\Delta-1)^{i-1}$ length-$h$ augmenting paths. Every invocation matches the other endpoint of at least one such path. Furthermore, it can be shown that no new augmenting paths with at most $h$ edges are created. Therefore, after $t_i$ invocations, there is no augmenting path with $h$ edges or fewer.
\begin{theorem}\label{thm:pos-m-strong}
For any $\Delta \ge 1$ and $\varepsilon > 0$, there is a local algorithm with approximation factor ${1+\varepsilon}$ for maximum matching in $2$-coloured graphs. \qed
\end{theorem}
\section*{Acknowledgements}
This work was supported in part by the Academy of Finland, Grants 116547, 118653 (ALGODAN), and 132380, by Helsinki Graduate School in Computer Science and Engineering (Hecse), and by the Foundation of Nokia Corporation.
\def\small{\small}
|
\section{Introduction}
The sublattice of finite index formed by the points of intersection of a lattice and a rotated copy of the same lattice is called a coincidence site lattice or CSL. It was Friedel
in 1911 who first recognized the use of CSLs in describing and classifying grain boundaries in crystals \cite{Fr}. Since then, CSLs have proven to be an indispensable tool in the
study of grain boundaries and interfaces \cite{KW, R,Bo,P}. This prompted various authors to examine the CSLs of several lattices including cubic and hexagonal ones
\cite{GBW,G,G2}.
The discovery of quasicrystals triggered a renewed interest in CSLs. This led to the analysis of CSLs from a more mathematical point of view. Known results
for lattices were again considered and reformulated so that they may be readily extended to aperiodic situations. This was necessary since the first stage in
solving the coincidence problem for quasicrystals involved calculating the coincidence site modules (CSMs) of the underlying translation modules, such as
modules with 5, 8, 10, and 12-fold symmetry (see \cite{PBR,B} and references therein, see also \cite{W,WL,OWL,WL2}). Hence, coincidences of lattices and
modules in dimensions $d\leq 4$ were investigated in \cite{B,Z,Z2,BZ,BGHZ}. Recent results include the decomposition of coincidence isometries of lattices and
modules in Euclidean $n$-space as a product of at most $n$ coincidence reflections \cite{Zo,H} and the relationship between the sets of coincidence and
similarity isometries of lattices and modules \cite{Gl,Gl2}.
The mathematical treatment of the coincidence problem is very often restricted to linear coincidence isometries, that is, rotations and improper rotations,
whereas isometries containing a translational part are ignored -- as we did in the first two paragraphs above. Nevertheless, general (affine) isometries are
important in crystallography. Indeed, the situation where one shifts the two component crystals against each other has been investigated in \cite{GC,F} and
references therein. It was shown that these shifts are needed to minimize the grain boundary energy, thus they are often referred to as ``rigid
relaxations''. However, some authors claim that minimizing the energy may require shifts that destroy all coincidence sites.
Even though the idea of introducing a shift after applying a linear coincidence isometry has already been dealt with in the physical literature, not much can be
found in the mathematical literature where a systematic treatment of the subject is still missing. Thus, we now aim to generalize the notion of a CSL and CSM, respectively, that is,
we investigate the possible intersections of two lattices (modules) that are related by an isometry. To simplify the discussion, we restrict our attention here to the lattice case
though most of the results also work in the module case. In fact, some steps in the general direction have been made in \cite{PBR}. There, the authors have considered
coincidence rotations around certain points which are not lattice (module) points. For instance, they determined the set of coincidence rotations about the center
of a Delauney cell of the square lattice and calculated the corresponding indices.
In this paper we discuss a related and special case: the coincidence problem for shifted lattices. That is, after translating the lattice by some vector and
upon rotation of this shifted lattice (with respect to the origin), we consider its intersection with the shifted lattice. This should be useful in the context of bicrystallography
\cite{GP, PV}. Similar to the approach in \cite{PBR, B}, we start our investigation with the square lattice and provide solutions that, when modified appropriately, also apply to
planar modules.
The purpose of this paper is to shed further light on the geometry of CSLs. It is beyond the scope of this paper to discuss the actual grain boundary energy, which would require
considering the actual Hamiltonians. In particular, the paper is not intended to determine which translations are the most favourable ones in terms of energy.
\section{The coincidence problem for lattices}
Let $\Gamma\subseteq \mathbb{R}^d$ be a $d$-dimensional lattice and $R\in O(d)$. We say that $R$ is a \emph{(linear) coincidence isometry} of $\Gamma$ if $\Gamma(R):=\Gamma\cap R\Gamma$ is a sublattice
of finite index in $\Gamma$ and we call $\Gamma(R)$ a \emph{coincidence site lattice} (CSL) of $\Gamma$. The \emph{coincidence index} of a coincidence isometry $R$, denoted by $\Sigma(R)$, is
given by $\Sigma(R)=[\Gamma:\Gamma(R)]=\frac{\operatorname{vol}(\Gamma(R))}{\operatorname{vol}(\Gamma)}$. Geometrically, $\Sigma(R)$ is equal to the ratio of the volume of a fundamental domain of
$\Gamma$ with the volume of a fundamental domain of $\Gamma(R)$.
We denote the set of (linear) coincidence isometries of $\Gamma$ by $OC(\Gamma)$ and the set of coincidence rotations of $\Gamma$, that is, $OC(\Gamma)\cap SO(d)$, by $SOC(\Gamma)$. The set $OC(\Gamma)$
forms a group having $SOC(\Gamma)$ as a subgroup \cite{B}.
We summarize here the known results for the square lattice $\mathbb{Z}^2$ (see \cite{PBR} for details). The group of coincidence rotations of $\mathbb{Z}^2$ is $SOC(\mathbb{Z}^2)=
SO(2,\mathbb{Q})$, that is, the coincidence rotations of $\mathbb{Z}^2$ are the special orthogonal matrices having rational entries. In determining the structure of this group, the
square lattice is identified with the ring of Gaussian integers $\Gamma=\mathbb{Z}[i]$, where $i=\sqrt{-1}$, embedded in the set of complex numbers $\mathbb{R}[i]=\mathbb{C}$. In
this setting, a coincidence rotation $R$ by an angle of $\theta$ corresponds to multipication by the complex number $e^{i\theta}$, where
\begin{equation}\label{coincrot}
e^{i\theta}=\varepsilon\cdot \prod_{p\equiv 1(4)}{\left(\frac{\omega_p}{\overline{\omega_p}}\right)}^{n_p}
\end{equation}
with $n_p\in\mathbb{Z}$ and only a finite number of $n_p\neq 0$, $\varepsilon$ is a unit in $\mathbb{Z}[i]$, $p$ runs over the rational primes $p\equiv 1\imod{4}$,
and $\omega_p$, and its complex conjugate $\overline{\omega_p}$ are the Gaussian prime factors of $p=\omega_p\cdot\overline{\omega_p}$. If we denote by $z$ the numerator of
$e^{i\theta}$, that is,
\begin{equation}\label{num}
z=\prod_{\stackrel{p\equiv 1(4)}{n_p>0}}{\omega_p}^{n_p}\cdot\prod_{\stackrel{p\equiv 1(4)}{n_p<0}}{{\left(\overline{{\omega_p}}\right)}^{\,-n_p}},
\end{equation}
then the coincidence index of $R$ is the number theoretic norm of $z$, that is, $\Sigma(R)=N(z)=z\cdot\overline{z}$, and the CSL obtained from $R$, $\Gamma(R)$, is the principal ideal
$(z):=z\mathbb{Z}[i]$. Consequently, the group of coincidence rotations of the square lattice is given by $SOC(\mathbb{Z}^2)=SOC(\Gamma)\cong C_4\times\mathbb{Z}^{(\aleph_0)}$, where
$C_4$ is the cyclic group of order 4 with generator $i$, and $\mathbb{Z}^{(\aleph_0)}$ is the direct sum of countably many infinite cyclic
groups each of which is generated by $\omega_p/\overline{\omega_p}$ with $p\equiv 1\imod{4}$.
Every coincidence isometry $T\in OC(\Gamma)\setminus SOC(\Gamma)$ can be written as $T=R\cdot T_r$, where $R\in SOC(\Gamma)$ and $T_r$ is the reflection along the real axis (complex conjugation).
Here, $\Sigma(T)=\Sigma(R)$ and $\Gamma(T)=\Gamma(R)$. Finally, $OC(\mathbb{Z}^2)=OC(\Gamma)\cong SOC(\Gamma)\rtimes C_2$ (semi-direct product), where $C_2$ is the cyclic group of order 2
generated by $T_r$.
The possible coincidence indices and the number of CSLs for a given index $m$ may be described by means of a generating function. Let $\hat{f}(m)$ be the number of coincidence
rotations of $\Gamma$ and $f(m)$ be the number of CSLs of $\Gamma$, for a given index $m$. Then $\hat{f}(m)=4f(m)$, where the factor 4 stems from the fact that there are four symmetry
rotations. The function $f(m)$ is multiplicative (that is, $f(1)=1$ and $f(mn)=f(m)f(n)$ when $m$, $n$ are relatively prime), and $f(p^r)=2$ for primes $p\equiv 1\imod{4}$ whereas
$f(p^r)=0$ for primes $p\equiv 2,3\imod{4}$, where $r\in\mathbb{N}$. We write the generating function for $f(m)$ as a Dirichlet series $\Phi(s)$ given by
\begin{alignat*}{2}
\Phi(s)&=\sum_{m=1}^{\infty}{\frac{f(m)}{m^s}}=\prod_{p\equiv 1(4)}{\frac{1+p^{-s}}{1-p^{-s}}}\\
&=1+\tfrac{2}{5^s}+\tfrac{2}{13^s}+\tfrac{2}{17^s}+\tfrac{2}{25^s}+\tfrac{2}{29^s}+\tfrac{2}{37^s}+
\tfrac{2}{41^s}+\tfrac{2}{53^s}+\tfrac{2}{61^s}+\tfrac{4}{65^s}+\tfrac{2}{73^s}+\ldots.
\end{alignat*}
\section{Coincidences of shifted lattices}
We now turn our attention to lattices $\Gamma$ in $\mathbb{R}^d$ that are shifted by some vector $x\in\mathbb{R}^d$ and we look at intersections of the form $(x+\Gamma)\cap R(x+\Gamma)$, where
$R\in O(d)$. We remark that we actually only need to consider values of $x$ in a fundamental domain of $\Gamma$.
An $R\in O(d)$ is said to be a \emph{(linear) coincidence isometry of the shifted lattice} $x+\Gamma$ if $(x+\Gamma)\cap R(x+\Gamma)$ is a sublattice of $x+\Gamma$ of finite index and we also call
$(x+\Gamma)\cap R(x+\Gamma)$ a CSL of the shifted lattice $x+\Gamma$. We denote the set of all coincidence isometries of $x+\Gamma$ by $OC(x+\Gamma)$. The following theorem characterizes the set
$OC(x+\Gamma)$ and relates the CSLs of $x+\Gamma$ with the CSLs of $\Gamma$ \cite{LZ}.
\begin{theorem}\label{cor3}
Let $\Gamma$ be a lattice in $\mathbb{R}^d$ and $x\in\mathbb{R}^d$.
\begin{enumerate}
\item $OC(x+\Gamma)=\set{R\in OC(\Gamma):Rx-x\in\Gamma +R\Gamma}$
\item If $R\in OC(x+\Gamma)$ with $Rx-x=t+Rs$ for some $t,s\in\Gamma$, then \[(x+\Gamma)\cap R(x+\Gamma)=x+(t+\Gamma(R)).\]
\end{enumerate}
\end{theorem}
Theorem \ref{cor3} tells us that a CSL of $x+\Gamma$ obtained from $R\in OC(x+\Gamma)$ is just a translate of the CSL $\Gamma(R)$ in $\Gamma$. Consequently, $[x+\Gamma: (x+\Gamma)\cap R(x+\Gamma)]=\Sigma(R)$
which means that shifting the lattice does not give rise to new values of coincidence indices. In addition, we see that $OC(x+\Gamma)$ is a subset of $OC(\Gamma)$. The set
$OC(x+\Gamma)$ is non-empty because the identity $1\in OC(x+\Gamma)$. Also, $OC(x+\Gamma)$ is closed under inverses, that is, $R^{-1}\in OC(x+\Gamma)$ whenever $R\in OC(x+\Gamma)$. However, given
$R_1$, $R_2\in OC(x+\Gamma)$, $R_2R_1$ is not necessarily in $OC(x+\Gamma)$. In fact, $OC(x+\Gamma)$ is not a group in general \cite{LZ}.
\section{The coincidence problem for the shifted square lattice}
From this point onwards, we take $\Gamma=\mathbb{Z}[i]$, the square lattice viewed as the ring of Gaussian integers, and $x\in \mathbb{C}$. From \eqref{coincrot} and \eqref{num}, we
see that we can associate each $R\in SOC(\Gamma)$ to $(z,\varepsilon)$, and we will write this as $R(z,\varepsilon)$. That is, $R(z,\varepsilon)\in SOC(\Gamma)$ stands for multiplication by the complex number
$\varepsilon\frac{z}{\overline{z}}$. We may assume that $\frac{z}{\overline{z}}$ is reduced, that is, $z$ and $\overline{z}$ have no factors in common. In addition, we shall simply set
$z=1$ whenever $R(z,\varepsilon)\in P(\Gamma)$, where $P(\Gamma)$ denotes the point group of $\Gamma$.
Let $SOC(x+\Gamma):=OC(x+\Gamma)\cap SO(d)$. We start with the following lemma.
\begin{lemma}\label{lem3}
Let $\Gamma=\mathbb{Z}[i]$, $x\in\mathbb{C}$, $R=R(z,\varepsilon)\in SOC(\Gamma)$, and $T=RT_r$.
\begin{enumerate}
\item $R\in SOC(x+\Gamma)$ if and only if $(\varepsilon z-\overline{z})x\in\mathbb{Z}[i]$
\item $T\in OC(x+\Gamma)$ if and only if $\varepsilon z\overline{x}-\overline{z}x\in\mathbb{Z}[i]$
\end{enumerate}
\end{lemma}
\begin{proof}
Recall that $\Gamma$ is a principal ideal domain. Since $\varepsilon$ is a unit in $\Gamma$ and $z$, $\overline{z}$ are relatively prime,
\[\Gamma+R\Gamma=\Gamma+\varepsilon\frac{z}{\overline{z}}\Gamma=\frac{1}{\overline{z}}(\overline{z}\Gamma+z\Gamma)=\frac{1}{\overline{z}}\gcd(z,\overline{z})\Gamma=\frac{1}{\overline{z}}\Gamma.\]
By Theorem \ref{cor3}, $R\in SOC(x+\Gamma)\Leftrightarrow Rx-x\in\Gamma+R\Gamma\Leftrightarrow \varepsilon\frac{z}{\overline{z}}x-x\in\frac{1}{\overline{z}}\Gamma\Leftrightarrow (\varepsilon
z-\overline{z})x\in\mathbb{Z}[i].$ Similarly, $\Gamma+T\Gamma=\frac{1}{\overline{z}}\Gamma$. Applying again Theorem \ref{cor3}, we obtain the second statement.
\end{proof}
We now obtain the following results about $SOC(x+\Gamma)$ and $OC(x+\Gamma)$.
\begin{theorem}\label{SOCgroup}
If $\Gamma=\mathbb{Z}[i]$ and $x\in\mathbb{C}$ then $SOC(x+\Gamma)$ is a subgroup of $SOC(\Gamma)$.
\end{theorem}
\begin{proof}
We have already mentioned that $1\in SOC(x+\Gamma)$ and $SOC(x+\Gamma)$ is closed under inverses. Let $R_j(z_j,\varepsilon_j)\in SOC(x+\Gamma)$ for $j=1,2$ and $g=\gcd(z_1,\overline{z_2})$. By
Lemma \ref{lem3}, $(\varepsilon_jz_j-\overline{z_j})x\in\mathbb{Z}[i]$ for $j=1,2$. Write $z_1=h_1g$, $\overline{z_2}=\overline{h_2}g$, and hence, $h_1$ and $\overline{h_2}$ are
relatively prime. This means that $R_1R_2$ corresponds to $(h_1h_2,\varepsilon_1\varepsilon_2)$ so that $R_1R_2\in SOC(x+\Gamma)$ if
$\left(\varepsilon_1\varepsilon_2h_1h_2-\overline{h_1}\overline{h_2}\,\right)x\in\mathbb{Z}[i]$ from Lemma \ref{lem3}. Now,
\begin{alignat*}{2}
(\varepsilon_1\varepsilon_2h_1h_2-\overline{h_1}\overline{h_2}\,)x&=\frac{1}{g}(\varepsilon_1\varepsilon_2z_1h_2-\overline{h_1z_2}\,)x\\
&=\frac{1}{g}[(\varepsilon_1\varepsilon_2z_1h_2-\varepsilon_2h_2\overline{z_1})+(\varepsilon_2\overline{h_1}z_2-\overline{h_1z_2}\,)]x\\
&=\frac{1}{g}[\varepsilon_2h_2\underbrace{(\varepsilon_1z_1-\overline{z_1})x}_{\in\;\mathbb{Z}[i]}+\overline{h_1}\underbrace{(\varepsilon_2z_2-\overline{z_2})x}_
{\in\;\mathbb{Z}[i]}]\in \frac{1}{g}\Gamma.
\end{alignat*}
Similarly, we also obtain that $(\varepsilon_1\varepsilon_2h_1h_2-\overline{h_1}\overline{h_2}\,)x\in \frac{1}{\overline{g}}\,\Gamma$. Hence,
\[(\varepsilon_1\varepsilon_2h_1h_2-\overline{h_1}\overline{h_2}\,)x\in\frac{1}{g}\Gamma\cap\frac{1}{\overline{g}}\Gamma=\frac{1}{g\overline{g}}(g\Gamma\cap\overline{g}\Gamma)
=\frac{1}{g\overline{g}}\lcm{(g,\overline{g})}\Gamma=\mathbb{Z}[i]\]
since $g$, $\overline{g}$ are relatively prime.
\end{proof}
For $OC(x+\Gamma)$, the situation is more complicated. One can show the following results (see \cite{LZ}).
\begin{theorem}\label{prop5}
Let $\Gamma=\mathbb{Z}[i]$ and $x\in\mathbb{C}$.
\begin{enumerate}
\item The set $OC(x+\Gamma)$ is a subgroup of $OC(\Gamma)$ if and only if for any $T_1$, $T_2\in$ \mbox{$OC(x+\Gamma)\setminus SOC(x+\Gamma)$}, $T_1T_2\in SOC(x+\Gamma)$.
\item If $OC(x+\Gamma)$ contains a reflection $T\in P(\Gamma)$ then $OC(x+\Gamma)$ is a subgroup of $OC(\Gamma)$. Also, $OC(x+\Gamma)=SOC(x+\Gamma)\rtimes\langle T\rangle$, where $\langle
T\rangle=\set{1,T}\cong C_2$ is the group generated by $T$.
\item Suppose $OC(x+\Gamma)$ does not contain a reflection $T\in P(\Gamma)$. If $RT_r\in OC(x+\Gamma)$ where $R(z,\varepsilon_1)\in SOC(\Gamma)$ then for any unit $\varepsilon_2$,
$R_2=R_2(z,\varepsilon_2)\notin SOC(x+\Gamma)$.
\end{enumerate}
\end{theorem}
When computing for $OC(x+\Gamma)$, we see from Theorem \ref{prop5} that it is convenient to determine whether there is a reflection $T\in P(\Gamma)$ that is in $OC(x+\Gamma)$. If such a
reflection $T$ exists, then $OC(x+\Gamma)$ is a group and it is the semi-direct product of $SOC(x+\Gamma)$ and $\langle T\rangle$. Otherwise, we need to check if $RT_r\in OC(x+\Gamma)$
only for those reflections $RT_r$ for which $R(z,\varepsilon)\in SOC(\Gamma)$ and $R'=R'(z,\varepsilon')\notin SOC(x+\Gamma)$ for all unit $\varepsilon'$ holds.
\section{Specific examples}
For the rest of the discussion, we shall assume that $R(z,\varepsilon)\in SOC(\Gamma)$. The following theorem solves completely the case when $x$ has an irrational component
\cite{LZ}.
\begin{theorem} Let $x=a+bi\in\mathbb{C}$. If $a$ or $b$ is irrational then $OC(x+\Gamma)$ is a group of at most two elements. In particular, if
\begin{enumerate}
\item $a$ is irrational and $b$ is rational then
$OC(x+\Gamma)=\left\{\begin{aligned}
\langle T_r\rangle &\;\text{if }2b\in\mathbb{Z}\\
\set{1} &\;\text{otherwise}.
\end{aligned}\right.\;$
\item $a$ is rational and $b$ is irrational then
$OC(x+\Gamma)=\left\{\begin{aligned}
\langle T\rangle &\;\text{if }2a\in\mathbb{Z}\\
\set{1} &\;\text{otherwise},
\end{aligned}\right.$\\ where $T$ is the reflection along the imaginary axis.
\item both $a$ and $b$ are irrational, and
\begin{enumerate}
\item $a$, $b$ are rationally independent then $OC(x+\Gamma)=\set{1}$.
\item $a=\frac{p_1}{q_1}+\frac{p_2}{q_2}b$ where $p_j$, $q_j\in\mathbb{Z}$, $p_j$ and $q_j$ are
relatively prime (for $j=1,2$) with
\begin{enumerate}
\item $p_2q_2$ even, then $OC(x+\Gamma)=\left\{\begin{aligned}
\langle R&T_r\rangle &&\text{if }q_1|2q_2\\
\{&1\} &&\text{otherwise},
\end{aligned}\right.$\\where $R=R(p_2+q_2i,1)\in SOC(\Gamma)$.
\item $p_2q_2$ odd, then $OC(x+\Gamma)=\left\{\begin{aligned}
\langle R&T_r\rangle &&\text{if }q_1|q_2\\
\{&1\} &&\text{otherwise},
\end{aligned}\right.$\\where $R=R(\frac{p_2+q_2}{2}-\frac{p_2-q_2}{2}i, i)\in SOC(\Gamma)$.
\end{enumerate}
\end{enumerate}
\end{enumerate}
\end{theorem}
\begin{ex}
\begin{enumerate}
\item[]
\item Suppose $x=\frac{1}{\sqrt{2}}+\frac{1}{\sqrt{3}}i$. We immediately see that we cannot write $\frac{1}{\sqrt{2}}=c+d\frac{1}{\sqrt{3}}$
where $c,d\in\mathbb{Q}$ since $\sqrt{2}\notin\mathbb{Q}\left(\sqrt{3}\,\right)$. Hence, $OC(x+\Gamma)=\set{1}$.
\item Let $x=\sqrt{2}-\frac{\sqrt{2}}{2}i$. We have $\sqrt{2}=\frac{0}{1}+\left(\frac{-2}{1}\right)\left(-\frac{\sqrt{2}}{2}\right)$, and since $1|(2\cdot 1)$,
$OC(x+\Gamma)=\langle RT_r\rangle$ where $R=R(-2+i,1)\in SOC(\Gamma)$.
\end{enumerate}
\end{ex}
It only remains to consider the case when both $a$ and $b$ are rational. We now consider $x=a+bi\in\mathbb{Q}(i)$ and write $x=\frac{p}{q}$ where $p$, $q\in\mathbb{Z}[i]$,
and $p$, $q$ are relatively prime (in $\mathbb{Z}[i]$). The following lemma tells us that $SOC(x+\Gamma)$ depends only on the denominator $q$ of $x$.
\begin{lemma}\label{prop8}
Let $\Gamma=\mathbb{Z}[i]$, $x=\frac{p}{q}\in\mathbb{Q}(i)$ where $p$, $q\in\mathbb{Z}[i]$, with $p$, $q$ relatively prime, and $R(z,\varepsilon)\in SOC(\Gamma)$. Then
$R\in SOC(x+\Gamma)$ if and only if $q$ divides $\varepsilon z-\overline{z}$. Furthermore, $SOC(x+\Gamma)=SOC(\frac{1}{q}+\Gamma)$.
\end{lemma}
\begin{proof}
We know from Lemma \ref{lem3} that if $R\in SOC(x+\Gamma)$ then $(\varepsilon z-\overline{z})x=\frac{(\varepsilon z-\overline{z})p}{q}\in\mathbb{Z}[i]$. Since $p$ and $q$ are relatively
prime, $q|(\varepsilon z-\overline{z})$. The second statement follows from the first.
\end{proof}
\begin{lemma}\label{lemfunddom}
Suppose $\Gamma=\mathbb{Z}[i]$ and $x\in\mathbb{C}$. If $x'=Qx$ for some $Q\in P(\Gamma)$ then \[OC(x'+\Gamma)=Q[OC(x+\Gamma)]Q^{-1}.\]
\end{lemma}
\begin{proof}
Since $SOC(\Gamma)$ is a normal subgroup of $OC(\Gamma)$, $R\in SOC(\Gamma)$ if and only if $QRQ^{-1}\in SOC(\Gamma)$. Thus, if $R\in SOC(\Gamma)$ then it follows from Theorem \ref{cor3} that
\begin{alignat*}{2}
R\in OC(x'+\Gamma)&\Leftrightarrow Rx'-x'\in\Gamma+R\Gamma\\
&\Leftrightarrow Q(Q^{-1}RQx-x)\in Q(\Gamma+Q^{-1}RQ\Gamma)\\
&\Leftrightarrow Q^{-1}RQx-x\in \Gamma+Q^{-1}RQ\Gamma\\
&\Leftrightarrow Q^{-1}RQ\in OC(x+\Gamma)\\
&\Leftrightarrow R\in Q[OC(x+\Gamma)]Q^{-1}.\qedhere
\end{alignat*}
\end{proof}
Recall that we only need to consider values of $x=a+bi$ in a fundamental domain of $\Gamma$. A fundamental domain of $\Gamma$ is $\set{a+bi\in\mathbb{C}:-\frac{1}{2}\leq
a,b<\frac{1}{2}}$ (see Figure \ref{funddom}). Observe that every point $x'$ in the chosen fundamental domain can be written as $x'=Qx$ where $Q\in P(\Gamma)$ and
$x\in\set{a+bi\in\mathbb{C}: 0\leq b\leq a\leq\frac{1}{2}}$ (a fundamental domain of the symmetry group of $\Gamma$ which is a crystallographic group of type $p4m$). Hence, it follows
from Lemma \ref{lemfunddom} that we only need to compute $OC(x+\Gamma)$ for values of $x=a+bi$, where $0\leq b\leq a\leq\frac{1}{2}$ (see Figure \ref{funddom}).
\begin{figure}[ht]
\setlength{\unitlength}{1.5in}
\begin{center}
\begin{picture}(1.3,1.3)(-0.65,-0.65)
\linethickness{0.15pt}
\put(-0.65,0){\vector(1,0){1.3}}
\put(0,-0.65){\vector(0,1){1.3}}
\put(0.005,-0.60){-$\frac{1}{2}$}
\put(-0.6,-0.1){-$\frac{1}{2}$}
\put(0.02,0.55){$\frac{1}{2}$}
\put(0.52,-0.1){$\frac{1}{2}$}
\put(-0.5,-0.5){\line(1,0){1}}
\put(-0.5,-0.5){\line(0,1){1}}
\multiput(-0.5,0.5)(0.105,0){10}{\line(1,0){0.055}}
\multiput(0.5,0.5)(0,-0.105){10}{\line(0,-1){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\end{picture}
\end{center}
\caption{A fundamental domain of $\Gamma$, or unit cell, and a fundamental domain of the symmetry group of $\Gamma$ (black triangle)}\label{funddom}
\end{figure}
\begin{lemma}\label{lemgroup}
If $\Gamma=\mathbb{Z}[i]$ and $x=a+bi\in\mathbb{C}$, then $OC(x+\Gamma)$ is a subgroup of $OC(\Gamma)$ if one of the following conditions is satisfied: $a\in\frac{1}{2}\mathbb{Z}$,
$b\in\frac{1}{2}\mathbb{Z}$ or $a\pm b\in\mathbb{Z}$. Furthermore, $OC(x+\Gamma)=SOC(x+\Gamma)\rtimes \langle RT_r\rangle$ where $R=R(1,\varepsilon)\in SOC(\Gamma)$, and
\[\varepsilon=\left\{\begin{aligned}
1 & \;\text{if }b\in\tfrac{1}{2}\mathbb{Z}\\
-1 & \;\text{if }a\in\tfrac{1}{2}\mathbb{Z}\\
i & \;\text{if }a-b\in\mathbb{Z}\\
-i & \;\text{if }a+b\in\mathbb{Z}.
\end{aligned}\right.\]
\end{lemma}
\begin{proof}
Consider the reflection $RT_r\in P(\Gamma)$. By Lemma \ref{lem3}, $RT_r\in OC(x+\Gamma)$ if and only if $\varepsilon\overline{x}-x\in\mathbb{Z}[i]$. The result follows by applying Theorem
\ref{prop5}.
\end{proof}
In particular, for values of $a$ and $b$ for which $0\leq b\leq a\leq \frac{1}{2}$, $OC(x+\Gamma)$ is a subgroup of $OC(\Gamma)$ when $a=\frac{1}{2}$, $b=0$, or $a=b$ (the boundaries of the
triangle in Figure \ref{funddom}).
Before looking at some examples, we note that given $R(z,\varepsilon)\in SOC(\Gamma)$, we have
\begin{equation}\label{ezmcz}
\varepsilon z-\overline{z}=\left\{\begin{aligned}
2\,&\Im(z) &&\;\text{if }\varepsilon=1\\
-2\,&\Re(z) &&\;\text{if }\varepsilon=-1\\
-[\Re(z)+&\Im(z)](1-i) &&\;\text{if }\varepsilon=i\\
-i[\Re(z)-&\Im(z)](1-i) &&\;\text{if }\varepsilon=-i\;.
\end{aligned}\right.
\end{equation}
In addition, we see from \eqref{coincrot} and \eqref{num} that $\Re(z)$ and $\Im(z)$ are relatively prime and of different parity (that is, one is odd and the other is even).
We will also exhibit the number of possible coincidence rotations and CSLs obtained with given index $m$ of the shifted lattice $x+\Gamma$ by means of generating functions. We shall
denote by $\hat{f}_x(m)$ the number of coincidence rotations of $x+\Gamma$ of index $m$, and $f_x(m)$ the number of CSLs of $x+\Gamma$ of index $m$.
\begin{ex} $x=\frac{1}{2}+\frac{1}{2}i=\frac{1}{1-i}$
\setlength{\unitlength}{1.25in}
\begin{center}
\begin{picture}(0.65,0.80)(0,-0.15)
\linethickness{0.10pt}
\put(0,0){\vector(1,0){0.65}}
\put(0,0){\vector(0,1){0.65}}
\put(-0.1,0.48){$\frac{1}{2}$}
\put(0.48,-0.15){$\frac{1}{2}$}
\multiput(0,0.5)(0.105,0){5}{\line(1,0){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\put(0.5,0.5){\circle*{0.08}}
\put(0.6,0.5){\makebox(0,0){$x$}}
\end{picture}
\end{center}
The denominator of $x$ is $q=1-i$ and we see from \eqref{ezmcz} that $q|(\varepsilon z-\overline{z})$ for all $z$, $\varepsilon$. Lemmas \ref{prop8} and \ref{lemgroup} implies that
$SOC(x+\Gamma)=SOC(\Gamma)\cong C_4\times{\mathbb{Z}}^{(\aleph_0)}$ and $OC(x+\Gamma)=OC(\Gamma)\cong SOC(x+\Gamma)\rtimes C_2$. Clearly, $OC(x+\Gamma)$ is a subgroup of $OC(\Gamma)$ in this case.
Also, $\hat{f}_x(m)=\hat{f}(m)$ and $f_x(m)=f(m)$.
These results agree with the results obtained in the Appendix of \cite{PBR}
(just shift the center of the Delaunay cell into the origin).
We also note here that $(S)OC(x+\Gamma)=(S)OC(\Gamma)$ if and only if $x=\frac{m}{2}+\frac{n}{2}i$, where $m$, $n$ are odd integers.
\end{ex}
\begin{ex} $x=\frac{1}{2}$
\setlength{\unitlength}{1.25in}
\begin{center}
\begin{picture}(0.65,0.80)(0,-0.15)
\linethickness{0.10pt}
\put(0,0){\vector(1,0){0.65}}
\put(0,0){\vector(0,1){0.65}}
\put(-0.1,0.48){$\frac{1}{2}$}
\put(0.46,-0.18){$\frac{1}{2}$}
\multiput(0,0.5)(0.105,0){5}{\line(1,0){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\put(0.5,0){\circle*{0.08}}
\put(0.57,0.07){\makebox(0,0){$x$}}
\end{picture}
\end{center}
The denominator of $x$ is $q=2$. Since the sum of $\Re(z)$ and $\Im(z)$ is odd, we obtain from \eqref{ezmcz} that for all $z$, $q|(\varepsilon z-\overline{z})$ if and only if
$\varepsilon=\pm 1$. Hence, by Lemma \ref{prop8}, \[SOC(x+\Gamma)=\set{R(z,\varepsilon)\in SOC(\Gamma):\varepsilon=\pm 1}\cong C_2\times{\mathbb{Z}}^{(\aleph_0)}.\] From Lemma \ref{lemgroup},
$OC(x+\Gamma)=SOC(x+\Gamma)\rtimes\langle T_r\rangle$ and is a subgroup of $OC(\Gamma)$ of index 2. In this case, we have $f_x(m)=f(m)$ but $\hat{f}_x(m)=2f_x(m)$.
\end{ex}
\begin{ex} $x_0=\frac{1}{3}$ and $x_1=\frac{1}{3}+\frac{1}{3}i$
\setlength{\unitlength}{1.25in}
\begin{center}
\begin{picture}(0.65,0.80)(0,-0.15)
\linethickness{0.10pt}
\put(0,0){\vector(1,0){0.65}}
\put(0,0){\vector(0,1){0.65}}
\put(-0.1,0.48){$\frac{1}{2}$}
\put(0.46,-0.18){$\frac{1}{2}$}
\multiput(0,0.5)(0.105,0){5}{\line(1,0){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\put(0.33,0){\circle*{0.08}}
\put(0.33,-0.1){\makebox(0,0){$x_0$}}
\put(0.33,0.33){\circle*{0.08}}
\put(0.23,0.33){\makebox(0,0){$x_1$}}
\end{picture}
\end{center}
Both $x_0$ and $x_1$ have denominator $q=3$. It is easy to see that if both rational integers $m$, $n$ are not divisible by $q$, then either $m+n$ or $m-n$ is divisible by
$q$. Hence, by \eqref{ezmcz}, there is a unique $\varepsilon$ so that $q|(\varepsilon z-\overline{z})$ for all $z$. We conclude from Lemma \ref{prop8} that
$SOC(x_j+\Gamma)\cong{\mathbb{Z}}^{(\aleph_0)}$ for $j=0,1$. In addition, by Lemma \ref{lemgroup}, $OC(x_j+\Gamma)=SOC(x_j+\Gamma)\rtimes\langle RT_r\rangle$, where $R=R(1,i^j)\in
P(\Gamma)$, and $OC(x_j+\Gamma)$ is a subgroup of $OC(\Gamma)$ of index 4. Finally, we have $\hat{f}_{x_j}(m)=f_{x_j}(m)=f(m)$.
\end{ex}
\begin{ex} $x_0=\frac{1}{5}$, $\frac{2}{5}$ and $x_1=\frac{1}{5}+\frac{1}{5}i$, $\frac{2}{5}+\frac{2}{5}i$
\setlength{\unitlength}{1.25in}
\begin{center}
\begin{picture}(0.65,0.8)(0,-0.15)
\linethickness{0.10pt}
\put(0,0){\vector(1,0){0.65}}
\put(0,0){\vector(0,1){0.65}}
\put(-0.1,0.48){$\frac{1}{2}$}
\put(0.46,-0.15){$\frac{1}{2}$}
\multiput(0,0.5)(0.105,0){5}{\line(1,0){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\put(0.2,0){\circle*{0.08}}
\put(0.4,0){\circle*{0.08}}
\put(0.2,0.2){\circle{0.08}}
\put(0.4,0.4){\circle{0.08}}
\put(0.75,0.4){\circle*{0.08}}
\put(0.75,0.2){\circle{0.08}}
\put(0.88,0.4){\makebox(0,0){$x_0$}}
\put(0.88,0.2){\makebox(0,0){$x_1$}}
\end{picture}
\end{center}
We have the denominator $q=5$ for $x_j$, $j=0$, $1$. By considering each possible combination of $\Re(z)$ and $\Im(z)$ modulo $q$, it can be verified that for all $z$ with
$5\nmid N(z)$, there is a unique $\varepsilon$ such that $R(z,\varepsilon)\in SOC(\frac{1}{5}+\Gamma)$. This means that $SOC(x_j+\Gamma)\cong{\mathbb{Z}}^{(\aleph_0)}$ for $j=0,1$ by Lemma
\ref{prop8}. Also, it follows from Lemma \ref{lemgroup} that $OC(x_j+\Gamma)$ is a subgroup of $OC(\Gamma)$ and $OC(x_j+\Gamma)=SOC(x_j+\Gamma)\rtimes\langle RT_r\rangle$, where
$R=R(1,i^j)\in P(\Gamma)$. Furthermore, $\hat{f}_{x_j}(m)=f_{x_j}(m)$ where the Dirichlet series generating function for $f_{x_j}(m)$ is given by
\begin{alignat*}{2}
\Phi_{x_j}(s)&=\sum_{m=1}^{\infty}{\frac{f_{x_j}(m)}{m^s}}=\prod_{\stackrel{p\equiv 1(4)}{p\neq 5}}\frac{1+p^{-s}}{1-p^{-s}}\\
&=1+\tfrac{2}{13^s}+\tfrac{2}{17^s}+\tfrac{2}{29^s}+\tfrac{2}{37^s}+\tfrac{2}{41^s}+\tfrac{2}{53^s}+\tfrac{2}{61^s}+\tfrac{2}{73^s}
+\tfrac{2}{89^s}+\tfrac{2}{97^s}+\tfrac{2}{101^s}+\tfrac{2}{109^s}+\tfrac{2}{113^s}+\\
&\quad\tfrac{2}{137^s}+ \tfrac{2}{149^s}+\tfrac{2}{157^s}+\tfrac{2}{169^s}+\tfrac{2}{173^s}+\tfrac{2}{181^s}+
\tfrac{2}{193^s}+\tfrac{2}{197^s}+\tfrac{4}{221^s}+\tfrac{2}{229^s}+\ldots.\,.
\end{alignat*}
\end{ex}
\begin{ex} $x=\frac{2}{5}+\frac{1}{5}i=\frac{i}{1+2i}$
\setlength{\unitlength}{1.25in}
\begin{center}
\begin{picture}(0.65,0.80)(0,-0.15)
\linethickness{0.10pt}
\put(0,0){\vector(1,0){0.65}}
\put(0,0){\vector(0,1){0.65}}
\put(-0.1,0.48){$\frac{1}{2}$}
\put(0.46,-0.18){$\frac{1}{2}$}
\multiput(0,0.5)(0.105,0){5}{\line(1,0){0.055}}
\thicklines
\put(0,0){\line(1,0){0.5}}
\put(0.5,0){\line(0,1){0.5}}
\put(0,0){\line(1,1){0.5}}
\put(0.4,0.2){\circle*{0.08}}
\put(0.3,0.2){\makebox(0,0){$x$}}
\end{picture}
\end{center}
Since the denominator $q=1+2i$ of $x$ does not divide $1-i$, we see from \eqref{ezmcz} that $q|(\varepsilon z-\overline{z})$ if and only if $5|(\varepsilon z-\overline{z})$. Hence,
$SOC(x+\Gamma)=SOC(\frac{1}{5}+\Gamma)\cong{\mathbb{Z}}^{(\aleph_0)}$ by Lemma \ref{prop8}. Applying Theorem \ref{prop5}, if $RT_r\in OC(x+\Gamma)$ where $R(z,\varepsilon)\in SOC(\Gamma)$ then
$5|N(z)$ because $OC(x+\Gamma)$ does not contain a reflection in $P(\Gamma)$. Observe that given $z$ with $5|N(z)$, we must find either $1+2i$ or $1-2i$ (and not both) in the
factorization of $z$ into primes in $\mathbb{Z}[i]$. If $(1-2i)|z$ then $z\overline{x}\in\mathbb{Z}[i]$ and $\varepsilon z\overline{x}-\overline{z}x=\varepsilon
z\overline{x}-\overline{z\overline{x}}\in\mathbb{Z}[i]$, $\forall\varepsilon$. Lemma \ref{lem3} then implies that
\[OC(x+\Gamma)=SOC(x+\Gamma)\cup\set{RT_r: R(z,\varepsilon)\in SOC(\Gamma)\text{ with }(1-2i)|z}.\]
Here, we have an example of a set $OC(x+\Gamma)$ that is not a group. Indeed, let $T_k=R_kT_r\in OC(x+\Gamma)\setminus SOC(x+\Gamma)$ where $R_k=R_k(z,\varepsilon_k)\in SOC(\Gamma)$, for $k=1,2$,
with $\varepsilon_1\neq\varepsilon_2$. We obtain that $T_1T_2$ is not the identity with $T_1T_2=R_1{R_2}^{-1}\in P(\Gamma)$ which means that $T_1T_2\notin OC(x+\Gamma)$. By Theorem \ref{prop5},
$OC(x+\Gamma)$ is not a subgroup of $OC(\Gamma)$.
The Dirichlet series generating function for $f_{x}(m)$ is given by
\begin{alignat*}{2}
\Phi_{x}(s)&=\sum_{m=1}^{\infty}{\frac{f_{x}(m)}{m^s}}=\frac{1}{1-5^{-s}}\cdot\prod_{\stackrel{p\equiv 1(4)}{p\neq 5}}\frac{1+p^{-s}}{1-p^{-s}}\\
&=1+\tfrac{1}{5^s}+\tfrac{2}{13^s}+\tfrac{2}{17^s}+\tfrac{1}{25^s}+\tfrac{2}{29^s}+\tfrac{2}{37^s}+\tfrac{2}{41^s}+\tfrac{2}{53^s}+\tfrac{2}{61^s}+\tfrac{2}{65^s}+
\tfrac{2}{73^s}+\ldots\,.
\end{alignat*}
Also, if we denote by $\hat{F}_x(m)$ the number of (linear) coincidence isometries of $x+\Gamma$ of index $m$, we obtain that
\[\hat{F}_{x}(m)=\left\{\begin{aligned}
f_{x}(m) &\;\text{if } 5 \nmid\, m\\
4\,f_{x}(m) &\;\text{if } 5\,|\, m.\\
\end{aligned}\right.\]
Hence, the Dirichlet series generating function for $\hat{F}_{x}(m)$ is given by
\begin{alignat*}{2}
\Psi_{x}(s)&=\sum_{m=1}^{\infty}{\frac{\hat{F}_{x}(m)}{m^s}}=\frac{1+3\cdot 5^{-s}}{1-5^{-s}}\cdot\prod_{\stackrel{p\equiv 1(4)}{p\neq 5}}\frac{1+p^{-s}}{1-p^{-s}}\\
&=1+\tfrac{4}{5^s}+\tfrac{2}{13^s}+\tfrac{2}{17^s}+\tfrac{4}{25^s}+\tfrac{2}{29^s}+\tfrac{2}{37^s}+\tfrac{2}{41^s}+\tfrac{2}{53^s}+\tfrac{2}{61^s}+\tfrac{8}{65^s}+
\tfrac{2}{73^s}+\ldots\,.
\end{alignat*}
\end{ex}
\section{Conclusion and outlook}
We have seen that the coincidence isometries of a shifted lattice are also coincidence isometries of the original lattice. Moreover, the CSLs of the shifted lattice are merely translations of CSLs of the original lattice. Thus, no new values of coincidence indices $\Sigma$ are obtained by shifting the lattice, and some $\Sigma$-values even disappear or their multiplicity is reduced.
The coincidences of a shifted square lattice were examined in this paper by identifying the lattice with the ring of Gaussian integers. The problem was completely solved for the case when
the shift consists of an irrational component. For the remaining case, that is, when the shift may be written as a quotient of two Gaussian integers that are relatively prime, one needs
to compute the set of coincidence rotations for each possible denominator via some divisibility condition. Partial results are given here and in \cite{LZ} on how to obtain the coincidence isometries and indices for any given denominator and corresponding numerator. General results in this direction will depend on the arithmetic of the Gaussian integers.
It should be emphasized that the order of rotation and translation of a lattice is in general not interchangeable. In this paper, we compare the shifted lattice with its rotated copy, that is, the translation (say $x$) comes first before rotation. This corresponds to the situation where the lattice is first rotated by $R$ and is shifted afterwards by the vector $Rx-x$, which is equivalent to a rotation of the lattice about a different point ($-x$), thus keeping at least one point ($-x$) fixed. In particular, the CSLs of a shifted lattice are shifted copies of the intersection of a lattice with a rotated, then translated version of the same lattice (see \cite{LZ}).
The next step is to extend these results to planar modules by identifying the modules with rings of cyclotomic integers (\cite{PBR}). General results for lattices will of course also hold in three dimensions, but the approach in this case will not be via complex numbers but via quaternions.
Finally, it is expected that the ideas behind the study of a shifted lattice may be applied to crystals where there is more than one atom per primitive unit cell, as described in \cite{GP,PV}. A related mathematical problem is the following: Suppose the coincidence problem for a sublattice (of finite index) of a given lattice has already been solved. What can be deduced about the coincidence indices of the the original lattice? A possible approach to answer this question involves looking at the coincidences of the corresponding cosets, which are just shifted copies of the sublattice.
\subsection*{Acknowledgements}
The authors are grateful to the referee for his valuable remarks on the manuscript. M. Loquias would like to thank the Deutscher Akademischer Austausch Dienst (DAAD) for financial support during his stay in Germany. This work was supported by the German Research Council (DFG), within the CRC~701.
|
\section{Introduction and motivation}
Predictions of inflation seem to be in excellent agreement with the
CMB data~\cite{2009-Komatsu.etal-ApJS}.
However, it is still unclear what is the nature of the field which drives
inflation.
Historically, a canonical scalar field has been the preferred candidate
for inflaton, but in recent years, also a non-canonical scalar field,
dubbed as $k$-inflaton, was considered as serious alternative
mechanisms to drive inflation~\cite{1999-Armendariz-Picon.etal-PLB}.
\par
Both scenarios lead to nearly scale invariant power-spectra with
negligible running and hence can not be distinguished (or ruled out)
from the current CMB data~\cite{2009-Komatsu.etal-ApJS}.
The future missions, including PLANCK~\cite{2005-Planck-ScientificProgrammeof},
hold promise in ruling our either of these two scenarios by looking at
the non-Gaussianity of the primordial spectra, but
quantifying non-Gaussianity requires one to go beyond linear
order~\cite{1997-Bruni.etal-CQG,2003-Maldacena-JHEP,2005-Seery.Lidsey-JCAP}.
\par
There are four different approaches in the literature to study
cosmological perturbations:
{\tt 1)}~solving Einstein's equations
order-by-order~\cite{1963-Lifshitz.Khalatnikov-AdP};
{\tt 2)}~the covariant approach based on a general frame vector
$u^{\alpha}$~\cite{1966-Hawking-ApJ};
{\tt 3)}~the Arnowitt-Deser-Misner (ADM) approach based on the normal
frame vector $n^{\alpha}$~\cite{1980-Bardeen-PRD};
{\tt 4)}~the reduced action approach~\cite{1980-Lukash-ZhETF}.
In the case of linear perturbations, it was shown that all of
these four approaches lead to identical equations of motion.
However, to our knowledge, a complete analysis has not been done in the
literature for higher-order perturbations.
\par
In this talk, to illustrate the problems that may occur at
higher-order, and not to get bogged-down with the gauge issues, we
consider a simple situation:
we freeze all metric perturbations
and focus on the perturbations of a minimally-coupled,
generalised scalar field $\phi$, whose Lagrangian density is given
by~\cite{1999-Armendariz-Picon.etal-PLB}
{\small
\begin{equation}
\label{eq:genscal}
{\cal L} = P(X, \phi) \ , \qquad \mbox{where} \qquad 2\,X = \nabla^{\alpha}
\phi\,\nabla_{\alpha} \phi \ .
\end{equation} }
More precisely, we will only consider linear perturbations of the
scalar field,
{\small \begin{equation}
\label{eq:def-Perphi}
\phi(t, {\bf x})=\phi_{_0}(t) + \delta^{(1)}\phi(t, {\bf x})
\ ,
\end{equation}
}
about the 4-dimensional FRW~background, while expanding all the
dependent quantities, like $X$ and stress tensor, up to second order,
and highlight the main differences in these approaches.
For this purpose, it is convenient to compare the ratio
{\small
\begin{equation}
\label{eq:cs-def}
c_{\rm s}^2 = \frac{\mbox{coefficients of~} (^{(0)}\!\!\dot{\varphi}_{,i}/a)^2}
{\mbox{coefficients of~}
^{(0)}\!\!\ddot{\varphi}^2}
\ ,
\end{equation}
}
in the components of the stress~tensor and related quantities.
Since $c_{\rm s}^2$ is dimensionally the square of a speed,
we will refer to this ratio as the ``speed of propagation''.
\section{Key results}
Below, we will provide the key results and, for details, we refer the
reader to Ref.~\cite{2009-Appignani.etal-Arx}.
\\[5pt]
\noindent
{\tt Perturbed tensor and ADM approach:}
For an arbitrary scalar field Lagrangian, $\delta^{^{(2)}}\! T_{00}$ may represent
an unstable perturbation; only under very special conditions, most notably
for the canonical scalar field, the {\it effective speed of propagation\/} of
$\delta^{^{(2)}}\! T_{00}$ (and $\delta^{^{(2)}}\! T_{ii}$) are the same as that of the standard
definition~\cite{1999-Garriga.Mukhanov-PLB}
{\small
\begin{equation}
c^2_{^{(0)}\!\!\dot{\varphi}} = \frac{P_{_X}^{^{(0)}}}{P_{_X}^{^{(0)}} + P_{_{XX}}^{^{(0)}}\,\dot\phi_{_0}^2}
\end{equation}
}
Using~\eqref{eq:cs-def}, we can define speeds related with the
propagation of density perturbations in the background frame from
$\delta^{^{(2)}}\! T_{00}$ and $\delta^{^{(2)}}\! T_{ii}$, respectively,
{\small
\begin{equation}
\label{eq:csT}
c_{0}^2= \strut\displaystyle\frac{P_{_X}^{^{(0)}} - P_{_{XX}}^{^{(0)}}\, \dot\phi_{_0}^2} {P_{_X}^{^{(0)}}
+4\,P_{_{XX}}^{^{(0)}}\,\dot\phi_{_0}^2+ P_{_{XXX}}^{^{(0)}}\, \dot\phi_{_0}^4} \, , \quad c_{\parallel}^2
= c^2_{\delta \phi} \, .
\end{equation}
}
Note that these velocities become imaginary indicating that the
perturbations may be unstable.
The nature of this instability is easily understood in analogy
with classical mechanics:
when $c_{^{(0)}\!\!\dot{\varphi}}^2$ is negative, the system resembles an inverted
harmonic oscillator and, no matter how small $\delta \phi$,
it will rapidly run away from the background solution $\phi_{_0}$ and from the
perturbative regime.
\\[5pt]
\noindent
{\tt Covariant approach:}
In the fluid frame, the energy density exhibits the same kind of instability
as the perturbed stress-tensor $\delta^{^{(2)}}\! T_{00}$, also for the canonical scalar field.
In particular, the perturbations turn out to be unstable because of the negative
sign of the spatial momentum contribution.
Only under special conditions, the {\it effective speed of propagation\/}
of the energy density and pressure perturbations are equal.
Using~\eqref{eq:cs-def}, the speed of propagation for density and pressure
perturbations in the fluid frame are
{\small
\begin{equation}
c_{\rho}^2 = - \frac{P_{_X}^{^{(0)}} + P_{_{XX}}^{^{(0)}}\,\dot\phi_{_0}^2} {P_{_X}^{^{(0)}}
+4\,P_{_{XX}}^{^{(0)}}\,\dot\phi_{_0}^2 + P_{_{XXX}}^{^{(0)}}\, \dot\phi_{_0}^4} \, , \quad
c_{p}^2=c_{\parallel}^2=c_{\delta\phi}^2
\ ,
\end{equation}
}
which differ from Eq.~\eqref{eq:csT}.
\\[5pt]
\noindent
{\tt Symmetry reduced approach:}
The basic idea here is to perturb the action about the FRW~background,
up to second (or higher) order, and reduce it so that the perturbations are
described in terms of a single gauge-invariant variable.
The second order canonical Hamiltonian is identical to the stress-tensor
for the canonical scalar field, but differs for general non-canonical fields.
This implies that $\delta^{^{(2)}}\! T_{00}$ and the canonical Hamiltonian
$\delta^{^{(2)}}\!\mathcal{H}$ may become unstable under different conditions.
\par
\section{Discussion}
The first question is why the canonical Hamiltonian, perturbed
stress~tensor and super-Hamiltonian coincide for a canonical scalar
field, but not for general scalar field Lagrangians.
To go about answering this question, it is necessary to look at the four
approaches we have employed from a different perspective.
In the first two approaches -- perturbed stress~tensor and covariant
approach -- we perturb the general expression for the scalar field
stress~tensor and obtain its second order contribution $\delta^{^{(2)}}\! T_{00}$.
In the last two approaches -- ADM~formulation and symmetry-reduced
action -- we expand the action to second order in the perturbation and
obtain the super-(canonical) Hamiltonian of the corresponding
perturbed action.
While the super-Hamiltonian $\delta^{^{(2)}}\! H$ is identical to $\delta^{^{(2)}}\! T_{00}$,
the canonical Hamiltonian $\delta^{^{(2)}}\!\mathcal{H}$ differs.
\par
Under what condition the super-Hamiltonian $\delta^{^{(2)}}\! H$, the canonical
Hamiltonian $\delta^{^{(2)}}\!\mathcal{H}$ for non-canonical scalar fields are
identical?
They are all identical provided the time-variation of
background quantities (like $P_{_X}^{^{(0)}}$ and $P_{_{XX}}^{^{(0)}}$) can be neglected.
(For the canonical scalar field, these functions are indeed constant
and the perturbed quantities therefore coincide.)
Although such an approximation may be valid for specific non-canonical
fields, they fail for some of the known fields like Tachyon,
DBI~\cite{2009-Appignani.etal-Arx}.
\par
R.~C.~is supported by the INFN grant BO11 and S.~S.~was supported by
the Marie Curie Incoming International Grant IIF-2006-039205.
|
\section{Introduction}
One of the most interesting experiments in the history of physics
is the double-slit experiments, which infers to the quantum
mechanical nature of the particles. The technological developments
in producing low dimensional high mobility charge carrier systems,
enabled experimentalists to re-do the double-slit experiments
considering nanostructures. In the experiments performed at
cryogenic temperatures and considering a two dimensional electron
system (2DES), the phase and the transmission amplitude were
measured
simultaneously~\cite{Yacoby95:abinter1,Heiblum05:abinter}. The
findings of these and consequent experiments activated a huge
number of theoreticians to understand the physics underlying the
abrupt phase
changes~\cite{Hacken95:phase,Imry00:towards,Hacken2001:all,Theresa07:lapses,Theresa07:crossover,Imry07:lapses},
for a comprehensive review we suggest the reader to check
especially Ref.~\cite{Hacken2001:all} and the references given
thereby. In particular, G. Hackenbroich \textit{et. al}
investigated the effect of shape deformation of a parabolic
quantum dot (QD), in the absence of Coulomb interaction, and
showed that the degeneracy due to the symmetry of the QD is
lifted, however, for the deformed QD it is still possible to
obtain density of states in broad energy intervals, which are
large compared to the single particle level spacings $\delta$. The
interplay between the level width $\Gamma$ and $\delta$ is used to
give an explanation to the observed phase
anomalies~\cite{Theresa07:lapses,Imry07:lapses}. On the other
hand, the effect of interactions was included by M. Stopa by
solving the related Poisson and Schr\"odinger equations
self-consistently within a Hartree-Fock type mean field
approximation, however, its influence on the phase was left
unresolved. We should also note that, in these calculations a
rather simplified QD geometry was investigated compared to the
experiments.
This work aims to provide numerical support to the theories which
rely on the formation of a wide state at certain QD geometries. We obtain the potential profile of the real samples by
solving the Poisson equation in 3D using fast Fourier
transformation,
iteratively~\cite{Weichselbaum03:056707,Sefa08:prb}. In our
calculations, we consider the sample geometry presented in
Ref.~\cite{Heiblum05:abinter}. The next step is to obtain the
energy eigenstates and values for the calculated effective
potential. We solve the Schr\"odinger equation by diagonalizing
the single-particle Hamiltonian implementing the finite difference
techniques.
\section{Theory}
Here, we investigate the single particle
eigen-energies and eigenfunctions of the reduced 2D Hamiltonian
\begin{eqnarray} H=\frac{\bf{p}^2}{2m^*}+V(x,y), \end{eqnarray} where $\bf{p}$ is the
momentum operator in 2D, $m^*$ is the effective mass ($=0.067m_e$
in GaAs) and $V(x,y)=V_G(x,y)+V_D(x,y)+V_l(x,y)+V_H(x,y)$ is the
mean field potential composed of gates, leads, donors and Hartree
terms, respectively.
\newline
In physics, WKB approximation is one of the most frequently applied approximation to solve Schrdinger's equation.
The transmission amplitude $W_n(a,b)$ is calculated at the
barrier along the classical turning points (a,b), via \begin{eqnarray}
W_n(a,b)(E)=\frac{e^{\xi(E)}}{1+\frac{1}{4}e^{\xi(E)}},\nonumber \\
{\xi(E)}=-2\int_{a}^{b}dx\sqrt{\frac{2m}{\hbar^2}(V(x)-E)}.\end{eqnarray}
It is known that the transmission amplitude depend
almost linearly to the energy of the incoming state, assuming
plane waves and within the WKB approximation~\cite{Stopa96:selfcTF}, which we utilize likewise in the following to calculate transport through the QD.
We proceed our work by considering the real
geometry and the potential profile calculated within the
self-consistent Thomas-Fermi-Poisson (TFP) theory. In the
following section we first discuss the limitations of such a mean
field approximation and compare our method with the existing
calculation schemes in the high electron occupation regime, \emph{i.e.}
$N\gtrsim100$. We show that, the single particle energy states and
energies can be well described in this regime considering TFP
theory.\begin{figure}{\centering
\includegraphics[width=.7\linewidth]{fig1.pdf}
\caption{ \label{figure:figure1} The self-consistent potential plotted
in 3D as a function of the lateral coordinates. The lengths are in
units of effective Bohr radius and energy is normalized with the
effective Rydberg energy. The inset depicts the sample geometry,
where S stands for source lead and D stands for the drain lead.
Coupling of the QD to the leads is manipulated by changing the
applied potential $V_2$.}}
\end{figure}
\section{Results and Discussion}
The calculation of the electrostatic potential considering real
sample geometries together with the electron-electron (e-e)
interaction is a challenging issue. Since such a calculation
cannot be done analytically for almost all the cases, usually
numerical techniques are deployed. It is clear that for "more than
a few" electron regime ($N>10$) exact diagonalization methods are
either impossible or very costly in terms of computational effort.
\begin{figure}
{\centering
\includegraphics[width=.7\linewidth]{fig2.pdf}
\caption{ \label{figure:figure2} (a-g) Selected eigenfunctions residing
at the $E=1.96$ plateau (indicated by the horizontal line in i)
calculated for the color plotted potential (h), together with the
energy spectrum near $E_F$. The state $=160$ present a slight
asymmetry in coupling to source lead compared to the drain. The
asymmetric potential distribution with in the dot is visible.}}
\end{figure}
It is favorable to use a mean field approximation to describe the
(e-e) interactions, which is questionable in the "less than a few"
electron regime. The commonly used approach to determine the bare
confinement potential generated by the gates is the "frozen
charge" approximation~\cite{Davies95:4504}, which takes into
account properly the gate pattern and the effect of the spacer
between the gates and the 2DES. Since, it is not self-consistent
this approximation cannot account for the induced charges on the
metallic gates defining the QD. The effects resulting from the
induced charges and donor layer can be handled by solving the 3D
Poisson equation self-consistently. Almost a decade ago M. Stopa
introduced a very effective numerical scheme to describe the
electrostatics of such samples~\cite{Stopa96:selfcTF}, including
the e-e interactions either using a full Hartree, i.e. solving the
Poisson and Schr\"odinger equations self-consistently, or
considering Thomas-Fermi approximation (TFA). The
exchange-correlation interaction was accounted by a local density
approximation (LDA) using the density functional theory (DFT). It
was shown that the TFA is powerful enough to describe the
electrostatic potential even if the electrons are fully depleted
in some regions of the sample~\cite{Stopa96:selfcTF}.
Here, we stay in the TFA to calculate the electrostatic properties
of the real sample geometry using the algorithm developed by A.
Weichselbaum \textit{et.
al}~\cite{Andreas03:potential,Andreas:06}, which implements an
efficient grid relaxation technique to solve the 3D Poisson
equation. This approach was shown to be reliable to obtain the
potential profiles in the "more than a few" electron regime
considering QDs and quantum point contacts~\cite{Sefa08:prb}. The
next step in our calculation scheme is to obtain the single
particle energies and states, which we do same as described in the
previous section.
Figure.~\ref{figure:figure1} presents the calculated potential profile for
the sample geometry measured in Ref.~\cite{Heiblum05:abinter}. We
apply negative voltages to the gates shown in the inset. The upper
and lower two gates (denoted by red areas) are kept at the same
potential $V_2$, whereas the center gate (left black) and the
plunger gate (right black) are biased with a fixed voltage $V_1$.
Here, we consider a unit cell of 440x440 nm$^2$ spanned by 128x128
mesh matrix to calculate the self-consistent potential. The
surface potential is fixed to -0.75 V pinning the Fermi energy at
the mid gap. The 2DES is some 100 nm below the surface followed by
a thick GaAs layer. To achieve numerical convergence and satisfy
the open boundary conditions 3 mesh points of dielectric material
is assumed at all boundaries. In Figure.~\ref{figure:figure1} fixed
voltages of $V_1=-1.5$ V and $V_2=-2.2$ V are applied, the bulk
electron density is estimated to be $3\times10^{11}$ cm$^{-2}$
corresponding to $E_F\approx12.75$ meV, with the given density,
the number of electrons in the dot $N$ is similar to 200.
Figure.~\ref{figure:figure2} presents the calculated single particle wave
functions as a function of spatial coordinates, together with the
potential counter plot and the corresponding eigen energies versus
the state number. We show the states residing at the energy
\textit{plateau }, which lay in the close vicinity of $E_F$
(depicted by the horizontal solid line in Figure.~\ref{figure:figure2}i).
The states shown at the upper panel present the chaotic behavior,
whereas the first two states of
the mid panel are the non-propagating states. At $n=165$ a
resonant channel is observed, meanwhile the highest state shown
presents the chaotic behavior. These results show that,
qualitatively, transport through state 165 is much probable
compared to the others sitting at the same plateau. Although the
single particle energy eigenvalues are close to each other a
single channel is in charge of transport. At these gate voltages,
the QD is loosely defined as one can see that it is possible to
find an electron also at the left side of the actual QD. This
situation is changed by applying a higher negative potential to
the central and the plunger gates, $V_1=-2.0$ V. However, the QD
potential is not rotationally symmetric even if one neglects the
gates, since the center gate is geometrically different from the
plunger gate.
Now, we turn back our attention to the level width $\Gamma$ in a
qualitative manner. As we have discussed $\Gamma$ becomes
meaningful if one also considers both the source and drain leads.
\begin{figure}
{\centering
\includegraphics[width=.7\linewidth]{fig3.pdf}
\caption{ \label{figure:figure3} The transmission coefficients
calculated at the barrier. The turning points are obtained from
the self-consistent potential at the center of the barrier, where
the energy of the incoming wave cuts along.}}
\end{figure}
At this point it is useful to look at the transmission probability
$W_n(a,b)(E)$, in Fig.~\ref{figure:figure3} we show the this quantity as
a function of energy of the incoming wave calculated within the
WKB at various gate voltages, $V_1$, $V_2$. We see that, when the
upper and the lower gates are biased with small potentials, the
transmission increases linearly. This linearity changes if one
applies higher voltages to the barriers, however, for higher
energies the linearity is recovered. Such an observation leads us
to conclude that, essentially the
probability distribution determines the level widths, which may
become asymmetric considering transport at different energies.
To summarize, we have calculated the self-consistent electrostatic
potential exploiting the smooth variation of the bare potential
within the TFA. Next we obtained the single particle eigenstates
and energies considering a real sample geometry and crystal
structure. We found, similar to the Ref.~\cite{Imry00:towards}, that some single
particle levels bunch and present a energy plateau while changing
the state number. It was observed that, within these plateaus, not
all the states contribute to the transport since the overlap of
the dot wave functions and lead wave functions simply vanish. More
interestingly, we found that at intermediately high energies, the
wave functions are coupled to at least one of the leads much
stronger than the ones in their close energy vicinity. This
result, we believe, supports the phenomenological models, which
attribute the abrupt change of the phase lapses to
electron-electron interactions.
\ack
We would like to thank Jan von Delft for introducing us the ``phase lapses'' problem and motivating us to perform numerical calculations. Moty Heiblum is also acknowledged for his enlightening discussions. This work is partially supported by TUBiTAK, under grant no:109T083.
\section*{References}
|
\section{Introduction and some preliminary results}
In recent times, Minkowski Geometry (i.e., the geometry of finite dimensional, real Banach spaces; see \cite{tho}) became again an important research field. Strongly related to Banach Space Theory, Finsler Geometry and Classical Convexity, it is permanently enriched by new results in the spirit of Discrete and Computational Geometry, of Operations Research, Location Science, and further applied disciplines.
In addition, the most examined concepts of it naturally connect to Physics, Functional Analysis, and non-Euclidean Geometries.
We will not introduce basic notions and terminology of this field going beyond our purpose; for its fundaments the reader is referred to the monograph \cite{tho} and to the surveys \cite{martini 1} and \cite{martini 2}.
The present paper refers to {\em bisectors} in (finite dimensional normed or) \emph{ Minkowski spaces}, i.e., to collections of points which have, in each case, the same distance (with respect to the corresponding norm) to two given points ${\bf x}$, ${\bf y}$ of these spaces. Note that bisectors in Minkowski spaces play an essential role in Discrete and Computational Geometry, mainly in view of constructing (generalized) Voronoi diagrams, and also for motion planning with respect to translations; see, e.g., the surveys \cite{aurenhammer} and \cite{martini 2}.
In some previous papers on this topic (see \cite{gho1}, \cite{gho2}, and \cite{gho3}),
\'A. G. Horv\'ath proved that if the unit ball of a Minkowski space is strictly convex, then every bisector is a topological
hyperplane (see Theorem 2 in \cite{gho1}, or \cite{martini 2}). On the other hand, Example 3 in \cite {gho1} shows that
strict convexity does not follow from the fact that all bisectors are topological hyperplanes.
In these papers, the connections between shadow boundaries of the
unit ball and bisectors (regarding the direction ${\bf x}-{\bf y}$ or the points {\bf x} and {\bf y}, respectively) in Minkowski spaces are investigated. The author was sure
that the following statement is true: \emph{A bisector is a
topological hyperplane if and only if the corresponding shadow
boundary is a topological $(n-2)$-dimensional sphere}. However,
the respective conjecture was proved only in the three-dimensional case (Theorem
2 and Theorem 4 in \cite{gho2}).
In \cite{gho3}, some further topological observations on shadow boundaries are discussed, e.g., that they are
compact metric spaces containing $(n-2)$-dimensional closed,
connected subsets separating the boundary of $K$. \'A. G. Horv\'ath also investigated
the manifold case, and he proved (using an approximation theorem for
cell-like mappings) that the shadow boundary is homeomorphic to an
$(n-2)$-dimensional sphere. Consequently this result (if
the bisector is a homeomorphic copy of $R^{n-1}$, then the shadow
boundary is a topological $(n-2)$-sphere) confirms the
first direction of the above mentioned conjecture.
Independently, H. Martini and S. Wu \cite{martini-wu} introduced and investigated the concept of radial projection of bisectors. Strongly using the central symmetry of Minkowskian balls, they proved some interesting results on radial projections of bisectors.
Theorem 2.6 in \cite{martini-wu} says that the shadow boundary is a subset of the closure of such a radial projection, and
Theorem 2.9 there refers to the converse statement. If for a point ${\bf x}$ from the boundary of the unit ball there exists a point ${\bf z}$, unique except for the sign, such that ${\bf x}$ is orthogonal to ${\bf z}$ in the sense of Birkhoff (see below), then ${\bf z}$ is a point of the radial projection of the bisector corresponding to ${\bf x}$ and $-{\bf x}$.
In the present paper we introduce the concept of bounded representation of bisectors, which yields a useful combination of the notions of bisector, shadow boundary, and radial projection. We prove that the topological properties of the radial projection (in higher dimensions) do not determine the topological properties of the bisector. More precisely, the manifold property of the bisector does not imply the manifold property of the radial projection (see our Example below). The situation is different with respect to the bounded representation of the bisector. Namely, if one of them is a manifold, then the other is also. More precisely, if the bisector is a manifold of dimension $n-1$, then its bounded representation is homeomorphic to a closed $(n-1)$-dimensional ball $B^{n-1}$ (i.e., it is a cell of dimension $n-1$). And conversely, if the bounded representation is a cell, then the closed bisector is also (Theorem 1).
We will also present new approaches to higher dimensional analogues of several theorems given in \cite{martini-wu}. By our new terminology, we will rewrite and reprove Theorems 2.6, 2.9, and 2.10 from that paper.
\section{Basic notions and radial projections of bisectors}
Let $K$ be a compact, convex sets with nonempty interior (i.e., a {\em convex body}) in the $n$-dimensional Euclidean space $E^n$ which, in addition, is centred at the origin $O$. Then the $(n-1)$-dimensional boundary bd $K$ of $K$, in the following also denoted by $S$, can be interpreted as the {\em unit sphere} of an $n$-dimensional (real Banach or) \emph{Minkowski space} $M^n$ with norm $\|\cdot \|$, i.e.,
$$
S:=\{{\bf x}\in M^n\mbox{ : } \|{\bf x}\|=1\}.
$$
It is well known that there are different types of orthogonality in Minkowski spaces. In particular, for ${\bf x},{\bf y}\in M^n$ we say that ${\bf x}$ is {\em Birkhoff orthogonal} to ${\bf y}$ if $\|{\bf x}+t{\bf y}\|\geq \|{\bf x}\|$ for all $t\in \mathbb{R}$, denoted by ${\bf x}\bot _B{\bf y}$ (see \cite{birkhoff}); and {\bf x} is {\em isosceles orthogonal} to ${\bf y}$ if $\|{\bf x}+{\bf y}\|=\|{\bf x}-{\bf y}\|$, denoted by ${\bf x}\bot _I{\bf y}$ (cf. \cite{james}). The {\em shadow boundary} $S(K,{\bf x})$ of $K$ with respect to the direction {\bf x} is the intersection of $S$ and all supporting lines of $K$ having direction ${\bf x}$.
Given a point ${\bf x}\in
S$, the \emph{bisector} of $-{\bf x}$ and ${\bf x}$, denoted by $B(-{\bf x},{\bf x})$, consists of all those vectors {\bf y}
which are isosceles orthogonal to ${\bf x}$ with respect to the Minkowski norm generated by $K$. The \emph{radial projection} $P({\bf x})$ of this bisector consists of those points ${\bf y}$ of $S$ for which there is a positive real value $t$ such that $t{\bf y}\in B(-{\bf x},{\bf x})$.
We remark that, in the relative topology of $S$, $P({\bf x})$ can either be closed or open; this can be easily seen in the cases of the Euclidean and of the maximum norm. Thus, for
topological investigations in higher dimensions we suggest the extension of the definition of $B(-{\bf x},{\bf x})$ to ideal points by a limit property.
\begin{defi}
Consider the compactification of $E^n$ to a closed ball $B^n$ by the set of ideal points ${\bf x}_\infty $ ($-{\bf x}_\infty\not ={\bf x}_\infty$). We say that ${\bf y}_\infty :=\infty {\bf y}\in B(-{\bf x},{\bf x})$ if there is a sequence $(t_i{\bf y}_i)\in B(-{\bf x},{\bf x})$ for which $\lim\limits_{i\rightarrow \infty} {\bf y}_i={\bf y}$. We call the points of the original bisector {\em ordinary points} and the new points {\em ideal points}, respectively.
\end{defi}
By this definition $P({\bf x})$ could be closed as we can see in our first statement. Let $P({\bf x})^l$ be the collection of those points ${\bf y}$ of $S$ for which $$
\|t{\bf y}+{\bf x}\|< \|t{\bf y}-{\bf x}\|
$$
holds, for all real $t\geq 0$. Let $P({\bf x})^r$ denote the image of $P({\bf x})^l$ under reflection at the origin.
\begin{proposition}
In the described way, $S$ is decomposed into three disjoint sets: $P({\bf x})$, $P({\bf x})^l$, and $P({\bf x})^r$. $P({\bf x})$ is an
at least $(n-2)$-dimensional closed (and therefore compact) set in $S$ which is connected for $n \geq 3$, the
sets $P({\bf x})^l$ and $P({\bf x})^r$ are arc-wise connected components of their union.
\end{proposition}
\noindent {\bf Proof: } By Theorem 5.1 of \cite{martini-wu}, $P({\bf x})$ is connected for $n\geq 3$. We prove that it is also closed with respect to the relative topology of the boundary of the unit ball. To see this, consider a convergent
sequence $({\bf y}_i$) in $P({\bf x})$ having the limit {\bf y}. For any $i$ there is a new sequence of points $({\bf y}^j_i)$ such that for every pair $\{i,j\}$ there are $t_j\in \mathbb{R}^+$ and ${\bf x}_i^j\in B(-{\bf x};{\bf x})$ such that $ (t^j_i{\bf y}^j
_i) ={\bf x}^j_i$. (For an ordinary point the mentioned sequence can be regarded as a constant one.) It is clear that for the diagonal sequence $({\bf y}^i_i)$ we have
$$
\lim\limits_{i\rightarrow \infty}{\bf y}^i_i = {\bf y},
$$
implying that {\bf y} is also in $P({\bf x})$.
\begin{figure}
\centerline{\includegraphics[scale=0.8]{rp3.eps}}
\caption{Vectors used in the proof of Proposition 1}
\end{figure}
The continuity property of the norm function implies that all points of $S$ belong to precisely one
of the three mentioned sets. Thus the first statement is clear, and the union of $P({\bf x})^l$ and $P({\bf x})^r$
is open with respect to the topology of $S$.
Observe once more that $P({\bf x})^l$ and $P({\bf x})^r$ are images of each other regarding reflection at the origin. Furthermore, they are arc-wise connected sets. To prove this, consider the following inequality for
an element {\bf y} of $P({\bf x})^r$:
$$
\|({\bf y}-t({\bf y}-{\bf x})-{\bf x}\| = (1-t)\|{\bf y}-{\bf x}\| < (1-t)\|{\bf y} + {\bf x}\| = \|({\bf y}-t({\bf y}-{\bf x})) + {\bf x}- 2t{\bf x}\|,
$$
where $0 \leq t \leq 1$ is an arbitrary parameter. The point ${\bf z}_t := ({\bf y}-t({\bf y}-{\bf x}))+{\bf x} = (1-t){\bf y}+(1+t){\bf x}$
is on the right half-line, starting with the point $(1-t)({\bf y} + {\bf x}) ={\bf z}_t- 2t{\bf x}$ and being parallel to the
vector ${\bf x}$, meaning that its norm is larger than the norm of the point ${\bf z}_t - 2t{\bf x}$ (see Fig. 1).
Thus
$$
\|{\bf z}_t\|\geq \|{\bf z}_t -2t{\bf x}\|,
$$
and so
$$
\|({\bf y}-t({\bf y}-{\bf x}))-{\bf x}\| < \|({\bf y}- t({\bf y}-{\bf x})) + {\bf x}\|.
$$
A consequence of this inequality is that the arc of $S$ connecting the respective endpoints of the
vectors ${\bf y}$ and ${\bf x}$ belongs to the set $P({\bf x})^r$. Thus every two points of $P({\bf x})^r$ can be connected by
an arc, as we stated.
Now, with respect to the topology of their union, they are connected components. This means
that both of them are also open with respect to the topology of $S$. Thus $P({\bf x})$ separates $S$. By
a theorem of P. S. Aleksandrov (Theorem 5.12 in vol. I of \cite{alexandrov}) we get that the topological dimension of
$P({\bf x})$ is at least $n-2$.
\hfill $\Box$
\noindent {\bf Remark: }
If we identify the opposite points of $S$, then we get an $(n-1)$-dimensional projective space $P$ dissected into two parts, one of them open and the other one closed, respectively. The open part contains the points of the identified sets $P({\bf x})^r$ and $P({\bf x})^l$, while the closed one contains the identified point-pairs of $P({\bf x})$.
\section{Bounded representation of the bisector}
\begin{figure}
\centerline{\includegraphics[scale=0.8]{rp.eps}}
\caption{Connection between the parameters $t$ and $t_{\bf z}$.}
\end{figure}
We define now an important mapping of the bisector.
Let ${\bf z}$ be a point of $B(-{\bf x},{\bf x})$. If it is an ordinary point, then there is a unique
value $1 < t_{\bf z} < \infty$ for which ${\bf z}\in (t_{\bf z}S + x) \cap (t_{\bf z}S-x)$. Let $\Phi : B(-{\bf x},{\bf x})\longrightarrow K$ denote
the mapping which sends {\bf z} into $\Phi({\bf z}) = \frac{1}{t_{\bf z}}{\bf z}$. For ordinary points the mapping $\Phi $ of the bisector
is continuous. We extend $\Phi $ to the ideal points by the following rule: The image of an ideal point is its radial projection. Denote the
image set of $\Phi $ (with respect to this extended mapping) by $\Phi(B(-{\bf x},{\bf x}))$. We will call this set \emph{the bounded representation of
the bisector}. We now have a connection between the concept of bisector and metrical properties of $K$.
\begin{lemma}
The bounded representation of the bisector is the union of the shadow boundary of $K$ and the locus of the midpoints of the chords of K parallel to {\bf x}.
\end{lemma}
\noindent {\bf Proof: }
For an ordinary point {\bf z} of the bisector we have $1 \leq t_{\bf z} < \infty$, and thus the norm of
$$
\frac{1}{t_{\bf z}}{\bf z} =\frac{1}{2}\left( \frac{1}{t_{\bf z}}({\bf z}-{\bf x}) +\frac{1}{t_{\bf z}}
({\bf z} + {\bf x})\right)
$$
is less or equal to 1.
If it is equal to 1, then the point $\frac{1}{t_{\bf z}}{\bf z}$ is a point of a horizontal segment (parallel to {\bf x}) of the boundary and thus a point of the shadow boundary, and the set of all points corresponding to the value $t_{\bf z}$ yields a horizontal segment of $S$. If now $t\geq t_{\bf z}$, the points of the bounded representation corresponding to this value $t$ form
another segment containing the segment of $t_{\bf z}$. Thus the directions determined by the points of the segment of $t_{\bf z}$ are ideal points of the bisector, proving that the points of the shadow boundary are images of certain ideal points.
\begin{figure}
\centerline{\includegraphics[scale=0.4]{rp1.eps}}
\caption{Bounded representation of the bisector}
\end{figure}
In the other case the obtained point is the midpoint of that chord whose endpoints are $\frac{1}{t_{\bf z}}({\bf z}-{\bf x})\in S$ and $\frac{1}{t_{\bf z}}({\bf z}+{\bf x})\in S$, respectively.
Now, by the definition of ideal points, the continuity of the mapping is clear. In fact, we have to check that the image of a point of the bisector with large norm is close to the boundary $S$ of $K$. Since, by definition, $t_{\bf z}$ is equal to $\|{\bf z}-{\bf x}\|$, we have the two inequalities
$$
1\geq \|\frac{1}{t_{\bf z}}{\bf z}\|=\frac{\|{\bf z}\|}{\|{\bf z}-{\bf x}\|}=\frac{1}{\|\frac{{\bf z}}{\|{\bf z}\|}-\frac{{\bf x}}{\|{\bf z}\|}\|}\geq \frac{1}{1+\frac{\|{\bf x}\|}{\|{\bf z}\|}},
$$
showing that for ${\bf z}$ with large norm its bounded representation is close to $S$. To visualize the proof, we show in Fig. 2 the bisector and its bounded representation in a
two-dimensional space.
\hfill $\Box$
For example, by \cite{james2} we get from this lemma immediately the following
\begin{corollary}
{\em The bounded representation of the bisector $B({\bf x},-{\bf x})$ with respect to any point {\bf x} from the unit sphere of a Minkowski space is contained in an $(n-1)$-subspace if and only if the Minkowski space is Euclidean.}
\end{corollary}
It is easy to see that the shadow boundary $S(K,{\bf {\bf x}})$ is the set of points of $S$ which are orthogonal to the vector ${\bf x}$ in the sense of Birkhoff, and thus some results on Birkhoff orthogonality in \cite{martini-wu} can be extended to higher dimensions. These are described in the following.
Theorem 2.6 in \cite{martini-wu} says that the shadow boundary is a subset of the closure of the radial projection. A consequence of the concept of bounded representation of the bisector is the fact that such theorems can be extended to higher dimensions. To prove this, we observe that the radial projection is exactly the radial projection of the bounded representation of the bisector, implied by Lemma 1. So it contains the shadow boundary, extending Theorem 2.6 of \cite{martini-wu}.
Theorem 2.9 in \cite{martini-wu} refers to the converse statement. If for a point ${\bf x}$ of $S$ there exists a unique point ${\bf z}$ (except for
the sign) such that ${\bf x}$ is orthogonal to ${\bf z}$ in the sense of Birkhoff, then ${\bf z}$ is a point of the radial projection of the bisector corresponding to ${\bf x}$. If we denote the sharp points of the shadow boundary as those points of $K$ which are unique in their carrying supporting line of $K$ parallel to ${\bf x}$, we can say that the sharp points of the shadow boundary corresponding to the direction of ${\bf x}$ belong to the radial projection of the bisector of ${\bf x}$. This is also a consequence of our present definitions and Lemma 1.
Let now $n=2$, and assume that ${\bf z}$ is the radial projection of an ideal point of the bisector of ${\bf x}$. Then it belongs to the shadow boundary of ${\bf x}$ implying that it is Birkhoff orthogonal to ${\bf x}$. Thus, if we consider a boundary point {\bf z} of the radial projection which is not Birkhoff orthogonal to ${\bf x}$, this is a projective image of an ordinary point of the bisector. Thus ${\bf z}$ is a point of the radial projection in the classical sense, too. This proves Theorem 2.10 in \cite{martini-wu}
The main aim of this paper is to prove the following theorem on the topology of the bounded representation of the bisector.
\begin{theorem}
If the bisector is a manifold of dimension $n-1$ with boundary, then its bounded representation is homeomorphic to the $(n-1)$-dimensional closed ball $B^{n-1}$. Conversely, if the bounded representation is a topological ball of dimension $n-1$, then the bisector is of the same type. Furthermore, its relative interior (which is the set of its ordinary points) is a topological hyperplane of dimension $n-1$.
\end{theorem}
\noindent {\bf Proof: } Assume that the bisector is a manifold of dimension $n-1$ with boundary. Then an ordinary point has a relatively open $(n-1)$-dimensional neighborhood in the bisector, and thus there are interior points. On the other hand, there is no ideal point which could be in the relative interior of the bisector implying that the set of ordinary points of the bisector is a manifold of dimension $n-1$. Hence our assumption implies that the shadow boundary $S(K,{\bf x})$ is a manifold of dimension $n-2$. In fact, from Theorem 5 and Theorem 4 in \cite{gho3} we get that the shadow boundary is also a topological manifold of dimension $n-2$. Theorem 2 says that it is homeomorphic to $S^{n-2}$. On the other hand, the set $C$ of midpoints of correspondingly directed chords containing interior points of $K$ is always homeomorphic to the positive part $S^+$ of the boundary $S$ of $K$, determined by the shadow boundary. Thus it is homeomorphic to $\mathbb{R}^{n-1}$. Finally we observe that the boundary of the latter set $C$ is the shadow boundary itself, showing that the bounded representation of the bisector is homeomorphic to $B^{n-1}$, as we stated.
We remark that the converse statement is true if and only if the manifold property of the bounded representation can be extended to the bisector. This is clear for the points mapping to the interior of $K$, but it is not evident for other points of the bisector. The problem is that the pre-images of a point of the shadow boundary could form a point or a half-line, respectively. Thus $\Phi $ is not an injective (but, of course, a surjective) continuous mapping. Clearly, both of the two sets (the bisector and its bounded representation) are continua, i.e., compact, connected Hausdorff ($T_2$) spaces. Moreover, the points and half-lines are cell-like sets; thus $\Phi$ is a cell-like mapping. Restricting $\Phi$ to the ideal point of the bisector, we get a bijective mapping onto the shadow boundary. We prove that the set of ideal points is compact in the bisector. Of course, the ideal points of the bisector give a proper part $I$ of $S^{n-1}$ bounding the topological ball $B^n$. Hence this point set can be regarded as a subset of an $(n-1)$-dimensional Euclidean space $\mathbb{R}^{n-1}$. (We can consider ${\bf x}_\infty$ as the center of a stereographic projection.) Its clear that $I$ is bounded. It is also closed by its definition, and so it is compact by the Heine-Borel theorem on compact sets in $\mathbb{R}^{n-1}$ (see, e.g., p. 9 in \cite{knopp}). On the other hand, the shadow boundary can also be regarded as an $(n-2)$-sphere embedded into a Euclidean $(n-1)$-space, because ${\bf x}$ is not a point of it. A continuous and bijective mapping from a compact set of $\mathbb{R}^{n-1}$ into $\mathbb{R}^{n-1}$ is a homeomorphism (see again \cite{knopp}). Thus the ideal points of the bisector give a topological $(n-2)$-dimensional sphere.
Now we prove that the ordinary points of the bisector are, with respect to its relative topology, interior points of it. We remark that it is trivial for a point ${\bf z}\in B(-{\bf x},{\bf x})$ if $\Phi ({\bf z})$ is an interior point of $K$, because $\Phi $ (by its definition) is a homeomorphism on the collection of such points onto the interior of the bounded representation of the bisector. Thus it is also relatively open with respect to the bisector, and this part of the bisector is a topological manifold, homeomorphic to $\mathbb{R}^{n-1}$.
Let now $\Phi ({\bf z})$ belong to the shadow boundary. Since it is a topological sphere of dimension $n-2$, there is a cell of dimension $n-2$ (a homeomorphic copy of a closed ball of dimension $n-2$), namely $Z$, containing $\Phi ({\bf z})$ in its interior. The pre-image $\Phi^{-1}(\mbox{int }B)$ of the interior $\mbox{int }B$ of $B$ is (by the continuity of $\Phi $) open with respect to the topology of the bisector and contains ${\bf z}$. Thus it has also an interior point with respect to the topology of the bisector.
Finally we observe that from the compactness of $B$ the existence of an $\varepsilon$ follows for which the set
$$
\{{\bf v} \mbox{ : } \|{\bf z}\|-\varepsilon \leq\|{\bf v}\| \leq \|{\bf z}\|-\varepsilon \mbox{, }{\bf v}\in \Phi^{-1}(B)\}
$$
is a closed cone (truncated by two parallel surfaces) containing ${\bf z}$ in its interior. Since the interior of this body is homeomorphic to $\mathbb{R}^{n-1}$, we get that the set of ordinary points is a manifold of dimension $n-1$. In the proof of Theorem 5 from \cite{gho3} it is shown that if the ordinary points of the bisector yield an $(n-1)$-manifold, then it is homeomorphic to $\mathbb{R}^{n-1}$, and Theorem 6 there establishes that it is a topological hyperplane. Thus we proved that the closed bisector is a cell of dimension $n-1$ whose interior can be embedded in the $n$-dimensional Euclidean space in a standard (unknotted) way, as we stated.
\hfill $\Box$
In higher dimensions the topology of the radial projection does (in contrast to the bounded representation) not simplify by the simplification of the topology of the bisector. Finally we give an example showing that it is possible that the bisector is a manifold with boundary of dimension $n-1$, but $P({\bf x})$ is not a manifold.
\begin{figure}
\centerline{\includegraphics[scale=0.8]{rp2.eps}}
\caption{The radial projection is not a manifold}
\end{figure}
\begin{example} Take a cartesian coordinate system in the Euclidean space, with the respective coordinate-axes $x,y$ and $z$, and vectors ${\bf x}$ and ${\bf z}$ having coordinates $(1,0,0)$ and $(0,0,1)$, respectively. Consider the example from \cite{gho1} in which the body $K$ is the convex hull of two half-circles with parallel diameters and in symmetric position with respect to the origin, such that their affine hulls are parallel to the plane $x=0$ (see Fig. 4). Besides the two half circles, the ruled part of the boundary of their convex hull contains four conic surface parts and two opposite triangles. The bounded representation of the bisector corresponding to the direction orthogonal to the planes of the half-circles is homeomorphic to a plane. It can be obtained in the following way: Cut the surface of $K$ by the segments parallel to ${\bf x}$ into two parts. The described half-disks do no longer belong to the surface. Apply an affinity to these two parts, by the ratio $\frac{1}{2}$, with direction orthogonal to the end planes. Finally glue these parts together at their common vertical segment $[-{\bf z},{\bf z}]$. It is clear that the central projection of this ruled surface from the midpoint of $[-{\bf z},{\bf z}]$ is the union of those curves which are the intersections of the planes through $[-{\bf z},{\bf z}]$ and the respective points of $S$. The obtained four parts are conic surfaces separated by $-{\bf z}$, ${\bf z}$ and the two opposite points of the half-circles lying on the horizontal plane $z=0$. Of course, in this case the radial projection is not a topological manifold.
\end{example}
|
\section{Introduction}
New X-ray data from high-throughput missions such as
{\it XMM-Newton} and {\it Suzaku}, along with the high spectral
resolution afforded by the {\it Chandra}
High Energy Transmission Grating
(HETG) has led to the discovery of several
important spectral signatures in Active Galactic Nuclei (AGN) that
were not detectable using previous missions.
A joint {\it XMM-Newton}/{\it Chandra} observation of NGC 3516
\citep{turner02a} revealed the first detection of narrow line emission
at 5.6 and 6.2\,keV (the latter was able to be resolved from the
strong Fe K$\alpha$ line owing to the relatively high spectral
resolution afforded by the HEG grating). While the line energies
matched those expected for ionized species of Cr and Mn, the line
strengths exceeded what might be expected from illumination of
material with cosmic abundance ratios. One explanation considered was
enhancement of Cr and Mn abundances from spallation of Fe. However,
this was initially disfavored as the line ratios did not agree well
with those predicted by \citet{skibo97} for spallation of disk gas
\citep{turner02a}. The alternative possibility, of Doppler-shifted
components of Fe emission, led to a suggestion of emitting hotspots on
the accretion disk surface - with an expectation of observable shifts
in line flux and energy as the hotspot traverses its orbit. Further
examples of the phenomenon, dubbed `transient Fe lines', soon came
from observations of other AGN, with many lines reported in the
5-6\,keV regime \citep[e.g.][]{yaqoob03a,turner04a}.
In this paper we report the highly significant detection of an emission
line at 5.44\,keV in the rest-frame of the nearby
narrow line Seyfert 1 type AGN
NGC\,4051. The systemic redshift of the host galaxy is
$z=0.0023$ that yields a
distance 9.3\,Mpc for $H_0=\,74$\,km\,s$^{-1}$\,Mpc$^{-1}$ assuming
the redshift
to arise from the Hubble flow. However, for such nearby galaxies
a more reliable estimation of distance can be obtained
from use of the Tully-Fisher relation, giving
a distance of 15.2\,Mpc \citep{russell04a} in this case,
which we adopt in this paper.
NGC\,4051 was observed by {\em Suzaku}\ in 2005 and 2008. The 2005 observations
have previously been described and analyzed by \citet{terashima09a}, who showed
that the X-ray source was highly variable on short timescales, and that
a spectral model including variable
partial-covering absorption was required to fit the data.
Here we make a joint analysis of the 2005 observation together with the new
2008 data in which we concentrate specifically on the detection of the narrow emission
line and its variability. We then discuss the possible origin of the line in the
context of several popular models.
\section{Observations}
The {\em Suzaku}\ X-ray Imaging Spectrometer \citep[XIS][]{koyama07}
instrument comprises four X-ray telescopes \citep{mitsuda07} each
with a CCD in the focal plane. XIS CCDs 0,2,3 are configured to be
front-illuminated (FI) and provide useful data over $\sim 0.6-10.0$\,keV
with energy resolution FWHM $\sim 150\,$eV at 6\,keV. XIS 1 is a
back-illuminated CCD and has an enhanced soft-band response (down to
0.2\,keV) but lower area at 6\,keV than the FI CCDs as well as a larger
background level at high energies, consequently this detector was not used in our analysis.
{\em Suzaku}\ also carries a non-imaging, collimated Hard X-ray Detector \citep[HXD][]{takahashi07}
whose PIN instrument provides useful data over 15-70\,keV for bright AGN.
Our analysis used all available {\it Suzaku} observations of NGC 4051
comprising data from 2005 Nov 10-13 (observation identifier
700004010) and 2008 Nov 6-12 (703023010) and 23-25 (703023020)
as summarized in Table~1.
The data were reduced using v6.4.1 of {\sc HEAsoft} and screened to
exclude: i) periods during and within 500 seconds of the South Atlantic
Anomaly (SAA), ii) with an Earth elevation angle less than 10$^\circ$ and
iii) with cut-off rigidity $>6$ GeV. The source was observed at the nominal
center position for the XIS during 2008 and at the nominal center position for the
HXD during 2005. The FI CCDs were in $3 \times 3$ and $5
\times 5$ edit-modes, with normal clocking mode. For the XIS we
selected events with grades 0,2,3,4, and 6 and removed hot and
flickering pixels using the SISCLEAN script. The spaced-row charge
injection (SCI) was used. The XIS products were extracted from
circular regions of 2.9\arcmin \, radius with background spectra from
a region of the same size, offset from the source (avoiding the
calibration sources at the edges of the chips). The response and
ancillary response files were created using {\sc xisrmfgen v2007 May}
and {\sc xissimarfgen v2008 Mar}.
NGC 4051 is too faint to be detected in the HXD GSO instrument, but
was detectable in the PIN. For the analysis we used the model ``D''
background \citep{fukazawa09a}. As the PIN background rate is strongly variable around
the orbit, we first selected source data to discard events within
500s of an SAA passage, we also rejected events with day/night elevation
angles $> 5^\circ$. The time filter resulting from the screening
was then applied to the background events model file to give PIN
model-background data for the same time intervals covered by the
on-source data. As the background events file was generated using ten
times the actual background count rate, an adjustment to the
background spectrum was applied to account for this factor.
{\sc hxddtcor v2007 May} was run to apply the deadtime correction to the
source spectrum. To take into account the cosmic X-ray background
\citep{boldt87,gruber99} {\sc xspec} v 11.3.2ag was used to generate a
spectrum from a CXB model \citep{gruber99} normalized to the $34^\prime
\times 34^\prime$ {\it Suzaku} PIN field of view, and combined with the PIN
instrument background file to create a total background file. The mean
exposure times for XIS are given in Table~1. The exposure times for the PIN were
112, 204 and 59\,ks for 2005 Nov 10, 2008 Nov 6 and 2008 Nov 23, respectively.
Spectral fits used data from {\sc XIS} 0, 2 and 3, for the 2005 data.
As use of XIS2 was discontinued after a charge leak was discovered
in Nov 2006, the 2008 analysis used only XIS 0 and 3. In this paper,
XIS data were
fit over $3.0-10$\,keV. PIN data were fit
simultaneously with the XIS, in the range 15-50\,keV.
In the spectral analysis, the {\sc pin} model flux was increased by a factor
1.16 for 2008 data and 1.18 for 2005 data, which are the appropriate adjustments for the instrument
cross-calibration at those epochs of the observation.
XIS data were binned at the HWHM instrumental resolution while PIN data were binned
to be a minimum of 5$\sigma$ above the background level for the
spectral fitting.
\section{Spectral Fitting}
\label{sec:spectralfitting}
The mean {\em Suzaku}\ count rate over 0.5-10\,keV during the low state of 2005 was
0.45 XIS count\,s$^{-1}$ per FI XIS and 0.04 PIN count\,s$^{-1}$. During the
high-state of 2008 Nov 6 it was 2.06 XIS count\,s$^{-1}$ per FI XIS
and 0.07 PIN count\,s$^{-1}$ (Table~1). The count rates recorded correspond to an observed
2-10\,keV flux range $0.87- 2.42 \times 10^{-11} {\rm erg\, cm^{-2} s^{-1}}$
and a 10-50\,keV flux range $2.44 - 3.63 \times 10^{-11} {\rm erg\, cm^{-2} s^{-1}}$.
The 2005 data represent a historically low state, as noted by
\citet{terashima09a} and during 2008 the source was close to the
historical average flux level. The 2008 Nov 6 {\it Suzaku} observation was
accompanied by a contemporaneous HETG exposure that reveals a wealth of emission and
absorption features from several zones of ionized gas. The HETG
analysis will be described in detail in another paper \citep{lobban10a}.
\subsection{Fitting the PCA Offset Component}
To assist in modeling the full-band {\it Suzaku} data, we first
performed a decomposition of the events from all three {\it Suzaku}
observations using Principal Components Analysis using the SVD method
and code of \citet{miller07a}. A full description of the PCA
decomposition of these data is reported by \citet{miller09b}. In summary, PCA is a mathematical decomposition
of the data into orthogonal 'eigenvectors'. The dominant variations
are ascribed to eigenvectors of the lowest orders.
In the case of NGC~4051 the data were found to be
well-described by a steady ``offset''
component having a hard spectrum (Figure ~\ref{fig:pca} and see \citealt{miller09b}) while the
first order variable component, eigenvector 1, is
consistent with an absorbed power-law of
constant slope whose variations in intensity dominate the spectral
variability of the source (consistent with the analysis of
\citealt{terashima09a}). The physical origin of the offset component is
of great interest. The low-state spectrum of NGC 4051 is composed of
this component plus some contribution by eigenvector 1. Fitting the
lowest flux subset of the 2005 data, \citealt{terashima09a} found the
hard component to be best described using a partially-covered
reflection component. Our analysis of the PCA offset component
provides another way to probe the lowest flux levels of the source
and, while we find an acceptable fit using a combination of reflection
(modeled using {\sc pexrav}) plus a contribution from the powerlaw
continuum ($\Gamma=2.3$, fixed from fitting eigenvector\,1), both
emission components require absorption by a complex of ionized
gas. Fortunately the multiple layers of ionized gas can be well
constrained using HETG and LETG data \citep{collinge01a,steenbrugge09a,lobban10a},
reducing the
degeneracy of the {\it Suzaku} fit. In addition to these components,
Fe\,K$\alpha$ line emission is evident in the offset spectrum at a
fitted energy $E=6.398\pm0.023$\,keV having a width
$\sigma=0.045^{+0.037}_{-0.045}$\,keV and flux $n=1.90 \pm 0.26 \times
10^{-5}$ photons cm$^{-2}\, {\rm s}^{-1}$. (Throughout this paper,
errors are quoted at 90\% confidence for the appropriate number of
interesting parameters in the fit.) The resulting fit statistic was
$\chi^2=130/153\, d.o.f$.
Intriguingly, the normalization of the reflection continuum
is extremely high relative to that of the primary continuum. The
inconsistency of reflection and continuum strengths may
indicate that the partial-covering absorption comprises a larger
part of this component than currently modeled. The relatively high
flux of the hard component was also noted by \citet{terashima09a}
based upon fitting the 2005 data, and is fully explored by \citet{lobban10a}.
In addition to the prominent narrow Fe\,K line, a strong excess of
line emission is evident in the PCA offset spectrum, at 5.44 keV
(Figure~\ref{fig:pca}). No significant residuals
appear in eigenvector 1 at that energy \citep{miller09b}. The new line is thus consistent
with having an origin
associated with the hard offset component and neutral component of Fe\,K$\alpha$ emission.
\subsection{Fitting the 2008 Contemporaneous Chandra and Suzaku data}
\label{sec:hetg_suz}
As the 2008 Nov 6 epoch provides contemporaneous HEG data for the {\it Suzaku} observation, we
fit those spectra together over 3-8\,keV, combining the superior energy-resolution of HEG
and the good statistical quality of the FI XIS units (0,3) to obtain the best possible
constraints on the width and energy of the neutral component of
Fe\,K$\alpha$. A joint fit to those data using a simple powerlaw plus single Gaussian line
showed residual excesses in the 5 - 6 keV band in {\it Suzaku} XIS data
(Figure~\ref{fig:cts_res2008}).
We refit, adding to the model a single component of absorption, a Gaussian representation of the
K$\beta$ line fixed at a rest-energy of 7.05\,keV with a flux linked to be
13.5\% of the K$\alpha$ component \citep[e.g.][]{leahy93a,palmeri03a},
a narrow absorption line detected
at an observed energy of 7.1\,keV (as discovered by
\citealt{pounds04c} and confirmed by \citealt{terashima09a}) plus a component to model
the narrow line emission evident at 6.62\,keV in HEG data \citep{lobban10a}.
The fit yielded
a line energy
$E=6.410 \pm 0.015$\,keV, width $\sigma=50^{+26}_{-33}$ eV and normalization
$n=1.90\pm 0.395 \times 10^{-5} {\rm photons\, cm^{-2}\, s^{-1}}$.
The equivalent width for the Fe\,K$\alpha$
line component is 95\,eV against the total continuum level observed
during 2008.
The constraint on
the width of the Fe K$\alpha$ line is equivalent to a FWHM velocity-broadening of
$5540_{-3664}^{+2846}$\,km\,s$^{-1}$ (90\,percent confidence interval).
Assuming the true line centroid energy to be 6.40\,keV, the
90\% confidence constraint
on energy from the combined HEG/XIS data
limits the bulk velocity of this emitter to
$250 \ga v \ga -1200$\,km\,s$^{-1}$ where negative velocity denotes outflow.
As the PCA decomposition indicates the presence of a weak broad component of
Fe K$\alpha$ on eigenvector 1 \citep{miller09b},
we tried adding a second component of Fe K$\alpha$ emission to the model,
with the energies of both line components linked. This addition improved the fit-statistic by
$\Delta \chi^2=5$, giving
the same line
energy as for the single component model, with $\sigma_1 =1^{+23}_{-1}$ eV and normalization
$n_1=1.12\pm 0.32 \times 10^{-5} {\rm photons\, cm^{-2}\, s^{-1}}$ (EW$=41$ eV),
$\sigma_2 =145^{+70}_{-52}$ eV and normalization
$n_2=1.07\pm 0.45 \times 10^{-5} {\rm photons\, cm^{-2}\, s^{-1}}$ (EW$=43$ eV).
Although the two-component parameterization of Fe\,K$\alpha$
offers only a marginal improvement to the fit, it is of interest with regard to consistency with the
PCA decomposition, which indicates the weak broad component of Fe K$\alpha$
emission to be present on eigenvector 1.
In subsequent fits we fix the Fe K$\alpha$ model line width at
$\sigma=50$ eV obtained from the simple fit,
while bearing in mind the possible more complex solution for the line profile.
As the PCA is consistent with a common origin for Fe K$\alpha$ emission and
the line at 5.44 keV, and as the widths of the two lines are consistent (and as there is no
other useful constraint available for the indeterminate line width) hereafter we fixed
both to $\sigma=50$ eV in spectral fitting.
\subsection{Fitting the 2005 Suzaku data}
\label{sec:linesig}
To assess the significance of the line at 5.44 keV we returned to direct
fitting of the 2005 data, where the source was observed to be at a low
flux and where PCA indicates that the line would have the highest
equivalent width. We fit the summed 2005 data from the FI XIS units
0, 2 and 3 using an absorbed continuum model, applying it to the
3-10\,keV band for the purpose of examining the data for features
of interest. A Gaussian line was included in the fit, to account for
the Fe K$\alpha$ emission, fixed at the energy and width found using
HEG/XIS. As evident in Figure ~\ref{fig:cts_res} the mean spectrum
from 2005 data shows an excess of counts at 5.44\,keV with additional structure
evident around 6 \,keV. We performed a more complex
fit to properly assess the line strengths and significance.
We added to the model a Gaussian representation of the
Fe K$\beta$ line fixed at a rest-energy of 7.05\,keV with a flux linked to be
13.5\% of the K$\alpha$ component. The model also included
the narrow absorption line detected
at an observed energy of 7.1\,keV as discovered by
\citet{pounds04c} and confirmed by \citet{terashima09a}, plus a component to model
the narrow line emission evident at 6.62\,keV \citep{lobban10a}.
The fit yielded $\chi^2=176/101\, dof$ and
the strongest unmodeled feature remains
evident as an excess of counts at 5.44\, keV.
Addition of a Gaussian line ($\sigma=50$\, eV) reduced the fit-statistic
by $\Delta \chi^2= 32$, yielding
$E=5.44 \pm0.03$\,keV, flux $n=5.03^{+2.02}_{-2.01} \times 10^{-6}$
photons cm$^{-2}{\rm s^{-1}}$ and equivalent width
$46 \pm 16$ eV. Addition of a second line of the same width
yielded a further improvement
$\Delta \chi^2= 34$, $E=5.95 \pm 0.05$\,keV, line flux
$n=5.05^{+2.05}_{-1.95} \times 10^{-6}$ photons cm$^{-2}{\rm s^{-1}}$,
with an equivalent width of $44^{+17}_{-16} $ eV and a final fit statistic
$\chi^2=117/97\, d.o.f.$
In this fit the flux and equivalent width of the Fe K$\alpha$ line were
$n=1.59 \pm 0.23 \times 10^{-5}$ and $195 \pm 24$ eV respectively.
To further examine the source behavior at low flux levels we
isolated the lowest flux subset of the 2005 data using an
intensity filter based upon the 0.5-10 keV band count rate,
taking data below a threshold of 0.47 ct s$^{-1}$ (per FI XIS).
The intensity cut effectively removed the periods of
source flaring from consideration. This filtering thus resulted in a
very-low-state spectrum that had a mean 2--10\,keV flux
$6.1 \times 10^{-12} {\rm erg\, cm^{-2} s^{-1}}$. Again, fitting a simple
absorbed continuum model over 3--10\,keV leaves two strong
positive residual features between 5--6\,keV (Figure ~\ref{fig:vloeeuf})
confirming the prominence of the 5.44\, keV line at low flux levels and indicating the possible presence of
emission at 5.95\, keV.
\subsection{The Full Model}
To assess the full properties of the source and investigate the robustness of the line detections
to the continuum model used, we then fit over 0.75-50 keV (the full band of the {\it Suzaku} data).
We simultaneously fit
spectra from the three {\it Suzaku} observations. The 2008
HETG exposure (that overlapped the 2008 Nov 6 {\it Suzaku}
observation) shows a wealth of absorption and emission lines that can
be modeled using three zones of ionized gas \citep{lobban10a}: those
soft-band absorbers
were fixed in the {\it Suzaku} fit, after which two
additional gas layers were required to accurately model the source
across the full {\it Suzaku} bandpass. The fit also included an Fe K$\alpha$ emission
line fixed at a width of $\sigma=50$\, eV as before.
Finally, an absorbed ionized
reflector was included to account for the curvature in the soft band.
The full model is detailed in Table~2 and the fit is illustrated in
Figure~\ref{fullspec}. The marked spectral
variability observed (Figure~\ref{fullspec}) can be accounted for by
allowing variations in the covering fraction of one absorbing layer
(having $N_H \sim 10^{23} {\rm cm^{-2}}, {\rm log}\, \xi \sim 0.18$),
a solution that is similar to other well-studied AGN of this class
\citep[e.g.][]{pounds04c,risaliti07a,miller08a,turner08a}. This fit
yielded $\chi^2=947/547\, d.o.f.$ with the overall curvature modeled
well across all epochs and the residual $\chi^2$ contributed mainly by
some unmodeled spectral features. The addition of a line ($\sigma=50$\, eV)
constrained to lie in the range 3 - 8 keV yielded a detection at 5.44$\pm0.03$ keV
and improved the fit statistic by $\Delta \chi^2=20$; addition of another
line under the same constraints gave 5.98$\pm0.05$ keV and
$\Delta \chi^2=9$ improvement to the fit.
We conclude that the high significance for the detection of the 5.44\, keV line is robust to the
continuum used while the significance of the detection of a line at 5.95 keV is sensitive to
the continuum form assumed.
\subsection{A Critical Examination of the Reality of the Lines}
Before proceeding any further with the modeling,
we performed several checks to be assured that the new line is
attributable to the AGN and is significant.
First, to confirm that the line is not an artifact of poor background subtraction,
we examined the background spectrum and calibration source data.
First we note that the background comprises just 2.5\% of the total count
rate in the 5-7 keV band for the XIS
spectral data when the source is at its lowest flux level, during 2005 and
1.3\% when the source is brighter, during 2008.
Regarding the line at 5.44\,keV, we found there to be no
line emission evident at that energy in the background spectrum. Fitting the background spectra from
the three {\it Suzaku} observations we obtained an upper limit (90\% confidence) for the
flux of a line at 5.44\, keV in the background spectrum
$n < 4.14 \times 10^{-8}$ photons cm$^{-2}{\rm s^{-1}}$, i.e. $< 1$\% of the detected line flux.
The lack of a feature at comparable
flux or equivalent width in the background data also rules out an origin of the 5.44\,keV line
as a detector feature. We then examined each XIS independently, and found the 5.44\,keV
line to be significantly detected in each XIS unit independently, further verifying the reality of
the reported feature (Figure ~\ref{fig:in3xis}).
Regarding the more tentative line at 5.95\,keV: the XIS chip does include a Mn calibration source
emitting a line at 5.9\,keV for the purpose of calibration of the XIS energy scale. The calibration
source is located at the chip edge and the counts from that source are obviously
conservatively excluded from any source and background
extraction cells for scientific analysis. However, we do detect a weak
Mn K$\alpha$ emission in the background spectrum. The Mn contamination is
at a level $< 10\%$ of the measured line strength in the AGN spectrum. As the source
extraction cell is further from the Mn calibration source than the background cell and
therefore subject to less contamination,
we estimate the Mn line contamination from the calibration source to be $< 10$\% for the 2005 spectrum.
In spectral fitting, the greatest concern for the Mn line is the possible contamination by the weak broadened Fe K$\alpha$
emission evident in the fit to eigenvector 1 \citep{miller09b} which may lead to the strength of the Mn
line being over-estimated from simple fits. Because of the various
issues associated with a clean measurement of any line
at 5.95\, keV and the sensitivity to the continuum form assumed
we concentrate only on the strong line detection at 5.44\, keV.
To confirm the significance of the line
we performed Monte Carlo simulations using the method described in
\citet{porquet04a} and in \citet{markowitz06a}. We took the
null hypothesis to be that the
spectrum is simply an absorbed power-law continuum with
parameters derived from fitting the broad-band data but allowing for the
statistical uncertainty on the continuum parameters and including the
narrow Fe K$\alpha$ line
whose presence is well-established in this source
\citep{lobban10a}.
We used the {\sc xspec} command {\it fakeit} to create 3000 fake
{\it Suzaku} spectra with photon statistics expected
from the 2005 exposure, assuming the same instruments to be operational as
for the actual observation. The simulated data were grouped to the
HWHM energy-resolution of the instruments, the same as the observational data.
Following the procedure used to test the real data for
the presence of a narrow line, we fitted each fake spectrum to obtain the
values of $\Delta \chi^2$ obtained from statistical fluctuations
in the data.
To map the distribution of $\Delta \chi^2$ across
the simulated spectra, each simulated spectrum was fitted over the
3-10 keV energy range, stepping through using energy bins whose centers were increased
in increments
of 100 eV. The line energy was allowed to be free within each energy bin tested
and the value of $\Delta \chi^2$ was recorded at each point in
the spectrum. This method makes no assumptions about the energy at which a line
might be detected and over the course of the testing, all energies are tested for the presence
of a line.
When we fit the simulated data we are testing whether we
can produce, from statistical fluctuations, a contribution to $\chi^2$ at the same or greater level
as found in the actual data at any energy in the range of interest.
The fits to the simulated data yield a
distribution of $\Delta \chi^2$ for
comparison with the actual data.
We found the most extreme statistical fluctuation to
yield a $\Delta \chi^2$ contribution of 19.3 in the 3-10\,keV band
in the set of simulated data (i.e. in 3000 simulations no false line appears at any energy contributing
$\Delta \chi^2=32$ as found in 2005 data).
Thus the probability of satisfying the null hypothesis is
$p < 3.3 \times 10^{-4}$ for the line found at 5.44\, keV.
\subsection{Application of a Disk Hotspot Model}
As an alternative to modeling using individual Gaussian
lines, we fitted the data using a single disk line from a narrow annulus, such that the red horn of such a line
might explain the peak at 5.44\, keV. We assumed
the system to contain a non-rotating black hole, that the line is Fe K$\alpha$ emission from neutral
material (at 6.4\,keV), that the emissivity pattern across the disk can be described
by $r^{-q}$ where the emissivity index q=-2.5 and that the hotspot
exists over a narrow annulus of width $\Delta r=1r_g$. We
used the same baseline model as for testing the Gaussian lines
(section~\ref{sec:linesig}).
To fit the 5.44\, keV peak as the red Doppler horn of a disk-line
requires an emitting radius 21$\pm 3 r_g$
(where $r_g = GM_{BH}/c^2$ denotes gravitational radii)
in a low inclination system
with ${27^{+1}_{-3}}^\circ$ with
$\chi^2=128/95\, d.o.f$.; the corresponding blue horn is
then predicted to lie at 6.62\,keV
and the data are consistent with that model.
While there is evidence for line emission in the 6.5-7\,keV regime, such lines
are commonly observed as emission from ionized species of Fe,
and thus the identification of
emission blue-ward of 6.4\,keV is currently ambiguous.
\subsection{Examination of the Line Variability}
\label{sec:linevar}
Fits to the mean spectra from each observation have shown consistent
line fluxes for Fe K$\alpha$ and the line at 5.44\, keV
and these show an equivalent width that appears to have
changed as the source flux varied.
Figure~\ref{fig:hiloratio} shows the ratio of data in the Fe K$\alpha$ regime to a common local
continuum fit; the fact that the source spectrum is steeper at high flux
is reflected in the systematics of the residuals. To confirm the
significance of the change in equivalent width we fit the 2008 data with a model that fixed the
line equivalent width for the feature at 5.44\,keV to that found during 2005;
after refitting this resulted in a worse fit with $\Delta \chi^2=142$.
Alternatively, fixing the flux of the new lines at the 2005 values and refitting,
we found $\Delta \chi^2=0$, indicating that the line fluxes may be consistent
with lines of constant flux across the data.
As the limits on line flux provide the potential to
distinguish between models, we tested all available high-quality data
in a self-consistent manner. Taking the model from
section~\ref{sec:linesig} we fixed the line energy and width to
specifically test the constancy of the 5.44\, keV line.
In addition to testing the mean spectrum from each {\it Suzaku}
observation, we sub-divided the 2005 and 2008 exposures to sample the
line more finely in flux.
For 2005 an intensity selection was made on the 0.5-10 keV count rate
at 0.47 ct s$^{-1}$ per XIS as before. For 2008 intensity
selections were made $ < 2.0 $ ct s$^{-1}$/XIS ({\it i1}), 2.0 - 3.0 ct s$^{-1}$/XIS ({\it i2}) and
$ > 3.0$ ct s$^{-1}$/XIS ({\it i3}). The results are shown in Table~3
where both the mean fits for each observation are tabulated, as well as the
intensity-selected results. As noted previously, the line is required
at a high level of confidence in
the 2008 Nov 6 data (which is more sensitive to the features than the
Nov 23 data, owing to the long exposure and high flux state of the source).
Table~3 shows line flux along with improvements in the fit-statistic and
equivalent width for each fit. Note that the mean observation
fits in Table~3 show
slightly tighter constraints in line flux for 2005 compared to the values noted
in section~\ref{sec:linesig} because the initial fits to 2005 data had the
line energy left free.
The data are consistent with the 5.44\, keV line existing
at the same flux level throughout the {\it Suzaku} observations considered in Table~3 and the high significance
of the line detection across several time slices of data provides compelling evidence for the reality of
the line; conclusively
ruling out the possibility of the line detection being a statistical fluctuation in the spectral data.
Considering the three independent {\it Suzaku} detections of the line, as shown in Table~3 {\it Suzaku}
observations 1,2 and 3 yield improvements to the fit $\Delta \chi^2=32,19,19$ respectively and
a probability of all three detections being false is $p < 3 \times 10^{-11}$.
We repeated the fits with the line energy allowed to be free and found
the fitted energy to be consistent with 5.44\,keV and
that the line flux had not been significantly affected by freezing
the energy.
To extend the test for line flux variations we reduced and fit the
{\it XMM-Newton} spectra obtained during 2001 and 2002. We
followed the standard reduction method for the pn data, as
detailed by \citet{ponti06a} and found the data to be consistent with the presence of a line at the same
flux as found using {\it Suzaku}.
An independent analysis of the {\it XMM} data by \citet{demarco09a}
reported the presence of an excess of counts in the 5.4-6.2 keV band for NGC 4051, compared to their
parameterization of the local continuum. Examination of the {\it XMM} data shows an excess of
emission at $\sim 6$\, keV in the {\it XMM} spectra and so tentatively supports the possibility
of that additional energy-shifted line in this source.
{\it BeppoSax}
also observed NGC 4051, finding the source to be in a very low flux
state during 1998 as reported by \citet{guainazzi98b}. We extracted
and fit the archived {\it BeppoSax} Medium Energy Concentrator
spectrum from units 2 and 3 (combined) and tested for the presence of
the line. However, the data did not yield a significant detection of
and the upper limits on line flux were very loose $\sim 4 \times 10^{-5} {\rm photons\, cm^{-2} s^{-1}}$ for a
line at 5.44\,keV.
Given the poor constraint obtained we do not consider those data any further.
Combining the results from {\it Suzaku} and {\it XMM}
observations (specifically, Table~3 lines 1,2,4,5,7-9 \& 10)
and comparing the data to a constant model yields
$\chi^2=4.7/7\, d.o.f.$ Further to this test, we split the data by time
instead of intensity and repeated the test, reaching the same conclusion,
i.e. that the data are consistent with a
line of constant flux over a timescale of several years and over large
changes in observed continuum flux. The limits on measured line fluxes mean we
can rule out line variability greater than a factor $\sim 2$ in the line flux
sampled on these timescales, across the baseline time
period considered.
\section{Discussion}
While numerous claims exist in the literature for emission lines at
unexpected energies
\citep[see][for a review]{turner09a},
the reality of the lines has been questioned by some.
\citet{vaughan08a} considered 38 published results on transient emission
and absorption lines reported in the literature: those authors find a
linear relationship between the fitted feature strength and its
uncertainty. \citet{vaughan08a} noted that observations with more signal apparently
reveal weak lines but do not show tightly constrained strong lines
which should sometimes also show up by chance. The conclusion of the
\cite{vaughan08a} literature review was that there is a publication
bias in reporting of these results, and that many of the reported
detections are merely statistical fluctuations. This question has now
been addressed with two systematic analyses of
samples of AGN. \citet{tombesi10a} present a study of the occurrence
and reality of energy-shifted absorption lines in a sample of AGN
finding a deviation from the linear relationship of line equivalent
width (EW) and uncertainty found by \citet{vaughan08a} in the sense that
their distribution showed more significant detections of lines in
sources studied.
\citet{tombesi10a} compare their absorption
line measurements with those of the Fe K$\alpha$ emission in the same
sources and show that these two sets of measurements
follow the same distribution in the EW/uncertainty plane.
\citet{tombesi10a} conclude that the
absorption line detections are generally not statistical fluctuations
and that the absence of well-constrained detections of strong features
in the \citet{vaughan08a} analysis may be due in part to a limit on
the ability of current X-ray instruments to detect such lines. It
also would appear likely that long observations of bright sources may
not have been proposed or approved early in the {\it XMM} and
{\it Suzaku} missions and so the sample of observations completed to date
is likely biased against those that would have shown
tightly constrained energy-shifted lines: such a bias may explain at least some of
the effect discussed by \citet{vaughan08a}. Another recent study by
\citet{demarco09a} undertook a systematic analysis of a sample of
bright Seyfert 1 galaxies and confirmed many of the individual
detections of energy-shifted emission lines claimed in the literature,
supporting the general reality of the phenomenon by consideration of
the statistics of the sample results as a whole. With conflicting
views in the literature it is clear that the detection of significant
new examples
of the energy-shifted line phenomenon is very important at this time.
\subsection{Hard X-ray Line Emission in NGC 4051}
\label{sec:hardxrayline}
NGC 4051 can be modeled using a powerlaw continuum covered by multiple zones of gas, several in the
column density range $10^{23} - 10^{24} {\rm cm^{-2}}$,
that impart emission and absorption features to the X-ray spectrum. The marked spectral variability
with observed flux can be explained by changes in covering of the powerlaw continuum by one of the high-column absorbers, as found for
other similar AGN \citep[e.g.][]{pounds04c,risaliti07a,miller08a,turner08a}.
Significant line emission has been observed at 5.44\,keV in
the 2005 {\it Suzaku} observation of NGC 4051. Comparison of
low and high-state X-ray data for NGC 4051 shows that
the newly-discovered line appears prominent
in the low-flux state along with the narrow component of Fe K$\alpha$ emission from neutral gas, as expected if the observed
low state is simply those times when a relatively
large fraction of the
continuum is suppressed by absorption.
The most recent measurements indicate NGC 4051 to have a
black hole mass $ M_{BH} = 1.73^{+0.55}_{-0.52} \times 10^6 M_{\rm \odot}$
and a radius for the $H\beta$ broad line region (BLR)
$R_{BLR}= 1.87^{+0.54}_{ -0.50} $ light days \citep{denney09a}.
The $H\beta$ FWHM in the rms spectrum of \citeauthor{denney09a} is
$1034 \pm 41$\,km\,s$^{-1}$ (although those authors used the velocity
dispersion rather than the FWHM in their mass estimate).
If we assume the same geometrical correction factor between line width
and circular velocity as those authors to scale the respective FWHM measurements,
the line width of Fe K$\alpha$ is indicative of an origin at
$r \simeq 1.87(1034/5540)^2 \simeq 0.065$\,light days, or $2 \times 10^{14}$\,cm,
with a 90 percent confidence range
$8.6 \times 10^{13}- 1.7 \times 10^{15}$\,cm. As we have scaled to the optical reverberation
results, this radius estimate is secure provided the X-ray and optical line-emitting regions
have similar structure and orientation with respect to the observer, but the large uncertainty
is dominated by the uncertainty in the line width measurement.
Further to the measurement error is the uncertainty
as to whether, as suggested by the form of eigenvector 1, the Fe K$\alpha$ emission has both broad and narrow
components, in which case we would have to conclude that we are seeing contributions from
regions both within and outside of the region noted above.
Given the limited signal-to-noise in the regime of Fe K$\alpha$ in the HEG data
this question remains open with current data.
In the single component model for Fe K$\alpha$ emission, spectral fitting to the HEG plus {\it Suzaku} data
yields an Fe K$\alpha$ emission line of flux
$n=1.90 \pm 0.40 \times 10^{-5}$ photons cm$^{-2}$\,s$^{-1}$.
The strength of the Fe K$\alpha$ line emission, normalized to the illuminating continuum,
can be used to set limits on the reprocessing gas in which it arises.
Following \citet{yaqoob10a} and correcting for the continuum slope found here, we
estimate an efficiency for the production of line photons (defined as the ratio of
line flux to incident flux above the ionization edge)
to be
$x_{\scriptscriptstyle {\rm FeK}\alpha} \sim 0.018$. The line measured here sets a lower
limit on the column
density of the emitting region
$N_H > 10^{24}{\rm cm^{-2}}$ and in the
toroidal reprocessor model suggests a global covering factor
$\sim 0.9$ with an approximately face-on view down the pole of the structure.
As it is difficult to determine the intrinsic continuum strength from the observed continuum
the interpretation of these data in the context of the toroidal model is subject to some
uncertainty that in turn, leaves the derived global covering factor uncertain by a factor of a few.
Combining the minimum column density of the line-of-sight gas
with the radial constraints and knowledge of the gas having a high covering fraction
yields a mass estimate $\ga 4 \times 10^{-4}$M$_{\odot}$ for the gas emitting the Fe K$\alpha$ line,
assuming the nominal radius of emission of $2 \times 10^{14}$\,cm.
The 90\,percent confidence uncertainty in the distance translates into a confidence region on the
mass lower limit of $5 \times 10^{-5} - 0.1 $\,M$_\odot$.
If a torus is not the true geometry of the gas then, of course,
the covering factor could be different to this value.
Further to the uncertainties mentioned, the line may be be comprised
of contributions from two regions. Taking instead the two-component
fit to the Fe K$\alpha$ profile then the column and/or global covering
requirements are reduced for each of the two emitting regions.
The PCA decomposition \citep{miller09b} suggests a link between
the origin of the Fe K$\alpha$ line and that of the
line at 5.44\,keV.
Observation of similar lines in other AGN has motivated
discussion of several possible origins, including spallation and hotspot
emission from the accretion disk.
The solution found in the context of the disk hotspot model
suggests that the emitting radius
is between $18-24\,r_g$. The orbital timescales at
these radii are $\sim 4-6$ ks for
18-24 $r_g$ for the black
hole mass considered here. The observation of a steady line flux over
2005 - 2008 provides a constraint on the disk hotspot hypothesis as
the three year baseline is equivalent to $\sim 16,000-24,000$ orbits about the
black hole over the radial range of interest (and tens of orbits
just considering the line persistence within the 2005 observation).
Theoretical modeling indicates that hotspot events are not expected to
last longer than a few orbital timescales at these small radii
\citep{karas01a} and that a given hotspot would suffer measurable flux
and energy changes as the material spirals in \citep{dovciak04a}.
However, persistent steady lines could arise from a special radius in
the disk or other rotating reprocessor without any constraining
expectation of flux or energy variability. If the special radius is
interpreted as the truncation radius of the inner disk this
places the innermost edge of
the disk at 18-24$r_g$. The inner edge of the disk may emit more strongly
than the rest of the structure if it is inflated due to radiation
pressure as in the advection-dominated accretion flow scenario.
Assuming the bolometric luminosity to be $L_{bol}= 10^{43} {\rm erg\, s^{-1}}$
(\citealt{vasudevan09a}, corrected to the Tully-Fisher distance 15.2\,Mpc, \citealt{russell04a}) and
assuming a radiative efficiency $\eta^{\rm BOL} = 0.05$ the mass accretion rate is estimated to be
$\dot{M} \simeq 0.0035$\,M$_\odot$year$^{-1}$, $\sim 10$\% of the Eddington accretion rate.
For such high accretion rates
the transition from thin disk to an advection-dominated accretion flow
may occur anywhere up to a radius of $\sim 2000 r_g$ \citep{narayan98a}
and so our fitted radius is consistent with such a scenario.
If features
on the truncated edge of the disk are steady in flux, as found
for the line in NGC 4051,
then Galactic binaries would also be expected to show
such lines in the low flux state and the absence of such would disfavor the
truncated disk origin for the lines. While similar
lines have not been reported to date for Galactic
black hole binaries \citep[see][for a review]{done07b},
constraints on such lines have not yet been explored in that source class,
leaving this an open question.
In an alternative model, the fact that the specific energy of the
line is coincident with K$\alpha$ emission from neutral Cr (5.4\,keV)
prompts a renewed interest in the
spallation of Fe as a mechanism for enhancing otherwise weak lines, especially
since PCA is consistent with
a common origin for the new line and the neutral
component of Fe. In a companion paper, \citet{turner09c}, explore in
detail the spallation interpretation of the new result, extending
the work of \citet{skibo97} in the light of new understanding about
the environs of active nuclei. \citet{turner09c} find the observed
abundance enhancement to be high and that these extreme enhancement effects are
most likely to be achieved in gas out of the plane of the accretion
disk. In such a picture, the timescale for spallation may be as
short as a few years if the cosmic ray output is comparable to the
bolometric output of the nucleus. \citet{turner09c} also estimate the
expected radio and $\gamma-ray$ flux from the proposed spallation
process in NGC 4051 and find predictions to be consistent with current
flux measurements in those bands.
\section{Conclusions}
Our analysis of {\em Suzaku}\ data from NGC 4051 taken during 2005 and 2008
has revealed line emission at 5.44\,keV in the
rest-frame of the galaxy. We have established the reality of the
line at $>99.9\%$ confidence in data from 2005, supported by Monte Carlo
simulations that show the probability of the line being a statistical
fluctuation is $p < 3.3 \times 10^{-4}$. The possibility of the line arising from
a statistical fluctuation in the spectral data has been firmly ruled out
by establishing its detection in time-sliced data. Further to this,
we have confirmed the
line to be evident in all three XIS units independently, and established
that the observed line is inconsistent with arising from the X-ray background.
The source spectrum varies with flux, and the low state is dominated
by a hard spectral form with the new line plus the neutral component
of Fe K$\alpha$ emission superimposed upon that, suggestive of a
common origin for both. The line has an equivalent width during
2005 of about 45 eV while the Fe K$\alpha$ line is measured at 195 eV. These
reprocessed signatures show up prominently in the source low state when the
continuum is suppressed by the highest covering fraction of absorption.
As disk hotspot emission would
be expected to vary in flux and energy over relatively short
timescales, the limits on line variability disfavor this particular origin although
the data remain consistent with emission from a special location such as the
innermost radius of the accretion disk. The alternative picture, that
the line is Cr {\sc i} K$\alpha$ emission following
spallation of Fe is also found to be a good explanation
of the data: that possibility and its implications are explored in a
companion paper.
\acknowledgments
TJT acknowledges NASA grant NNX08AL50G. LM acknowledges STFC grant number PP/E001114/1.
We are grateful to the anonymous referee whose comments significantly
improved this manuscript: we also thank
the {\em Suzaku}\ operations team for performing this observation
and providing software and calibration for the data analysis.
This research has also made use of data obtained from the
High Energy Astrophysics Science
Archive Research Center (HEASARC), provided by NASA's Goddard Space
Flight Center.
\bibliographystyle{apj}
|
\section{Introduction}
Image registration is an essential operation in a variety of medical imaging applications including disease diagnosis, longitudinal studies, data fusion, image segmentation, image guided therapy, volume reconstruction, pathology detection, and shape measurement~(\cite{nonmedsurvey,survey,survey2}). It is the process of finding a geometric transformation between a pair of scenes, the \textit{source scene} and the \textit{target scene}, such that the similarity between the transformed source scene (\textit{registered source}) and target scene becomes optimum.
There are many challenges in the registration of medical images. Among these, those that stem from the artifacts associated with images include the presence of noise, interpolation artifacts, intensity non-uniformities, and intensity non-standardness. Although considerable research has gone into addressing the effects of noise~(\cite{holden}), interpolation~(\cite{reg_interp, reg_interp2, interp_likar}), and non-uniformity in image registration~(\cite{reg_nu}), little attention has been paid to study the effects of image intensity standardization/non-standardness in image registration. This aspect constitutes the primary focus of this paper.
MR image intensities do not possess a tissue specific numeric meaning even in images acquired for the same subject, on the same scanner, for the same body region, by using the same pulse sequence~(\cite{std, std_numeric}). Not only a registration algorithm needs to capture both large and small scale image deformations, but it also has to deal with global and local image intensity variations. The lack of a standard and quantifiable interpretation of image intensities may cause the geometric relationship between homologous points in MR images to be affected considerably. Current techniques to overcome these differences/variations fall into two categories. The first class of methods uses intensity modelling and/or attempts to capture intensity differences during the registration process. The second group constitutes post processing methods that are independent of registration algorithms. Notable studies that have attempted to solve this problem within the first class are~(\cite{demons,elastic,ash}). While global intensity differences are modelled with a linear multiplicative term in~(\cite{ash}), local intensity differences are modelled with basis functions. In~(\cite{elastic}), a locally affine but globally smooth transformation model has been developed in the presence of intensity variations which captures intensity variations with explicitly defined parameters. In~(\cite{demons}), intensities of one image are mapped into those of another via an adaptive transformation function. Although incorporating intensity modelling into the registration algorithms improves the accuracy, simultaneous estimation of intensity and geometric changes can be quite difficult and computationally expensive.
The papers that belong to the second group of methods are~(\cite{std, std_numeric, std_new, std_nu, std_mtr}) in which a two-step method is devised for standardizing the intensity scale in such a way that for the same MRI protocol and body region, similar intensities achieve similar tissue meaning. The methods transform image intensities non-linearly so that the variation of the overall mean intensity of the MR images within the same tissue region across different studies obtained on the same or different scanners is minimized significantly. Furthermore, the computational cost of these methods is considerably small in comparison to methods belonging to the first class. Once tissue specific meanings are obtained, quantification and image analysis techniques, including registration, segmentation, and filtering, become more accurate.
The non-standardness issue was first demonstrated in~(\cite{std}) where a method was proposed to overcome this problem. The new variants of this method are studied in~(\cite{std_new}). Numerical tissue characterizability of different tissues is achieved by standardization and it is shown that this can significantly facilitate image segmentation and analysis in~(\cite{std_numeric, zhuge_seg}). Combined effects of non-uniformity correction and standardization are studied in~(\cite{std_nu}) and the sequence of operations to produce the best overall \textit{image quality} is studied via an interplaying sequence of non-uniformity correction and standardization methods. In~(\cite{std_gscale}), an improved standardization method based on the concept of generalized scale is presented. In~(\cite{std_mtr}), the performance of standardization methods is compared with the known tissue characterizing property of magnetization transfer ratio (MTR) imaging and it is demonstrated that tissue specific intensities may help characterizing diseases.
The motivation for the research reported in this paper is the preliminary indication in~(\cite{bagci07}) of the potential impact that intensity standardization may have on registration accuracy. Currently no published study exists that has examined how intensity non-standardness alone may affect registration. The goal of this paper is, therefore, to study the effect of non-standardness on registration in isolation. Toward this goal, first intensity non-uniformities are corrected in a set of images, and subsequently, they are standardized to yield a ``clean set'' of images. Different levels of non-standardness are then introduced artificially into these images which are then subjected to known levels of affine deformations. The clean set is also subjected to the same deformations. The deformed images with and without non-standardness are separately registered to clean images and the differences in their registration accuracy are quantified to express the influence of non-standardness. The underlying methods are described in Section II and the analysis results are presented in Section III. Section IV presents some concluding remarks.
\section{Methods }
\label{sec:methods}
\subsection{Notations and Overview}
\label{subsec:term}
We represent a 3D image, called \textit{scene} for short, by a pair $\mathcal{C}=(C,f)$ where $C$ is a finite 3D array of voxels, called \textit{scene domain} of $\mathcal{C}$, covering a body region of the particular patient for whom image data $\mathcal{C}$ are acquired, and $f$ is an intensity function defined on $C$, which assigns an integer intensity value to each voxel $\nu \in C$. We assume that $f(\nu) \geq 0$ for all $\nu \in C$ and $f(\nu)=0$ if and only if there are no measured data for voxel $\nu$.
In dealing with standardization issues, the body region and imaging protocol need to be specified. All images that are analyzed for their dependence on non-standardness for registration accuracy are assumed to come from the same body region $B$ and acquired as per the same acquisition protocol $P$. The non-standardness phenomenon is predominant mainly in MR imaging. Hence, all image data sets considered in this paper pertain to MRI. However, the methods described here are applicable to any modality where this phenomenon occurs (such as radiography and electron microscopy).
There are six main components to the methods presented in this paper: (1) intensity non-uniformity correction, referred to simply as \textit{correction} and denoted by an operator $\kappa$; (2) intensity standardization denoted by an operator $ \psi $; (3) an affine transformation of the scene, denoted by $T$ used for the purpose of creating mis-registered scenes; (4) introduction of artificial intensity non-standardness denoted by the operator $\overline{\psi}$; (5) an affine scene transformation that is intended to register a scene with its mis-registered version; (6) evaluation methods used to quantify the goodness of scene registration.
Super scripts $c,s,\overline{s},t$ and $r$ are used to denote, respectively, the scenes resulting from applying correction, standardization, introduction of non-standardness, mis-registration, and registration operations to a given scene. Examples:
$\mathcal{C}^c=\kappa\left( \mathcal{C}\right) ; \mathcal{C}^{cs}=\kappa\psi\left( \mathcal{C}\right); \mathcal{C}^{cs\overline{s}}=\overline{\psi}\left( \mathcal{C}^{cs}\right); \mathcal{C}^{cs\overline{s}t}=T\left( \mathcal{C}^{cs\overline{s}}\right)$. When a registration operation $T_r$ is applied to a scene $\mathcal{C}$, the target scene to which $\mathcal{C}$ is registered will be evident from the context. The same notations are extended to sets of scenes. For example, if $\chi$ is a given set of scenes for body region $B$ and protocol $P$, then $\chi^{cs\overline{s}}=\overline{\psi}\left(\chi^{cs} \right) $, where $\chi^{cs}=\kappa\psi\left(\chi \right)$
Our approach to study the effect of non-standardness on registration is as follows:
\noindent (S1) Take a set $\chi$ of scenes, pertaining to a fixed $B$ and $P$, but acquired from different subjects in routine clinical settings.
\noindent (S2) Apply correction followed by standardization to the scenes in $\chi$ to produce the set $\chi^{cs}$ of \textit{clean scenes}. $\chi^{cs}$ is as free from non-uniformities, and more importantly, from non-standardness, as we can practically make. As justified in~(\cite{std_nu}), the best order and sequence of these operations to employ in terms of reducing non-uniformities and non-standardness is $\kappa$ followed by $\psi$. This is mainly because any correction operation introduces its own non-standardness.
\noindent (S3) Apply different known levels of non-standardness to the scenes in $\chi^{cs}$ to produce the set $\chi^{cs\overline{s}}$.
\noindent (S4) Apply different known levels of affine deformations $T$ to the scenes in $\chi^{cs\overline{s}}$ to form the scene set $\chi^{cs\overline{s}t}$. Apply the same deformations to the clean scenes in the set $\chi^{cs}$ to create $\chi^{cst}$. In this manner for any scene $\mathcal{C}^{cs}\in\chi^{cs}$, we have the same scene after applying some non-standardness and the same deformation $T$, namely $\mathcal{C}^{cs\overline{s}t}$.
\noindent (S5) Register each scene $\mathcal{C}^{cs}\in\chi^{cs}$ to $\mathcal{C}^{cst}\in\chi^{cst}$ and determine the required affine deformation $T_s$ (the subscript s indicates ``standardized"). Similarly register each $\mathcal{C}^{cs}\in\chi^{cs}$ to $\mathcal{C}^{cs\overline{s}t}\in\chi^{cs\overline{s}t}$ and determine affine deformation $T_{ns}$ (ns for ``not standardized") needed.
\noindent (S6) Analyze the deviations of $T_s$ and $T_{ns}$ from the true applied transformation $T$ over all scenes and as a function of the applied level of non-standardness and affine deformations.
In the rest of this section, steps S1-S6 are described in detail.
\noindent \textit{S1: Data Sets}\\
Two separate sets of image data (i.e., two sets $\chi$) are used in this study, both brain MR images of patients with Multiple Sclerosis, one of them being a T2 weighted acquisition, and the other, a proton density (PD) weighted set, with the following acquisition parameters: Fast Spin Echo sequence, 1.5T GE Signa scanner, TR=2500 \textit{msec}, voxel size 0.86x0.86x3 $mm^3$. Each of the two sets is composed of 10 scenes. Since the two data sets for each patient are acquired in the same session with the same repetition time but by capturing different echos $(TE=18 msec, 96 msec)$, the T2 and PD scenes for each patient can be assumed to be in registration.
\noindent \textit{S2. Non-uniformity Correction, Standardization}\\
For non-uniformity correction, we use the method based on the concept of local morphometric scale called g-\textit{scale}~(\cite{mada}). Built on fuzzy connectedness principles, the g-\textit{scale} at a voxel $\nu$ in a scene $\mathcal{C}$ is the largest set of voxels fuzzily connected to $\nu$ in the scene such that all voxels in this set satisfy a predefined homogeneity criterion. Since the g-\textit{scale} set represents a partitioning of the scene domain $C$ into fuzzy connectedness regions by using a predefined homogeneity criterion, resultant g-\textit{scale} regions are locally homogeneous, and spatial contiguity of this local homogeneity is satisfied within the g-\textit{scale} region.
g-\textit{scale} based non-uniformity correction is performed in a few steps as follows. First, g-\textit{scale} for all foreground voxels is computed. Second, by choosing the largest g-\textit{scale} region, background variation is estimated. Third, a correction is applied to the entire scene by fitting a second order polynomial to the estimated background variation. These three steps are repeated iteratively until the largest g-\textit{scale} region found is not significantly larger than the previous iteration's largest g-\textit{scale} region.
Standardization is a pre-processing technique which maps non-linearly image intensity gray scale into a standard intensity gray scale through a training and a transformation step. In the training step, a set of images acquired for the same body region $B$ as per the same protocol $P$ are given as input to \textit{learn} histogram-specific parameters. In the transformation step, any given image for $B$ and $P$ is standardized with the estimated histogram-specific landmarks obtained from the training step. In the data sets considered for this study, $B=head$ and $P$ represents two different protocols, namely T2 and PD. The training and transformation steps are done separately for the two protocols.
The basic premise of standardization methods is that, in scenes acquired for a given $\left\langle B, P\right\rangle $, certain tissue-specific landmarks can be identified on the histogram of the scenes. Therefore, by matching the landmarks, one can standardize the gray scales. Median, mode, quartiles, and deciles, and intensity values representing the mean intensity in each of the largest few g-\textit{scale} regions have been used as landmarks. Additionally, to handle outliers, a ``low'' and ``high'' intensity value (selected typically at 0 and 99.8 percentiles) are also selected as landmarks.
In the training step, the landmarks are identified for each training scene specified for $\left\langle B, P\right\rangle $ and intensities corresponding to the landmarks are mapped into an assumed standard scale. The mean values for these mapped landmark locations are computed. In the transformation step, the histogram of each given scene $\mathcal{C}$ to be standardized is computed, and intensities corresponding to the landmarks are determined. Sections of the intensity scale of $\mathcal{C}$ are mapped to the corresponding sections of the standard scale linearly so that corresponding landmarks of scene $\mathcal{C}$ match the mean landmarks determined in the training step. (The length of the standard scale is chosen in such a manner that the overall mapping is always one-to-one and no two intensities in $\mathcal{C}$ map into a single intensity on the standard scale.) Note that the overall mapping is generally not a simple linear scaling process but, indeed, a non-linear (piece-wise linear) operation; see~(\cite{std,std_new}) for details. In the present study, standardization is done separately for T2 and PD scenes.
\noindent \textit{S3. Applying Non-standardness}\\
To artificially introduce non-standardness into a \textit{clean scene} $\mathcal{C}^{cs}=\left( C, f^{cs}\right) $, we use the idea of the inverse of the standardization mapping described in~(\cite{std_nu}). A typical standardization mapping is shown in Figure~\ref{img:mapping}. In this figure, only three landmarks are considered - ``low'' and ``high'' intensities $(p1$ and $p2)$ and the median $(\mu)$ corresponding to the foreground object. There are two linear mappings: the first from $[p_{1},\mu]$ to $[s_1,\mu_s]$ and the second from $[\mu, p_{2}]$ to $[\mu_s, s_2]$. $\left[ s_1, s_2\right] $ denotes the standard scale. The horizontal axis denotes the non-standard input scene intensity and vertical axis indicates the standardized output scene intensity. In inverse mapping, since $\mathcal{C}^{cs}$ has already been standardized, the vertical axis can be considered as the input scene intensity, $f^{cs}(\nu)$, and the horizontal axis can be considered as the output scene intensity, $f^{cs\overline{s}}(\nu)$, where mapping the \textit{clean scene} through varying the slopes $m_1$ and $m_2$ results in non-standard scenes. By using the values of $m_1$ and $m_2$
within the range of variation observed in initial standardization mappings of corrected scenes, the non-standard scene intensities can be obtained by
\begin{equation}
\label{eq:inversemapping}
f^{cs\overline{s}}(\nu)= \left\{ \begin{array}{ll}
\lceil \frac{f^{cs}(\nu)}{m_1} \rceil &
\textrm{if $f^{cs}(\nu) \leq \mu_s$}\\
\lceil \frac{\left( f^{cs}(\nu)-\mu_s \right)}{m_2} + \mu \rceil &
\textrm{if $f^{cs}(\nu) > \mu_s,$}
\end{array} \right.
\end{equation}
where $\lceil . \rceil$ converts any number y$\in \Re$ to the closest integer Y, and $\mu_s$ denotes the median intensity on the standard scale.
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=8cm]{invmap.eps}
\end{tabular}
\end{center}
\caption{The standardization transformation function for inverse mapping with the various parameters shown.\label{img:mapping}}
\end{figure}
In order to keep the number of registration experiments manageable, this simple model was used which involves only two variables $m_1$ and $m_2$. Even so, as described later, this study entails nearly 20,000 registration experiments.
\noindent \textit{S4. Applying Affine Deformations}\\
All components of the affine transformation - rotations about all three axes and translation, scaling, and shear in all three directions - are taken into account in creating scene sets $\chi^{cst}$ and $\chi^{cs\overline{s}t}$.
\noindent \textit{S5. Scene Registration}\\
The algorithm that determines the affine transformation matrix by minimizing the sum of squared scene intensity differences as described in~(\cite{ash}) is used in this step. A separate transformation matrix is found for registering each $\mathcal{C}^{cs}$ to $\mathcal{C}^{cst}$, resulting in $T_s$, and $\mathcal{C}^{cs}$ to $\mathcal{C}^{cs\overline{s}t}$, resulting in $T_{ns}$.
\noindent \textit{S6. Evaluation}\\
Two types of tests were carried out - \textit{accuracy} and \textit{consistency}. The goal of the accuracy test was to determine how close the recovered registration transformations are to the known true transformations. The aim of the consistency test was to check how the observed accuracy behavior would consistently occur when different accuracy tests are conducted. In each test, two transformations $T_s$ and $T_{ns}$ are compared by using the methodology that is described in~(\cite{reg_udupa}) and summarized below. Let $\mathcal{C}_{T2}\in\chi_{T2}$ and $\mathcal{C}_{PD}\in\chi_{PD}$ be the T2 and PD scenes of any particular patient. Let $T_{s,x}$ and $T_{ns,x}$, $x\in\left\lbrace T2, PD\right\rbrace $, be the transformations obtained in Step S5 by matching $\mathcal{C}_x^{cs}$ to $\mathcal{C}_x^{cst}$ and $\mathcal{C}_x^{cs}$ to $\mathcal{C}_x^{cs\overline{s}t}$, respectively. In the \textit{accuracy} tests, $T_{s,x}$ and $T_{ns,x}$ are compared with the true (known) transformation $T$ over all levels of non-standardness and deformations that were employed. Both $T_{s,x}$ and $T_{ns,x}$ are expected to be the same as $T$. In the \textit{consistency} tests, $T_{s, T2}$ with $T_{s, PD}$ and $T_{ns, T2}$ with $T_{ns, PD}$ are compared over all levels of non-standardness and deformations that were applied, and they are expected to be equal because PD and T2 scenes of the same patient are already in registration as described in S1. We measure the error between two transformations (in both the above scenarios) by the root-mean-squared error (RMSE) for the eight corner voxels of the box that approximately bounds the head, i.e., the volume of interest in our application.
In the accuracy test, for a given level of applied non-standardness and affine deformation, we get 20 pairs of RMSE values, each pair indicating how close $T_{s, x}$ and $T_{ns, x}$ are to the true transformation $T$. A paired t-test is conducted to compare the accuracy of the two transformations based on RMSE values. The outcome of this test will be that either of the two transformations is more accurate than the other or there is no significant difference between the two (throughout we use $P \leq 0.05$ to indicate statistical significance). The set of all levels of applied deformations is divided into three groups - small, medium, large. For each of 8 levels of applied non-standardness and under each of these three groups, the number of occurrences of wins (w), losses (l) and non-significant differences (n) are counted for $T_{ns}$ over $T_s$. The number of wins and losses is normalized to get values in the range $[0, 1]$: $W_x=\frac{w}{w+l+n}$, $L_x=\frac{l}{w+l+n}$. A particular configuration of wins and losses can be identified by a point with coordinates $\left( W_x, L_x\right) $ in a win-loss triangle as in Figure~\ref{img:merit}. Large values of $L_d$ and small values of $W_d$ indicate that the performance point is closer to the point $\left( 1, 0\right) $ of all-wins. The following metric is used to express the ``goodness'' value of the configuration.
\begin{equation}
\gamma = \frac{L_d}{W_d}=\sqrt{\frac{\left( 1-L_x\right)^2 +W_x^2}{\left( 1-W_x\right)^2 +L_x^2}}.
\end{equation}
\begin{figure}[b]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=8cm]{merit.eps}
\end{tabular}
\end{center}
\caption{Mapping procedure for ``goodness'' value in normalized win-loss $W_x-L_x$ plane. \label{img:merit}}
\end{figure}
The procedure in the consistency test is similar to the above except that we have 10 pairs of RMSE values to compare and these values are obtained not by using the known true transformation but by using $T_{S,PD}$ for $T_{S,T2}$ and $T_{NS,PD}$ for $T_{NS,T2}$.
\section{Experimental Results}
\label{sec:results}
\subsection{Implementation Details}
\subsubsection{Correction and Standardization}
These operations are carried out by using the 3DVIEWNIX software~(\cite{3dviewnix}). Based on the experiments in~(\cite{std, std_numeric, std_new}), minimum and maximum percentile values are set to $pc_{1}=0$ and $pc_2=99.8$, respectively. In the standard scale, $s_1$ and $s_2$ are set to $s_1=1$ and $s_2=4095$. Figure~\ref{img:cs} shows the original, corrected, and standardized (after correction) of two PD and two T2-weighted slices taken from two different studies in the first, second and third rows, respectively. The gain in the similarity of resulting image intensities for similar tissue types obtained can be readily seen.
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=10 cm]{correctionandstandardization2.eps}
\end{tabular}
\end{center}
\caption{Two slices from PD (first and second columns) and two slices from T2-weighted scenes (third and fourth columns) selected from two different studies before correction and standardization are displayed at default windows in the first row. Corresponding slices of the g-scale corrected scenes are shown in the second row. Clean scenes obtained after the standardization process of the corresponding corrected scenes are displayed at a standard window in the third row. \label{img:cs}}
\end{figure}
\subsubsection{Adding known levels of non-standardness}
We combine eight different ranges of the slopes $m_1$ and $m_2$ to introduce small, medium, and large scale non-standardness into the scenes. This means that, for each \textit{clean scene}, we obtain eight scenes, one of which is the default \textit{clean scene} itself, two scenes consisting of small scale non-standardness, two scenes consisting of medium scale non-standardness, and three scenes consisting of large scale non-standardness. The ranges of applied non-standardness are given in Table~\ref{table:slopes}. We have arrived at these values by examining the training part of the standardization process through computing the ranges of the slopes $m_1$ and $m_2$ that are utilized in standardizing the corrected scenes. Figures~\ref{img:pd} and~\ref{img:t2} illustrate the process of introducing known levels of non-standardness into the \textit{clean} slices of a PD and a T2-weighted scene utilized in our study. In both figures, the first display shows the original clean slice and the rest show the resulting non-standard slices.
\begin{table}[h]
\caption{Description of the different range of the slopes $m_1,m_2$ for introducing artificial non-standardness\label{table:slopes}}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline \hline
function & Range & Description \\ \hline \hline
$\overline{\psi}_1$ & $\left\lbrace 0.9 \leq m_1,m_2 \leq 1.5 \right\rbrace$ & \multirow{2}{*}{Small Scale} \\
$\overline{\psi}_2$ & $\left\lbrace 0.6 \leq m_1,m_2 \leq 0.9 \right\rbrace$ & \\ \hline \hline
$\overline{\psi}_3$ & $\left\lbrace 1.5 \leq m_1,m_2 \leq 2.0 \right\rbrace$ & \multirow{2}{*}{Medium Scale}\\
$\overline{\psi}_4$ & $\left\lbrace 2.0 \leq m_1,m_2 \leq 2.4 \right\rbrace$ & \\ \hline \hline
$\overline{\psi}_5$ & $\left\lbrace 2.4 \leq m_1,m_2 \leq 2.7 \right\rbrace$ & \multirow{3}{*}{Large Scale} \\
$\overline{\psi}_6$ & $\left\lbrace 2.7 \leq m_1,m_2 \leq 3,0 \right\rbrace$ & \\
$\overline{\psi}_7$ & $\left\lbrace 3.0 \leq m_1,m_2 \leq 3.3 \right\rbrace$ & \\ \hline \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=7 cm]{wipd8different.eps}
\end{tabular}
\end{center}
\caption{First image in the first row is a slice of a clinical PD weighted clean scene of the brain. The other slices are obtained by adding the 7 different levels of non-standardness into the clean scene (7 different levels of non-standardness are $\overline{\psi_1}$ to $\overline{\psi_7}$). All images are displayed at the fixed gray level window chosen for the clean scene. \label{img:pd}}
\end{figure}
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=7 cm]{wit8different.eps}
\end{tabular}
\end{center}
\caption{First image in the first row is a slice of a clinical T2 weighted clean scene of the brain. The other slices are obtained by adding the 7 different levels of non-standardness into the clean scene. (7 different levels of non-standardness are $\overline{\psi_1}$ to $\overline{\psi_7}$). All images are displayed at the fixed gray level window chosen for the clean scene. \label{img:t2}
\end{figure}
\subsubsection{Applying known amounts of deformation}
Three different rotations (0, medium, and large angle), three translations (0, medium, and large displacement), three levels of scaling (0, medium, and large), and three levels of shearing (0, medium, and large) are combined to introduce 81 different known levels of deformation such that for all non-zero transformations, all three directions/axes are involved. Table~\ref{table:defval} summarizes the amount of these transformations used for each axis in the three different groups.
\begin{table}[h]
\caption{The amount of deformations corresponding to different groups of transformations. \label{table:defval}}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline \hline
Transformation Type & Zero & Medium & Large \\ \hline \hline
Translation & $0$ pixels & $5$ pixels & $20$ pixels \\ \hline
Rotation & $0^o$& $2^o$ & $6^o$\\ \hline
Scaling & $1$ & $1.05$ & $1.15$\\ \hline
Shearing & $0$ & $0.01$ & $0.05$\\ \hline \hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Registration}
We use the scene based affine registration method made available in the SPM software~(\cite{ash}). The algorithm determines the affine transformation matrix that optimally registers the two scenes by optimizing the sum of squared differences between scene intensities.
\subsection{Results}
For each of the 20 clean scenes in $\chi_{T2}^{cs} \bigcup \chi_{PD}^{cs}$, considering the 7 different levels of non-standardness together with one level of standardness (i.e. total of 8 levels) and 3 x 3 x 3 x 3 = 81 different levels of mis-registration, there will be 8 x 81 = 648 scenes. Thus, in the accuracy test, there will be 20 x 648 = 12,960 registration experiments. In the consistency test, similarly, there will be 6,480 additional registration experiments. These additional experiments can be considered a validation for accuracy tests because they show how consistent the accuracy of the registration experiments are by using the fact that T2 and PD scenes are in registration.
The results of the comparison experiments are reported in Tables~\ref{table:goodness} and~\ref{table:goodness2} for accuracy and consistency tests, respectively, for 7 sets of non-standard scenes with respect to the registration performance of \textit{clean scenes}. The tables summarize the effectiveness of the registrations for each type of deformation recovered. The goodness values indicate that the ability to recover known deformations from transformed scenes is lower if intensity variations between source and target scenes are large. The goodness value $\gamma < 1$ for scenes with non-standardness indicates that the registration between \textit{clean scenes} outperforms registration between scenes with certain levels of non-standardness and this is true only if $W_x<L_x$.
\begin{table}[h]
\begin{center}
\caption{Comparison of methods for Accuracy. The Goodness values $\gamma$ are listed. Type of non-standardness are indicated by $\overline{\psi_1}, ..., \overline{\psi_7}$, and the type of affine deformations are indicated by small, medium, and large in the columns. \label{table:goodness}}
\begin{tabular}{|c|c|c|c|c|}
\hline Type of Non-Standardness & Small & Medium & Large & Total \\
\hline $\overline{\psi}_1 $ & 1 & 0.8222 & 0.6562 & 0.7811\\
\hline $\overline{\psi}_2 $ & 0.9400 & 0.8305 & 0.6167 & 0.7716\\
\hline $\overline{\psi}_3 $ & 0.9369 & 0.7751 & 0.6309 & 0.7651 \\
\hline $\overline{\psi}_4 $ & 0.9318 & 0.7004 & 0.6048 & 0.7565 \\
\hline $\overline{\psi}_5 $ & 0.8806 & 0.6004 & 0.5622 & 0.6254 \\
\hline $\overline{\psi}_6 $ & 0.7565 & 0.5511 & 0.5341 & 0.5881 \\
\hline $\overline{\psi}_7 $ & 0.7447 & 0.5901 & 0.5051 & 0.5819 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\begin{center}
\caption{Comparison of methods for Consistency. The Goodness values $\gamma$ are listed. Type of non-standardness are indicated by $\overline{\psi_1}, ..., \overline{\psi_7}$, and the type of affine deformations are indicated by small, medium, and large in the columns. \label{table:goodness2}}
\begin{tabular}{|c|c|c|c|c|}
\hline Type of Non-Standardness & Small & Medium & Large & Total \\
\hline $\overline{\psi}_1 $ & 1.3427 & 0.9289 & 0.7423 & 1\\
\hline $\overline{\psi}_2 $ & 1.1491 & 1 & 0.8039 & 0.9530\\
\hline $\overline{\psi}_3 $ & 1 & 0.8039 & 0.7423 & 0.8417 \\
\hline $\overline{\psi}_4 $ & 0.8622 & 0.7447 & 0.5434 & 0.7062 \\
\hline $\overline{\psi}_5 $ & 0.9289 & 0.6722 & 0.5023 & 0.6934 \\
\hline $\overline{\psi}_6 $ & 0.8636 & 0.6890 & 0.5023 & 0.6722 \\
\hline $\overline{\psi}_7 $ & 1.0752 & 0.7097 & 0.2998 & 0.6468 \\ \hline
\end{tabular}
\end{center}
\end{table}
A strong possible reason for the better performance of \textit{clean scenes} with respect to the non-standard scenes is that structural information for the same subject in different non-standardness levels is not the same. Therefore, correlations of the intensity values for each structure in the scenes may not reach the optimum to which the registration algorithm converges. Registration parameters are obtained through maximizing the similarity of two scenes, and it is well demonstrated in Tables~\ref{table:goodness} and~\ref{table:goodness2} that correlation of the intensities is maximum when each structure in the scene has fixed tissue specific meaning.
Another possible reason is that the relationship between voxel intensities may be non-linear. Since the introduction of non-standardness is itself a non-linear process, the similarity function is likely to be affected by this situation in the form of local fluctuations which may even lead to not only less accurate registration results but also to the situation of the optimization process getting locked at local optima. The opposite situation may happen as well, especially for large scale deformations; the registration algorithm may easily fail regardless of the standardization level of the scenes (see Figure~\ref{img:uclu} (a) for a failing example of \textit{clean scenes}). Local fluctuations in the similarity measure due to non-standardness may lead to different optimum points depending on the degree of non-standardness, some of which may even improve registration, as shown in Figure~\ref{img:uclu} (b). Although the registration algorithm did not get stuck in the latter case, the accuracy of the registration quality was not high especially in terms of translation parameters. In order to cope with possible failing examples in registration, we ran the registration algorithm with proper initial estimation of the transformation matrix rather than the default identity transformation matrix used in all experiments. Figure~\ref{img:uclu} (c) shows the registered and transformed \textit{clean scenes} overlaid where fuzziness in gray scale images demonstrates the misalignment. Compared to the registration of non-standard scene in Figure~\ref{img:uclu} (a), the performance of the registration of \textit{clean scene} in this example is still better when the initial estimation of transformation matrix is given as input to the registration process.
\begin{figure}[h]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=12 cm]{uclu3.eps}
\end{tabular}
\end{center}
(a) \quad \quad \quad \quad \quad \quad \quad (b) \quad \quad \quad \quad \quad \quad \quad \quad \quad (c)
\caption{(a) An example of registration failure for \textit{clean scene} with a large deformation. (b) An example for registration success for non-standard scene for the same amount of deformation as in the example in (a). (c) If a proper initialization matrix is given as input to the registration algorithm, clean scene registration performance becomes better than the example for non-standard scene. \label{img:uclu}}
\end{figure}
Based on the fact that similarity of a pair of registered \textit{clean scenes} is higher than the similarity of a pair of registered non-standard scenes, it can be deduced that substantially improved uniformity of tissue meaning between two scenes of the same subject being registered improves registration accuracy. Our experimental results demonstrate that scenes are registered better whenever the same tissues are represented by the same intensity levels.
We note that, in both tables, most of the entries are less than 1. This indicates that in both accuracy and consistency tests, the standardized scene registration task wins over the registration of non-standard scenes. Table~\ref{table:goodness} on its own does not convey any information about what the actual accuracies in the winning cases are, or about whether the win happens for T2 scenes only, PD only, or for both. The fact that a majority of the corresponding cells in these tables both indicate wins suggests that accuracy-based wins happen for both T2 and PD scenes. Conversely, a favorable $\gamma$ value in Table~\ref{table:goodness2} does not convey any information about whether the high consistency indicated also signals accuracy. Thus, accuracy and consistency are to some extent independent factors, and they together give us a more complete picture of the influence of non-standardness on registration.
\section{Concluding Remarks}
\label{sec:conc}
We described a controlled environment for determining the effects of intensity standardization on registration tasks in which the best image quality (\textit{clean scene}) was established by the sequence of correction operation followed by standardization. We introduced several different levels of non-standardness into the \textit{clean scenes} and performed nearly 20,000 registration experiments for small, medium and large scale deformations. We compared the registration performance of \textit{clean scenes} with the performance of scenes with non-standardness and summarized the resulting goodness values. From overall accuracy and consistency test results in Tables~\ref{table:goodness} and~\ref{table:goodness2}, we conclude that intensity variation between scenes degrades registration performance. Having tissue specific numeric meaning in intensities maximizes the similarity of images which is the essence of the optimization procedure in registration. Standardization is therefore strongly recommended in the registration of images of patients in any longitudinal and follow up study, especially when image data come from different sites and different scanners of the same or different brands.
In this paper, we introductorily addressed the problem of the potential influence of intensity non-standardness on registration. This is indeed a small segment of the much larger full problem: Unlike the specific intra-modality (or intra protocol) registration task considered here, there are many situations in which the source and the target images may be from different modalities or protocols (e.g., CT to MRI, PET to MRI, and T1 to T2 registration etc.), and each such situation may have its own theme of non-standardness. Further, these themes may depend on the body region, the scanner, and its brand. We determined that a full consideration of these aspects was just beyond the scope of this paper. Since the sum of squared differences is one of the most appropriate similarity metrics for intra-modality registration, we focused on this metric in our study. But, clearly, more studies of this type in the more general settings mentioned above are needed.
Thus far, we controlled the computational environment via two factors: standardization and correction. A third important factor, noise, can be also embedded into the framework. It is known that correction itself introduces non-standardness into the scenes and it also enhances noise. Investigating the interrelationship between correction and noise suppression algorithms and determining the proper order for these operations has been studied recently~(\cite{montillo}). A question immediately arises as to how standardization affects registration accuracy for different orders of correction and noise filtering. Based on the study in~(\cite{montillo}), we may conclude that non-uniformity correction should precede noise suppression and that standardization should be the last operation among the three to obtain best image quality. However, it remains unclear as to how a combination of deterministic methods (standardization and correction) affects a random phenomenon like noise. It is thus important to study these three phenomena in the future on their own or in relation to how they may influence the registration process, especially in multi-center studies wherein data come from different scanners and brands of scanners.
\section{Acknowledgement}
This paper is presented in SPIE Medical Imaging - 2010. The complete version of this paper is published in Elsevier Pattern Recognition Letters, Vol(31), pp.315--323, 2010.
This research is partly funded by the European Commission Fp6 Marie Curie Action Programme (MEST-CT-2005-021170). The second author's research is funded by an NIH grant EB004395
|
\section*{Acknowledgements}
This work was supported by the National Nature Science Foundation of China (Grant No. 10874071).
|
\section{Basic Idea, Purpose, and Methods of Realistic Modeling}
\label{sec.intro}
The oxide materials can be rather complex.
Nevertheless, in many cases their
electronic and magnetic properties are
controlled by a small group of states located near the
Fermi level and well isolated from the rest of the spectrum.
A typical example of the electronic structure of YVO$_3$, obtained in the
local-density approximation (LDA), is shown in Fig.~\ref{fig.YVO3LDA}.
\begin{figure}[!h]
\centering
\includegraphics[width=3.0in]{figure1.eps}
\caption{(Left panel) Electronic structure of orthorhombic
YVO$_3$ in LDA. The shaded area shows the contributions of the
$3d$ states of V. (Right panel) Enlarged behavior of $t_{2g}$ bands
computed from LMTO basis functions (solid curves)
and downfolded bands (dot-dashed curves).
The corresponding bands in the left panel are shown by arrows.
The
Fermi level is at zero energy.}
\label{fig.YVO3LDA}
\end{figure}
In this case, the active states are
twelve $t_{2g}$ bands of predominantly V-character, which are separated by finite energy
windows from other states, both from below and from above.
Such a division of the entire electronic structure into the ``active'' and
``inactive'' (or low- and high-energy)
states opens a formal way for combining the first-principles
calculations, based on the density-functional theory (DFT),
with the many-body treatment of some effective model, formulated rigorously in the restricted Hilbert
space of ``active'' states.
This is the main idea of realistic
modeling of complex oxide materials.
The purpose of this project is twofold:
(1)
To incorporate the physics of Coulomb correlations,
which is greatly oversimplified
in conventional LDA;
(2)
To provide a transparent physical picture for electronic and magnetic properties of complex compounds.
In this sense, the realistic modeling can be regarded as supplementary approach to more conventional
first-principles electronic structure calculations, which are currently on the rise.
Thus, the first step of the project is construction of the low-energy model
(typically, the multiorbital Hubbard model) for the states near the Fermi level:
\begin{equation}
\hat{\cal{H}} = \sum_{{\bf RR}'} \sum_{\alpha \beta}
t_{{\bf RR}'}^{\alpha \beta}\hat{c}^\dagger_{{\bf R}\alpha}
\hat{c}^{\phantom{\dagger}}_{{\bf R}'\beta} +
\frac{1}{2}
\sum_{\bf R} \sum_{\alpha \beta \gamma \delta} U_{\alpha \beta
\gamma \delta} \hat{c}^\dagger_{{\bf R}\alpha} \hat{c}^\dagger_{{\bf R}\gamma}
\hat{c}^{\phantom{\dagger}}_{{\bf R}\beta}
\hat{c}^{\phantom{\dagger}}_{{\bf R}\delta},
\label{eqn.ManyBodyH}
\end{equation}
which would include the effect of other (``inactive'') states
in the definition of
the model parameters of the Hamiltonian (\ref{eqn.ManyBodyH}). All these parameters
are derived totally from the ``first-principles'' on the basis of DFT.
We would like to emphasize that,
although the derivations are
inevitably based on some approximations,
we do not use any adjustable parameters apart from these
approximations.
The procedure of constructing the
model Hamiltonian was described in details in the review article \cite{review2008}. Briefly,
in order to derive
the one-electron part ($t_{{\bf RR}'}^{\alpha \beta}$), we use the generalized
downfolding method. For the isolated low-energy bands, this procedure is exact, as it is
clearly seen in Fig.~\ref{fig.YVO3LDA} from the comparison of the
original band structure, obtained in the linear muffin-tin orbital method (LMTO) \cite{LMTO}, and the one after
the downfolding. The parameters of screened Coulomb interactions
($U_{\alpha \beta \gamma \delta}$) are typically obtained by combining the
constrained DFT technique \cite{Dederichs} with the random-phase approximation \cite{Ferdi04}.
The latter is very efficient for treating the screening of correlated electrons by
themselves \cite{PRL05}, which can dramatically reduce the effective Coulomb interactions in the
low-energy bands. For example, the bare Coulomb repulsion between $3d$ electrons is
about
20-25 eV. However, for the low-energy bands in solids this value is
typically reduced till 2-4 eV~\cite{review2008}.
Once the model is constructed, it can be solved by using various many-body techniques.
In the present work, we typically start with the mean-field Hartree-Fock (HF) approximation, and
take into account the correlation interactions by considering the
perturbation theory expansion near the HF ground state.
This procedure may be justified, if the degeneracy of the ground state is lifted
by the crystal distortions \cite{review2008}.
Below, we present examples of realistic modeling for two types of
transition-metal oxides.
\section{Spin-Orbital-Lattice Coupling in vanadates $R$VO$_3$}
\label{sec.AVO3}
The vanadates $R$VO$_3$ (where $R$ is the three-valent, typically rare-earth, element)
have attracted a considerable experimental and theoretical attention.
All these compounds crystallize in the distorted perovskite structure.
For example, considered in the present work YbVO$_3$ can have orthorhombic and
monoclinic modifications, realized at $T$$=$ 15 and 75 K, respectively. However, a relatively
small change of the lattice parameters may cause a dramatic reconstruction
of electronic, and associated to it, magnetic structure. For example, at $T$$=$ 15 and 75 K
YbVO$_3$ forms the so-called G- and C-type antiferromagnetic (AFM) structure,
respectively (Fig.~\ref{fig.YbVO3OO}).
\begin{figure}[!h]
\centering
\includegraphics[width=1.5in]{figure2a.eps} \includegraphics[width=1.5in]{figure2b.eps}
\caption{Distribution of the charge densities associated with the occupied
$t_{2g}$ orbitals (the orbital ordering) realized in the orthorhombic (left)
and monoclinic (right) phase of YbVO$_3$ in the HF approximation.
Different magnetic sublattices associated with two opposite
directions of spins in the AFM structure are shown by different colors.}
\label{fig.YbVO3OO}
\end{figure}
The phenomenon called ``spin-orbital-lattice coupling'' and is typically regarded as the
test for various theories of the electronic structure. First attempts of realistic
modeling of $R$VO$_3$ are summarized in~\cite{review2008}. After that we have
performed a systematic study for the entire series $R$VO$_3$
($R$$=$ La, Ce, Pr, Nd, Sm, Gd, Tb, Ho, Yb, Lu, and Y), using the experimental crystal structure
available in the literature~\cite{RVO3_structure1,RVO3_structure2}. For all considered compounds,
the effective Hubbard model was constructed for $t_{2g}$ bands (Fig.~\ref{fig.YVO3LDA}) in the basis of three
Wannier orbitals per each V-site.
The details will be
published elsewhere. Here we present an example of the ``canonical behavior'' realized in
YbVO$_3$, where the crystal distortion quenches the orbital structure in some
particular configuration (Fig.~\ref{fig.YbVO3OO}),
which uniquely defines the type of the magnetic ground state,
basically via Goodenough-Kanamori rules. In this case, once the degeneracy of
$t_{2g}$-levels is lifted
by the crystal distortion, the correct type of the magnetic ground state can be successfully reproduced
already at the level of HF approximation. For example, the stabilization energy is clearly the largest
for the experimentally observed G- and C-type AFM states when YbVO$_3$ crystallizes in the
orthorhombic and monoclinic structure, respectively (Fig.~\ref{fig.YbVO3E}).
\begin{figure}[!h]
\centering
\includegraphics[width=3.0in]{figure3.eps}
\caption{Stabilization energies of the main AFM states in YbVO$_3$
relative to the ferromagnetic state as obtained in the HF approximation (dark blue area)
and after taking into account the correlation interactions in the second order of
perturbation theory (light blue area) and in the t-matrix theory (hatched area).}
\label{fig.YbVO3E}
\end{figure}
Moreover, the correct AFM ground state is additionally stabilized by correlation interactions,
which are taken into account via perturbation-theory expansion near the HF solutions. We have considered
two such technique for the total energy~\cite{JETP07}: one is the regular second-order perturbation theory and
the other one is the t-matrix theory.
Both of them provide very consistent explanation for YbVO$_3$, although the
correlation energies obtained in the t-matrix theory for the AFM states are
systematically smaller due to the higher-order correlation effects, which are
included to the t-matrix, but not to the second-order perturbation theory.
\section{Inversion Symmetry Breaking in Manganites $R$MnO$_3$}
\label{sec.multiferroicity}
The multiferroic manganites, such as BiMnO$_3$ and TbMnO$_3$,
are currently under very intensive investigation. The multiferroicity means that
the magnetic order in certain system without the inversion symmetry coexists with some finite ferroelectric polarization.
The coupling of these two order parameters
provides a unique opportunity to control the magnetic properties by applying the
electric field and vice versa. From this point of view, the most interesting is the situation,
where the inversion symmetry is broken by
the magnetic degrees of freedom.
In order to study the microscopic origin of the inversion symmetry breaking in manganites, we have
constructed the effective model in the basis of
three $t_{2g}$ and two $e_g$ orbitals at each Mn-site \cite{JPSJ,JETP09}.
The main results can be summarized as follows.
The mechanism of the inversion symmetry breaking is related to the
behavior of interatomic magnetic interactions,
which depends on the distribution of occupied
$e_g$ orbitals (Fig.~\ref{fig.OrbitalOrderManganites}).
\begin{figure}[!h]
\centering
\includegraphics[width=3.0in]{figure4.eps}
\caption{Typical distribution of the charge densities associated with the occupied
$e_g$ orbitals (the orbital ordering) realized in the monoclinic (BiMnO$_3$, left)
and orthorhombic (LaMnO$_3$, right) manganites
with the notations of main magnetic interactions.}
\label{fig.OrbitalOrderManganites}
\end{figure}
Although
the orbital ordering differs dramatically in monoclinic and
orthorhombic manganites, the basic idea is rather generic: the nearest-neighbor interactions $J_{NN}$
always compete with some long-range AFM interactions $J_{LR}$. The existence of $J_{LR}$ is
related to the fact that, due to the screening, the on-site Coulomb repulsion $U$
is not particularly large (about 2.2 eV in manganites). Therefore, other mechanisms
of magnetic interactions, besides conventional
superexchange (being of the order of $1/U$), can be also operative. These mechanisms,
which are of the higher orders than $1/U$, together with the form of the orbital ordering,
naturally explain the large values of $J_{LR}$
in some particular bonds \cite{JPSJ}. When $J_{LR}$ becomes larger than $J_{NN}$,
they can lead to the
formation of complex magnetic structures with the broken inversion symmetry.
In monoclinic BiMnO$_3$, this is
the collinear AFM structure, where each of the two sublattices (1,2) and (3,4) are
coupled antiferromagnetically
(see Fig.~\ref{fig.OrbitalOrderManganites} for the notations of Mn-sites)~\cite{JETP09}.
In TbMnO$_3$, the ground state is the incommensurate spin-spiral,
where the directions of spins in the ${\bf ab}$-planes vary as
${\bf e}_{\bf R}$$=$$(\cos {\bf qR},\sin {\bf qR}, 0)$, and the spin moments between the
planes are coupled antiferromagnetically. Such a magnetic structure can be easily calculated
by using the generalized Bloch theorem, which combines the lattice translations with the
spin rotations~\cite{Sandratskii}. The total energy minimum
corresponds to the propagation along the orthorhombic ${\bf b}$-axis, in agreement with the
experiment. However, the obtained value $q_b$$=$$0.675$ (in units $\pi/b$, Fig.~\ref{fig.TbNbO3Eq}),
is substantially larger than
experimental $q_b$$=$ $0.25$~\cite{Arima} and $0.28$~\cite{Kimura}, reported for
magnetic structures with the moments lying in the ${\bf ab}$- and ${\bf bc}$-plane,
respectively.
\begin{figure}[!h]
\centering
\includegraphics[width=3.0in]{figure5.eps}
\caption{Dependence of the total energy
(measured in meV per one formula unit)
on the homogeneous spin-spiral
vector ${\bf q}$ obtained in the Hartree-Fock approximation for TbMnO$_3$ without
the relativistic spin-orbit coupling.}
\label{fig.TbNbO3Eq}
\end{figure}
In order to resolve this discrepancy,
we consider the relativistic spin-orbit interaction (SOI). In BiMnO$_3$,
the SOI is responsible for the spin canting away from the collinear AFM structure
and formation of the net ferromagnetic moment. Thus, the ferroelectric response in BiMnO$_3$, caused by the
hidden AFM order, coexists with the ferromagnetism and can be controlled by the magnetic field~\cite{JETP09}.
The magnetic interactions of the relativistic origin play an important role also in TbMnO$_3$~\cite{Mochizuki}.
In the search for the true magnetic ground state of TbMnO$_3$, we have investigated several magnetic structures with
different periodicity. The structure corresponding to the lowest energy is shown in Fig.~\ref{fig.TbNbO3SO}.
\begin{figure}[!h]
\centering
\includegraphics[width=3.0in]{figure6.eps}
\caption{Distribution of magnetic moments in the ${\bf ab}$-plane
obtained in the HF approximation with SOI
for TbMnO$_3$. The angles formed by Mn-moments in the bonds Mn-O-Mn are specified by numbers.}
\label{fig.TbNbO3SO}
\end{figure}
The periodicity of this structure
is described by $q_b$$=$$0.25$, in agreement with the experiment~\cite{Arima}. Nevertheless,
it is no longer the
uniform spin-spiral.
It would be interesting to check our finding experimentally.
\section{Conclusions}
\label{sec.conclusions}
The realistic modeling combines the accuracy and predictable power of first-principles electronic structure calculations with the flexibility and insights of the model analysis. The first applications of this method are very encouraging. We hope that these ideas will continue to develop to become a powerful tool for theoretical analysis of complex oxide materials and other strongly correlated systems.
\section*{Acknowledgment}
The work is partly supported by Grant-in-Aid for Scientific
Research (C) No. 20540337 and
in Priority Area ``Anomalous Quantum Materials''
from the
Ministry of Education, Culture, Sport, Science and Technology of
Japan.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\label{sec:introduction}
The discovery of frequent patterns (\emph{motifs}) in biological
sequences has attracted wide interest in recent years, due to the
understanding that sequence similarity is often a necessary condition
for functional correlation. Among other applications, motif discovery
proves an important tool for identifying regulatory regions and
binding sites in the study of functional genomics. From a
computational point of view, a major complication for the discovery of
motifs is that they may feature some sequence variation without loss
of function. The discovery process must therefore target
\emph{approximate motifs}, whose occurrences are similar but not
necessarily identical. Approximate motifs are often modeled through
the use of the \emph{don't care} character in certain positions, which
is a wild card matching all characters of the alphabet, called \emph{solid
characters} \cite{Parida08}.
Finding interesting approximate motifs is computationally
challenging. As the number of don't cares increases
and/or the minimum frequency threshold decreases, the output may
explode combinatorially, even if the discovery targets only maximal
motifs---a subset of the motifs which implicitly represents the
complete set. Moreover, even when the final output is not too large,
partial data during the inference of target motifs might lead to
memory saturation or to extensive computation during the intermediate
steps.
\sloppy
A large body of literature in the last decade has dealt with
efficient motif discovery
\cite{Parida00,ApostolicoP04,PisantiCGS05,ApostolicoT07a,Ukkonen07,MorrisNU08,ArimuraU08,ApostolicoT08,ApostolicoCP09},
and an excellent survey of known results can be found in the book
\cite{Parida08}. In order to alleviate the computational burden of
motif extraction and to limit the output to the most promising or interesting discoveries, some works combine the traditional use of a
frequency threshold with restrictions on the flexibility of the
extracted motifs, often captured by limitations on the number of
occurring don't cares.
In a recent work, Apostolico et al.~\cite{ApostolicoCP09} study the
extraction of \emph{extensible motifs}, comprising standard don't
cares and extensible wild cards. The latter are spacers of variable
length that can take different size (within pre-specified limits) in
each occurrence of the motif. An efficient tool, called \textsc{varun}, is
devised in \cite{ApostolicoCP09} for extracting all maximal extensible
motifs (according to a suitable notion of maximality defined in the
paper) which occur with frequency above a given threshold $\sigma$ and
with upper limits $D$ on the length of the spacers. \textsc{varun}\ returns
the extracted motifs sorted by decreasing z-score, a widely adopted
statistical measure of interestingness. The authors demonstrate the
effectiveness of their approach both theoretically, by proving that
each maximal motif features the highest z-score within the class of
motifs it represents, and experimentally, by showing that
the returned top-scored motifs comprise biologically relevant ones
when run on protein families and \textsc{dna}\ sequences.
A slightly more general way of limiting the number of don't cares in a
motif has been explored in \cite{RigoutsosF98}. The authors define $\langle L,W
\rangle$ motifs, for $L \leq W$, where at least $L$ solid characters
must occur in each substring of length $W$ of the motif. They propose a
strategy for extracting $\langle L,W \rangle$ motifs which are also
maximal, although their notion of maximality is not internal to the
class of $\langle L,W \rangle$ motifs. As a consequence, the algorithm
is not complete, since it disregards all those $\langle L,W \rangle$
motifs that are subsumed by a maximal non-$\langle L,W \rangle$ one.
\paragraph*{\bf Our results.}
Our work focuses on the discovery of \emph{rigid motifs}, which
contain blocks of solid characters (solid blocks) separated by one
or more don't cares. We propose a more general approach for
controlling the number of don't cares in rigid motifs. Specifically,
we introduce the notion of \emph{dense motif}, a frequent pattern
where the fraction of solid characters is above a given threshold. Our
density notion is more flexible and general than the one considered in
\cite{Parida08,ApostolicoCP09}, since it allows for arbitrarily long
runs of don't cares as long as the fraction of solid characters in the
pattern is above the threshold. We define a natural notion of
\emph{maximality} for dense patterns and devise an efficient
algorithm, called \textsc{madmx}\ (pronounced \emph{Mad Max}), which performs
complete \textsc{ma}ximal \textsc{d}ense \textsc{m}otif
e\textsc{x}traction from an input sequence, with respect to
user-specified frequency and density thresholds.
The key technical result at the core of our extraction strategy is a
closure property which affords the complete generation of all maximal
dense motifs in a breadth-first fashion, through an
\emph{apriori}-like strategy \cite{AgrawalS94-1}, starting from a
relatively small set of solid blocks, and then repeatedly applying a
suitable combining operator, called \emph{fusion}, to pairs of
previously generated motifs. In this fashion, our strategy avoids the
generation and consequent storage of intermediate patterns which are
not in the output set, which ensures time and space complexities
polynomial in the combined size of the input and the output.
We performed a number of experiments on \textsc{madmx}\ to assess the
biological significance of maximal dense motifs and to compare \textsc{madmx}\
against its most recent and close competitor \textsc{varun}. For the first objective, we
used \textsc{madmx}\ to extract maximal dense motifs from a number of human
\textsc{dna}\ fragments. We compared the output set against those in
\texttt{RepBase} \cite{repbase}, the largest repository of repetitive
patterns for eukaryotic species, using \textsc{repeatmasker}\ \cite{repeatmasker},
a popular tool for masking repetitive \textsc{dna}. The experiments show that
all of our returned motifs are occurrences of patterns in
\texttt{RepBase}, and \emph{fully} characterize the family of \textsc{sine}/\textsc{alu}\
repeats (and partially the \textsc{line}/\textsc{l1}\ family). This provides evidence
that the notion of density, when applied to rigid motifs, captures
biological significance.
Next we compared the z-score performance of \textsc{madmx}\ and \textsc{varun}. We ran
both algorithms on several families of \textsc{dna}\ fragments, limiting
\textsc{varun}\ to the generation of rigid motifs and setting the parameters
so as to obtain comparable output sizes, with motifs listed by
decreasing z-score. The experiments show that the top-$m$
highest-ranking motifs returned by \textsc{madmx}\ almost always feature
higher z-scores than the corresponding top-$m$ ones returned by \textsc{varun},
even for large values of $m$, with only a modest increase in running
time, which may be partly due to the fact that coding of \textsc{madmx}\ is yet
to be optimized. In fairness, we must remark that \textsc{varun}\ deals also
with extensible motifs while \textsc{madmx}\ only targets rigid motifs.
The paper is organized as follows. In Section~\ref{sec:prelim} several
technical definitions and properties of motifs with don't cares are
given. Section~\ref{sec:maximal} proves the closure property at the
base of \textsc{madmx}\ and provides a high-level description of the
algorithm. In Section~\ref{sec:exp}, the experimental validation of
\textsc{madmx}\ is presented.
\section{Preliminary Definitions and Properties}
\label{sec:prelim}
Let $\Sigma$ be an alphabet of $m$ characters and let $s = s[0] s[1]
\dots s[n-1] $ be a string of length $n$ over $\Sigma$. We use $s[i
\dots j ]$ to denote the substring $s[i]$ $s[i+1]$ $\cdots$ $s[j]$ of
$s$, for $i\leq j$. Characters in $\Sigma$ are also called \emph{solid
characters}. We use $\circ \not\in \Sigma$ to denote a distinguished
character called \emph{wild card} or \emph{don't care} character. Let
$\epsilon$ denote the empty string. A \emph{pattern} $x$ is a string
in $\{\epsilon\} \cup \Sigma \cup \Sigma(\Sigma \cup \{ \circ \})^*
\Sigma$. However, whenever necessary, we will assume that patterns are
implicitly padded to their left and right with arbitrary sequences of don't care
characters.
Given two patterns $x, y$ we say that $y$ is \emph{more
specific} than $x$, and write $x \preceq y$, iff for every $i \geq 0$
either $x[i]=y[i]$ or $x[i] = \circ$.
Given two patterns $x, y$ we say that $x$ \emph{occurs in $y$ at
position $\ell$} iff $x \preceq y[\ell \ldots \ell+|x|-1]$: we also
say that $y$ \emph{contains}~$x$.
For a string $s$, the \emph{location list} ${\cal L}_x$ of a pattern
$x$ in $s$ is the complete set of positions at which $x$ occurs in
$s$. We refer to $f(x) = |{\cal L}_x|$ as the \emph{frequency} of
pattern $x$ in $s$. (Note that $f(\epsilon)=n$.) As in
\cite{Ukkonen07}, the \emph{translated representation} of the location
list ${\cal L}_x=\{l_0,l_1,l_2,\dots,l_k\}$ is $\tau{(\cal
L}_x)=\{l_1-l_0,l_2-l_0,\dots,l_k-l_0\}$. Given two patterns $x,y$,
we say that $y$ \emph{subsumes} $x$ in $s$ if $f(x)=f(y)$ and $y$
contains $x$. As a consequence, if $y$ subsumes $x$ then $\tau({\cal
L}_x)=\tau({\cal L}_y)$. A pattern $x$ is \emph{maximal} if it is
not subsumed by any other pattern $y$. (We observe that this notion of
maximality coincides with that of \cite{PisantiCGS05}.) Given a
pattern $x$, its \emph{maximal extension} ${\cal M}(x)$ is the maximal
pattern that subsumes $x$, which can be shown to be unique
\cite{PisantiCGS05}.
In what follows, we call \emph{solid block} a string in $\Sigma^{+}$
and a \emph{don't care block} a string in $\circ^{+}$. Furthermore,
given a pattern $x$, $\mbox{dc}(x)$ denotes the number of don't care
characters contained in $x$.
\begin{definition}
The \emph{density} $\delta(x)$ of $x$ is:
$\delta(x)=1-\mbox{dc}(x)/|x|$. Given a (density) threshold $\rho$,
$0 <\rho \leq 1$, we say that a pattern $x$ is \textit{dense} if
$\delta(x) \geq \rho$.
\end{definition}
Note that a solid block is a dense pattern with respect to every
threshold $\rho$.
It is reasonable to concentrate the attention on dense patterns that
are not subsumed by any other dense pattern, since they are the most interesting dense representatives
in the equivalence classes induced by ``sharing'' the same translated representation; these representatives are defined below.
\begin{definition}\label{def:maxdense}
A dense pattern $x$ is a \emph{maximal dense pattern} in $s$ if it is
not subsumed by any other dense pattern $x' \neq x$.
\end{definition}
Observe that a maximal dense pattern $x$ needs not be a maximal
pattern in the general sense, since ${\cal M}(x)$ might be a nondense
pattern. However, every dense pattern $x$ is subsumed by \emph{at
least} one maximal dense pattern. In fact, all of the maximal dense
patterns that subsume $x$ are dense substrings of ${\cal M}(x)$, namely,
those that contain $x$ and are not substrings of any other dense
substring of ${\cal M}(x)$. We want to stress that there might be several
maximal dense patterns that subsume~$x$. As an example, for
$\rho=2/3$, the dense pattern $x=\mathtt{B}$ in the string
$S=\mathtt{AdBeCfAgBhC}$ is subsumed by maximal dense patterns
$\mathtt{A} \circ \mathtt{B}$ and $\mathtt{B} \circ \mathtt{C}$, while
${\cal M}(x) = \mathtt{A} \circ \mathtt{B} \circ \mathtt{C}$ is not dense.
\begin{definition}
Given a frequency threshold $\sigma$ and a density threshold $\rho$, a
pattern $x$ is a \emph{dense maximal motif} in $s$ if $x$ is a maximal
dense pattern in $s$ with respect to $\rho$, and $f(x)\geq \sigma$.
A dense maximal motif for $\rho=1$ is also referred to as
\emph{maximal solid block}.
\end{definition}
\noindent
\textbf{Problem of interest.}
We are given an input string $s$, a frequency threshold $\sigma$,
and a density threshold $\rho$. Find all the maximal dense motifs in $s$.
\smallskip
In the rest of the paper, we will omit referencing
the input string $s$ when clear from the context. An important
property of maximal dense patterns, which we will exploit in our
mining strategy, is that all of their solid blocks are maximal solid
blocks. This property is stated in the following proposition whose
proof, omitted for brevity, extends a similar result holding for
arbitrary maximal patterns \cite{Ukkonen07,Pisanti02}.
\begin{proposition}\label{maximalpatterns}
Let $x$ be a maximal dense pattern with respect to a density threshold
$\rho$, and let $b=x[i \ldots j]$ be a solid block
in $x$ such that $x[i-1]=x[j+1]=\circ$ and $j \geq i$. Then, $b$ is a maximal solid block.
\end{proposition}
\section{An Algorithm for MAximal Dense Motif eXtraction}
\label{sec:maximal}
In this section we describe our algorithm, called \textsc{madmx}\
(pronounced \emph{Mad Max}), for \textsc{ma}ximal \textsc{d}ense
\textsc{m}otif e\textsc{x}traction. The algorithm adopts a
breadth-first
\emph{apriori}-like strategy \cite{AgrawalS94-1}, similar in spirit to the
one developed in \cite{ApostolicoCP09}, using maximal solid blocks
as building blocks by Proposition~\ref{maximalpatterns}. \textsc{madmx}\ operates by repeatedly
combining together, in a suitable fashion, pairs of maximal dense
motifs, and extracting from the combinations
less frequent maximal dense motifs.
A key notion for the algorithm, underlying the aforementioned
combining operations, is the \textit{fusion} of characters/patterns.
\begin{definition}
Given three characters $c,c_1,c_2 \in \Sigma \cup \{\circ\}$, we say that $c$ is the fusion of $c_1$ and $c_2$, and write $c = c_1 \bigtriangledown c_2$,
if one of the following holds:
\begin{enumerate}
\item $c = c_1 = c_2$;
\item $c_1 = \circ$, $c = c_2 \neq \circ$;
\item $c = c_1 \neq \circ$, $c_2 = \circ$.
\end{enumerate}
\end{definition}
The above notion of fusion generalizes to patterns as follows.
\begin{definition}
Given three patterns $x,y,z$ and an integer $d$, we say that $z$ is
the \emph{$d$-fusion of $x$ and $y$}, and write $z=x
\bigtriangledown_d y$, if $z$ can be obtained by removing the leading and
trailing don't care characters from the pattern $m$ defined as $m[i]
= x[i+d] \bigtriangledown y[i]$, for all indices $i$.
\end{definition}
\iffalse
Note that if $d>|x|$ we have $x \bigtriangledown_d y = x\circ^{d'}y$
for a certain $d'>0$, while if $d< - |y|$ we have $x
\bigtriangledown_d y = y\circ^{d''}x$ for a certain $d''>0$.
\fi
The breadth-first strategy adopted by our algorithm crucially
relies on the following theorem, which highlights the structure of dense
motifs:
\begin{theorem}\label{theorem:completeness}
Let $x$ be a maximal dense motif with $\mbox{dc}(x)>0$. Then:
\begin{enumerate}
\item[(a)] there exists a maximal solid block $b$ in $x$
such that ${\cal M}(x)={\cal M}(b)$, or
\item[(b)] there exist two maximal dense motifs $y_1,y_2$ such that:
\begin{itemize}
\item ${\cal M}(x)={\cal M}(y_1 \bigtriangledown_d y_2)$, for some $d$;
\item there are two maximal solid blocks $b_1,b_2$ in $x$ and an
integer $\hat{d}>0$ such that $b_1$ is a maximal solid block in
$y_1$, $b_2$ is a maximal solid block in $y_2$, and $b_1
\circ^{\hat{d}} b_2$ is contained in $y_1\bigtriangledown_dy_2$;
\item $f(x) < \min\{f(y_1),f(y_2)\}$;
\end{itemize}
\end{enumerate}
\end{theorem}
For the proof of Theorem~\ref{theorem:completeness} we need to define
another type of pattern combination, namely the operation of
\textit{merge} between two patterns, which is similar to the one
introduced in \cite{PisantiCGS05}. Given two characters $c_1,c_2$,
we define the operator $\oplus$ between them such
that $c_1 \oplus c_2 = \circ$, if $c_1 \neq c_2$, and
$c_1 \oplus c_2 = c_1 = c_2$, otherwise.
\begin{definition}
Given two patterns $x,y$ and an integer $d$, the $d$-merge of $x$ and
$y$ is the pattern $z = x \oplus_d y$ which can be obtained by removing
all leading and trailing don't cares from the pattern $m$
defined as $m[i] = x[i+d] \oplus y[i]$ for all $i$.
\end{definition}
We want to stress the difference between the notions of merging and
fusion: the merge of two patterns $x,y$ is always well defined and more
general than $x,y$, while the fusion of $x,y$ may not exist and, if it does,
is more specific than $x,y$.
For the proof of Theorem~\ref{theorem:completeness} we also need the
property established by the following lemma.
\begin{lemma}\label{lem:maxmax}
Let $x$ and $y$ be maximal patterns, and $d$ be an integer such that $z = x
\oplus_d y \neq \epsilon$. Then $z$ is a maximal pattern. Moreover, if
$z \neq x$ (resp., $z \neq y$) then $f(z) > f(x)$ (resp., $f(z) >
f(y)$).
\end{lemma}
\begin{proof}
First we prove that $z$ is maximal. By contradiction, suppose that
this is not the case. Then, there exists a position $i$ such that
$z[i]=\circ$ and we can replace the $\circ$ with a solid character $c$
without decreasing the frequency of the pattern. (Note that the
position of the substitution can be to the left of the first character
in $z$ or to the right of the last character in $z$.) Since $x$ and
$y$ are more specific than $z$, to every occurrence of $x$ and $y$ in
the string corresponds an occurrence of $z$. Hence, every occurrence of
$x$ (resp., $y$) in the string, contains $c$ in its $i+d$th (resp.,
$i$th) position. Therefore, by maximality of $x$ and $y$, it must be
$z[i]=x[i+d]=y[i]=c$, which is a contradiction. The relations between
the frequencies of $x, y$ and $z$ follow trivially by their
maximality. \qed
\end{proof}
We are now ready to prove the theorem.
\begin{proof}[Theorem~\ref{theorem:completeness}]
Given a pattern $x$ and two nonnegative integers $i \leq j$, we let
$x^*[i\dots j]$ denote the pattern obtained by removing all the
leading and trailing don't care characters from $x[i\dots j]$. Since
$x$ is a maximal dense pattern and $dc(x)>0$, it is easy to see that
there exist two dense patterns $x_1,x_2$ and an integer $d>0$ such
that $x = x_1 \circ^d x_2$, hence there exists an index
$s_1>0$ such that $x^*[0 \dots s_1 - 1]$ and $x^*[s_1+1 \dots |x|-1]$
are dense. We call these two patterns the \textit{level-1
decomposition} of $x$ (observe that many such decompositions may
exist). Also, we let $\ell_1 = 0$ and $r_1 =
|x|-1$. Now, consider the following iterative process:
\begin{enumerate}
\item\label{step1} If in the level-$i$ decomposition of $x$ both
$x^*[\ell_i\dots s_i-1]$ and $x^*[s_i+1\dots r_i]$ have frequency
strictly greater than $f(x)$, \textit{or} at least one of
$x^*[\ell_i\dots s_i-1]$ and $x^*[s_i+1\dots r_i]$ is a solid block
with frequency equal to $f(x)$, then terminate;
\item Otherwise, let $y = x^*[\ell_{i+1}\dots r_{i+1}]$ be (an
arbitrary) one of $x^*[\ell_i\dots s_i-1]$ or $x^*[s_i+1\dots r_i]$
which is not a solid block and has frequency equal to $f(x)$. Since
$y$ is dense, there exists an index $s_{i+1}$, $\ell_{i+1} <
s_{i+1} < r_{i+1}$ such that $x^*[\ell_{i+1} \dots s_{i+1}-1]$ and
$x^*[s_{i+1}+1 \dots r_{i+1}]$ are both dense. Call these two
patterns the level-$(i+1)$ decomposition of $x$. Set $i = i+1$ and go to Step~\ref{step1}.
\end{enumerate}
Assume that the decomposition process ends by finding a solid block
$b$ that is a solid block in $x$ and has $f(b)=f(x)$. Then, ${\cal M}(b) =
{\cal M}(x)$ and the theorem follows. Otherwise, at the last level $j$ of
the decomposition, we have that $f(x) < \min\left\{f(x^*[\ell_j\dots
s_j-1]), f(x^*[s_j+1\dots r_j])\right\} $. In this latter case, as
explained in Section~\ref{sec:prelim} (after
Definition~\ref{def:maxdense}), we can determine two maximal dense
patterns $y_1,y_2$ such that $y_1$ contains $x^*[\ell_j\dots s_j-1]$,
$y_2$ contains $x^*[s_j+1\dots r_j]$, and with ${\cal M}(y_1) =
{\cal M}(x^*[\ell_j\dots s_j-1])$ and ${\cal M}(y_2) = {\cal M}(x^*[s_j+1\dots
r_j])$. Since $f(y_1) = f(x^*[\ell_j\dots s_j-1])$ and $f(y_2) =
f(x^*[s_j+1\dots r_j])$, we have that $f(x) <
\min\left\{f(y_1),f(y_2)\right\}$. Observe that by construction there
must exist two solid blocks $b_1,b_2$ in $x$ and an integer $\hat{d}$
such that $b_1$ is a solid block in $y_1$, $b_2$ is a solid block in
$y_2$, and $b_1 \circ^{\hat{d}} b_2$ is a sequence of two solid blocks
in $x$. In fact, $b_1$ (resp., $b_2$) is the last (resp., the first) solid
block of $x^*[\ell_j\dots s_j-1]$ (resp., $x^*[s_j+1\dots r_j]$).
Next, we show that there exists a $d$ such that the $d$-fusion $y_1
\bigtriangledown_d y_2$ is well defined, contains $b_1 \circ^{\hat{d}}
b_2$, and ${\cal M}(y_1 \bigtriangledown_d y_2) = {\cal M}(x)$. We proceed as
follows. Let us ``align'' ${\cal M}(x)$ and $y_1$ so to match the occurrences
of $b_1$ in both patterns. Then, for a certain integer
$p$, ${\cal M}(x)[i+p]$ corresponds to $y_1[i]$. Assume, for the sake
of contradiction, that there exists an index $j$ such that
${\cal M}(x)[j+p]$ is not more specific than $y_1[j]$. Then,
Lemma~\ref{lem:maxmax} implies that $z = {\cal M}(x) \oplus_p {\cal M}(y_1) \neq
{\cal M}(y_1)$, which contains $x^*[\ell_j\dots s_j-1]$, is maximal and has
frequency strictly greater than $f(y_1)$, which is impossible because
we have chosen $y_1$ such that ${\cal M}(x^*[\ell_j\dots s_j-1]) =
{\cal M}(y_1)$ and therefore $f(x^*[\ell_j\dots s_j-1])= f(y_1)$. Therefore,
${\cal M}(x)$ contains $y_1$. A similar argument shows that
${\cal M}(x)$ contains $y_2$.
Since $y_1$ and $y_2$ are contained in ${\cal M}(x)$,
there must exist a $d$ such that $y_1 \bigtriangledown_d y_2$
is well defined and can be aligned with ${\cal M}(x)$ in such a way to match
the blocks $b_1$ and $b_2$ of $y_1$ and $y_2$ with the corresponding
blocks in ${\cal M}(x)$. Moreover, ${\cal M}(x)$ contains
$y_1 \bigtriangledown_d y_2$, hence $f(y_1 \bigtriangledown_d y_2)
\geq f({\cal M}(x)) = f(x)$. However, since $y_1 \bigtriangledown_d y_2$
contains both $x^*[\ell_j\dots s_j-1]$ and $x^*[s_j+1\dots r_j]$,
it contains also $x^*[\ell_j\dots r_j]$, which, by the decomposition
process, has frequency equal to $f(x)$. Therefore,
$f(y_1 \bigtriangledown_d y_2) \leq f(x)$, and the theorem follows since
$f(y_1 \bigtriangledown_d y_2) =f(x)$. \qed
\end{proof}
In essence, Theorem~\ref{theorem:completeness} guarantees that we can
find any maximal dense motif $x$ either within ${\cal M}(b)$, for some
maximal solid block $b$, or by $d$-fusing two higher-frequency maximal
dense motifs $y_1,y_2$, for some $d$, finding
$z={\cal M}(y_1\bigtriangledown_d y_2)$ and then possibly ``trimming'' $z$
on both sides to obtain $x$.
\begin{figure}[t]
\begin{center}
\begin{algorithm}[H]
\begin{small}
\SetKw{Whilef}{while}
\SetKw{Dof}{do}
\SetKw{Thenf}{then}
\SetKw{Forf}{for}
\SetKw{Iff}{if}
\caption{\textsc{madmx} \label{alg:maxdense}}
\KwIn{String $s$, frequency threshold $\sigma$, density threshold $\rho$}
\KwOut{Maximal dense motifs}
\BlankLine
$\mathit{previous} \leftarrow \emptyset$, $\mathit{current} \leftarrow \emptyset$, $\mathit{next} \leftarrow \emptyset$ \;
$\mathit{blocks}$ $\gets$ maximal solid blocks of $s$ with frequency $\geq \sigma$\; \label{codemax:init}
\Forf {\bf each} $b \in \mathit{blocks}$ \Dof \label{codemax:forinit}\\
\Indp
find ${\cal M}(b)$ \;
\label{codemax:2} ${\cal DM}\leftarrow$ extractMaximalDense(${\cal M}(b)$)\;
\Forf {\bf each} $x \in {\cal DM}$ \Dof {$\mathit{current} \leftarrow \mathit{current} \cup \{x\}$\; \label{codemax:forend}}
\Indm
\Whilef $\mathit{current} \neq \emptyset$ \Dof \label{codemax:whileinit}\\
\Indp
\Forf {\bf each} $x_1 \in \mathit{current}$ \Dof \label{codemax:1}\\
\Indp
\Forf {\bf each} $x_2 \in \mathit{previous} \cup \mathit{current}$ \Dof\label{codemax:1a} \\
\Indp
\Forf {\bf each} $d$ s.t. $z = x_1
\bigtriangledown_d x_2$ is a valid fusion \Dof \label{codemax:fusion}\\
\Indp
find ${\cal M}(z)$\;
\label{codemax:3} ${\cal DM} \leftarrow$ extractMaximalDense(${\cal M}(z)$)\;
\Forf {\bf each} $x \in {\cal DM}$ \Dof\\
\Indp
\lIf{$f(x) \geq \sigma$ and $x \notin \mathit{previous} \cup \mathit{current} $}{$\mathit{next} \leftarrow \mathit{next} \cup \{x\}$\;}
\Indm
\Indm
\Indm
\Indm
$\mathit{previous} \gets \mathit{previous} \cup \mathit{current}$\;
$\mathit{current} \gets \mathit{next}$; $\mathit{next}\gets\emptyset$\; \label{codemax:whileend}
\Indm
\Return $\mathit{previous}$\;
\end{small}
\end{algorithm}
\end{center}
\caption{Pseudocode of algorithm \textsc{madmx}.} \label{code:maxdense}
\end{figure}
\sloppy Algorithm~\textsc{madmx}, whose pseudocode is reported in
Figure~\ref{code:maxdense}, implements the strategy inspired by
Theorem~\ref{theorem:completeness}. It employs three (initially empty)
sets \emph{previous}, \emph{current}, and \emph{next}. In
Line~\ref{codemax:init}, the algorithm first stores the maximal solid
blocks $b$ in $s$ for the given frequency in the set \emph{blocks}
(see Section~\ref{sec:prelim}). Then, it extracts all of the
appropriate maximal dense motifs from ${\cal M}(b)$ in
Lines~\ref{codemax:forinit}--\ref{codemax:forend}, using the function
extractMaximalDense, as implied by
Theorem~\ref{theorem:completeness}(a). Finally,
Lines~\ref{codemax:whileinit}--\ref{codemax:whileend} implement the
strategy as implied by Theorem~\ref{theorem:completeness}(b).
(In Line~\ref{codemax:fusion} a $d$-fusion $y_1\bigtriangledown_d y_2$ is considered \emph{valid}
if it satifies the second property of Theorem~\ref{theorem:completeness}(b).)
An important issue for the efficiency of \textsc{madmx}\ is that it needs to
compute the exact frequency of each generated pattern. For what
concerns the fusion operation of two patterns $x_1, x_2$ in
Line~\ref{codemax:fusion}, observe that a simple computation on the
pairs $(\ell_1,\ell_2)\in {\cal L}_{x_1} \times {\cal L}_{x_2}$ is
sufficient to yield the frequencies of all the valid fusions of two
patterns. However, given $z =x_1\bigtriangledown_d x_2$, for a
maximal dense pattern $w$ which does not contain $z$ in its entirety,
we can only conclude that $f(w) \geq f(z)$. We then label the motifs
for which the exact frequencies are known as \emph{final}, and those
for which only a lower bound to their frequencies is known as
\emph{tentative}, and update the lower bounds and the labels during
the execution of the algorithm. Whenever the set \emph{current}
contains no final motifs, we can label as final the tentative motif in
\emph{current} with the highest lower bound to its frequency, and
continue with the generation. The proof of the correctness of this
assumption and further details on the implementation of the algorithm
will be provided in the full version of this extended abstract.
A crude upper bound on the running time of \textsc{madmx}\ can be derived by
observing that, for each pair of dense maximal motifs in output, the
time spent during all the operations concerning that pair is (naively)
$\BO{n^3}$, where $n$ is the length of the input string. If $P$
patterns are produced in output, the overall time complexity is
$\BO{n^3 P^2}$.
\section{Experimental Validation of MADMX}
\label{sec:exp}
We developed a first, non-optimized, implementation of \textsc{madmx}\ in \texttt{C++}
also including an additional feature which eliminates, from the set of
initial maximal solid blocks, those shorter than a given threshold
$min_\ell$. The purpose of this latter heuristics is to speed up motif
generation driving it towards the discovery of (possibly) more
significant motifs, with the exclusion of spurious, low-complexity
ones. (The code is available for download at {\tt
http://www.dei.unipd.it/wdyn/?IDsezione=4534}.)
We performed two classes of experiments to evaluate how significant is
the set of motifs found using our approach. The first class of
experiments, described in Section~\ref{sub:eval-repbase}, compares our
motifs with the known biological repetitions available in
\texttt{RepBase} \cite{repbase}, a very popular genomic database. The
second class of experiments, described in
Section~\ref{sub:eval-z-score}, aims at comparing the motifs extracted
by \textsc{madmx}\ with those extracted by \textsc{varun}\ using the same $z$-score
metric employed in \cite{ApostolicoCP09} for assessing their relative
statistical significance.
\subsection{Evaluating significance by known biological repetitions}
\label{sub:eval-repbase}
\texttt{RepBase} \cite{repbase} is one of the largest repositories of
prototypic sequences representing repetitive \textsc{dna}\ from different
eukaryotic species, collected in several different ways.
\texttt{RepBase} is used as a reference collection for masking and
annotation of repetitive \textsc{dna}\ through popular tools such as
\textsc{repeatmasker}\ \cite{repeatmasker}. \textsc{repeatmasker}\ screens an input
\textsc{dna}\ sequence $s$ for simple repeats and low complexity portions, and
interspersed repeats using \texttt{RepBase}. Sequence comparisons are
performed through Smith-Waterman scoring. \textsc{repeatmasker}\ returns a detailed
annotation of the repeats occurring in $s$, and a modified version of
$s$ in which all of the annotated repeats are masked by a special
symbol (\texttt{N} or \texttt{X}). With the current version of \texttt{RepBase}, on average,
almost 50\% of a human genomic \textsc{dna}\ sequence will be masked by the
program \cite{repeatmasker}.
Most of the interspersed repeats found by \textsc{repeatmasker}\ belong to the
families called \textsc{sine}/\textsc{alu}\ and \textsc{line}/\textsc{l1}: the former are \emph{Short
INterspersed Elements} that are repetitive in the \textsc{dna}\ of
eukaryotic genomes (the Alu family in the human genome); the latter
are \emph{Long Interspersed Nucleotide Elements}, which are typically
highly repeated sequences of 6K--8K bps, containing \textsc{rna}
polymerase II promoters. The \textsc{line}/\textsc{l1}\ family forms about 15\% of the
human genome.
We have conducted an experimental study using \textsc{madmx}\ and \textsc{repeatmasker}\ on
\emph{Human Glutamate Metabotropic Receptors} \textsc{hgmr}~1\ (410277 bps) and
\textsc{hgmr}~5\ (91243 bps) as input sequences. We have downloaded the
sequences from the March 2006 release of the UCSC Genome database
(\texttt{http://genome.ucsc.edu}).
\noindent \textsc{repeatmasker}\ version was open-3.2.7, sensitive mode, with the
query species assumed to be homologous; it ran using
\texttt{blastp} version 2.0a19MP-WashU, and \texttt{RepBase} update
20090120.
The experiments to assess the biological significance of the maximal
dense motifs extracted by \textsc{madmx}\ involved three separate stages. In
the first stage, we ran \textsc{repeatmasker}\ on the input sequences \textsc{hgmr}~1\ and
\textsc{hgmr}~5, searching for interspersed repeats using \texttt{RepBase}. One
of the output files (\texttt{.out}) of \textsc{repeatmasker}\ contains the list of found repeats,
and provides, for each occurrence, the substring $s[i \ldots j]$ of
the input sequence $s$ which is locally aligned with (a substring of) the
repeat.
In the second stage, we ran \textsc{madmx}\ on the same DNA sequences, with
density threshold $\rho=0.8$, frequency threshold $\sigma = 4$, and
$\min_\ell = 15$. In order to filter out simple repeats and low
complexity portions, which are dealt with by \textsc{repeatmasker}\ without
resorting to \texttt{RepBase}, we modified \textsc{madmx}\ eliminating
periodic maximal solid blocks (with short periods), which are the
seeds of simple repeats. Then, we identified the occurrences of the
motifs returned by \textsc{madmx}\ in the input sequences, using \textsc{repeatmasker}\ as
a pattern matching tool (i.e., replacing \texttt{RepBase} with the set
of motifs returned by \textsc{madmx}\ as the database of known repeats). The
underlying idea behind this use of \textsc{repeatmasker}\ was to employ the same
local alignment algorithms, so to make the comparison fairer.
In the third stage, we cross-checked the intervals associated with the
occurrences of the \texttt{RepBase} repeats against those associated
with the occurrences of our motifs. Surprisingly, \textsc{madmx}\ was able to
identify and characterize \emph{all} of the intervals of the known
\textsc{sine}/\textsc{alu}\ repeats in \textsc{hgmr}~1\ and \textsc{hgmr}~5\ (respectively, 56 repeats plus an
extra unclassified for \textsc{hgmr}~1, and 20 plus an extra unclassified for
\textsc{hgmr}~5). The remaining occurrences of the motifs permitted to identify
29 repeats out of 78 of the \textsc{line}/\textsc{l1}\ family in \textsc{hgmr}~1. (A more detailed
account of the whole range of experiments conducted using
\textsc{repeatmasker}\ and the data sets by Tompa et el.\mbox{} and Sandve et al.\mbox{}
will be provided in the full version.)
\subsection{Evaluating significance by statistical z-score ranking}
\label{sub:eval-z-score}
The z-score is the measure of the distance in standard deviations of
the outcome of a random variable from its expectation. Consider a
\textsc{dna}\ sequence $s$ of length $n$ as if it was generated by a
stationary, i.i.d. source with equiprobable symbols; an
approximation to the z-score for a
motif of length $m$ that contains $c$ solid characters and appears $f$
times in $s$ is given by $Z = \frac{f - (n-m+1) \times
p}{\sqrt{(n-m+1) \times p \times (1-p)\,}}$, where $p=(1/4)^c$.
This metric was used in \cite{ApostolicoCP09} to assess the
significance of the motifs extracted by \textsc{varun}\ and to rank them in
the output.
We employed the code for \textsc{varun}\ provided by the authors to extract
the rigid motifs from the \textsc{dna}\ sequences analyzed in
\cite{ApostolicoCP09}. We then ran \textsc{madmx}\ on the same sequences
using the same frequency parameters, and setting the minimum density
threshold $\rho$ in such a way to obtain a comparable yet smaller
output size. In this fashion, we tested the ability of \textsc{madmx}\ to
produce a succinct yet significant set of motifs, by virtue of its
more flexible notion of density.
The results are shown in Table~\ref{tab:madmx&varun}. For \textsc{varun}\ we
used $D=1$, thus allowing at most one don't care between two solid
characters, and ran \textsc{madmx}\ with $min_\ell=1$, so to obtain the
\emph{complete} family of maximal dense motifs. In the table, there
is a row of the table for each sequence (identified in the first
column). Each sequence, whose total length is reported in the second
column, is obtained as the concatenation of a number of smaller
subsequences, reported in the third column. On each sequence, both
tools were run with the same frequency threshold $\sigma$, and the
table reports for both the output size in terms of the number of
motifs returned and the execution time in seconds. Also, for \textsc{madmx}, the table
reports the density threshold $\rho$ used in each experiment.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l|c|c|c||c|c||c|c|c||c|c|c|c|c|}
\hline
& & & & \multicolumn{2}{|c||}{{\textsc{varun}}} &
\multicolumn{3}{|c||}{{\textsc{madmx}}} & \multicolumn{5}{|c|}{best top-$m$ z-scores} \\
\cline{5-14}
name & length & \# & $\sigma$ & \,$|$output$|$\, & time & $\rho$ & \,$|$output$|$\, & time & $m$=10 & $m$=50 & $m$=100 & $m^*$ & $\hat{m}$\\
\hline
\hline
\texttt{ace2} & 500 & 1 & 2 & 1866 & 3s & 0.7 & 1762 & 18s & 10 & 50 & 100 & 1571 & 1067\\
\hline
\texttt{ap1} & 500 & 1 & 2 & 1555 & 1s & 0.7 & 1304 & 5s & 10 & 50 & 100 & 392 & 13 \\
\hline
\texttt{gal4} & 3000 & 6 & 4 & 9764 & 12s & 0.67 & 7606 & 67s & 10 & 49 & 99 & 16 & 16\\
\hline
\texttt{gal4}$^{(*)}$ & 3000 & 6 & 4 & 9764 & 12s & 0.65 & 11733 & 191s & 10 & 50 & 100 & 9764 & 301\\
\hline
\texttt{uasgaba} & 1000 & 2 & 2 & 4586 & 30s & 0.70 & 4194 & 90s & 10 & 50 & 100 & 175 & 175\\
\hline
\end{tabular}
\end{center}
\caption{Results of the comparison with \textsc{varun}.}\label{tab:madmx&varun}
\end{table}
For each experiment, we compared the best top-$m$ z-scores, with
$m=10, 50$, and $100$, as follows. Note that, in general, the top-$m$
motifs found by \textsc{madmx}\ and \textsc{varun}\ differ. Thus, we let $z_{M}^{i}$
(resp., $z_{V}^{i}$) be the z-score of the $i$th motif in
decreasing z-score order obtained
by \textsc{madmx}\ (resp., \textsc{varun}). For each $m$, the table reports how many times it
was $z_{M}^{i} \geq z_{V}^{i}$, for $1 \leq i \leq m$. Also, column
$m^*$ (resp., column $\hat{m}$) gives the maximum $m$ such that
$z_{M}^{i} \geq z_{V}^{i}$ (resp., $z_{M}^{i} > z_{V}^{i}$) for every
$1 \leq i \leq m$.
Even when \textsc{madmx}\ is calibrated to yield a slightly smaller output,
the quality of the motifs extracted, as measured by the z-score, is
higher than those output by \textsc{varun}. Indeed, for sequences
\texttt{ace2} and \texttt{uasgaba} a very large prefix of the
top-ranked motifs extracted by \textsc{madmx}\ features strictly greater
z-scores of the corresponding top-ranked ones extracted by \textsc{varun}. In
fact, for all of the four sequences, at least the thirteen top-ranked
motifs enjoy this property. To shed light on the slightly worse
performance of \textsc{madmx}\ on \texttt{gal4}, we re-ran \textsc{madmx}\ with a
different density threshold, so to obtain a slightly larger output
(see row \texttt{gal4}$^{(*)}$). In this case, the top-$301$ motifs
extracted by \textsc{madmx}\ have z-score strictly greater than the
corresponding motifs extracted by \textsc{varun}, while the execution time
remains still acceptable.
For all runs, the top z-score of a motif discovered by \textsc{madmx}\ is
considerably higher than the one returned by \textsc{varun}. Specifically, on
\texttt{ace2} our best z-score is 387\,763 vs.\mbox{} 12\,027 of
\textsc{varun}; on \texttt{ap1}, we have 12\,027 vs.\mbox{} 1\,490; on
\texttt{gal4} it is 75 vs.\mbox{} 28; on \texttt{gal4}$^{(*)}$ it is
150 vs.\mbox{} 28; on \texttt{uasgaba} we have 134\,532 vs.\mbox{}
67\,059. This reflects the high selectivity of \textsc{madmx}, which is to be
attributed mostly to adoption of a more flexible density constraint.
We must remark that \textsc{madmx}\ (in its current nonoptimized
version) is slower than \textsc{varun}, but it still runs in time
acceptable from the point of view of a user. To further investigate the
tradeoff between execution time and significance of the discovered
motifs, we repeated the experiments running \textsc{madmx}\ with $\min_\ell=2$
and $\rho=0.65$, for all sequences. The running time of \textsc{madmx}\ was
almost halved, while the small output produced still featured high
quality. In fact, for sequences \texttt{ace2}, \texttt{ap1}, and \texttt{uasgaba} the top-$100$ motifs extracted by \textsc{madmx}\ have z-score greater or equal than the corresponding ones returned by \textsc{varun}.
We also have attempted a comparison between \textsc{varun}\ and \textsc{madmx}\
on longer sequences (such as \textsc{hgmr}~1) at higher frequencies
(since, unfortunately, \textsc{varun}\ does not seem to be able to handle low
frequencies on very long sequences). Even allowing a higher number of don't cares between solid characters ($D=2$) for the
motifs of \textsc{varun},
all of the top-$m$ z-scores featured by the motifs extracted by \textsc{madmx}\
are greater than or equal to the
corresponding scores in the ranking of \textsc{varun}, with $m$ reaching the size of
\textsc{varun}'s output.
In fairness, we remark that \textsc{varun}\ was designed to work at its best
on protein sequences, while \textsc{madmx}'s main target are
\textsc{dna}\ sequences. Hence, these two tools should be regarded as
complementary. Moreover, \textsc{varun}\ has the advantage of retrieving
flexible motifs, while \textsc{madmx}\ focuses only on rigid ones.
\paragraph{\bf Acknowledgments}
The authors wish to thank Alberto Apostolico and Matteo Comin for providing
the code and giving valuable insights on \textsc{varun}, Ben Raphael for
suggesting the use of \textsc{repeatmasker}, and Roberta Mazzucco and Francesco
Peruch for coding \textsc{madmx}.
\bibliographystyle{plain}
|
\section{Introduction}
A major challenge in particle physics is to
understand the pattern of fermion masses and mixing angles. With the discovery of neutrino
oscillations flavor has become even more mysterious since the nearly tri-bimaximal mixing
strongly differ from the quark sector. The minimal supersymmetric standard model (MSSM)
does not provide insight into the flavor problem by contrast the generic MSSM contains
even new sources of flavor and chirality violation, stemming from the
supersymmetry-breaking sector which are the sources of the so-called supersymmetric
flavor problem. The origin of these flavor-violating
terms is obvious: In the standard model (SM) the quark and lepton Yukawa
matrices are diagonalized by unitary rotations in flavor space and the
resulting basis defines the mass eigenstates. If the same
rotations are carried out on the squark fields of the MSSM, one obtains
the super--CKM/PMNS basis in which no tree--level FCNC couplings are present.
However, neither the $3\times 3$ mass terms $m_{\tilde{Q}}^{2}$, $m_{\tilde{u}}^{2} $, $m_{\tilde{d}}^{2} $, $ m_{\tilde{L}}^{2}$ and $m_{\tilde{e}}^{2} $ of the left--handed
and right--handed sfermions nor the tri-linear Higgs--sfermion--sfermion
couplings are necessarily diagonal in this basis.
The tri-linear $\overline{Q}H_d {A}^d d_R $, $\overline{Q}H_u {A}^u u_R $ and $\overline{L}H_d {A}^l e_R $ terms induce
mixing between left--handed and right--handed sfermions after the Higgs
doublets $H_d$ and $H_u $ acquire their vacuum expectation values (vevs)
$v_d$ and $v_u$, respectively.
In the current era of precision flavor physics stringent bounds on these
parameters have been derived from FCNC processes in the quark and in the lepton sector,
by requiring that the
gluino--squark loops and chargino--sneutrinos/neutralino--slepton loops do not exceed the
measured values of the considered observables
\cite{Hagelin:1992tc,Gabbiani:1996hi,Ciuchini:1998ix,Borzumati:1999qt,Becirevic:2001jj,
Arganda:2005ji,Masiero:2004js,Foster:2005wb,Ciuchini:2007cw,Masiero:2008cb,Ciuchini:2007ha,Crivellin:2009ar,Altmannshofer:2009ne}.
However, in \cite{Crivellin:2008mq,Crivellin:2009pa} it is
shown that all flavor violation in the quark sector can solely originate from trilinear
SUSY breaking terms because all FCNC bounds are satisfied for $M_{SUSY}\geq500\rm{GeV}$.
Dimensionless quantities are commonly defined in the mass insertion parametrization as:
\begin{equation}
\delta_{IJ}^{f\,XY} = \frac{\left(\Delta m^2_F\right)^{IJ}_{XY}}{\sqrt{m_{\tilde{f}_{IX}}^2 m_{\tilde{f}_{JY}}^{2}}}.
\label{delta}
\end{equation}
In \eq{delta} $I$ and $J$ are flavour indices running from $1$ to $3$, $X,Y$ denote the chiralities $L$ and $R$, $\left(\Delta m^2_F\right)^{IJ}_{XY}$ with $\,F = U,D,L$ is the off-diagonal element of the sfermion mass matrix (see Appendix \ref{sec:appendixDiag}) and $m_{\tilde{f}_{IX}^2}$, $m_{\tilde{f}_{JY}}^2$ are the corresponding diagonal ones. In this article we are going to complement the analysis of \cite{Crivellin:2008mq} with respect to three important points:
\begin{itemize}
\item Electroweak correction are taken into account.
Therefore, we are able to constrain also the flavor-violating and chirality-changing terms
in the lepton sector.
\item The constraints on the flavor-diagonal mass
insertions $\delta^{u,d,l\,LR}_{11,22}$ are obtained from the requirement that the
corrections should not exceed the measured masses. This has already been done in the
seminal paper of Gabbiani et al. \cite{Gabbiani:1996hi}. We improve this calculation by
taking into account QCD corrections and by using the up-to-date values of the fermion
masses.
\item The leading chirally-enhanced two-loop corrections
are calculated. As we will see, this allows us to constrain the elements
$\delta^{f\,RL}_{13}$ (and $\delta^{d\,RL}_{23}$), if at the same time, also
$\delta^{f\,LR}_{13}$ ($\delta^{d\,LR}_{23}$) is different from zero.
\end{itemize}
Our paper is organized as follows: In Sec.~\ref{sec:RenormalizationAllg}
we study the impact of chirally enhanced parts of the self-energies for quarks and leptons
on the fermion masses and mixing matrices (CKM matrix and PMNS matrix).
First, we introduce the general formalism in Sec.~\ref{sec:generalformalism}
and then specify to the MSSM with non-minimal sources of flavor violation in
Sec.~\ref{sec:selfenergies} where we compute the chirally enhanced parts of the
self-energies
for quarks and leptons taking into account also the leading two-loop corrections.
Sec.~\ref{sec:Numerics} is devoted to the numerical analysis. Finally we conclude in
Sec.~\ref{sec:Conclusions}.
\begin{figure}[t!]
\includegraphics[width=1\linewidth]{Wellenfunktionsrenormierung.eps}
\caption{ Flavor-valued wave-function renormalization.
}\label{fig:Wellenfunktionsrenormierung}
\end{figure}
\section{Finite renormalization of fermion masses and mixing matrices}\label{sec:RenormalizationAllg}
We have computed the finite renormalization of the CKM matrix by SQCD
effects in Ref. \cite{Crivellin:2008mq,Crivellin:2009sd} and of the PMNS matrix in Ref.
\cite{Girrbach:2009uy}. In this section we compute the finite renormalization of fermion
masses and mixing angles induced through one-particle irreducible flavor-valued
self-energies beyond leading-order. We first consider the general case and then specify to
the MSSM.
\begin{figure}[t,h]
\includegraphics[width=0.45\textwidth]{2-loop-self-energy.eps}
\caption{ One-particle irreducible two-loop self-energy constructed
out of two one-loop self energies with $I\neq J\neq K$.
}\label{fig:2-loop-self-energy}
\end{figure}
\subsection{General formalism}\label{sec:generalformalism}
In this section we consider the general effect of one-particle irreducible self-energies. It is possible to decompose any self-energy in its chirality-changing and its chirality-flipping parts in the following way:
\begin{equation}
\renewcommand{\arraystretch}{1.8}
\begin{array}{l}
\Sigma _{IJ}^f(p) = \left( {\Sigma _{IJ}^{f\;LR}(p^2) + p\hspace{-0.44em}/\hspace{0.06em}\Sigma _{IJ}^{f\;RR}(p^2) } \right)P_R \\
\phantom{\Sigma _{IJ}^f(p) = }+ \left( {\Sigma _{IJ}^{f\;RL}(p^2) + p\hspace{-0.44em}/\hspace{0.06em}\Sigma _{IJ}^{f\;LL}(p^2) } \right)P_L\,.
\end{array}
\label{self-energy-decomposition}
\end{equation}
Note that chirality-changing parts $\Sigma _{IJ}^{f\;LR}$ and $\Sigma _{IJ}^{f\;RL}$ have mass dimension 1, while $\Sigma _{IJ}^{f\;LL}$ and $ \Sigma _{IJ}^{f\;RR}$ are dimensionless.
With this convention the renormalization of the fermion masses is given by:
\begin{widetext}
\begin{equation}
m_{f_I }^{\left( 0 \right)} \to m_{f_I }^{\left( 0 \right)} + \Sigma _{II}^{f\;LR}(m_{f_I }^{2}) + \frac{1}{2}m_{f_I }^{} \left( {\Sigma _{II}^{f\;LL}(m_{f_I }^{2}) + \Sigma _{II}^{f\;RR}(m_{f_I }^{2}) } \right) + \delta m_{f_I } = m_{f_I }^{\text{phys}}\,.
\label{massrenormalization}
\end{equation}
\end{widetext}
If the self-energies are finite, the counter-term $\delta m_{f_I }$ in
\eq{massrenormalization}
is zero in a minimal renormalization scheme like $\overline{\text{MS}}$. In the following
we choose this minimal scheme for two reasons: First, $A$-terms are theoretical quantities
which are not directly related to physical observables. For such quantities it is always
easier to use a minimal scheme which allows for a direct relation between theoretical
quantities and observables. Second, we consider the limit in which the light fermion
masses and CKM elements are generated radiatively. In this limit it would be unnatural to
have tree-level Yukawa couplings and CKM elements in the Lagrangian which are canceled by
counter-terms as in the on-shell scheme.
The self-energies in \eq{self-energy-decomposition} do not only renormalize
the fermion masses. Also a rotation $1+\Delta U^{f\;L}_{IJ}$ in flavor-space which has to
be applied to all external fields is induced through the diagram in
Fig.~\ref{fig:Wellenfunktionsrenormierung}:
\begin{widetext}
\begin{align}
\Delta U_{IJ}^{f\;L} &= \frac{1}{m_{f_J}^2 -m_{f_I}^2 }\left( m_{f_J } ^2 \Sigma _{IJ}^{f\;LL} \left( {m_{f_I } ^2 } \right) + m_{f_J } m_{f_I } \Sigma _{IJ}^{f\;RR} \left( {m_{f_I}^2} \right) + m_{f_J } \Sigma _{IJ}^{f\;LR} \left( {m_{f_I}^2} \right) + m_{f_I} \Sigma _{IJ}^{f\;RL} \left( {m_{f_I}^2 } \right) \right)\;\;\;{\rm{for}}\; I \ne J,\nonumber \\
\Delta U_{II}^{f\;L} & = \frac{1}{2}{\mathop{\rm Re}\nolimits} \left[ {\Sigma _{II}^{f\;LL} \left( {m_{f_I } ^2 } \right) + 2m_{f_I } \Sigma _{II}^{f\;LR\prime} \left( {m_{f_I } ^2 } \right) + m_{f_I } ^2 \left( {\Sigma _{II}^{f\;LL\prime} \left( {m_{f_I } ^2 } \right) + \Sigma _{II}^{f\;RR\prime} \left( {m_{f_I } ^2 } \right)} \right)} \right].
\label{WFR}
\end{align}
\end{widetext}
The prime denotes differentiation with respect to the argument. The flavor-diagonal part arises from the truncation of flavor-conserving self-energies. \eq{massrenormalization} and \eq{WFR} are valid for arbitrary one-particle irreducible self-energies.
\subsection{Self-energies in the MSSM}\label{sec:selfenergies}
Self-energies with supersymmetric virtual particles are of special importance
because of a possible chiral enhancement which can lead to order-one corrections.
In this section we calculate the chirally enhanced (by a factor $\frac{A_f^{IJ}}{M_{SUSY}Y_f^{IJ}}$ or $\tan\beta$) parts of the fermion self-energies in the MSSM. Therefore it is only necessary to evaluate the diagrams at vanishing external momentum.
We choose the sign of the self-energies $\Sigma$ to be equal to the sign of
the mass, e.g. calculating a self-energy diagram yields $-i\Sigma$. Then, with the
conventions given in the Appendix \ref{sec:appendix}, the gluino contribution to the quark
self-energies is given by:
\begin{align}
\Sigma_{q_{IL}-q_{JR}}^{\tilde{g}}=&-\sum_{i=1}^{6}\frac{m_{\tilde{g}}}{16\pi^2}
\left(\Gamma_{q_{JR}}^{\tilde{g}\tilde{q}_{i}}\right)^*\Gamma_{q_{IL}}^{\tilde{g}\tilde{q}
_{i}}B_{0}(m_{\tilde{g}}^{2},m_{\tilde{q}_i}^{2})\\
=& \frac{\alpha_s}{2 \pi}C_F\sum_{i = 1}^6 m_{\tilde{g}} W_Q^{(J+3)i*}
W_Q^{Ii}B_{0}(m_{\tilde{g}}^{2},m_{\tilde{q}_i}^{2})
\end{align}
and for the neutralino and chargino contribution to the quark self-energy we receive:
\begin{align}
\Sigma_{d_{IL}-d_{JR}}^{\tilde{\chi}^{0}}=&-\sum_{i=1}^{6}\sum_{j=1}^{4}\frac{m_{\tilde{\chi}_{j}^{0}}}{16\pi^2}\Gamma_{d_{JR}}^{\tilde{\chi}_{j}^{0}\tilde{d}_{i}*}\Gamma_{d_{IL}}^{\tilde{\chi}_{j}^{0}\tilde{d}_{i}}B_{0}(m_{\tilde{\chi}_{j}^{0}}^{2},m_{\tilde{d}_i}^{2})
\\ \Sigma_{d_{IL}-d_{JR}}^{\tilde{\chi}^{\pm}}=&-\sum_{i=1}^{6}\sum_{j=1}^{2}\frac{m_{\tilde{\chi}_{j}^{\pm}}}{16\pi^2}\Gamma_{d_{JR}}^{\tilde{\chi}_{j}^{\pm}\tilde{u}_{i}*}\Gamma_{d_{IL}}^{\tilde{\chi}_{j}^{\pm}\tilde{u}_{i}}B_{0}(m_{\tilde{\chi}_{j}^{\pm}}^{2},m_{\tilde{u}_i}^{2})
\end{align}
The self-energies in the up-sector are easily obtained by interchanging $u$ and $d$.
We denote the sum of all contribution as:
\begin{equation}
\Sigma_{IJ}^{q\,LR} = \Sigma_{q_{IL}-q_{JR}}^{\tilde{g}} +
\Sigma_{q_{IL}-q_{JR}}^{\tilde{\chi}^{0}} + \Sigma_{q_{IL}-q_{JR}}^{\tilde{\chi}^{\pm}}
\end{equation}
Note that the gluino contribution are dominant in the case of non-vanishing $A$-terms, since they involve the strong coupling constant.
In the lepton case, neutralino--slepton and chargino--sneutrino loops contribute the
non-decoupling self-energy $\Sigma_{IJ}^{\ell\,LR}$. With the convention in the Appendix
\ref{sec:appendix} the self-energies are given by:
\begin{align}
\Sigma_{\ell_{IL}-\ell_{JR}}^{\tilde{\chi}^{\pm}}=&-\sum_{j=1}^2\sum_{k=1}^3\frac{m_{
\tilde {
\chi}_{j}^{\pm}}}{16\pi^{2}}\Gamma_{\ell_{JR}}^{\tilde{\chi}_{j}^{\pm}\tilde{\nu}_{k}*}
\Gamma_{\ell_{IL}}^{\tilde{\chi}_{j}^{\pm}\tilde{\nu}_{k}}B_{0}(m_{\tilde{\chi}_{j}^{\pm}}
^ { 2 },m_{\tilde{\nu}_{k}}^{2}),\\
\Sigma_{\ell_{IL}-\ell_{JR}}^{\tilde{\chi}^{0}}=&-\sum_{i=1}^{6}\sum_{j=1}^{4}\frac{m_{
\tilde {
\chi}_{j}^{0}}}{16\pi^{2}}\Gamma_{\ell_{JR}}^{\tilde{\chi}_{j}^{0}\tilde{\ell}_{i}*}
\Gamma_ { \ell_ {
IL}}^{\tilde{\chi}_{j}^{0}\tilde{\ell}_{i}}B_{0}(m_{\tilde{\chi}_{j}^{0}}^{2},m_{\tilde{
\ell } _i } ^{2}).
\end{align}
Again, we denote the sum of all contribution as:
\begin{equation}
\Sigma_{IJ}^{\ell\,LR} = \Sigma_{\ell_{IL}-\ell_{JR}}^{\tilde{\chi}^{0}} +
\Sigma_{\ell_{IL}-\ell_{JR}}^{\tilde{\chi}^{\pm}}.
\end{equation}
With $I=J$ we arrive at the flavor-conserving case.
This can lead to significant quantum corrections to fermion masses, but except for the
gluino, the pure bino ($\propto g_1^2$) and the negligible small bino-wino mixing
($\propto g_1 g_2$) contribution, they are proportional to tree-level Yukawa couplings.
However, if the light fermion masses are generated radiatively from chiral
flavor-violation in the soft SUSY-breaking terms, then the Yukawa couplings of the first
and second generation even vanish and the latter effect is absent at all. Radiatively
generated fermion mass terms via soft tri-linear $A$-terms corresponds to the upper bound
found from the fine-tuning argument where the correction to the mass is as large as the
physical mass itself. This fine-tuning argument is based on 't~Hooft's naturalness
principle: A theory with small parameters is natural if the symmetry is enlarged when
these parameters vanish. The smallness of the parameters is then protected against large
radiative corrections by the concerned symmetry. If such a small parameter, e.g. a fermion
mass, is composed of several different terms there should be no accidental large
cancellation between them. We will derive our upper bounds from the condition that the
SUSY corrections should not exceed the measured value.
If we restrict ourself to the case with vanishing first and second generation tree-level Yukawa couplings, the off-diagonal entries in the sfermion mass matrices stem from the soft tri-linear terms. Thus we are left with $\delta^{f\,LR}_{IJ}$ only.
In the mass insertion approximation with only LR insertion the flavor violating self-energies simplifies. For the gluino (neutralino) self-energies which are relevant for our following discussion for the quark (lepton) case we get:
\begin{align}
\Sigma_{q_{IX}-q_{JY}}^{\tilde{g}}& = \frac{2\alpha_s}{3 \pi} M_{\tilde{g}} m_{\tilde{q}_{JY}}m_{\tilde{q}_{IX}}\delta_{IJ}^{q\,XY}\, C_0\left(M_1^2, m_{\tilde{q}_{JY}}^2,m_{\tilde{q}_{IX}}^2 \right),\\
\Sigma_{\ell_{IX}-\ell_{JY}}^{\tilde{B}}& = \frac{\alpha_1}{4 \pi} M_1
m_{\tilde{\ell}_{JY}}m_{\tilde{\ell}_{IX}}\delta_{IJ}^{\ell\,XY}\, C_0\left(M_1^2,
m_{\tilde{\ell}_{JY}}^2,m_{\tilde{\ell}_{IX}}^2 \right).
\end{align}
Since the sneutrino mass matrix consists only of a LL block,
there are no chargino diagrams in the lepton case with LR insertions at all.
Since the SUSY particles
are known to be much heavier than the five lightest quarks it is possible to evaluate the
one-loop self-energies at vanishing external momentum and to neglect higher terms which
are suppressed by powers of $m_{f_I}^2/M_{SUSY}^2$. The only possibly sizable decoupling
effect concerning the $W$ vertex renormalization is a loop-induced right-handed $W$
coupling (see \cite{Crivellin:2009sd}). Therefore \eq{self-energy-decomposition}
simplifies to
\begin{equation}
\Sigma_{IJ}^{f\;(1)} = {\Sigma _{IJ}^{f\;LR\;(1)} } P_R + {\Sigma _{IJ}^{f\;RL\;(1)} } P_L
\end{equation}
at the one-loop level (indicated by the superscript (1)). In this approximation the self-energies are always chirality changing and contribute to the finite renormalization of the quark masses in \eq{massrenormalization} and to the flavor-valued wave-function renormalization in \eq{WFR}. At the one-loop level we receive the well known result
\begin{equation}
m_{f_I }^{\left( 0 \right)} \to m_{f_I }^{\left( 1 \right)}= m_{f_I}^{(0)} + \Sigma _{II}^{f\;LR\;(1)}
\label{masse-1-loop}
\end{equation}
for the mass renormalization in the $\overline{\text{MS}}$ scheme. According to \eq{WFR} the flavor-valued rotation which has to be applied to all external fermion fields is given by:
\begin{widetext}
\begin{equation} \renewcommand{\arraystretch}{2.5}
\Delta U^{f\;L\;(1)} = \left( {\begin{array}{*{20}c}
0 & {\dfrac{{m_{f_2 } \Sigma _{12}^{f\;LR\;(1)} + m_{f_1 } \Sigma _{12}^{f\;RL\;(1)} }}{{m_{f_2 } ^2 - m_{f_1 } ^2 }}} & {\dfrac{{m_{f_3 } \Sigma _{13}^{f\;LR\;(1)} + m_{f_1 } \Sigma _{13}^{f\;RL\;(1)} }}{{m_{f_3 } ^2 - m_{f_1 } ^2 }}} \\
{\dfrac{{m_{f_1 } \Sigma _{21}^{f\;LR\;(1)} + m_{f_2 } \Sigma _{21}^{f\;RL\;(1)} }}{{m_{f_1 } ^2 - m_{f_2 } ^2 }}} & 0 & {\dfrac{{m_{f_3 } \Sigma _{23}^{f\;LR\;(1)} + m_{f_2} \Sigma _{23}^{f\;RL\;(1)} }}{{m_{f_3 } ^2 - m_{f_2 } ^2 }}} \\
{\dfrac{{m_{f_1 } \Sigma _{31}^{f\;LR\;(1)} + m_{f_3 } \Sigma _{31}^{f\;RL\;(1)} }}{{m_{f_1 } ^2 - m_{f_3 } ^2 }}} & {\dfrac{{m_{f_2 } \Sigma _{32}^{f\;LR\;(1)} + m_{f_3 } \Sigma _{32}^{f\;RL\;(1)} }}{{m_{f_2 } ^2 - m_{f_3 } ^2 }}} & 0 \\
\end{array}} \right) .
\label{deltaU1}
\end{equation}
\end{widetext}
The corresponding corrections to the right-handed wave-functions are obtained by simply exchanging $L$ with $R$ and vice versa in \eq{deltaU1}. Note that the contributions of the self-energies $\Sigma _{IJ}^{f\;RL\;(1)}$ with $J>I$ are suppressed by small mass ratios. Therefore, the corresponding off-diagonal elements of the sfermion mass matrices cannot be constrained from the CKM and PMNS renormalization. However, since we treat, in the spirit of Ref. \cite{Logan:2000iv}, all diagrams in which no flavor appears twice on quark lines as one-particle irreducible, chirally-enhanced self-energies can also be constructed at the two-loop level (see Fig.~(\ref{fig:2-loop-self-energy})):
\begin{widetext}
\begin{equation} \renewcommand{\arraystretch}{3}
\begin{array}{l}
\Sigma _{IJ}^{f\;RR\;\left( 2 \right)} \left( {p^2 } \right) = \sum\limits_{K \ne I,J} \dfrac{{\Sigma _{IK}^{f\;RL\;(1)} \Sigma _{KJ}^{f\;LR\;(1)} }}{{p^2 - m_{f_K } ^2 }} ,\;\;\;\;\;\;\;\;\;\;\,
\Sigma _{IJ}^{f\;LL\;\left( 2 \right)} \left( {p^2 } \right) = \sum\limits_{K \ne I,J} {\dfrac{{\Sigma _{IK}^{f\;LR\;(1)} \Sigma _{KJ}^{f\;RL\;(1)} }}{{p^2 - m_{f_K } ^2 }}} ,\\
\Sigma _{IJ}^{f\;LR\;\left( 2 \right)} \left( {p^2 } \right) = \sum\limits_{K \ne I,J} {m_{f_K } \dfrac{{\Sigma _{IK}^{f\;LR\;(1)} \Sigma _{KJ}^{f\;LR\;(1)} }}{{p^2 - m_{f_K } ^2 }}} ,\;\;\;\;
\Sigma _{IJ}^{f\;RL\;\left( 2 \right)} \left( {p^2 } \right) = \sum\limits_{K \ne I,J} {m_{f_K } \dfrac{{\Sigma _{IK}^{f\;RL\;(1)} \Sigma _{KJ}^{f\;RL\;(1)} }}{{p^2 - m_{f_K } ^2 }}} . \\
\end{array}
\end{equation}
Therefore, the chiral-enhanced two-loop corrections to the masses and the wave-function renormalization are given by:
\begin{equation}\renewcommand{\arraystretch}{1.8}
\left( {\begin{array}{*{20}c}
{m_{f_1 }^{\left( 0 \right)} } \\
{m_{f_2 }^{\left( 0 \right)} } \\
{m_{f_3 }^{\left( 0 \right)} } \\
\end{array}} \right) \to \left( {\begin{array}{*{20}c}
{m_{f_1 }^{\left( 0 \right)} + \Sigma _{11}^{f\;LR\;(1)} - \dfrac{{\Sigma _{12}^{f\;LR\;(1)} \Sigma _{21}^{f\;LR\;(1)} }}{{m_{f_2 } }} - \dfrac{{\Sigma _{13}^{f\;LR\;(1)} \Sigma _{31}^{f\;LR\;(1)} }}{{m_{f_3 } }}} \\
{m_{f_2 }^{\left( 0 \right)} + \Sigma _{22}^{f\;LR\;(1)} - \dfrac{{\Sigma _{23}^{f\;LR\;(1)} \Sigma _{32}^{f\;LR\;(1)} }}{{m_{f_3 } }}} \\
{m_{f_3 }^{\left( 0 \right)} + \Sigma _{33}^{f\;LR\;(1)} } \\
\end{array}} \right),
\label{equ:massRen2}
\end{equation}
\begin{equation}\renewcommand{\arraystretch}{2}
\Delta U_L^{f\;\left( 2 \right)} = \left( {\begin{array}{*{20}c}
{ - \dfrac{{\left| {\Sigma _{12}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_2 } ^2 }} - \dfrac{{\left| {\Sigma _{13}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_3 } ^2 }}} & { - \dfrac{{\Sigma _{13}^{f\;{\rm{LR}}\;(1)} \Sigma _{32}^{f\;{\rm{LR}}\;(1)} }}{{m_{f_2 } m_{f_3 } }}} & {\dfrac{{\Sigma _{12}^{f\;{\rm{LR}}\;(1)} \Sigma _{23}^{f\;{\rm{RL}}\;(1)} }}{{m_{f_3 }^2 }}} \\
{\dfrac{{\Sigma _{23}^{f\;{\rm{RL}}\;(1)} \Sigma _{31}^{f\;{\rm{RL}}\;(1)} }}{{m_{f_2 } m_{f_3 } }}} & { - \dfrac{{\left| {\Sigma _{23}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_3 } ^2 }} - \dfrac{{\left| {\Sigma _{12}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_2 } ^2 }}} & {\dfrac{{\Sigma _{21}^{f\;{\rm{LR}}\;(1)} \Sigma _{13}^{f\;{\rm{RL}}\;(1)} }}{{m_{f_3 }^2 }}} \\
{\dfrac{{\Sigma _{32}^{f\;{\rm{RL}}\;(1)} \Sigma _{21}^{f\;{\rm{RL}}\;(1)} }}{{m_{f_2 } m_{f_3 } }}} & { - \dfrac{{\Sigma _{31}^{f\;{\rm{RL}}\;(1)} \Sigma _{12}^{f\;{\rm{LR}}\;(1)} }}{{m_{f_2 } m_{f_3 } }}} & { - \dfrac{{\left| {\Sigma _{13}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_3 } ^2 }} - \dfrac{{\left| {\Sigma _{23}^{f\;LR\;(1)} } \right|^2}}{{2m_{f_3 } ^2 }}} \\
\end{array}} \right),
\label{WFR2}
\end{equation}
\end{widetext}
where we have neglected small mass ratios. In the quark case, we already know about the
hierarchy of the self-energies from our fine-tuning argument. In this case \eq{WFR2} is
just necessary to account for the unitarity of the CKM matrix \cite{Crivellin:2008mq}.
However, the corrections to $m_{f_{1} }^{\left( 0 \right)}$ in \eq{equ:massRen2} can be
large. For this reason we can also constrain
$\Sigma _{31}^{f\;LR\;(1)}$ with 't~Hooft's naturalness criterion if at the same time
$\Sigma _{13}^{f\;LR\;(1)}$ is different from zero.
\section{Numerical Analysis}\label{sec:Numerics}
In this section we are going to give a complete numerical evaluation of the all possible constraints on the
SUSY breaking sector from 't~Hooft's naturalness argument. This criterion is applicable
since we gain a flavor symmetry \cite{Crivellin:2008mq} if the light fermion masses are
generated radiatively. Therefore the situation is different from e.g. the little hierarchy
problem, where no additional symmetry is involved.
First of all, it is important to note that all off-diagonal elements of the fermion mass matrices have to be smaller than the average of their assigned diagonal elements
\begin{equation}
\left(\Delta m^2_F\right)^{IJ}_{XY}<\sqrt{m_{\tilde{f}_{IX}}^2 m_{\tilde{f}_{JY}}^2},
\end{equation}
since otherwise one sfermion mass eigenvalue is negative. We note that in
Ref. \cite{Gabbiani:1996hi} this constraint is not imposed.
All constraints in this section are
non-decoupling since we compute corrections to the Higgs-quark-quark coupling which
is of dimension 4. Therefore, our constraints on the soft-supersymmetry-breaking
parameters do not vanish in the limit of infinitely heavy SUSY masses but rather converge
to a constant \cite{Crivellin:2008mq}. However, even though $\delta^{f\;LR}_{IJ}$ is a
dimensionless parameter it does not only involve SUSY parameter. It is also proportional
to a vacuum expectation and therefore scales like $v/M_{\rm{SUSY}}$. Thus, our constraints
on $\delta^{f\;LR}_{IJ}$ do not approach a constant for $M_{\rm{SUSY}}\to\infty$ but
rather get stronger. Similar effects occur in Higgs-mediated FCNC processes which decouple
like $1/{M_{\rm{Higgs}}^2}$ rather than $1/M_{\rm{SUSY}}^2$
\cite{Hall:1993gn,Hamzaoui:1998nu,Banks:1987iu}.
However, Higgs-mediated
effects can only be induced within supersymmetry in the presence of non-holomorphic terms
which are not required for our constraints. An example of a non-decoupling Higgs-mediated
FCNC process is the observable $R_K=\Gamma\left(K\to
e\nu\right)/\Gamma\left(K\to\mu\nu\right)$ that is currently analyzed by the
NA62-experiment. In this case Higgs contributions can induce
deviations from lepton flavor universality \cite{Masiero,Masiero:2008cb,JG}.
\subsection{Constraints on flavor-diagonal mass insertions at one loop}
\begin{figure}
\includegraphics[width=0.7\linewidth]{delta-u-11.eps}
\includegraphics[width=0.7\linewidth]{delta-u-22.eps}
\includegraphics[width=0.7\linewidth]{delta-d-11.eps}
\includegraphics[width=0.7\linewidth]{delta-d-22.eps}
\caption{Constraints on the diagonal mass insertions $\delta_{11,22}^{u,d\,LR}$ obtained
by
applying 't~Hooft's naturalness criterion.
}\label{fig:LRudcs}
\end{figure}
The diagonal elements of the $A$-terms can be constrained from the
fermion masses by demanding that $\Sigma^{f\;LR\;(1)}_{II}\leq m_{f_I}$ [see
\eq{masse-1-loop}]. The bounds on the flavor-conserving $A$-term for the up, charm, down
and strange quarks are shown in Fig.~(\ref{fig:LRudcs}) and the constraints from the
electron and muon mass are depicted in Fig.~(\ref{fig:LRemu}). The upper bound derived
from the fermion mass is roughly given by
\begin{equation}
\left|\delta_{II}^{q\,LR}\right|\lesssim \frac{3 \pi\, m_{q_I}(M_{SUSY})}{\alpha_s(M_{SUSY}) M_\text{SUSY}}
\label{quark}
\end{equation}
for quarks and
\begin{equation}
\left|\delta_{II}^{\ell\,LR}\right|\lesssim \frac{8 \pi m_{\ell_I}}{\alpha_1
M_\text{SUSY}}
\label{lepton}
\end{equation}
for leptons in the case of equal SUSY masses. In the lepton case \eq{lepton} can be further simplified, since we can neglect the running of the masses:
\begin{equation}
\begin{array}{c}
|\delta_{11}^{\ell\,LR}|\lesssim
0.0025\left(\frac{500\,\text{GeV}}{M_\text{SUSY}}\right),\\
\left|\delta_{22}^{\ell\,LR}\right|\lesssim
0.5\left(\frac{500\,\text{GeV}}{M_\text{SUSY}}\right)\,.
\end{array}
\end{equation}
\begin{figure}
\includegraphics[width = 0.7\linewidth]{PlotdeltaLR11.eps}
\includegraphics[width = 0.7\linewidth]{PlotdeltaLR22.eps}
\caption{Contraints on the diagonal mass insertion $\delta_{11,22}^{\ell\,LR}$ as a
function of $M_1$ and $m_{\tilde{e}}$, $m_{\tilde{\mu}}$. }\label{fig:LRemu}
\end{figure}
However, as already pointed out in Ref. \cite{Borzumati:1999sp} a muon mass
that is solely generated radiatively potentially leads to measurable contributions to the
muon anomalous magnetic moment. This arises from the same one-loop diagram as
$\Sigma^{\ell\;LR}_{22}$ with an external photon attached. Therefore, the SUSY
contribution is not suppressed by a loop factor compared to the case with tree-level
Yukawa couplings.
\subsection{Constraints on flavor-off-diagonal mass insertions from CKM and PMNS renormalization}\label{subsec:Numericoffdiagonal}
\subsubsection{CKM matrix}
A complete analysis of the constraints for the CKM renormalization was already carried out in Ref. \cite{Crivellin:2008mq}. The numerical effect of the chargino contributions is negligible at low $\tan\beta$ and the neutralino contributions amount only to corrections of about $5$\% of the gluino contributions. Therefore, we refer to the constraints on the off-diagonal elements $\delta^{q\;LR}_{IJ}$ given in Ref. \cite{Crivellin:2008mq}.
\subsubsection{Threshold corrections to PMNS matrix}
Up to now, we have only an upper bound for the matrix element $U_{e3} = \sin\theta_{13}
e^{-i\delta}$
and thus for the mixing angle $\theta_{13}$; the best-fit value is at or
close to zero: $\theta_{13} = 0.0^{+7.9}_{-0.0}$ \cite{Fogli:2008jxx}.
It might well be that it vanishes at tree level due to a particular
symmetry and obtains a non-zero value due to corrections. So we can ask the question if threshold corrections to the PMNS matrix could spoil the prediction $\theta_{13} = 0^{\circ}$ at the weak scale.
We demand the absence of fine-tuning for these corrections and therefore
require that the SUSY loop contributions do not exceed the value of
$U_{e3}$,
\begin{align}
\label{equ:finetuningUe3}
\left|\Delta U_{e3}\right| \leq\left| U_{e3}^\text{phys}\right| .
\end{align}
The renormalization of the PMNS matrix is described in detail in \cite{Girrbach:2009uy}, where the on-shell scheme was used. As discussed in Sec.~(\ref{sec:RenormalizationAllg}) we also use the $\overline{\text{MS}}$ scheme in this section. Then the physical PMNS matrix is given by:
\begin{align}
U^{\text{phys}} = U^{(0)} + \Delta U\,,
\end{align}
where $\Delta U$ should not be confused with the wave function renormalization $\Delta U^{f\,L}$. Then $\Delta U$ is given by
\begin{align}
\Delta U = \left(\Delta U^{\ell\,L}\right)^T U^{(0)}.\label{equ:PMNSequation}
\end{align}
Note that in contrast to the corrections to the CKM matrix, there is a
transpose in $\Delta U^{\ell\,L}$, because the first index of the PMNS matrix corresponds
to down-type fermions and not to up-type fermion as in the CKM matrix.
Only the corrections to the small element $U_{e3}$ can be sizeable, since all other elements are of order one.
If we set all off-diagonal element to zero except for
$\delta^{\ell\,LR}_{13}\neq 0$, we get
\begin{equation}
\renewcommand{\arraystretch}{2.4}
\begin{array}{l}
\Delta U_{e3} = \dfrac{ \Delta
U^{\ell\,L}_{31}U_{\tau3}^\text{phys}-U_{e3}^\text{phys}\left|\Delta
U^{\ell\,L}_{31}\right|^2}{1 + \left|\Delta U^{\ell\,L}_{11}\right|^2} \\
\phantom{\Delta U_{e3}}\approx -U_{\tau
3}^\text{phys}\dfrac{\Sigma_{31}^{\ell\,RL}}{m_{\tau}}.
\end{array}
\end{equation}
Note that here, in contrast to the renormalization of the CKM matrix,
the physical PMNS element appears. This is due to the fact that one has to solve the
linear system in \eq{equ:PMNSequation} as described in \cite{Girrbach:2009uy}.
By means of the fine-tuning argument
we can in principle derive upper bounds for $\delta_{13}^{\ell\,LR}$. The results depend
on the SUSY mass scale $M_\text{SUSY}$ and the assumed value for $\theta_{13}$.
Here, we consider the corrections stemming from flavor-violating $A$-terms to
the small matrix element $U_{e3}$. The $\delta_{13}^{\ell\,LL}$-contribution was already
studied in \cite{Girrbach:2009uy} with the result that they are negligible small. We also
made a comment about the $\delta_{13}^{\ell\,LR}$-contribution which is outlined in more
detail. Our
results depend on the overall SUSY mass scale, the value of $\theta_{13}$ and of
$\delta_{13}^{\ell\,LR}$.
In Fig.~(\ref{fig:PMNSMSUSY1000}) you can see the percentage deviation
of $U_{e3}$ through this SUSY loop corrections in dependence of $\delta_{13}^{\ell\,LR}$
(top) and $\theta_{13}$ (bottom) for $M_{\text{SUSY}} = 1000$ GeV. The constraints on
$\delta_{13}^{\ell\,LR}$ get stronger with smaller $\theta_{13}$ and with larger
$M_{\text{SUSY}}$.
In Fig.~(\ref{fig:th13vsdelta}) the excluded
$\left(\theta_{13},\delta_{13}^{\ell\,LR}\right)$-region is below the curves for different
$M_\text{SUSY}$ scales.
The derived bound can be simplified to
\begin{equation}
\left|\delta_{13}^{\ell\,LR}\right|\lesssim 0.2 \left(\frac{500\,
\text{GeV}}{M_\text{SUSY}}\right)\left|\theta_{13}\, \text{in degrees}\right|.
\end{equation}
Exemplarily, we get for reasonable SUSY masses of $M_\text{SUSY} = 1000$ GeV and
$\theta_{13} = 3^{\circ}$ an upper bound of $\left|\delta_{13}^{\ell\,LR}\right|\leq
0.3$.
The constraints on $\delta_{13}^{\ell\,LR}$ from $\tau\rightarrow e \gamma$ are of the
order
of $0.02$ \cite{Girrbach:2009uy} and in general better than our derived bounds if
$\theta_{13}$ is non-zero. As an important consequence, we note that $\tau\rightarrow e
\gamma$
impedes any measurable correction from supersymmetric loops to
$U_{e3}$ : E.g.\ for sparticle masses of 500 GeV we find $|\Delta
U_{e3}|\leq 10^{-3}$ corresponding to a correction to the mixing angle
$\theta_{13}$ of at most $0.06^{\circ}$. That is, if the DOUBLE CHOOZ
experiment measures $U_{e3}\neq 0$, one will not be able to ascribe
this result to the SUSY breaking sector. Stated positively,
$U_{e3}\gtrsim 10^{-3}$ will imply that at low energies the flavor
symmetries imposed on the Yukawa sector to motivate tri-bimaximal
mixing are violated. This finding confirms the pattern found in \cite{Girrbach:2009uy}
where
the product $\delta^{\ell\;LL}_{13}\delta^{\ell\;LR}_{33}$ has been studied instead of
$\delta^{\ell\;LR}_{13}$.
\begin{figure}
\includegraphics[width=.7\linewidth]{PlotProzentvsdeltaMSUSY1000MS.eps}
\hspace{.01\linewidth}
\includegraphics[width=.7\linewidth]{PlotProzentvsth13MSUSY1000MS.eps}
\caption{$|\Delta U_{e3}|/U_{e3}$ in \%. Top: as a function of $\delta_{13}^{\ell\,LR}$
for $M_{\text{SUSY}} = 1000$ GeV and different values of $\theta_{13}$ (green $1^{\circ}$;
blue: $3^{\circ}$; red: $5^{\circ}$ ). Bottom: as a function of $\theta_{13}$ for
$M_{SUSY} = 1000$ GeV and different values of $\delta_{LR}^{\ell\,13}$ (red:
$\delta_{13}^{\ell\,LR} = 0.5$; blue: $\delta_{13}^{\ell\,LR} = 0.3$; green:
$\delta_{13}^{\ell\,LR} = 0.1$) (both from top to bottom)}
\label{fig:PMNSMSUSY1000}
\end{figure}
\begin{figure}
\includegraphics[width=.7\linewidth]{Plotth13vsdeltaMSUSY3.eps}
\caption{The excluded $\left(\theta_{13},\delta_{13}^{\ell\,LR}\right)$-region is below
the curves for (from bottom to top) $M_\text{SUSY} = $ 500 GeV (red), 1000 GeV (blue),
2000 GeV (green) and 5000 GeV (yellow). The black dashed line denotes the future
experimental sensitivity to $\theta_{13} = 3^{\circ}$.}
\label{fig:th13vsdelta}
\end{figure}
\subsection{Constraints from two-loop corrections to fermion masses}
Combining two flavor-violating self-energies can have
sizable impacts on the light fermion masses according to \eq{equ:massRen2}. Requiring that
no large numerical cancellations should occur between the tree-level mass (which is absent
in the case of a radiative fermion mass) and the supersymmetric loop corrections we can
derive bounds on the products $\delta^{f\,LR}_{IK}\delta^{f\,LR}_{KI}$ which contain the
so far less constrained elements $\delta^{f\,LR}_{KI}$, $K>I$.
We apply the fine-tuning argument to the two-loop
contribution originating from flavor-violating $A$-terms, e.g. $\left|\Sigma_{11}^{f\,LR
(2)}\right|\leq m_{f_1}$. The bound $\Sigma_{11}^{f\,LR (2)} = m_{f_1}$ corresponds to a
100\% change in the fermion mass through supersymmetric loop corrections which is
equivalent to the case that the fermion Yukawa coupling vanishes.
The upper bound depends on the overall SUSY mass scale and is roughly given as
\begin{equation}
\left|\delta_{I3}^{q\,LR}\delta_{3I}^{q\,LR}\right|\lesssim \frac{9 \pi^2\, m_{q_I}m_{q_3}(M_{SUSY})}{(\alpha_s(M_{SUSY}) M_\text{SUSY})^2},\;\;I\neq3
\label{quark2}
\end{equation}
for quarks and
\begin{equation}
\left|\delta_{13}^{\ell\,LR}\delta_{31}^{\ell\,LR}\right|\lesssim \frac{64 \pi^2
m_{\ell_1}m_{\ell_3}}{(\alpha_1 M_\text{SUSY})^2}
\label{lepton2}
\end{equation}
for leptons. Again, \eq{lepton2} can be further simplified
\begin{equation}
\left|\delta_{13}^{\ell\,LR}\delta_{31}^{\ell\,LR}\right|\leq 0.021\left(\frac{500\,
\text{GeV}}{M_{\text{SUSY}}}\right)^2.
\end{equation}
The contributions proportional to $\delta_{13}^{f\,LR}\delta_{31}^{f\,LR}$ cannot be
important,
since these elements are already severely constrained by FCNC processes
\cite{Dittmaier:2007uw}. As studied in
Ref. \cite{Plehn:2009it}, single-top production involves the same mass
insertion $\delta_{31}^{u\,LR}$ which can also induce a right-handed $W$ coupling if at
the same time $\delta_{33}^{d\,LR}\neq 0$ \cite{Crivellin:2009sd}.
Therefore our bound can be used to place a constraint on this cross
section. Also the product $\delta_{23}^{u,\ell\,LR}\delta_{32}^{u,\ell\,LR}$ cannot be
constrained, since the muon and the charm are too heavy. However,
$\delta_{23}^{d\,LR}\delta_{32}^{d\,LR}$ can be constrained as shown in
Fig.~(\ref{fig:mstrange2loop}). Our results for the up, down, and electron mass are
depicted in Fig.~(\ref{fig:mup2loop}),(\ref{fig:mdown2loop}) and (\ref{fig:me2loop}). In
the quark case also the bounds from the CKM renormalization on $\delta_{13,23}^{q\,LR}$
are taken into account.
\section{Conclusions}\label{sec:Conclusions}
According to 't~Hooft's naturalness principle, the smallness of a quantity is linked to a
symmetry that is restored if the quantity is zero. The smallness of the Yukawa couplings
of the first two generations (as well as the small CKM elements involving the third
generation) suggest the idea that Yukawa couplings (except for the third generation) are
generated through radiative corrections
\cite{Weinberg:1972ws,Donoghue:1983mx,Borzumati:1999sp,Ferrandis:2004ri,Crivellin:2008mq,
Crivellin:2009pa}. It might well be that the chiral flavor symmetry is broken by soft
SUSY-breaking terms rather than by the trilinear tree-level Yukawa couplings.
We use 't~Hooft's naturalness criterion to constrain the chirality-changing mass insertion
$\delta_{IJ}^{u,d,\ell\,LR}$ from the mass and CKM renormalization. Therefore, we compute
the
finite renormalization of fermion masses and mixing angles in the MSSM, taking into
account the leading two-loop effects. These corrections are not only important, in order
to obtain a unitary CKM matrix, they are also numerically important for light fermion
masses. This allows us to constrain the product $\delta_{13}^{f\,LR}\delta_{31}^{f\,LR}$
(and $\delta_{23}^{d\,LR}\delta_{32}^{d\,LR}$) which is important, especially with respect
to the before unconstrained element $\delta_{13}^{u\,RL}$. All constraints given in this
paper are non-decoupling. This means they do not vanish in the limit of infinitely heavy
SUSY masses unlike the bounds from FCNC processes. Therefore our constraints are always
stronger than the FCNC constraints for sufficiently heavy SUSY (and Higgs) masses.
The PMNS renormalization is a bit more involved since the matrix is not
hierarchical. The radiative decay $\tau\to e \gamma$ severely limits the size of the loop
correction $\Delta U_{e3}$ to the PMNS element $U_{e3}$. In a previous paper we
have studied this topic for effects triggered by the product
$\delta_{13}^{\ell\,LL}\delta_{33}^{\ell\,LR}$ \cite{Girrbach:2009uy}. In this paper we
have
complemented that analysis by investigating $\delta_{13}^{\ell\,LR}$ instead. Assuming
reasonable slepton masses and noting that the Daya Bay neutrino experiment is only
sensitive to values of $\theta_{13}$ above $3^\circ$, we conclude that the threshold
corrections to $U_{e3}$ are far below the measurable limit. Consequently, if a symmetry at
a high scale imposes tri-bimaximal mixing, SUSY loop corrections cannot spoil this
prediction $\theta_{13} = 0$ at the weak scale. This is an important result for the proper
interpretation of a measurement of $\theta_{13}$. Thus if DOUBLE CHOOZ or Daya Bay
neutrino experiment will measure a non-zero $\theta_{13}$ then this is also true at a high
energy scale.
\begin{figure}
\includegraphics[width=.8\linewidth]{PlotRegionMSUSYscen4YukNull.eps}
\includegraphics[width=.8\linewidth]{PlotboundsMSUSYYukNull2.eps}
\caption{Results of the two-loop contribution to the electron mass.
Above: Region compatible with the naturalness principle for (from top to bottom)
$M_{\text{SUSY}} =$ 200 GeV (yellow), 500 GeV (green), 800 GeV (blue), 1000 GeV (red).
Bottom: Allowed range for $\delta_{13}^{\ell\,LR}\delta_{31}^{\ell\,LR}$ as a function of
$M_\text{SUSY}$.}
\label{fig:me2loop}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\linewidth]{delta-u-13-31.eps}
\includegraphics[width=0.8\linewidth]{delta-u-13-31-Produkt.eps}
\caption{Results of the two-loop contribution to the up quark mass. Above: Region compatible with the naturalness principle (100\% bound) for (from top to bottom) $M_{\text{SUSY}} =$ 500 GeV (yellow), 1000GeV (green), 1500 GeV (blue), 2000 GeV (red). Bottom: Allowed range for $\delta_{13}^{u\,LR}\delta_{31}^{u\,LR}$ as a function of $M_\text{SUSY}$.}
\label{fig:mup2loop}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\linewidth]{delta-d-13-31.eps}
\includegraphics[width=0.8\linewidth]{delta-d-13-31-Produkt.eps}
\caption{Results of the two-loop contribution to the down quark mass. Above: Region compatible with the naturalness principle for (from top to bottom) $M_{\text{SUSY}} =$ 500 GeV (yellow), 1000GeV (green), 1500 GeV (blue), 2000 GeV (red). Bottom: Allowed range for $\delta_{13}^{d\,LR}\delta_{31}^{d\,LR}$ as a function of $M_\text{SUSY}$.}
\label{fig:mdown2loop}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\linewidth]{delta-d-23-32.eps}
\includegraphics[width=0.8\linewidth]{delta-d-23-32-Produkt.eps}
\caption{Results of the two-loop contribution to the strange quark mass. Above: Region compatible with the naturalness principle (100\% bound) for (from top to bottom) $M_{\text{SUSY}} =$ 500 GeV (yellow), 1000GeV (green), 1500 GeV (blue), 2000 GeV (red). Bottom: Allowed range for $\delta_{23}^{d\,LR}\delta_{32}^{d\,LR}$ as a function of $M_\text{SUSY}$.}
\label{fig:mstrange2loop}
\end{figure}
\begin{acknowledgments}
We like to thank Ulrich Nierste for
helpful discussions and proofreading the article.
This work is supported by BMBF grants 05HT6VKB and
05H09VKF and by the EU Contract No.~MRTN-CT-2006-035482,
\lq\lq FLAVIAnet''. Andreas Crivellin and Jennifer Girrbach acknowledge the financial support by the State of
Baden-W\"urttemberg through \emph{Strukturiertes Promotionskolleg
Elementarteilchenphysik und Astroteilchenphysik}\ and the \emph{Studienstiftung des deutschen Volkes},
respectively.
\end{acknowledgments}
\begin{appendix}
\section{Conventions} \label{sec:appendix}
\subsection{Loop integrals}
For the self-energies, we need the following loop integrals:
\begin{align}
B_{0}(x,y) & = -\Delta - \frac{x}{x-y} \ln{\frac{x}{\mu^{2}}} -
\frac{y}{y-x} \ln{\frac{y}{\mu^{2}}},\\
& \textrm{with}\quad
\Delta = \frac{1}{\epsilon}-\gamma_{E} + \ln 4\pi. \nonumber
\\[3pt]
C_{0}(x,y,z) & = \frac{xy\ln{\frac{x}{y}} +
yz\ln{\frac{y}{z}}+ xz\ln{\frac{z}{x}}}{(x-y)(y-z)(z-x)} .
\end{align}
\subsection{Diagonalization of mass matrices and Feynman rules}\label{sec:appendixDiag}
For the vacuum expectation value we choose the normalization without the factor $\sqrt{2}$ and define the Yukawa couplings in the following way:
\begin{align}
v = \sqrt{v_{u}^{2}+v_{d}^{2}}=174\textrm{ GeV},\quad \tan\beta = \frac{v_u}{v_d},\\
m_{l} = -v_{d} Y_{l},\quad m_{d} = -v_{d} Y_{d},\quad m_{u} = v_{u} Y_{u}.
\end{align}
\subsubsection*{Neutralinos $\tilde{\chi}^{0}_{i}$}
In the following we mainly use the convention of \cite{Rosiek}.
\begin{align}
& \Psi^{0} = \left(\tilde{B},\tilde{W},\tilde{H}^{0}_{d},
\tilde{H}^{0}_{u}\right) ,\nonumber \\
& \mathcal{L}_{\tilde{\chi}^{0}_{mass}}
= -\frac{1}{2}(\Psi^{0})^\top M_{N}\Psi^{0} + \text{h.c.}
\nonumber
\\
\label{Neutralinomassenmatrix}
& M_{N} =
\begin{pmatrix}
M_{1} & 0 & -\frac{g_{1}v_{d}}{\sqrt{2}} &
\frac{g_{1}v_{u}}{\sqrt{2}}
\\
0 & M_{2} & \frac{g_{2}v_{d}}{\sqrt{2}} &
-\frac{g_{2}v_{u}}{\sqrt{2}}
\\
-\frac{g_{1}v_{d}}{\sqrt{2}} & \frac{g_{2}v_{d}}{\sqrt{2}}& 0&
-\mu
\\
\frac{g_{1}v_{u}}{\sqrt{2}} & -\frac{g_{2}v_{u}}{\sqrt{2}}& -\mu&
0
\end{pmatrix}.
\end{align}
$M_{N}$ can be diagonalised with an unitary transformation such that
the eigenvalues are real and positive.
\begin{equation}
\label{Neutralinodiag}
Z_{N}^\top M_{N}Z_{N}=M_{N}^{D} =
\begin{pmatrix}
m_{\tilde{\chi}_{1}^{0}} & &0 \cr &\ddots & \cr 0 & &
m_{\tilde{\chi}_{4}^{0}}
\end{pmatrix}.
\end{equation}
%
For that purpose, $Z_{N}^{\dagger} M_{N}^{\dagger} M_{N} Z_{N} = (M_{N}^{D})^{2}$ can be
used.
$Z_{N}$ consists of the eigenvectors of the Hermitian matrix $M_{N}^{\dagger} M_{N}$.
Then the columns can be multiplied with phases $e^{i\phi}$, such that
$Z_{N}^{T}M_{N}Z_{N}=M_{N}^{D}$ has positive and real diagonal elements.
\subsubsection*{Charginos $\tilde{\chi}^{\pm}_{i}$}
\begin{align}
&\Psi^{\pm} = \left( \tilde{W}^+,\, \tilde{H}_{u}^+,\,
\tilde{W}^{-},\, \tilde{H}_{d}^{-} \right), \nonumber \\
&\mathcal{L}_{\tilde{\chi}^{\pm}_{mass}} = -\frac{1}{2}
\left(\Psi^{\pm}\right)^\top M_{C} \Psi^{\pm} + \text{h.c.}
\nonumber
\\
& M_{C} =
\begin{pmatrix}
0 & X^\top \cr X & 0
\end{pmatrix}
, \qquad X =
\begin{pmatrix}
M_{2} & g_{2} v_{u} \cr g_{2} v_{d} & \mu
\end{pmatrix}
\label{Charginomassenmatrix}.
\end{align}
The rotation matrices for the positive and negative charged fermions differ, such that
\begin{equation}
Z_{-}^{T}XZ_{+} = \left(\begin{array}{cc}
m_{\tilde{\chi}_{1}} & 0 \\
0 & m_{\tilde{\chi}_{2}}
\end{array}
\right).
\end{equation}
\subsubsection*{Sleptons}
The sleptons $\tilde{L}_{2}^{I} = \tilde{e}_{IL}$ and $\tilde{R}^{I} =
\tilde{e}_{IR}^{+}$ mix
to six charged mass eigenstates $\tilde{L}_{i},\,i=1\dots6$:
\begin{align}
&\tilde{L}_{2}^{I} = W_{L}^{Ii*}\tilde{\ell}_{i}^{-} , \quad
\tilde{R}^{I} = W_{L}^{(I + 3)i}\tilde{\ell}_{i}^{ + } ,\nonumber\\
&W_{L}^{\dagger}
\begin{pmatrix}
(m_{L}^{2})_{LL}& (m_{L}^{2})_{LR} \cr
(m_{L}^{2})_{RL}^{\dagger}&(m_{L}^{2})_{RR}
\end{pmatrix}
W_{L}= \text{diag}\left(m_{\tilde{\ell}_{1}}^{2}, \ldots, m_{\tilde{\ell}_{6}}^{2}
\right)\nonumber ,
\end{align}
%
and the slepton mass matrix is composed of
\begin{eqnarray}
(m_{L}^{2})_{LL}^{IJ} & = & \frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)
\left(1-2 c_{W}^{2}\right)}{4s_{W}^{2} c_{W}^{2}}\delta_{IJ} +
v_{d}^{2}Y_{\ell_{I}}^{2}\delta_{IJ}\nonumber\\
&\;&+ (m_{\tilde{L}}^{2})_{IJ}^{T},\nonumber\\
(m_{L}^{2})_{RR}^{IJ} & = & -\frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)}{2 c_{W}^{2}}
\delta_{IJ}+v_{d}^{2}Y_{\ell_{I}}^{2}\delta_{IJ}+m_{\tilde{\overline{e}}IJ}^{2},
\nonumber\\
(m_{L}^{2})_{LR}^{IJ} & = & v_{u}\mu Y_{\ell}^{IJ*}+v_{d}A_{\ell}^{IJ*}.\nonumber
\end{eqnarray}
\subsubsection*{Lepton-slepton-neutralino coupling}
Feynman rule for incoming lepton $\ell_I$, outgoing neutralino and slepton
$\tilde{\ell}_{i}$:
\begin{align}
i\Gamma_{\ell_I}^{\tilde{\chi}_{j}^{0}\tilde{\ell}_{i}}=&i\underbrace{\left(\frac{
W_{ L} ^{Ii}}{\sqrt{2}}\left(g_{1}Z_{N}^{1j}+g_{2}Z_{N}^{2j}\right)+Y_{\ell_I}
W_{L}^{(I+3)i}Z_{N}^{3j}\right)}_{=
\Gamma_{\ell_{IL}}^{\tilde{\chi}_{j}^{0}\tilde{\ell}_{i}}}P_{L}\nonumber\\
&+i\underbrace{\left(-g_{1}\sqrt{2}W_{L}^{(I+3)i}Z_{N}^{1j*}+Y_{\ell_I}
W_{L}^{Ii}Z_{N}^{3j*}\right)}_{=
\Gamma_{\ell_{IR}}^{\tilde{\chi}_{j}^{0}\tilde{\ell}_{i}}}P_{R}.
\end{align}
\subsubsection*{Lepton-sneutrino-chargino coupling}
Feynman rule for incoming lepton $\ell_I$, outgoing chargino and sneutrino
$\tilde{\nu}_J$:
\begin{equation}
i\Gamma_{\ell_{I}}^{\tilde{\nu}_{J}\tilde{\chi}^{\pm}_{i}} =
-i\left(g_{2}Z_{ + }^{1i} P_{L} + Y_{\ell_{I}}Z_{-}^{2i*}
P_{R}\right)W_{\nu}^{IJ*} .\nonumber
\end{equation}
\subsubsection*{Down-squarks}
The down-squarks $\tilde{Q}_2^I =\tilde{d}_{IL} $ and $\tilde{D}^I = \tilde{d}_{IR}^{*}$ mix to six mass eigenstates $\tilde{d}_i,\,i=1\dots6$:
\begin{align}
&\tilde{Q}_{2}^{I} = W_{D}^{Ii*}\tilde{d}_{i}^{-} , \quad
\tilde{D}^{I} = W_{D}^{(I + 3)i}\tilde{d}_{i}^{ + } ,\nonumber\\
&W_{D}^{\dagger}
\begin{pmatrix}
(m_{D}^{2})_{LL}& (m_{D}^{2})_{LR} \cr
(m_{D}^{2})_{RL}^{\dagger}&(m_{D}^{2})_{RR}
\end{pmatrix},
W_{D}= \text{diag}\left(m_{\tilde{d}_{1}}^{2}, \ldots, m_{\tilde{d}_{6}}^{2} \right)\nonumber,
\end{align}
and the downs-squark mass matrix is composed of
\begin{eqnarray}
(m_{D}^{2})_{LL}^{IJ} & = & -\frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)\left(1+2 c_{W}^{2}\right)}{12s_{W}^{2} c_{W}^{2}}\delta_{IJ} \nonumber\\
&\;&+ v_{d}^{2}Y_{d_{I}}^{2}\delta_{IJ} + (m_{\tilde{Q}}^{2})_{IJ}^{T},\nonumber\\
(m_{D}^{2})_{RR}^{IJ} & = & -\frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)}{6 c_{W}^{2}}\delta_{IJ}+v_{d}^{2}Y_{d_{I}}^{2}\delta_{IJ}+m_{\tilde{\overline{d}}IJ}^{2},\nonumber\\
(m_{D}^{2})_{LR}^{IJ} & = & v_{u}\mu Y_{d}^{IJ*}+v_{d}A_{d}^{IJ*}.\nonumber
\end{eqnarray}
\subsubsection*{Up-squarks}
Finally, one has six up-squarks $\tilde{u}_i$ composed from fields $\tilde{Q}_1^I =\tilde{u}_{IL} $ and $\tilde{U}^I = \tilde{u}_{IR}^{I}$
\begin{align}
&\tilde{Q}_{1}^{I} = W_{U}^{Ii}\tilde{u}_{i}^{+} , \quad
\tilde{D}^{I} = W_{U}^{(I + 3)i*}\tilde{u}_{i}^{ - } ,\nonumber\\
&W_{U}^{T}
\begin{pmatrix}
(m_{U}^{2})_{LL}& (m_{U}^{2})_{LR} \cr
(m_{U}^{2})_{RL}^{\dagger}&(m_{U}^{2})_{RR}
\end{pmatrix},
W_{U}^*= \text{diag}\left(m_{\tilde{u}_{1}}^{2}, \ldots, m_{\tilde{u}_{6}}^{2} \right)\nonumber .
\end{align}
\begin{eqnarray*}
(m_{U}^{2})_{LL}^{IJ} & = & -\frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)\left(1-4 c_{W}^{2}\right)}{12s_{W}^{2} c_{W}^{2}}\delta_{IJ} \\ \nonumber
&\;&+ v_{u}^{2}Y_{u_{i}}^{2}\delta_{IJ} + (V m_{\tilde{Q}}^{2}V^\dagger)_{IJ}^{T},\nonumber\\
(m_{U}^{2})_{RR}^{IJ} & = & \frac{e^{2}\left(v_{d}^{2}-v_{u}^{2}\right)}{3 c_{W}^{2}}\delta_{IJ}+v_{u}^{2}Y_{u_{I}}^{2}\delta_{IJ}+m_{\tilde{\overline{u}}IJ}^{2},\nonumber\\
(m_{U}^{2})_{LR}^{IJ} & = & -v_{d}\mu Y_{u}^{IJ*}-v_{u}A_{u}^{IJ*}.\nonumber
\end{eqnarray*}
\subsubsection*{Quark-squark-gluino coupling}
Feynman rule for incoming quark $d_I,\,u_I$, outgoing gaugino and squark $\tilde{d}_{i},\,\tilde{u}_{i}$:
\begin{align}
i\Gamma_{d_I}^{\tilde{g}\tilde{d}_i} &= i g_s \sqrt{2}T^a\left(-W_D^{Ii}P_L + W_D^{(I+3)i}P_R\right),\\
i\Gamma_{u_I}^{\tilde{g}\tilde{u}_i} &= i g_s \sqrt{2}T^a\left(-W_U^{Ii*}P_L + W_U^{(I+3)i*}P_R\right).
\end{align}
\subsubsection*{Quark-squark-neutralino coupling}
Feynman rule for incoming quark $d_I,\,u_I$, outgoing neutralino and squark $\tilde{d}_{i},\,\tilde{u}_{i}$:
\begin{align}
i\Gamma_{d_I}^{\tilde{\chi}_j^0\tilde{d}_i} = & i \left(\frac{W_D^{Ii}}{\sqrt{2}}\left(-\frac{g_1}{3}Z_N^{1j} + g_2 Z_N^{2j}\right) + Y_{d_I}W_D^{(I+3)i}Z_N^{3j}\right) P_L\nonumber\\
& +i \left(-\frac{\sqrt{2}g_1}{3}W_D^{(I+3)i}Z_N^{1j*} + Y_{d_I} W_D^{Ii}Z_N^{3j*}\right) P_R\nonumber,\\
i\Gamma_{u_I}^{\tilde{\chi}_j^0\tilde{u}_i} = & i \left(\frac{W_U^{Ii*}}{\sqrt{2}}\left(-\frac{g_1}{3}Z_N^{1j} - g_2 Z_N^{2j}\right) - Y_{u_I}W_U^{(I+3)i*}Z_N^{4j}\right) P_L\nonumber\\
& +i \left(\frac{2\sqrt{2}g_1}{3}W_U^{(I+3)i*}Z_N^{1j*} - Y_{u_I} W_U^{Ii*}Z_N^{4j*}\right) P_R\nonumber.
\end{align}
\subsubsection*{Quark-squark-chargino coupling}
Feynman rule for incoming quark $d_I,\,u_I$, outgoing chargino and squark $\tilde{u}_{i},\,\tilde{d}_{i}$:
\begin{align}
i\Gamma_{d_I}^{\tilde{\chi}^{\pm}_j\tilde{u}_i} = & i \left(-g_2 W_U^{Ji*}Z_+^{1j} + Y_{u_J}W_U^{(J+3)i*}Z_+^{2j}\right)V^{JI} P_L\nonumber\\
& + i\left(- Y_{d_I}W_U^{Ji*}Z_-^{2j*}\right)V^{JI} P_R.,\nonumber\\
i\Gamma_{u_I}^{\tilde{\chi}^{\pm}_j\tilde{d}_i} = & i \left(-g_2 W_D^{Ji}Z_-^{1j} - Y_{d_J}W_D^{(J+3)i}Z_-^{2j}\right)V^{JI*} P_L\nonumber\\
& + i\left( Y_{u_I}W_D^{Ji}Z_+^{2j*}\right)V^{JI*} P_R.\nonumber
\end{align}
\end{appendix}
|
\section{Introduction}
Let $M$ be a non-empty closed subset of $\real^k$. A subset $C$ of $M$ is said to be \emph{$M$-convex} if $C$ is intersection of $M$ and a convex set in $\real^k$. The set $M$ endowed with the collection of $M$-convex subsets becomes a convexity space (see \cite{MR1234493}). Let $h(M)$ be the \emph{Helly number} of $M$, i.e. the minimal possible $h$ satisfying the following condition.
\begin{itemize}
\item[$(H)$] Every finite collection $C_1,\ldots,C_m$ ($m \ge h$) of $M$-convex sets, for which every sub-collection of $h$ elements has non-empty intersection, necessarily satisfies $C_1 \cap \cdots \cap C_m \ne \emptyset.$
\end{itemize}
If $h$ as above does not exist we set $h(M) := \infty.$ The main purpose of this note is to study Helly type results in spaces $M \subseteq \real^k$ paying special attention to \emph{mixed integer spaces}, i.e. sets of the form $M=\real^n \times \integer^d$, where $n, d \ge 0.$ For surveys of numerous Helly type results we refer to \cite{MR0157289}, \cite{MR1242986}, \cite{MR1228043} and the monograph \cite{MR1234493}.
It is known that $h(\real^n)=n+1,$ by the classical Helly Theorem (see \cite[Theorem\,1.1.6]{schneider}), and $h(\integer^d)=2^d$, by a result due to Doignon \cite[(4.2)]{MR0387090}, which was independently discovered by Bell \cite{Bell77} and Scarf \cite{MR0452678} (see also \cite[Theorem\,16.5]{schrijver} and \cite[p.\,176]{MR634767}). We obtain the following theorem.
\begin{theorem} \label{mixed:helly:bounds} Let $M \subseteq \real^k$ be non-empty and closed and let $n,d \ge 0$ be integers. Then
\begin{align}
h( \real^n \times M) & \le (n+1) h(M), \label{mixed:with:reals} \\
2^d h(M) & \le h( M \times \integer^d ), \label{mixed:with:integers} \\
h(\real^n \times \integer^d) & = (n+1) 2^d. \label{mixed:helly:eq}
\end{align}
\end{theorem}
Clearly, \eqref{mixed:helly:eq} is a direct consequence of \eqref{mixed:with:reals}, \eqref{mixed:with:integers} and the theorems of Helly and Doignon. Equality \eqref{mixed:helly:eq} is a common extension of $h(\real^n) = n+1$ and $h(\integer^d) = 2^d,$ and thus it can be viewed as Helly's theorem for mixed integer spaces.
We notice that in general \eqref{mixed:with:integers} cannot be improved to equality (and thus, the case $M=\real^n$, for which we derive the equality, is quite likely an exception). For showing this we define $M:=\{0,1,2,2.5\}$ so that $h(M)=2.$ It turns out that $h(M \times \integer) \ge 5 > 2 h(M).$ In fact, consider the set $A:= \{(0,0),(1,0),(1,1),(2,1),(2.5,2)\}$, see also Fig.~\ref{example}. Then the five sets $A \setminus \{a\},$ $a \in A,$ are $(M \times \integer)$-convex and do not satisfy $(H)$ for $h=4,$ though each four of them have non-empty intersection.
\begin{FigTab}{cc}{0.7mm}
\begin{picture}(70,65)
\put(0,0){\IncludeGraph{width=70\unitlength}{example.eps}}
\end{picture}
\\
\parbox[t]{0.85\textwidth}{\mycaption{Example for the case $h(M \times \integer^d) > 2^d h(M)$ for $d=1$\label{example}}}
\end{FigTab}
We wish to emphasize that theorems of Helly type are related, in a natural way, to the theory of linear inequalities. By this they also play a role in the theory of linear and integer programming, see for example \cite[Chapter\,7]{schrijver}, \cite[\S\,21]{MR1451876} and \cite{Clarkson95}. The following statement provides equivalent formulations of $h(M)$ in terms common for linear programming.
\begin{proposition} \label{helly:over:certificates}
Let $M \subseteq \real^k$ be non-empty and $h \ge 0$ be an integer. Then (H) is equivalent to each of the following two conditions.
\begin{itemize}
\item[$(A)$] For every collection of affine-linear functions $a_1,\ldots, a_m$ $(m \ge h)$ on $\real^k$ either $a_1(x) \ge 0,\ldots, a_m(x) \ge 0$ has a solution $x \in M$ or, otherwise, there exist $1 \le i_1,\ldots, i_h \le m$ such that the system $a_{i_1}(x) \ge 0, \ldots, a_{i_h}(x) \ge 0$ has no solution $x \in M.$
\item[$(B)$] For every collection of affine-linear functions $b_1,\ldots,b_m, c$ $(m \ge h-1)$ on $\real^k$ such that $\mu:=\sup \setcond{c(x)}{b_1(x) \ge 0,\ldots,b_m(x) \ge 0, \ x \in M}$ is finite there exist $1 \le i_1,\ldots,i_{h-1} \le m$ such that $\mu = \sup \setcond{c(x)}{b_{i_1}(x) \ge 0,\ldots,b_{i_{h-1}}(x) \ge 0, \ x \in M}.$
\end{itemize}
\end{proposition}
Equivalence $(H) \Longleftrightarrow (B)$ above is an extension of the result of Scarf \cite{MR0452678} (see also \cite{Todd1977} and \cite[Corollary\,16.5a]{schrijver}).
Since $h(\real^n \times \integer^d)$ is linear in $n,$ formula \eqref{mixed:helly:eq} is in correspondence with the polynomial solvability of linear mixed integer programs in the case that the number of integer variables is fixed, see~\cite{lenstra83}. In terms of linear inequalities, \eqref{mixed:helly:eq} states that, in the mixed integer case, the (largest) number of inequalities in the insolvability certificate doubles if we introduce another integer variable and increases by $2^d$ (where $d$ is the number of integer variables) if we introduce another real variable.
\newcommand{\fh}{\bar{h}}
\newcommand{\aff}{\rmcmd{aff}}
We also wish to discuss fractional Helly's theorems in spaces $M \subseteq \real^k$. The \emph{fractional Helly number} $\fh(M)$ of $M$ is defined to be the minimal $h$ such that for every $0<\alpha \le 1$ there exists $0<\beta=\beta(\alpha,M) \le 1$ such that every collection of $M$-convex sets $C_1,\ldots,C_n$ ($n \in \natur$) satisfies the following condition.
\begin{itemize}
\item[$(F)$] If $\bigcap_{i \in I} C_i \ne \emptyset$ for at least $\alpha \binom{n}{h}$ sets $I \subseteq \{1,\ldots,n\}$ of cardinality $h,$ then at least $\beta n$ elements of $C_1,\ldots,C_n$ have a common point.
\end{itemize}
If $h$ as above does not exist, we set $\fh(M)=\infty.$ It is known that $\fh(\real^d)=\fh(\integer^d)=d+1$, where the relation for $\real^d$ is due to Katchalski and Liu \cite{MR532152} and for $\integer^d$ due to B\'ar\'any and Matou\v{s}ek \cite{bar-mat-03}. B\'ar\'any and Matou\v{s}ek \cite[Remark at p.\,234]{bar-mat-03} point out that their arguments showing $\fh(\integer^d)=2^d$ do not use much of the geometry of $\integer^d$. Motivated by this observation, we prove the following extension.
\begin{theorem} \label{fh thm}
Let $M \subseteq \real^d$ be non-empty and closed and let $h(M)<\infty$. Then $\fh(M) \le d +1.$
\end{theorem}
Theorem~\ref{fh thm} is a common extension of the fractional Helly results from \cite{MR532152} and \cite{bar-mat-03}. Clearly, Theorem~\ref{fh thm} applies to mixed integer spaces, and by this the fractional Helly number of $\integer^d \times \real^n$ is $d+n+1.$ The fractional Helly Theorem also implies the so-called $(p,q)$-theorem for $p \ge q \ge d+1$ in spaces $M \subseteq \real^d$ with finite Helly number $h(M)$, see \cite[p.\,229]{bar-mat-03} and \cite{MR1921545}.
The proof of Theorem~\ref{fh thm} follows the ideas from \cite{bar-mat-03}. In particular, we need a colored Helly's theorem for spaces $M.$ Let $I_1,\ldots,I_{d+1}$ be some pairwise disjoint sets of cardinality $t.$ We introduce the set $$K^{d+1}(t):= \setcond{ \{i_1,\ldots,i_{d+1}\}}{i_j \in I_j \ \mbox{for} \ j =1,\ldots,d+1},$$ the so-called \emph{complete $(d+1)$-uniform $(d+1)$-partite hyperpgraph}. The foregoing notations are used in the following theorem.
\begin{theorem} \label{col h thm} {\upshape (Colored Helly's theorem for $M$-convex sets).}
Let $M \subseteq \real^d$ be non-empty and let $h(M)<\infty.$ Then for every $r \ge 2$ there exists an integer $t=t(r,M)$ with the following property. For every family of $M$-convex sets $C_i$, $i \in I_1 \cup \ldots \cup I_{d+1}$ for which $\bigcap_{j=1}^{d+1} C_{i_j} \ne \emptyset$ for every $\{i_1,\ldots,i_{d+1}\} \in K^{d+1}(t),$ there exists $j \in \{1,\ldots,d+1\}$ and $R \subseteq I_j$ with $\card{R}=r$ such that $\bigcap_{i \in R} C_i \ne \emptyset.$
\end{theorem}
We do not need to give a proof of Theorem~\ref{col h thm}. In fact, the proof in our case is identical to the proof for the case $M=\integer^d$ from \cite{bar-mat-03}, based on Lov\'asz's colored Helly theorem, the theorem of Erd\H{o}s and Simonovits on super-saturated hypergraphs, and combinatorial arguments which rely only on the fact that the Helly number of $\integer^d$ is finite.
As a consequence of \eqref{mixed:helly:eq} we can also estimate the Radon number of mixed integer spaces. For a non-empty $M \subseteq \real^d$ the \emph{Radon number} $r(M)$ of the space $M$ is defined to be the minimal $r$ such that for every $A \subseteq M$ with $\card{A} \ge r$ there exist disjoint sets $B, C \subseteq A$ with $M \cap \conv B \cap \conv C \ne \emptyset$. For $B$ and $C$ satisfying the previous condition the set $\{B,C\}$ is called a \emph{Radon partition} of $A$ in $M$. If $r$ as above does not exist we set $r(M) := \infty.$ By Radon's Theorem, $r(\real^n) = n+2$ (see \cite[Theorem\,1.1.5]{schneider}). It is known that $\Omega(2^d) = r(\integer^d) = O(d 2^d),$ see \cite{onn91}. The only known known value of $r(\integer^d)$ for $d \ge 2$ is $r(\integer^2) = 6$, see \cite{onn91} and \cite{MR2007959}. Furthermore, for $d=3$ one has $11 \le r(\integer^3) \le 17$ (see \cite{onn91} and \cite{MR2007959}).
\begin{theorem} \label{radon:bounds}
Let $M \subseteq \real^k$ be non-empty and closed and let $n, d\ge 0$ be integers. Then
\begin{align}
& r(M \times \integer^d) \ge (r(M) -1) 2^d + 1, \label{radon:with:integers} \\
(n+1) 2^d + 1 \le & \, r(\real^n \times \integer^d) \le (n+d) (n+1) 2^d - n - d + 2. \label{radon:mixed:integers}
\end{align}
\end{theorem}
We emphasize that the lower and upper bound in \eqref{radon:mixed:integers} are linear and quadratic in $n$, respectively. Thus, the exact asymptotics of $r(\real^n \times \integer^d)$ with respect to $n$ remains undetermined. Having \eqref{mixed:helly:eq} and \eqref{radon:mixed:integers}, it is suggestive to look for some type of Carath\'eodory's theorem in mixed integer spaces. However, the authors are currently not aware of any non-trivial notion of a Carath\'eodory number for $\real^n \times \integer^d.$
For dimension two $r(M)$ is uniquely determined by $h(M)$ with the only exceptional case $h(M)=4.$
\begin{theorem} \label{hel rad 2dim}
Let $M \subseteq \real^2$ be non-empty. If $h(M) \ne 4,$ then $r(M) = h(M)+1.$ For $h(M)=4,$ one has $r(M) \in \{5,6\}$.
\end{theorem}
It is not hard to see that for $h(M)=4,$ both cases $r(M)=5$ and $r(M)=6$ are possible. It is known that $r(\integer^2)=6,$ and it is not hard to verify that $h(\integer \times \real)=5.$ It would also be interesting to study the relationship between $r(M)$ and $h(M)$ for $M \subseteq \real^d$ and $d \ge 3.$
\section{Proofs}
In the proofs we use the standard terminology from the theory of polyhedra (see \cite{ziegler95}). A \emph{polyhedron} is the (possibly empty) intersection of finitely many closed half-spaces. Bounded polyhedra are said to be \emph{polytopes}. If $P$ is a polyhedron, then faces of $P$ having dimension $\dim P -1$ are called \emph{facets}. By $\conv$ we denote the convex hull operation.
We first give the proof of Proposition~\ref{helly:over:certificates} since it is used as auxiliary statement in our main results.
\begin{proof}[Proof of Proposition~\ref{helly:over:certificates}]
The implication $(H) \Longrightarrow (A)$ is trivial. Now, assuming that $(A)$ is fulfilled we derive $(H)$. Consider a collection $C_1,\ldots C_m$ ($m \ge h$) of $M$-convex sets such that every sub-collection of $h$ elements has non-empty intersection. For every $I \subseteq \{1,\ldots,m\}$ with $\card{I} = h$ we choose $p_I \in \bigcap_{i \in I} C_i.$ The polytope $$P_i := \conv \setcond{p_I}{I \subseteq \{1,\ldots,m\}, \ \card{I}=h, \ i \in I}$$ is a subset of $\conv C_i.$ Let $f_1,\ldots,f_{s}$ ($s \in \natur$) be affine-linear functions on $\real^n$ such that every $P_i$ can be given by $P_i = \setcond{x \in \real^n}{f_j(x) \ge 0 \ \text{for} \ j \in J}$ for an appropriate $J \subseteq \{1,\ldots,s\}$ and every $f_j$ is non-negative on some $P_i$. By construction, $\setcond{x \in M}{f_j(x) \ge 0 \ \text{for} \ j \in J} \ne \emptyset$ for every $J \subseteq \{1,\ldots,s\}$ with $\card{J}=h.$ Hence, in view of $(A)$,
$$
\emptyset \ne \setcond{x \in M}{f_1(x) \ge 0,\ldots,f_s(x) \ge 0} \subseteq P_1 \cap \cdots \cap P_m \cap M \subseteq C_1 \cap \cdots \cap C_m.
$$
Next, assuming that $(A)$ is fulfilled we derive $(B)$ by an argument analogous to the one given in \cite[Corollary\,16.5a]{schrijver}. Consider affine-linear functions $b_1,\ldots,b_m$ ($m \ge h-1$) such that the supremum $\mu$ in $(B)$ is finite. For every $t \in \natur$ the system $b_1(x) \ge 0,\ldots, b_m(x) \ge 0, c(x) \ge \mu+1/t$ has no solution $x \in M,$ and hence its subsystem consisting of $h$ inequalities inequalities has no solution in $M.$ Each such subsystem contains the inequality $c(x) \ge \mu+1/t.$ Thus, it follows that there exist $i_1,\ldots, i_{h-1}$ such that the system $b_{i_1}(x) \ge 0,\ldots, b_{i_{h-1}}(x) \ge 0, c(x) \ge \mu+1/t$ has no solution $x \in M$ for infinitely many $t.$ The latter obviously implies $(B).$
In order to show $(B) \Longrightarrow (A)$ we assume that $(A)$ is not fulfilled and derive that $(B)$ is not fulfilled, as well. Let $a_1,\ldots,a_m$ $(m \ge h)$ be affine-linear functions such that $\setcond{x \in M}{a_1(x) \ge 0,\ldots, a_m(x) \ge 0} = \emptyset$ and for every $I \subseteq \{1,\ldots,m\}$ such that $\card{I}=h$ there exists a $p_I \in \setcond{x \in M}{a_i(x) \ge 0 \ \mbox{for} \ i \in I}.$ Without loss of generality we may assume that every subsystem of $a_1(x) \ge 0,\ldots, a_m(x) \ge 0$ consisting of $m-1$ inequalities is solvable over $M.$ In fact, otherwise we can redefine the system $a_1(x) \ge 0,\ldots,a_m(x) \ge 0$ by passing to a proper subsystem. Consider $c(x) := a_m(x)$ and $b_j(x) := a_j(x)$ for $j \in \{1,\ldots,m-1\}.$ Choose affine-linear functions $b_m(x),\ldots,b_{m+k}(x)$ such that $\setcond{x \in \real^k}{b_m(x) \ge 0,\ldots,b_{m+k}(x) \ge 0}$ is a simplex which contains all $p_I$'s introduced above. Then
\[-\infty<\sup \setcond{c(x)}{b_1(x) \ge 0,\ldots,b_{m+k}(x) \ge 0, \ x \in M} < 0.\]
Furthermore, for all $1 \le i_1,\ldots,i_{h-1} \le m+k$ one has
\[\sup \setcond{c(x) }{b_{i_1}(x) \ge 0,\ldots, b_{i_{h-1}}(x) \ge 0, x \in M} \ge c(p_I) \ge 0\]
for every $I \subseteq \{1,\ldots,m\}$ satisfying $\card{I} = h$ and $\left(\{i_1,\ldots,i_{h-1}\} \cap \{1,\ldots,m-1\} \right) \cup \{ m \} \subseteq I.$ Hence $(B)$ is not fulfilled (for $m+k$ in place of $m$), and we are done.
\end{proof}
\begin{lemma} \label{helly:over:cert:lem}
Let $M \subseteq \real^k$ be non-empty and closed and let $h \in \natur,$ $h \ge k+1.$ Then $(H)$ is equivalent to the following condition.
\begin{itemize}
\item[$(A')$] For every choice of affine-linear functions $a_1,\ldots,a_m$ ($m \ge h$) on $\real^k$ such that the polyhedron
\begin{equation} \label{P:over:a}
P:= \setcond{x \in \real^k}{a_1(x) \ge 0,\ldots,a_m(x) \ge 0}
\end{equation}
satisfies the conditions:
\begin{enumerate}
\item \label{P:bounded} $P$ is bounded,
\item \label{P:full:dim} $P$ is $k$-dimensional,
\item \label{PdisjM} $P \cap M = \emptyset$,
\end{enumerate}
one necessarily has
\begin{equation} \label{certificate}
\setcond{x \in M}{a_{i_1}(x) \ge 0,\ldots,a_{i_h}(x) \ge 0} = \emptyset
\end{equation}
for some $1 \le i_1,\ldots,i_h \le m.$
\end{itemize}
\end{lemma}
\begin{proof}
The implication $(H) \Longrightarrow (A')$ is trivial. Now, assume that $(A')$ is fulfilled. We will show that $(A')$ also holds when we drop out Conditions~\ref{P:bounded} and \ref{P:full:dim} (which, in view of Proposition~\ref{helly:over:certificates}, yields the sufficiency). Assume that $P$ given by \eqref{P:over:a} satisfies Conditions~\ref{P:bounded} and \ref{PdisjM} but does not satisfy Condition~\ref{P:full:dim}. If $P= \emptyset,$ the existence of $i_1,\ldots,i_h$ satisfying \eqref{certificate} follows from Helly's theorem for $\real^k.$ Assume that $P \ne \emptyset.$ Then, employing the closedness of $M$ and compactness of $P$, we see that there exists an $\eps>0$ such that $P_\eps := \setcond{x \in \real^k}{a_1(x) + \eps \ge 0,\ldots, a_m(x) + \eps \ge 0}$ is a $k$-dimensional polytope with $P_\eps \cap M = \emptyset.$ Consequently, applying $(A')$ for the affine-linear functions $a_1(x) +\eps, \ldots, a_m(x)+ \eps,$ we obtain $$\setcond{x \in \real^k}{a_{i_1}(x) + \eps \ge 0,\ldots,a_{i_h}(x) + \eps \ge 0} \cap M = \emptyset$$ for some $1 \le i_1,\ldots,i_h \le m.$ The latter implies \eqref{certificate}. Thus, $(A')$ still holds when we drop out Condition~\ref{P:full:dim}. Take $P$ given by \eqref{P:over:a} which satisfies Condition~\ref{PdisjM} but does not satisfy Condition~\ref{P:bounded}. For $t \in \natur$ we introduce the polytope
$$Q_t := \setcond{x \in \real^k}{x \in P, \ \pm x_1+t \ge 0,\ldots, \pm x_k + t \ge 0}$$
where $x_1,\ldots,x_k$ are coordinates of $x$. Applying $(A')$ (with dropped out Condition~\ref{P:full:dim}) for the affine-linear functions $\pm x_i + t$ ($i \in \{1,\ldots,k\}$), $a_j(x)$ ($j \in \{1,\ldots,m\}$) that define $Q_t$ we find sets $I \subseteq \{1,\ldots,m\}$, $J^+, J^- \subseteq \{1,\ldots,k\}$ (a priori depending on $t$) such that $\card{I}+\card{J^+} + \card{J^-} \le h$ and
\begin{align}
\setcondbegin{x \in M}{} & \, a_i(x) \ge 0 \ \mbox{for} \ i \in I, \nonumber \\
& \setcondend{x_j+t \ge 0 \ \mbox{for} \ j \in J^+, \ -x_j+t \ge 0 \ \mbox{for} \ j \in J^-} = \emptyset. \label{IJ:cond}
\end{align}
Since the index sets $\{1,\ldots,m\}, \{1,\ldots,k\}$ are finite we can fix $I, J^-, J^+$ independent of $t$ and such that \eqref{IJ:cond} holds for infinitely many $t$'s. Then $\setcond{x \in M}{a_i(x) \ge 0 \ \mbox{for} \ i \in I} = \emptyset$, $\card{I} \le h,$ and the assertion follows.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{mixed:helly:bounds}] Inequalities \eqref{mixed:with:reals} and \eqref{mixed:with:integers} are trivial if $h(M) = \infty.$ Thus, we assume $h(M)< \infty.$ Furthermore, without loss of generality we assume that $M$ affinely spans $\real^k$ so that $h(M) \ge k+1.$ We derive \eqref{mixed:with:reals} with the help of Lemma~\ref{helly:over:cert:lem}. Consider arbitrary affine-linear functions $b_1,\ldots,b_s$ ($s \in \natur$) on $\real^n \times \real^k$ such that $$P:= \setcond{(x,y) \in \real^n \times \real^k}{b_1(x,y) \ge 0,\ldots, b_s(x,y) \ge 0}$$ is $(n+k)$-dimensional, bounded and $P \cap (\real^k \times M) = \emptyset$. Let $T$ be the canonical projection from $\real^n \times \real^k$ onto $\real^k.$ The $k$-dimensional polytope $T(P)$ can be represented by $$T(P) = \setcond{y \in \real^k}{a_1(y),\ldots, a_{m}(y) \ge 0},$$ where $a_1,\ldots,a_{m}$ are affine-linear functions on $\real^k$ such that for each $j \in \{1,\ldots,m\},$ the set $F_j:= \setcond{y \in T(P)}{a_j(y) = 0}$ is a facet of $T(P).$ Hence $G_j:=T^{-1}(F_j) \cap P$ is a face of $P$ of dimension at least $k -1$. Consequently, the cone $N_j$ of affine-linear functions $f(x,y)$ vanishing on $G_j$ and non-negative on $P$ has dimension at most $(n+k)-(k-1) = n+1.$ The cone $N_j$ is generated by those $b_i$, $i \in \{1,\ldots,s\}$, which vanish on some facet of $P$ that contains $G_j.$ The function $a_j(y)$, $y \in \real^k$, can also be viewed as an affine-linear function on $\real^n \times \real^k.$ Thus, by Carath\'edory's Theorem for convex cones (cf. \cite[\S\,7.7]{schrijver}) applied to the function $a_j(y)$ in the cone $N_j$, there exists $I_j \subseteq \{1,\ldots,s\}$ such that $\card{I_j} \le n+1$ and
\begin{equation} \label{a_j:over:b_i}
a_j(y) = \sum_{i \in I_j} \lambda_{i,j} b_i(x,y)
\end{equation}
for appropriate $\lambda_{i,j} \ge 0$ ($i \in I_j$). By the definition of $h(M),$ there exists $J \subseteq \{1,\ldots,m\}$ with $\card{J} \le h(M)$ such that $\setcond{y \in M}{a_j(y) \ge 0 \ \mbox{for} \ j \in J} = \emptyset.$ It follows that the system $b_i(x,y) \ge 0$ with $i \in \bigcup_{j \in J} I_j$ has no solution $(x,y) \in \real^n \times M$. This system consists of at most $(n+1) h(M)$ inequalities. Hence, in view of Lemma~\ref{helly:over:cert:lem}, we arrive at \eqref{mixed:with:reals}
Let us show \eqref{mixed:with:integers}. Let $h:=h(M)$ and $C_1,\ldots,C_h$ be $M$-convex sets such that $C_1 \cap \ldots \cap C_h = \emptyset$ but every sub-collection of $C_1,\ldots,C_h$ consisting of $h-1$ elements has non-empty intersection. Then the collection $(C_i \times \{j\}) \cup (C_i \times \{1-j\})$, where $i \in \{1,\ldots,h \}$, $j \in \{ 0,1\}$ consists of $2 h$ $(M \times \integer)$-convex sets, has empty intersection and the intersection over every of its proper sub-collections is non-empty. This shows the bound $h(M \times \integer) \ge 2 h(M).$ The general bound $h(M \times \integer^d) \ge 2^d h(M)$ is obtained by induction on $d.$
Equality \eqref{mixed:helly:eq} is a consequence of \eqref{mixed:with:reals}, \eqref{mixed:with:integers} and the theorems of Helly and Doignon.
\end{proof}
We remark that representation \eqref{a_j:over:b_i} can be viewed as Farkas type certificate of insolvability of a system of linear inequalities, see also \cite[\S\,13.1]{BertWeiBook} and \cite{CertificatesMixed2008} for related results.
Next we work towards the proof of Theorem~\ref{fh thm}. We show that the main tools in the proof of the fractional Helly theorem for $\integer^d$ given in \cite{bar-mat-03} can also be applied for sets $M \subseteq \real^d$ with a finite Helly number. First we obtain a weak form of the fractional Helly theorem.
\begin{theorem} \label{weak fh thm}
Let $M \subseteq \real^d$ be non-empty and closed and let $h(M)<\infty.$ Then $\fh(M) \le h(M).$
\end{theorem}
\begin{proof}
\newcommand{\II}{\mathcal{I}} We slightly adjust the proof from \cite[Proof of Theorem~2.5]{bar-mat-03}.
Let $h:=h(M).$
Consider $M$-convex sets $C_1,\ldots,C_n$ ($n \in \natur$). Let $\II$ be the collection of those $h$-element subsets $I$ of $\{1,\ldots,n\}$ for which $\bigcap_{i \in I} C_i \ne \emptyset.$ Assume that $\card{\II} \ge \alpha \binom{n}{h}$ for some $0 \le \alpha < 1.$ For every $I \in \II$ we choose $p_I \in \bigcap_{i \in I} C_i$ and introduce the polytopes $P_i := \conv \setcond{ p_I}{I \in \II, \ i \in I},$ $i=1,\ldots,n.$ By construction, $P_i \subseteq C_i$ for every $i,$ and $p_I \in \bigcap_{i \in I} P_i.$ If $S \subseteq \{1,\ldots,n\}$ we shall write $P_S:= \bigcap_{i \in S} P_i.$ It is known that for a given non-empty, compact convex set $K$ and almost all directions $u$, the direction $u$ is an outward normal of precisely one boundary point of $K$; for a precise formulation see \cite[Theorem\,2.2.9]{schneider}. Applying this result to the sets $\conv (P_S \cap M)$ with $S \subseteq \{1,\ldots,n\}$, we see that there exists an affine function $a$ such that $a$ is maximized in exactly one point $x_S$ on $P_S \cap M$ (as long as $P_S \cap M$ is non-empty).
\newcommand{\QQ}{\mathcal{Q}}
For every $I \in \II$ there exists an $(h-1)$-element subset $J=J(I)$ of $I$ such that $x_{J}=x_{I}$ is the unique point maximizing $a(x)$ for $x \in M \cap P_J.$ In fact, for $H_{I} := \setcond{x \in M}{a(x) > a(x_I)}$ the family $\setcond{P_i \cap M}{i \in I} \cup \{H_{I}\}$ has empty intersection. Therefore, by the definition of $h(M),$ some $h$-element subfamily of this family has empty intersection. The elements of this subfamily which do not coincide with $H_{I}$ determine $J.$
There are at most $\binom{n}{h-1}$ possible sets $J$ and at least $\alpha \binom{n}{h}$ different sets $I$. Thus, for a suitable $\beta=\beta(\alpha,h)>0$, some $J=:J^\ast$ is assigned to at least $\beta n$ different sets $I$. Each such $I$ has exactly one $i \not\in J^\ast$, and $x_{J^\ast}$ is a common point of these (at least $\beta n$ many) sets $P_i$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{fh thm}]
The proof is a consequence of Theorems~\ref{col h thm}, \ref{weak fh thm} and the Erd\H{o}s-Simonovits theorem on super-saturated hypergraphs; for details see \cite[p.\,232]{bar-mat-03}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{radon:bounds}]
We exclude the trivial case $r(M) = \infty.$ The inequality $r(M \times \integer) \ge 2 r(M) -1$ can be shown following the idea from \cite[proof of Proposition~2.1]{onn91} (see also \cite[pp.\,176-177]{MR1234493}). Consider a set $A \subseteq M$ with $\card{A} = r(M)-1$ such that $M \cap \conv B \cap \conv C = \emptyset$ for all disjoint $B, C \subseteq A.$ Then $(M \times \integer) \cap \conv B \cap \conv C = \emptyset$ for all disjoint $B, C \subseteq A \times \{0,1\}$. Thus $r(M \times \integer) \ge \card{A \times \{0,1\}} +1 = 2 r(M) -1.$ The general bound \eqref{radon:with:integers} is obtained by induction on $d.$
The lower bound in \eqref{radon:mixed:integers} is a direct consequence of \eqref{radon:with:integers} and Radon's Theorem. The upper bound in \eqref{radon:mixed:integers} follows from the known inequality $r(M) \le k (h(M)-1) + 2$ (see \cite{MR0514026}, \cite[p.\,169]{MR1234493}) and \eqref{mixed:helly:eq}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{hel rad 2dim}]
The case $h(M) = \infty$ is trivial. It is known that $r(M) \ge h(M)+1.$ For $h(M) \le 3$ it is easy to establish the inequality $r(M) \le h(M)+1.$ For the case $h(M)=4$ we show $r(M) \le 6$ following the idea from \cite[p.\,182]{MR2007959}. Assume the contrary, there exists a six-point set $A \subseteq M$ which does not possess a Radon partition in $M.$ Then $\conv A$ is a hexagon. We notice that any four of the six sets $A \setminus \{a\}, \ a \in A$ have non-empty intersection. Thus, by the definition of $h(M),$ $\emptyset \ne \bigcap_{a \in A} \conv (A \setminus \{a\}) \cap M = \conv \{a_1,a_3,a_5\} \cap \conv \{a_2, a_4, a_6\} \cap M,$ where $a_1,\ldots,a_6$ are consecutive vertices of $\conv A.$ Hence $\left\{ \{a_1,a_3,a_5\},\{a_2,a_4,a_6\} \right\}$ is a Radon partition of $A$ in the space $M,$ a contradiction. Now we consider the case that $h:=h(M) \ge 5$ and show that $r(M) \le h(M)+1.$ Let $A \subseteq M$ be a set of cardinality $h+1.$ If $A$ is not a vertex set of a convex polygon, it possesses a Radon partition. Assume that $A$ is a vertex set of a convex polygon. Then, by the definition of $h(M),$ we can choose
\begin{equation} \label{p in core}
p \in \bigcap_{a \in A} \conv (A \setminus \{a\}) \cap M.
\end{equation}
By Carath\'eodory's theorem (for $\real^2$) $p$ is in $\conv T$, for some three-element subset $T$ of $A.$ If $\conv T$ and $\conv A$ do not share edges, then taking into account \eqref{p in core} and the fact that $A$ is a vertex set of $\conv A$ we get $p \in \conv (A \setminus T) \cap \conv T \cap M.$ Consider the case that $\conv T$ and $\conv A$ share an edge. First notice that $\conv A$ and $\conv T$ cannot share two edges, since otherwise we would get a contradiction to \eqref{p in core}. We define $q, q_1, q_2, q_3, q_4$ such that $T = \{q, q_2, q_3 \},$ and $q_1,\ldots,q_4$ are consecutive vertices of $\conv A.$ Since $\card{A} \ge 6,$ for some $i=\{2,3\}$ the triangle $\conv \{q, q_{i+1},q_{i-1} \}$ does not share edges with $\conv A.$ Then $p \in \conv \{q,q_{i+1},q_{i-1} \}$, since otherwise one would get a contradiction to \eqref{p in core}. Hence for $T' := \{q, q_{i+1},q_{i-1} \}$, one has $p \in \conv (A \setminus T') \cap \conv T' \cap M,$ which shows that $r(M) \le h(M)+1.$
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
We resolve two open problems from \cite{Grunwald07} concerning
universal codes of the predictive plug-in type, also known as
``prequential'' codes. These codes were introduced independently by
Rissanen \cite{Rissanen84} in the context of MDL learning and by Dawid
\cite{Dawid84}, who proposed them as probability forecasting
strategies rather than directly as codes. Roughly, the plug-in codes
relative to parametric model ${\cal M} = \{ M_{\theta} \mid \theta \in
\Theta \}$ work by sequentially coding each outcome $x_i$ based on an
an estimator $\bar{\theta}_{i-1} = \bar{\theta}(x^{i-1})$ for all
previous outcomes $x^{i-1} = x_1, \ldots, x_{i-1}$, leading to
codelength (log loss) $- \lg M_{\bar{\theta}_{i-1}}(x_i)$, where $M_{\theta}$
denotes the probability density or mass function indexed by $\theta$. If we take
$\bar{\theta}_i = \hat{\theta}_i$ equal to the ML (maximum likelihood)
estimator, we call the resulting code the ``ML plug-in code''.
There are many papers about the redundancy and/or expected regret for
the ML plug-in codes, for a large variety of models including
multivariate exponential families, ARMA processes, regression models
and so on. Examples are \cite{Rissanen1986c,Gerencser1987, LiY00}. In all these papers the ML plug-in
code is shown to achieve an asymptotic expected regret or redundancy
of ${k\over2}\lg n + O(1)$, where $k$ is the number of parameters of the
model and $n$ is the sample size. This matches the behaviour of the
Shtarkov, Bayesian and two-part universal codes
and is optimal in several ways, see \cite{BarronRY98}; since the ML
plug-in codes are often easier to calculate than any of these other
three codes, this appears to be a strong argument for using them in
practical data compression and MDL-style model selection. Yet,
more recently \cite{DeRooijG05a,GrunwaldD05,DeRooijG06}, it was shown
that, at least for single-parameter exponential family models, when the data are generated i.i.d. $\sim P$, the
redundancy in fact grows as ${1\over2}\lg n\cdot{{\textnormal{var}}_P X\over{\textnormal{var}}_M
X}$, where $M$ is the distribution in ${\cal M}$ that is closest to
$P$ in Kullback-Leibler divergence, i.e. it minimizes $D(P\|M)$; a
related result for linear regression is in \cite{Wei1990}. In contrast
to the other cited works, \cite{DeRooijG05a,GrunwaldD05,DeRooijG06,Wei1990} do
not assume that $P \in {\cal M}$: the model may be {\em misspecified}.
Yet {\em if\/} $P\in{\cal M}$, then we have $M=P$ so that the redundancy
grows like it does in the other universal models. But when $M \neq P$,
the Shtarkov, Bayes and universal codes typically still achieve
asymptotic expected regret ${ 1 \over 2 } \lg n$, whereas the plug-in
codes behave differently. \cite{DeRooijG05a,DeRooijG06} show that this
leads to substantially inferior performance of the plug-in codes in
practical MDL model selection.
\subsection{The Two Open Problems/Conjectures}
In general, the estimator for ${\cal M}$ based on $x^{i-1}$ need not be an
element of the parametric model ${\cal M}$; for example, we may think
of the Bayesian predictive distribution as an estimator relative to
${\cal M}$, even though it is ``out-model'': rather than a single element of ${\cal
M}$, it is a mixture of
distributions in ${\cal M}$, each weighted by their posterior density
(see Section~\ref{sec:squashed} for an example). We may thus re-interpret Bayesian universal codes as
prequential codes based on ``out-model'' estimators. From now on, we
reserve the term ``prequential plug-{\em in\/} code'', abbreviated to
just ``plug-in code'', for codes based on ``{\em in}-model''
estimators, i.e. estimators required to lie within ${\cal
M}$. When we call a code just ``prequential'', it may be
sequentially constructed from either in-model or out-model
estimators.
\cite{GrunwaldD05} established a nonstandard redundancy, different
from $(k/2) \lg n$, only for ML and closely related plug-in
codes. \cite[Open Problem Nr. 2]{Grunwald07} conjectured that a
similar result should hold for {\em all\/} plug-in codes, even if they
are based on in-model estimators very different from the ML estimator:
the conjecture was that {\em no\/} plug-in code can achieve
guaranteed redundancy of $(k/2) \lg n$ if data are i.i.d. $\sim P$
and $P \neq M$. Our first main result, Theorem~\ref{thm:no_plugin}
below, shows that, essentially, this conjecture is true for general
one-parameter exponential families $(k=1)$. Specifically, the
redundancy can become much larger than $(1/2) \lg n$ if $P \not \in
{\cal M}$.
The second related conjecture \cite[Open Problem Nr. 3]{Grunwald07}
concerned the fact that for the normal location family with constant
variance $\sigma^2$, the Bayesian predictive distribution based on
data $x^{i-1}$ and a normal prior looks ``almost'' like an in-model
estimator for $x^{i-1}$, and hence the resulting code looks ``almost''
like a plug-in code: the Bayes predictive distribution is equal to the
normal distribution for $X_i$ with mean equal to the ML estimator
$\hat{\mu}(x^{i-1})$ but with a variance of order $\sigma^2 + O(1/n)$,
i.e. slightly larger than the variance $\sigma^2$ of
$P_{\hat{\mu}(x^{i-1})}$ (see Section~\ref{sec:squashed} for details).
Since the Bayesian predictive distribution does achieve the redundancy
$(1/2) \lg n$ even if $P \not \in {\cal M}$, this means that if ${\cal
M}$ is the normal location family, then there does exist an
``almost'' in-model estimator (i.e. a slight modification of the ML
estimator) that does achieve $(1/2) \lg n$ even if $P \not \in {\cal
M}$. Although this example does not extend straightforwardly to
other exponential families, \cite{Grunwald07} conjectured that there
should nevertheless be some general definition for ``almost'' in-model
estimators that achieve $(k/2) \lg n$ redundancy even if $P \not \in
{\cal M}$. Here we show that this conjecture is true, at least if
$k=1$: we propose the {\em slightly squashed\/} ML estimator, a modification of
the ML estimator that puts it slightly outside model ${\cal M}$, and
in Theorem~\ref{thm:robustML} we show that this estimator achieves
$(1/2) \lg n$ redundancy even if $P \not \in {\cal M}$. This result is
important in practice since, in contrast to the
Bayesian predictive distribution, the slightly squashed ML estimator is in
general just as easy to compute as the ML estimator itself.
\section{Notation and Definitions}
Throughout this text we use nats rather than bits as units of
information.
A sequence of outcomes $z_1,\ldots,
z_n$ is abbreviated to $z^n$. We write $E_{P}$ as a shorthand for $E_{Z \sim P}$,
the expectation of $Z$ under distribution $P$. When we consider a sequence of $n$
outcomes independently distributed $\sim P$, we use $E_{P}$ even as a shorthand
for the expectation of $(Z_1, \ldots, Z_n)$ under the $n$-fold product
distribution of $P$. Finally, $P(Z)$ denotes the probability mass function of
$P$ in case $Z$ is discrete-valued, and it denotes the density of $P$, in case
$Z$ takes its value in a continuum. When we write `density function of $Z$',
then, if $Z$ is discrete-valued, this should be read as `probability mass
function of $Z$'. Note however that in our second main result, Theorem~\ref{thm:robustML} we do not
assume that the data-generating distribution $P$ admits a density.
Let $\ensuremath{\mathcal Z}$ be a set of outcomes, taking values either in a finite or countable
set, or in a subset of $k$-dimensional Euclidean space for some $k \geq 1$. Let
$X: \ensuremath{\mathcal Z} \rightarrow {\mathbb{R}}$ be a random variable on $\ensuremath{\mathcal Z}$, and let
$\ensuremath{\mathcal X} = \{ x \in {\mathbb{R}} \; : \; \exists z \in \ensuremath{\mathcal Z}: X(z) = x\}$ be the
range of $X$. Exponential family models are families of distributions on
$\ensuremath{\mathcal Z}$ defined relative to a random variable $X$ (called `sufficient
statistic') as defined above, and a function $h: \ensuremath{\mathcal Z} \rightarrow [0,\infty)$.
Let $\textnormal{Z}(\eta) :=\int_{z\in\ensuremath{\mathcal Z}}e^{-\eta X(z)}h(z)dz$ (the integral to be
replaced by a sum for countable $\ensuremath{\mathcal Z}$), and $\ensuremath{\Theta_{\text{nat}}}
:=\{\eta\in{\mathbb{R}}:\textnormal{Z}(\eta)<\infty\}$.
\begin{definition}[Exponential family]\label{def:expfam}
The \emph{single parameter exponential family} {\rm \cite{BarndorffNielsen78}}
with {\em sufficient statistic\/} $X$ and {\em carrier\/} $h$ is the
family of distributions with densities $M_\eta(z) :={1\over
\textnormal{Z}(\eta)}e^{-\eta X(z)}h(z)$, where $\eta\in \ensuremath{\Theta_{\text{nat}}}$.
$\ensuremath{\Theta_{\text{nat}}}$ is called the \emph{natural parameter space}. The
family is called \emph{regular} if $\ensuremath{\Theta_{\text{nat}}}$ is an open interval
of ${\mathbb{R}}$.
\end{definition}
In the remainder of this text we only consider single parameter, regular
exponential families, but this qualification will
henceforth be omitted. Examples include the Poisson, geometric and multinomial
families, and the model of all Gaussian distributions with a fixed variance or
mean.
The statistic $X(z)$ is sufficient for $\eta$ \cite{BarndorffNielsen78}. This suggests
reparameterizing the distribution by the expected value of $X$, which is called
the \emph{mean value parameterization}. The function $\mu(\eta)=E_{M_\eta}[X]$
maps parameters in the natural parameterization to the mean value
parameterization. It is a diffeomorphism (it is one-to-one, onto, infinitely
often differentiable and has an infinitely often differentiable inverse)
\cite{BarndorffNielsen78}. Therefore the mean value parameter space $\ensuremath{\Theta_{\text{mean}}}$ is also an
open interval of ${\mathbb{R}}$. We write ${\cal M}=\{{M_\mu}\mid\mu\in\ensuremath{\Theta_{\text{mean}}}\}$ where
${M_\mu}$ is the distribution with mean value parameter $\mu$.
\commentout{We note that for some
models (such as Bernoulli and Poisson), the parameter space is usually given in
terms of the a non-open set of mean-values (e.g., $[0,1]$ in the Bernoulli case).
To make the model a regular exponential family, we have to restrict the set of
parameters to its own interior. Henceforth, whenever we refer to a standard
statistical model such as Bernoulli or Poisson, we assume that the parameter set
has been restricted in this sense.
}
We are now ready to define the plug-in universal model. This is a
distribution on infinite sequences $z_1, z_2, \ldots \in
\ensuremath{\mathcal Z}^{\infty}$, recursively defined in terms of the distributions
of $Z_{n+1}$ conditioned on $Z^n = z^n$, for all $n = 1,2 , \ldots$.
In the definition, we use the notation $x_i := X(z_i)$. Note that we
use the term ``model'' both for a single distribution (``plug-in
universal model'', a common phrase in information theory) and for a
family of distributions (``statistical model'', a common phrase in
statistics).
\begin{definition}[Plug-in universal model]
\label{def:preq}
Let ${\cal M}=\{{M_\mu}\mid\mu\in \ensuremath{\Theta_{\text{mean}}} \}$ be an exponential family with mean
value parameter domain $\ensuremath{\Theta_{\text{mean}}}$. Given ${\cal M}$, constant ${\bar\mu}_0 \in
\ensuremath{\Theta_{\text{mean}}}$ and a sequence of functions ${\bar\mu}(z^1),{\bar\mu}(z^2),\ldots$, such that
${\bar\mu}(z^n) =: {{\bar\mu}_n} \in \ensuremath{\Theta_{\text{mean}}}$, we define the
\emph{plug-in universal model} (or \emph{plug-in model} for short) $U$ by setting, for all $n$, all $z^{n+1}
\in \ensuremath{\mathcal Z}^{n+1}$: $$U(z_{n+1} \mid z^n) = M_{{{\bar\mu}_n}}(z_{n+1}),$$ where
$U(z_{n+1} \mid z^n)$ is the density/mass function of $z_{n+1}$ conditional on
$Z^n = z^n$. \end{definition}\smallskip
We usually refer to plug-in universal model in terms of the
codelength function of the corresponding plug-in universal code:
\begin{equation}
\label{eq:codelength}
L_U(z^n) = \sum_{i=0}^{n-1} L_U(z_{i+1} \mid z_i) =
\sum_{i=0}^{n-1} - \ln M_{{{\bar\mu}_i}}(z_{i+1}).
\end{equation}
The most important plug-in model is the ML (\emph{maximum
likelihood}) plug-in model, defined as follows:
\begin{definition}[ML plug-in model]
\label{def:preqb} Given ${\cal M}$ and
constants $x_0\in\ensuremath{\Theta_{\text{mean}}}$ and $n_0>0$, we define the
\emph{ML plug-in model} $\hat{U}$ by setting, for all $n$, all $z^{n+1}
\in \ensuremath{\mathcal Z}^{n+1}$:
$$\hat{U}(z_{n+1} \mid z^n) = M_{\hat{\mu}(z^n)}(z_{n+1}),$$
where
\begin{equation}
\label{eq:ML_estimator}
\hat{\mu}(z^n) = {{\hat\mu}_n} :=\frac{x_0 \cdot n_0+\sum_{i=1}^n
x_i}{n+n_0}.
\end{equation}
\end{definition}\smallskip
To understand this definition, note that for exponential families, for any
sequence of data, the ordinary maximum likelihood parameter is given by the
average $n^{-1} \sum x_i$ of the observed values of $X$ \cite{BarndorffNielsen78}. Here we
define our plug-in model in terms of a slightly modified maximum likelihood
estimator that introduces a `fake initial outcome' $x_0$ with multiplicity $n_0$
in order to avoid infinite code lengths for the first few outcomes (a well-known
problem sometimes called the ``inherent singularity'' of predictive coding
\cite{BarronRY98,Grunwald07}) and to ensure that the plug-in ML code
of the first outcome is well-defined. In practice we can take $n_0 = 1$ but our
result holds for any $n_0 > 0$.
\begin{definition}[Relative redundancy]
\label{def:red}
Following \cite{TakeuchiB98b,DeRooijG05a}, we define {\em relative redundancy\/} with respect to $P$ of a code $U$ that is
universal on a model ${\cal M}$, as:
\begin{equation}\label{eq:red}
{\mathcal{R}}_{U}(n) := E_P[L_U(Z^n)] - \inf_{\mu \in \ensuremath{\Theta_{\text{mean}}}}E_P [-
\ln {M_\mu}(Z^n)],
\end{equation}
where $L_U$ is the length function of $U$.
\end{definition}
We use the term \emph{relative redundancy} rather than just \emph{redundancy} to
emphasize that it measures redundancy relative to the element of the model that
minimizes the codelength rather than to $P$, which is not necessarily an element
of the model. From now on, we only consider $P$ under which the data
are i.i.d. Under this condition, let ${M_\mstar}$ be the element of
${\cal M}$ that minimizes KL divergence to $P$: $${\mu^*} := \arg \min_{{\mu} \in
\ensuremath{\Theta_{\text{mean}}}} D(P \| M_{\mu}) = \arg \min_{{\mu} \in \ensuremath{\Theta_{\text{mean}}}} E_{P} [ - \ln
{M_\mu}(Z)],$$
where the equality follows from the definition of KL divergence. If
${M_\mstar}$ exists, it is unique, and if $E_P[X] \in \ensuremath{\Theta_{\text{mean}}}$, then
$\mu^* = E_P[X]$ \cite[Ch. 17]{Grunwald07}, and the relative redundancy satisfies
\begin{equation}
\label{eq:redb}
{\mathcal{R}}_{U}(n) = E_{P}[L_U(Z^n)] -
E_{P}[- \ln {M_\mstar}(Z^n)].
\end{equation}
\section{First Result: Redundancy of Plug-In Codes}\label{sec:mainresult}
The three major types of universal codes, Bayes, NML and 2-part, achieve relative
redundancies that are (in an appropriate sense) close to optimal. Specifically,
under the conditions on ${\cal M}$ described above, and if data are
i.i.d. $\sim P$, then, under some mild conditions
on $P$, these universal codes satisfy:
\begin{equation}
\label{eq:bic}
{\mathcal{R}}_U(n) = \frac{1}{2} \ln n + O(1),
\end{equation}
(where the $O(1)$ may depend on $\mu$ and the universal code
used), whenever $P \in {\cal M}$ or $P \not \in {\cal M}$.
(\ref{eq:bic}) is the famous `$k$ over $2$ log $n$ formula' ($k=1$ in our case),
refinements of which lie at the basis of practical
approximations to MDL learning \cite{Grunwald07}.
While it is known that for $P \in {\cal M}$, the fourth major type of universal
code, the ML plug-in code, satisfies (\ref{eq:bic}) as well, it was
shown by
\cite{DeRooijG05a,GrunwaldD05} that when $P$ is not in the model, the ML
plug-in code
may behave suboptimally. Specifically, its relative redundancy satisfies:
\begin{equation}
\label{eq:redundancyML}
{\mathcal{R}}_{\hat{U}}(n) = \frac{1}{2} \frac{{\textnormal{var}}_P X}{{\textnormal{var}}_{M_{\mu^*}} X} \ln n +
O(1),
\end{equation}
and can be significantly larger than (\ref{eq:bic}), when the variance of $P$ is large.
In this paper, we show that not only the ML plug-in code, but
\emph{every} plug-in code may behave suboptimally, when $P \notin
{\cal M}$. In other words, modifying the ML estimator ${{\hat\mu}_n}$ or
introducing any other sequence of estimators ${{\bar\mu}_n}$, and
constructing the plug-in code based on that sequence will not help to
satisfy (\ref{eq:bic}). Thus the optimal redundancy can only be
achieved by codes outside ${\cal M}$, unless ${\cal M}$ is the
Bernoulli family (since we assume the data are i.i.d., in the
Bernoulli case we must have that $P \in {\cal M}$; but the Bernoulli
case is the {\em only\/} case in which we must have $P \in {\cal M}$).
Our main result, Theorem \ref{thm:no_plugin}, concerns the case in
which $P$ is itself a member of some exponential family ${\cal P}$, but
${\cal P}$ is in general different than ${\cal M}$. Then, the suboptimal
behavior of plug-in codes follows immediately as Corollary
\ref{coll:no_plugin}, stated further below.
\begin{theorem}
\label{thm:no_plugin}
Let ${\cal M} = \{{M_\mu}\mid\mu\in\ensuremath{\Theta_{\text{mean}}}\}$ and ${\cal P} = \{P_\mu \mid \mu \in \ensuremath{\Theta_{\text{mean}}}\}$ be single parameter exponential families with the same sufficient
statistic $X$ and mean-value parameter space $\ensuremath{\Theta_{\text{mean}}}$. Let $U$ denote
any plug-in model with respect to ${\cal M}$ based on the sequence
of estimators ${\bar\mu}_0,{\bar\mu}_1,{\bar\mu}_2,\ldots$.
Then, for Lebesgue almost all
${\mu^*} \in \ensuremath{\Theta_{\text{mean}}}$ (i.e. all apart from a Lebesgue measure zero set), for $X, X_1, X_2, \ldots$
i.i.d. $\sim
P_{\mu^*} \in {\cal P}$:
$$
\liminf_{n \to \infty} \frac{\mathcal{R}_U(n)}{\frac{1}{2} \ln n} \geq \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}} X}.
$$
\end{theorem}
\begin{proof}{\em (rough sketch; a detailed proof is in the Appendix)}
The proof is based on a theorem stated by Rissanen \cite{Rissanen86a} (see also
\cite{Grunwald07}, Theorem 14.2), a special case of which says the following.
Let $\Theta_0 \subset \ensuremath{\Theta_{\text{mean}}}$ be a closed, non-degenerate interval, ${\cal P}$ be defined as above, $P_{\mu}^{(n)}$ be a joint
distribution of $n$ outcomes generated i.i.d. from $P_\mu$, $Q$ be an arbitrary
probabilistic source, i.e. a distribution on infinite sequences $z_1, z_2, \ldots
\in \ensuremath{\mathcal Z}^{\infty}$, and let $Q^{(n)}$ be its restriction to the first $n$
outcomes. Define:
$g_n({\mu^*}) = \frac{D(P^{(n)}_{\mu^*} \|
Q^{(n)})}{\frac{1}{2} \ln n}.$
Then for Lebesgue almost all ${\mu^*} \in \Theta_0$, $\liminf_{n \to \infty}
g_n({\mu^*}) \geq 1$.
We apply Rissanen's theorem by constructing a source $Q$, specifying the conditional probabilities
$Q(z_{n+1}|z^{n}) := P_{{\bar\mu}_n}$, for every $n \geq 1$. We now have:
\begin{eqnarray}
\label{eq:KL_deriv}
D(P_{\mu^*}^{(n)} \| Q^{(n)})\!\!\!&=&\!\!\! \sum_{i=0}^{n-1} E_{P_{\mu^*}}
\left[ \ln P_{{\mu^*}}(Z_{i+1}) - \ln Q(Z_{i+1}|Z^{i}) \right] \nonumber \\
\!\!\!&=&\!\!\! \sum_{i=1}^{n-1} E_{P_{\mu^*}} \left[ D(P_{{\mu^*}} \|
P_{{{\bar\mu}_i}}) \right].
\end{eqnarray}
To see how (\ref{eq:KL_deriv}) is related to our case, let us first rewrite the redundancy in a more convenient form:
\begin{equation}
\label{eq:redundancy_convienient}
{\mathcal{R}}_U(n)=\sum_{i=0}^{n-1}\,E_{P_{\mu^*}}\left[D({M_\mstar}\parallel
M_{{\bar\mu}_i})\right].
\end{equation}
The derivation of (\ref{eq:redundancy_convienient}) make use of a standard result in the theory of exponential families and can be found e.g. in \cite{Grunwald07}.
Comparing (\ref{eq:KL_deriv}) and (\ref{eq:redundancy_convienient}),
we see that although in both expressions, the expectation is taken
with respect to $P_{\mu^*}$, (\ref{eq:KL_deriv}) is a statement about
KL divergence between the members of ${\cal P}$, while
(\ref{eq:redundancy_convienient}) speaks about the members of
${\cal M}$. The trick, which allows us to relate both expressions, is to
examine their second-order behavior. By expanding $D(P_{{\mu^*}} \|
P_{{{\bar\mu}_i}})$ into a Taylor series around ${\mu^*}$, we get: $$
D(P_{{\mu^*}} \| P_{{{\bar\mu}_i}}) \simeq 0 + \expdif{1}({\mu^*}) ({{\bar\mu}_i} -
{\mu^*}) + \frac{1}{2} \expdif{2}({\mu^*}) ({{\bar\mu}_i} - {\mu^*})^2,
$$
where we abbreviated $\expdif{k}(\mu) = \frac{d^k}{d\mu^k} D(P_{\mu^*}\|P_\mu)$. The term $\expdif{1}({\mu^*})$ is
zero, since $D({\mu^*} \| \mu)$ as a function of $\mu$ has its
minimum at $\mu={\mu^*}$ \cite{BarndorffNielsen78}. As is well-known \cite{BarndorffNielsen78}, for
exponential families the term $\expdif{2}(\mu)$ coincides precisely with the
Fisher information $I_{{\cal P}}(\mu)$ evaluated at $\mu$. Another standard result
\cite{BarndorffNielsen78} for the mean-value parameterization says
that for all $\mu$, $I_{{\cal P}}(\mu) = \frac{1}{{\textnormal{var}}_{P_\mu}
X}$. Therefore, we get $D(P_{{\mu^*}} \| P_{{{\bar\mu}_i}}) \simeq
\frac{1}{2} \frac{({{\bar\mu}_i} - {\mu^*})^2}{{\textnormal{var}}_{P_{\mu^*}} X}$, and
similarly, $D(M_{{\mu^*}} \| M_{{{\bar\mu}_i}}) \simeq \frac{1}{2}
\frac{({{\bar\mu}_i} - {\mu^*})^2}{{\textnormal{var}}_{M_{\mu^*}} X}$, so that
$
D(M_{{\mu^*}} \| M_{{{\bar\mu}_i}}) \simeq D(P_{{\mu^*}} \| P_{{{\bar\mu}_i}}) \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}} X}
$,
and using (\ref{eq:KL_deriv}) and (\ref{eq:redundancy_convienient}):
$$
\begin{array}{c}
{\mathcal{R}}_U(n) \simeq D(P_{\mu^*}^{(n)} \| Q^{(n)}) \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}} X}.
\end{array}
$$
The last step of the proof is to use Rissanen's theorem and conclude
that
$\liminf_{n \to \infty} \frac{{\mathcal{R}}_U(n)}{\frac{1}{2} \ln n}$
is equal to
$$
\begin{array}{c}
\liminf_{n \to \infty} \frac{D(P^{(n)}_{\mu^*} \|
Q^{(n)})}{\frac{1}{2} \ln n} \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}}
X}
\geq \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}} X},
\end{array}
$$
for Lebesgue almost all ${\mu^*} \in \Theta_0$, and thus for Lebesgue almost all ${\mu^*} \in \ensuremath{\Theta_{\text{mean}}}$.
\end{proof}
We now use Theorem~\ref{thm:no_plugin}
to show that the redundancy of plug-in codes is suboptimal for all
exponential families which satisfy the following very weak condition:
\begin{condition}\label{cnd:dispersion_model}
Let ${\cal M} = \{{M_\mu}\mid\mu\in\ensuremath{\Theta_{\text{mean}}}\}$ be a single parameter exponential
family with sufficient statistic $X$ and mean-value parameter space $\ensuremath{\Theta_{\text{mean}}}$.
We require that there exists another single-parameter exponential family ${\cal P} = \{P_\mu \mid \mu \in \ensuremath{\Theta_{\text{mean}}}\}$ with the same
mean-value parameter space as ${\cal M}$, but with strictly larger variance
than ${\cal M}$ for every $\mu \in \ensuremath{\Theta_{\text{mean}}}$.
\end{condition}
The Condition \ref{cnd:dispersion_model} is widely satisfied among known
exponential families. When $\ensuremath{\mathcal X} = [a,b]$, we define
$P_\mu$ to be a ``scaled'' Bernoulli model, by putting all probability mass on
$\{a,b\}$ in such a way that $E_{P_\mu} =\mu$. It is easy to show, that
such distribution has the highest variance among all distributions defined on $[a,b]$ with a given
mean value $\mu$; therefore ${\textnormal{var}}_{P_\mu}X > {\textnormal{var}}_{M_\mu}X$, unless ${\cal M}$ is
a ``scaled'' Bernoulli itself. When $\ensuremath{\mathcal X} = {\mathbb{R}}$, ${\cal P}$ can be chosen
to be a normal family with fixed, sufficiently large variance
$\sigma^2$. For $X=[0,\infty)$, ${\cal P}$ can be taken to be a gamma family with sufficiently
large scale parameter. When $\ensuremath{\mathcal X} = \{0,1,2,\ldots\}$, ${\cal P}$ can
be taken to be
negative binomial (with expected ``number of successes'' sufficiently small).
Thus, we see that for all commonly
used exponential families, except for Bernoulli, Condition \ref{cnd:dispersion_model} holds. On the
other hand if ${\cal M}$ is Bernoulli, Corollary \ref{coll:no_plugin}
is no longer relevant anyway, since then $P$ must lie in ${\cal M}$.
\begin{corollary}
\label{coll:no_plugin}
Let ${\cal M} = \{{M_\mu}\mid\mu\in\ensuremath{\Theta_{\text{mean}}}\}$ a single parameter exponential
family with sufficient statistic $X$ and mean-value parameter space
$\ensuremath{\Theta_{\text{mean}}}$, satisfying Condition \ref{cnd:dispersion_model}. Let $U$ denote
any plug-in model with respect to ${\cal M}$ based on any sequence
of estimators ${\bar\mu}_1,{\bar\mu}_2,\ldots$.
Then, there exists a family of distributions ${\cal P} = \{P_\mu \mid \mu \in
\ensuremath{\Theta_{\text{mean}}}\}$, such that for Lebesgue almost all
${\mu^*} \in \ensuremath{\Theta_{\text{mean}}}$, for $X, X_1, X_2, \ldots$ i.i.d. $\sim
P_{\mu^*}$:
$$
\liminf_{n \to \infty} \frac{\mathcal{R}_U(n)}{\ln n}
\geq \frac{1}{2} \frac{{\textnormal{var}}_{P_{\mu^*}} X}{{\textnormal{var}}_{M_{\mu^*}} X}
> \frac{1}{2},
$$
so that the set of ${\mu^*}$ for which $U$ achieves the regret $\frac{1}{2} \ln n
+ O(1)$ is a set of Lebesgue measure zero.
\end{corollary}
\begin{proof}
Immediate from Theorem \ref{thm:no_plugin}.
\end{proof}
\section{Second Result: Optimality of Squashed ML}
\label{sec:squashed}
We showed that every plug-in code, including the ML plug-in code,
behaves suboptimally for 1-parameter families ${\cal M}$ unless ${\cal M}$
is Bernoulli. This fact does not, however, exclude the possibility
that a small modification of the ML plug-in code, which puts the
predictions slightly outside ${\cal M}$, will lead to the optimal redundancy
(\ref{eq:bic}). An argument supporting this claim comes from
considering the Bayesian predictive distribution when ${\cal M}$ is the normal family with
fixed variance $\sigma^2$. In this case, the Bayesian code based on prior
$\mathcal{N}(\mu_0,\tau^2_0)$ has a simple form \cite{Grunwald07}:
$$U_{\text{Bayes}}(z_{n+1} \mid z^n) = f_{\mu_n,\tau_n^2+\sigma^2}(z_{n+1}),$$
where $f_{\mu,\sigma^2}$ is the density of normal distribution
$\mathcal{N}(\mu,\sigma^2)$,
$$
\begin{array}{c}\mu_n = \frac{(\sum_{i=1}^n x_i) + \frac{\sigma^2}{\tau_0^2}
\mu_0}{n+\frac{\sigma^2}{\tau_0^2}},
\quad \text{and} \quad
\tau_n^2 = \frac{\sigma^2}{n + \frac{\sigma^2}{\tau_0}}.\end{array} $$
Thus, the Bayesian predictive distribution is itself a Gaussian with
mean equal to the modified maximum likelihood estimator (with
$n_0 = \sigma^2/\tau_0^2$), albeit
with a slightly larger variance $\sigma^2 + O(1/n)$. This shows that
for the normal family with fixed variance, there exists an ``almost'' in-model code, which
satisfies (\ref{eq:bic}). This led \cite{Grunwald07} to conjecture
that something similar holds for general exponential families. Here we
show that this is indeed the case: we propose a simple modification of the ML plug-in universal
model, obtained by predicting $z_{n+1}$ using a slightly ``squashed'' version
$M'_{\hat{\mu}_n}$ of
the ML estimator $M_{\hat{\mu}_n}$, defined as:
$$
M'_{{{\hat\mu}_n}}(z_{n+1})
: = M_{{{\hat\mu}_n}}(z_{n+1}) \frac{1 +
\frac{1}{2n} I_{\cal M}({{\hat\mu}_n})(x_{n+1} - {{\hat\mu}_n})^2}{1 + \frac{1}{2n}},
$$
where
${{\hat\mu}_n}$ is defined as in (\ref{eq:ML_estimator}) and $I_{\cal M}(\mu)$ is the
Fisher information for model ${\cal M}$.
Note that $M'_{{{\hat\mu}_n}}(z_{n+1})(\cdot)$ represents a
valid probability density: it is non-negative due to
$I_{\cal M}({{\hat\mu}_n}) > 0$ (property of exponential families), and it is
properly normalized:
$$
\begin{array}{cc}
&\int_\ensuremath{\mathcal X} M'_{{{\hat\mu}_n}}(z_{n+1})(z) dz =
\left(1 + \frac{1}{2n} \right)^{-1} \bigg( \int_\ensuremath{\mathcal X} M_{{{\hat\mu}_n}}(z) dz \\
&+
\frac{1}{2n}I_{\cal M}({{\hat\mu}_n}) \int_\ensuremath{\mathcal X} (X(z) - {{\hat\mu}_n})^2 M_{{{\hat\mu}_n}}(z) dz
\bigg) = 1,
\end{array}
$$
where the final equality follows because for exponential families,
$I_{\cal M}(\mu) = ({\textnormal{var}}_{M_{\mu}}X)^{-1}$.
While $M' \not \in {\cal M}$, we have $D(M'_{\hat{\mu}_n} \|
M_{\hat{\mu}_n}) = O(1/n)$, i.e. $M'$ is ``almost'' in-model
estimator.
\begin{definition}[Squashed ML prequential model]
\label{def:preqc} Given ${\cal M}$,
constants $x_0\in\ensuremath{\Theta_{\text{mean}}}$ and $n_0>0$, we define the
\emph{slightly squashed ML prequential model} $U$ by setting, for all $n$, all
$z^{n+1} \in \ensuremath{\mathcal Z}^{n+1}$:
$$U(z_{n+1} \mid z^n) = M'_{{{\hat\mu}_n}}(z_{n+1}),$$
\end{definition}\smallskip
where $M'$ is the slightly squashed ML estimator as above.
The codelengths of the corresponding slightly squashed ML prequential code are
not harder to calculate than those of the ordinary
ML plug-in model and in some cases they are easier to calculate than the
lengths of the Bayesian universal code. On the other hand, we show
below that the slightly squashed ML code always achieves the optimal
redundancy, satisfying (\ref{eq:bic}).
\begin{theorem}
\label{thm:robustML}
Let $X, X_1, X_2, \ldots$ be i.i.d.$\sim P$, with $E_P[X] = {\mu^*}$.
Let ${\cal M}$ be a single parameter exponential family with sufficient
statistic $X$ and ${\mu^*}$ an element of the mean value parameter space. Let
$U$ denote the slightly squashed ML model with respect to ${\cal M}$. If
${\cal M}$ and $P$ satisfy Condition \ref{condition} below,
then:
\begin{equation}
\label{eq:main}
{\mathcal{R}}_U(n)= \frac{1}{2} \ln n + O(1).
\end{equation}
\end{theorem}
\begin{condition}\label{condition}
We require that the following holds both for $T:= X$ and $T:=- X$:
\begin{itemize}
\item If $T$ is unbounded from above then there is a
$k\in\{4,6,\ldots\}$ such that the first $k$ moments of $T$ exist under $P$,
that ${d^2\over d\mu^2}I_{\cal M}(\mu)=O\left(\mu^{k-4}\right)$, ${d^4 \over
d\mu^4} D(M_{\mu^*} \| {M_\mu}) = O(\mu^{k-6})$ and that either $I_{\cal M}(\mu)$ is
constant or $I_{\cal M}(\mu)=O\left(\mu^{k/2-3}\right)$.
\item If $T$ is bounded from above by a constant $g$ then ${d^2\over
d\mu^2}I_{\cal M}(\mu)$, ${d^4 \over
d\mu^4} D(M_{\mu^*} \| {M_\mu})$, and $I_{\cal M}(\mu)$ are polynomial in
$1/(g-\mu)$.
\end{itemize}
\end{condition}
The usefulness of Theorem \ref{thm:robustML} depends on the validity of
Condition \ref{condition} among commonly used exponential families. As
can be seen from Figure
\ref{fig:condition}, for some standard exponential families,
our condition applies whenever the fourth moment of $P$ exists.
\begin{proof}{\em (of Theorem~\ref{thm:robustML}; rough sketch --- a
detailed proof is in the Appendix)}
We express the relative redundancy of the slightly squashed
ML plug-in code $U$ by the sum of the relative redundancy of the ordinary
ML plug-in code $\hat{U}$ and the difference in expected codelengths between
$U$ and $\hat{U}$:
$$
\begin{array}{l}
{\mathcal{R}}_{U}(n) = E_{P}[L_U(Z^n)] -
E_{P}[- \ln {M_\mstar}(Z^n)] =\\
E_{P}[L_U(Z^n) - L_{\hat{U}}(Z^n)] + {\mathcal{R}}_{\hat{U}}(n)
= \\ E_{P}[L_U(Z^n) - L_{\hat{U}}(Z^n)] + \frac{1}{2}
\frac{{\textnormal{var}}_P X}{{\textnormal{var}}_{M_{\mu^*}} X} \ln n + O(1),
\end{array}
\vspace*{-3pt}
$$
where the last equality follows from (\ref{eq:redundancyML}).
We have:
$$
\begin{array}{l}
L_U(Z^n) - L_{\hat{U}}(Z^n) = \\
\sum_{i=0}^{n-1} \left(- \ln U(Z_{i+1} \mid Z_i) +
\ln {\hat{U}}(Z_{i+1} \mid Z_i) \right) = \\
\sum_{i=0}^{n-1} \left( \ln\left(1 + \frac{1}{2i}\right) -
\ln\left(1 + \frac{1}{2i} I_{\cal M}({{\hat\mu}_i})(X_{i+1} -
{{\hat\mu}_i})^2\right)\right) .
\end{array}
$$
Since $\ln\left(1+\frac{1}{2i}\right) = \frac{1}{2i} + O(i^{-2})$, we get
$\sum_{i=0}^{n-1} \ln\left(1 + \frac{1}{2i}\right) = \frac{1}{2} \ln n + O(1)
$. Denoting $V_i =\frac{1}{2i}
I_{\cal M}({{\hat\mu}_i})(X_{i+1} - {{\hat\mu}_i})^2$, we also get $\ln(1+V_i) = V_i +
O(i^{-2})$. Next, we consider $E_P[V_i]$:
$$
\begin{array}{l}
E_P[V_i] = \frac{1}{2i} E_P \left[
I_{\cal M}({{\hat\mu}_i})(X_{i+1} - {\mu^*} + {\mu^*} - {{\hat\mu}_i})^2 \right] = \\
\frac{1}{2i} E_P \left[I_{\cal M}({{\hat\mu}_i})\left( {\textnormal{var}}_{P_{\mu^*}}X + ({\mu^*} - {{\hat\mu}_i})^2
\right) \right]
= \\ \frac{1}{2i} \left({\textnormal{var}}_{P_{\mu^*}}X E_P \left[I_{\cal M}({{\hat\mu}_i})\right] + E_P
\left[I_{\cal M}({{\hat\mu}_i}) ({\mu^*} - {{\hat\mu}_i})^2\right] \right).
\end{array}
$$
The second term $E_P\left[I_{\cal M}({{\hat\mu}_i}) ({\mu^*} - {{\hat\mu}_i})^2\right]$ is
$O(i^{-1})$ as $E_P[({\mu^*} - {{\hat\mu}_i})^2] = O(i^{-1})$ and $E[I_{\cal M}({{\hat\mu}_i})]
= I_{\cal M}({\mu^*}) + O(i^{-1})$ (follows from expanding $I_{\cal M}({{\hat\mu}_i})$ up to
the first order around ${\mu^*}$). Similarly, the first term is
$({\textnormal{var}}_{P_{\mu^*}}X) I_{\cal M}({\mu^*}) + O(i^{-1})$. Thus, using
$I_{\cal M}({\mu^*}) = \frac{1}{{\textnormal{var}}_{M_{\mu^*}}X}$, we finally get: $$ E_P[-\ln(1+V_i)] = -E_P[V_i] + O(i^{-2}) = \frac{1}{2i} \frac{{\textnormal{var}}_{P_{\mu^*}}X}{{\textnormal{var}}_{M_{\mu^*}}X} + O(i^{-2}). $$ Taking all together, we see that the terms $\frac{{\textnormal{var}}_{P_{\mu^*}}X}{{\textnormal{var}}_{M_{\mu^*}}X}$ cancel and we finally get
$
R_U(n) = \frac{1}{2} \ln n + O(1).
$
Condition \ref{condition} is necessary to ensure that all Taylor
expansions above hold.
\end{proof}
\begin{figure}
\caption{Fisher information, its second derivative and a fourth derivative of
the divergence for a number of exponential families.
For the normal
distribution with fixed mean
we use mean 0 and the density of the squared outcomes is given as a function of the variance.
}
\begin{center}
\label{fig:condition}
\begin{footnotesize}
\begin{tabular}{l@{}c@{}c@{}c}
\hline
Distribution & $I(\mu)$ & $\frac{d^2}{d\mu^2}I(\mu)$ & ${d^4 \over
d\mu^4} D(M_{\mu^*} \| {M_\mu})$ \\
\hline \\[-3mm]
Bernoulli & $\frac{1}{\mu(1-\mu)}$ & $\frac{2}{\mu^3} + \frac{2}{(1-\mu)^3}$ &
$\frac{6{\mu^*}}{\mu^4} + \frac{6(1-{\mu^*})}{(1-\mu)^4}$\\
Poisson & $\frac{1}{\mu}$ & $\frac{2}{\mu^3}$ & $\frac{6{\mu^*}}{\mu^4}$\\
Geometric & $\frac{1}{\mu(\mu-1)}$ & $- \frac{2}{\mu^3} + \frac{2}{(1-\mu)^3}$ &
$\frac{6{\mu^*}}{\mu^4} - \frac{6({\mu^*}+1)}{(\mu+1)^4}$\\
Gamma (fixed $k$) & $\frac{k}{\mu^2}$ & $\frac{6k}{\mu^4}$ &
$-\frac{6k}{\mu^4} + \frac{24k{\mu^*}}{\mu^5}$\\
Normal (fixed mean) & $\frac{1}{2\mu^2}$ & $\frac{3}{\mu^4}$ &
$-\frac{3}{\mu^4} + \frac{12{\mu^*}}{\mu^5}$ \\
Normal (fixed variance) & $\sigma^2$ & $0$ & $0$\\
\hline
\end{tabular}
\vspace*{-20pt}
\end{footnotesize}
\end{center}
\end{figure}
\section{Future Work}
In future work, we hope to extend our results concerning the slightly squashed
ML estimator to the multi-parameter case and establish almost-sure variation
of Theorem ~\ref{thm:robustML}. We also plan to analyze the estimator
in the individual sequence framework, along the lines of
\cite{CesaBianchiLugosi06,Raginsky09}.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
Motivated by string theory considerations,
ADHM invariants of curves were introduced in \cite{modADHM} as an alternative
construction for the local stable pair theory of curves of Pandharipande and Thomas
\cite{stabpairs-I}. They have been subsequently generalized in \cite{chamberI}
employing a natural variation of the
stability condition. An important feature of this construction resides in its
compatibility with the Joyce-Song theory of generalized Donaldson-Thomas
invariants \cite{genDTI}. Explicit wallcrossing formulas
for ADHM invariants have been derived and proven in \cite{chamberII}
using Joyce theory \cite{J-I,J-II,J-III,J-IV} and Joyce-Song theory \cite{genDTI}.
The purpose of the present paper is to study a further generalization
of ADHM invariants allowing higher rank framing sheaves.
This generalization is motivated in part
by recent work of Toda \cite{ranktwo} and Stoppa \cite{ranktwoGW} on rank two
generalized Donaldson-Thomas invariants of Calabi-Yau threefolds.
In contrast to \cite{ranktwo}, \cite{ranktwoGW}, the invariants constructed
here count local objects with nontrivial D2-rank, in physics terminology.
Similar rank two Donaldson-Thomas invariants of Calabi-Yau threefolds
are defined and computed in \cite{AS-I, AS-II} using both wallcrossing and
direct virtual localization methods.
Local invariants with higher D6-rank are also interesting on physical grounds.
Explicit results for such invariants are required in order to test the OSV
conjecture \cite{Ooguri:2004zv} for magnetically charged black holes.
In particular, such results would be needed in order to extend the work
of \cite{Aganagic:2004js} to local D-brane configuration with nonzero D6-rank.
According to \cite{DM-split}, counting invariants with higher D6-rank are also
expected to determine certain subleading
corrections to the OSV formula \cite{Ooguri:2004zv}. Moreover,
walls of marginal stability for BPS states with nontrivial D6-charge in a local
conifold model have been studied from a supergravity point of view in \cite{JM}.
The construction presented below should be viewed as a rigorous mathematical
framework for the microscopic theory of such BPS states. A detailed comparison
will appear elsewhere.
From the point of view of six dimensional
gauge theory dynamics, the invariants constructed in this paper
can be thought of as a higher rank generalization of local Donaldson-Thomas
invariants of curves.
It should be noted however that they are not the same as the higher rank
local DT invariants defined in \cite{modADHM}, which, from a gauge theoretic
point of view, are Coulomb branch invariants (see also \cite{JM,Cirafici:2008sn} for a noncommutative
gauge theory approach.)
Instead, employing a different treatment of boundary conditions in the
six dimensional gauge theory, the approach presented below yields Higgs branch
invariants.
The geometric setup of the present construction is specified by a
triple ${\mathcal X}=(X,M_1,M_2)$
where $X$ is a smooth projective curve of $X$ over $\mathbb{C}$ of genus $g$, and
$M_1,M_2$ are line bundles on $X$ so that $M=M_1\otimes_X M_2$
is isomorphic to the anticanonical bundle $K_X^{-1}$.
The data ${\mathcal X}$ determines
an abelian category ${\mathcal C}_{\mathcal X}$ of quiver sheaves on $X$ constructed
in \cite[Sect 3]{chamberI}.
Section (\ref{adhmstab}) consists of a step-by-step construction of
counting invariants for objects of ${\mathcal C}_{\mathcal X}$ following \cite{genDTI}.
The required stability conditions, chamber structure and moduli stacks are
presented in sections (\ref{defbasic}), (\ref{chamber}), (\ref{moduli})
respectively. Some basic homological algebra results are provided in
section (\ref{extensions}). The construction is concluded in
section (\ref{invariants}). Given a stability parameter $\delta\in \mathbb{R}$
the geometric data ${\mathcal X}$
determines a function $A_\delta: \IZ^{\times 3}\to \IQ$,
which assigns to any triple $\gamma=(r,e,v)$ the virtual number of
$\delta$-semistable ADHM sheaves on $X$ of type $\gamma$.
This function is supported on $\IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$.
In physics terms, the integers $(r,e,v)$ correspond to D2, D0 and D6-brane
charges respectively. In the derivation of wallcrossing formulas, it is more
convenient to use the alternative notation $\gamma=(\alpha,v)$,
$\alpha =(r,e) \in \IZ\times\IZ$. Moreover, the invariants $A_\delta(\alpha,0)$
are manifestly independent on $\delta$, and will be denoted by $H(\alpha)$
since they are counting invariants for Higgs sheaves on $X$.
Note that for a fixed type $\gamma$ there is a finite set $\Delta(\gamma)\subset
\mathbb{R}$
of critical stability parameters dividing the real axis in stability chambers
(see lemma (\ref{chamberlemma}).
The invariants $A_\delta(\gamma)$ are constant when $\delta$ varies within
a stability chamber.
The chamber $\delta> \mathrm{max}\, \Delta(\gamma)$ will be referred to as the
asymptotic chamber, and the corresponding invariants will be also denoted by
$A_\infty(\gamma)$.
The main result of this paper is a wallcrossing formula for
$v=2$ ADHM invariants at a critical stability parameter $\delta_c>0$
of type $(\alpha, 2)$, for arbitrary $\alpha =(r,e)
\in \IZ_{\geq 1}\times \IZ$. Certain preliminary definitions will be needed
in the formulation of this result, as follows.
For any integer $l\in \IZ_{\geq 1}$, and any $v\in \{1,2\}$
let ${\mathcal {HN}}_-(\alpha,v,\delta_c,l,l-1)$ denote the set of ordered sequences
$((\alpha_i))_{1\leq i\leq l}$, $\alpha_i\in \IZ_{\geq 1}\times \IZ$,
$1\leq i\leq l$ satisfying the following conditions
\begin{equation}\label{eq:alphadecomp}
\alpha_1+\cdots +\alpha_l=\alpha
\end{equation}
and
\begin{equation}\label{eq:slopesA}
{e_1\over r_1} = \cdots = {e_{l-1}\over r_{l-1}}={e_l+v\delta_c\over r_l}
={e+v\delta_c\over r}
\end{equation}
For any integer $l\in \IZ_{\geq 2}$, let ${\mathcal {HN}}_-(\alpha,2,\delta_c,l,l-2)$
denote the set of ordered sequences
$((\alpha_i))_{1\leq i\leq l}$, $\alpha_i\in \IZ_{\geq 1}\times \IZ$,
$1\leq i\leq l$ satisfying condition \eqref{eq:alphadecomp},
\begin{equation}\label{eq:slopesB}
{e_1\over r_1} = \cdots = {e_{l-2}\over r_{l-2}}={e_{l-1}+\delta_c\over r_{l-1}}
={e_{l}+\delta_c\over r_l} = {e+2\delta_c\over r},
\end{equation}
and
\begin{equation}\label{eq:slopesC}
1/{r_{l-1}} < 1/{r_{l}}.
\end{equation}
Let $0<\delta_-<\delta_c<\delta_+$ be stability parameters so that there are no critical
stability parameters of type $(\alpha,2)$ in the intervals $[\delta_-,\ \delta_c)$,
$(\delta_c,\ \delta_+]$. For any triple $(\beta,v)$, $\beta\in
\IZ_{\geq 1}\times\IZ\times \IZ_{\geq 1}$, $v\in \{1,2\}$, the
invariants $A_{\delta_\pm}(\beta,v)$ will be denoted by
$A_\pm(\beta,v)$.
Then the following result holds
for $\delta_-,\delta_+$ sufficiently close to $\delta_c$.
\begin{theo}\label{wallcrossingthmA}
The $v=2$ ADHM invariants satisfy the following wallcrossing formula
\begin{equation}\label{eq:wallformulaG}
\begin{aligned}
& A_-(\alpha,2)-A_+(\alpha,2) = \\
&\mathop{\sum_{l\geq 2}}_{} {1\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l,l-1)}}_{}
A_+(\alpha_l,2)\prod_{i=1}^{l-1}f_2(\alpha_i)H(\alpha_i)\\
&-{1\over 2}\mathop{\sum_{l\geq 1}}_{} {1\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l+1,l-1)}}_{}
g(\alpha_{l+1},\alpha_{l}) A_+(\alpha_l,1)
A_+(\alpha_{l+1},1)
\prod_{i=1}^{l-1}f_2(\alpha_i)H(\alpha_i)\\
& +{1\over 2}\mathop{\sum_{(\alpha_1,\alpha_2)\in
{\mathcal{HN}}_-(\alpha,2,\delta_c,2,0)}}_{} \mathop{\sum_{l_1\geq 1}}_{}
\mathop{\sum_{l_2\geq 1}}_{} {1\over (l_1-1)!}{1\over (l_2-1)!}
\mathop{\sum_{(\alpha_{1,i})\in {\mathcal{HN}}_-(\alpha_1,1, \delta_c, l_1,l_1-1)}}_{}\\
&
\mathop{\sum_{(\alpha_{2,i})\in {\mathcal{HN}}_-(\alpha_2,1, \delta_c, l_2,l_2-1)}}_{}
g(\alpha_1,\alpha_2)A_+(\alpha_{1,l_1},1)A_+(\alpha_{2,l_2},1)
\prod_{i=1}^{l_1-1} f_1(\alpha_{1,i})H(\alpha_{1,i})
\prod_{i=1}^{l_2-1} f_1(\alpha_{2,i})H(\alpha_{2,i}) \\
\end{aligned}
\end{equation}
where
\[
\begin{aligned}
f_v(\alpha) & = (-1)^{v(e-r(g-1))}v(e-r(g-1)),\qquad v=1,2\\
g(\alpha_1,\alpha_2) & = (-1)^{e_1-e_2-(r_1-r_2)(g-1)}(e_1-e_2-(r_1-r_2)(g-1))\\
\end{aligned}
\]
for any $\alpha=(e,r)$ respectively $\alpha_i=(r_i,e_i)$, $i=1,2$,
and the sum in the right hand side of equation \eqref{eq:wallformulaG}
is finite.
\end{theo}
Theorem (\ref{wallcrossingthmA}) is proven in section (\ref{wallformula})
using certain stack function identities established in section (\ref{sfidentities}).
Formula \eqref{eq:wallformulaG} is shown to agree with the wallcrossing
formula of Kontsevich and Soibelman in section (\ref{KSsect}).
An application of theorem (\ref{wallcrossingthmA}) to genus zero invariants is
presented in section (\ref{genuszero}).
Consider the following generating functions
\begin{equation}\label{eq:partfctA}
Z_{{\mathcal X},v}(u,q) = \mathop{\sum_{r\geq 1}}_{}\mathop{\sum_{n\in\IZ}}_{}
u^rq^{n} A_\infty(r,n-r,v)
\end{equation}
where $v=1,2$.
Using the wallcrossing formula \eqref{eq:wallformulaG} and the comparison
result of section (\ref{KSsect}), the following closed formulas
are proven in section (\ref{genuszero}).
\begin{coro}\label{closedform}
Suppose $X$ is a genus $0$ curve
and $M_1\simeq {\mathcal O}_X(d_1)$, $M_2\simeq {\mathcal O}_X(d_2)$ where
$(d_1,d_2)=(1,1)$ or $(0,2)$.
Then
\begin{equation}\label{eq:almostproductA}
\begin{aligned}
& Z_{{\mathcal X},1}(u,q) = \prod_{n=1}^\infty(1-u(-q)^n)^{(-1)^{d_1-1}n}\\
& Z_{{\mathcal X},2}(u,q) = {1\over 4}\prod_{n=1}^\infty(1-uq^n)^{2(-1)^{d_1-1}n}
- {1\over 2}\sum_{\substack{r_1>r_2\geq 1,\ n_1,n_2\in \IZ \\ \mathrm{or}\
r_1=r_2\geq1,\ n_2>n_1 \\ \mathrm{or}\ r_1 \geq 1, \ n_1 \in \IZ, \ r_2=n_2=0 }} (n_1-n_2) (-1)^{(n_1-n_2)} \\
&\qquad \qquad \qquad \qquad A_{\infty}(r_1,n_1-r_1,1)
A_{\infty}(r_2,n_2-r_2,1) u^{r_1+r_2}q^{n_1+n_2}.\\
\end{aligned}
\end{equation}
\end{coro}
\begin{rema}
The computations in section (\ref{genuszero}) based on the Kontsevich-Soibelman wallcrossing formula can be generalized to invariants of arbitrary rank $v\geq 2$.
Then it follows that the rank $v$ invariants of local $(-1,-1)$ and
$(0,-2)$ curves are recursively determined by the
invariants of lower rank $1\leq v'\leq v$. The resulting formulas are quite complicated,
and will be omitted.
\end{rema}
{\it Acknowledgements}
We are very grateful to Greg Moore for comments and suggestions on the manuscript.
The work of D.-E. D. is supported in part by NSF grant PHY-0854757-2009.
WYC is supported by DOE grant DE-FG02-96ER40959.
\section{Higher rank ADHM invariants}\label{adhmstab}
\subsection{Definitions and basic properties}\label{defbasic}
Let $X$ be a smooth projective curve of genus $g\in \IZ_{\geq 0}$
over an infinite field $K$ of
characteristic $0$ equipped with
a very ample line bundle ${\mathcal O}_X(1)$.
Let $M_1,M_2$ be fixed line bundles on $X$ equipped with a fixed isomorphism
$M_1\otimes_X M_2\simeq K_X^{-1}$. Set $M=M_1\otimes_X M_2$.
For fixed data ${\mathcal X}=(X,M_1,M_2)$,
let ${\mathcal Q}_{{\mathcal X},s}$ denote the abelian category
of $(M_1,M_2)$-twisted coherent ADHM quiver sheaves. An object
of ${\mathcal Q}_{\mathcal X}$ is given by a collection ${\mathcal E}=(E,E_{\infty},
\Phi_{1}, \Phi_{2}, \phi,\psi)$
where
\begin{itemize}
\item $E, E_{\infty}$ are coherent ${\mathcal O}_X$-modules
\item $\Phi_{i}:E\otimes_X M_i \to E$, $i=1,2$ , $
\phi:E\otimes_X M_1\otimes_X M_2 \to E_{\infty}$, $
\psi:E_{\infty}\to E$ are morphisms of ${\mathcal O}_X$-modules
satisfying the ADHM relation
\begin{equation}\label{eq:ADHMrelation}
\Phi_1\circ(\Phi_2\otimes 1_{M_1}) - \Phi_2\circ (\Phi_1\otimes 1_{M_2})
+ \psi\circ \phi =0.
\end{equation}
\end{itemize}
The morphisms are natural morphisms of quiver sheaves i.e. collections
$(\xi,\xi_{\infty}):(E,E_{\infty}) \to
(E',E'_{\infty})$
of morphisms of ${\mathcal O}_X$-modules satisfying the obvious compatibility conditions
with the ADHM data.
Let ${\mathcal C}_{{\mathcal X}}$ be the full abelian
subcategory of ${\mathcal Q}_{{\mathcal X}}$ consisting of objects with
$E_{\infty} = V \otimes {\mathcal O}_X$, where $V$
is a finite dimensional
vector spaces over $K$ (possibly trivial.)
Note that given any two objects ${\mathcal E},{\mathcal E}'$
of ${\mathcal C}_{\mathcal X}$, the morphisms
$\xi_{\infty}: V\otimes {\mathcal O}_X\to V'\otimes {\mathcal O}_X$ must be of the form
$\xi_{\infty} = f\otimes 1_{{\mathcal O}_X}$, where $f:V\to V'$
is a linear map.
An object ${\mathcal E}$ of ${\mathcal C}_{{\mathcal X}}$ will be called locally free if $E$ is a coherent
locally free ${\mathcal O}_X$-module.
Given a coherent ${\mathcal O}_X$-module $E$ we will denote by $r(E)$, $d(E)$,
$\mu(E)$ the rank, degree, respectively slope of $E$ if $r(E)\neq 0$.
The type of an object ${\mathcal E}$ of ${\mathcal C}_{{\mathcal X}}$ is the collection
$(r({\mathcal E}), d({\mathcal E}), v({\mathcal E}))= (r(E),d(E),\mathrm{dim}(V)))\in
\IZ_{\geq 0}\times \IZ\times \IZ_{\geq 0}$.
An object of ${\mathcal O}_X$ will be called an ADHM sheaf in the following.
Throughout this paper, the integer $v({\mathcal E})$ will be called the rank of ${\mathcal E}$,
as opposed to the terminology used in \cite{modADHM,chamberI,chamberII},
where the rank of ${\mathcal E}$ was defined to be $r({\mathcal E})$.
Note that the objects of ${\mathcal C}_{{\mathcal X}}$ with $v({\mathcal E})=0$
form a full abelian category which is naturally equivalent
to the
abelian category of
Higgs sheaves
on $X$ with coefficient bundles $(M_1,M_2)$ (see for example
\cite[App. A]{chamberI} for brief summary of the relevant definitions.)
Let
$\delta\in \mathbb{R}$ be a stability parameter.
The $\delta$-degree of an object ${\mathcal E}$ of ${\mathcal C}_{\mathcal X}$ is defined by
\begin{equation}\label{eq:udeltadeg}
\mathrm{deg}_{\underline \delta}({\mathcal E}) = d({\mathcal E}) + \delta v({\mathcal E}).
\end{equation}
If $r({\mathcal E})\neq 0$, the $\delta$-slope of ${\mathcal E}$ is defined by
\begin{equation}\label{eq:udeltaslope}
\mu_\delta({\mathcal E}) = {\mathrm{deg}_{\delta}({\mathcal E})\over r({\mathcal E})}.
\end{equation}
\begin{defi}\label{udeltastability}
Let $\delta\in \mathbb{R}$ be a stability parameter.
A nontrivial object ${\mathcal E}$ of ${\mathcal C}_{\mathcal X}$ is
$\delta$-(semi)stable if
\begin{equation}\label{eq:udeltastabcondA}
r(E)\, \mathrm{deg}_{\delta}({\mathcal E}') \ (\leq) \ r(E')\, \mathrm{deg}_\delta({\mathcal E})
\end{equation}
for any proper nontrivial subobject $0\subset {\mathcal E}'\subset {\mathcal E}$.
\end{defi}
The following lemmas summarize some basic properties of $\delta$-semistable ADHM sheaves.
The proofs are either standard or very similar to those of \cite[Lemm. 2.4]{chamberI}, \cite[Lemm 3.7]{chamberI} and will be omitted.
\begin{lemm}\label{basicprop}
Suppose ${\mathcal E}$ is a $\delta$-semistable framed ADHM sheaf with $r({\mathcal E})>0$
for some $\delta \in \mathbb{R}$. Then
\begin{itemize}
\item[$(i)$] $E$ is locally free.
\item[$(ii)$] If $\delta>0$, there is no nontrivial
linear subspace $0\subset V' \subseteq V$ so that $\psi|_{V'\otimes
{\mathcal O}_X}$ is identically zero. Similarly, if $\delta<0$,
there is no proper linear subspace
$0\subseteq V'\subset V$ so that $\mathrm{Im}
(\phi)\subseteq V'\otimes {\mathcal O}_X$.
\item[$(iii)$]
If ${\mathcal E}$ is $\delta$-stable
any endomorphism of ${\mathcal E}$ in ${\mathcal C}_{{\mathcal X}}$ is either
trivial or an isomorphism. If the ground field $K$ is algebraically closed, the endomorphism
ring of ${\mathcal E}$ is canonically isomorphic to $K$.
\end{itemize}
\end{lemm}
\begin{lemm}\label{boundlemma}
For fixed $(r,e,v)\in \IZ_{>0}\times \IZ\times \IZ_{\geq 0}$
there is a constant $c\in \mathbb{R}$ (depending only on ${\mathcal X}$ and (r,e,v))
so that for any $\delta\in \mathbb{R}$,
any $\delta$-semistable framed
ADHM sheaf of type $(r,e,v)$ satisfies
\[
\mathrm{\mu_{\max}(E)} < c.
\]
In particular, the set of isomorphism classes of framed ADHM sheaves of fixed type
$(r,e,v)$
which are $\delta$-semistable for some $\delta\in \mathbb{R}$
is bounded.
\end{lemm}
Given a locally free ADHM sheaf ${\mathcal E}=(E, \Phi_1,\Phi_2,\phi,\psi)$ on $X$
of type $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$,
the data
\begin{equation}\label{eq:dualADHM}
\begin{aligned}
{\widetilde E} & = E^\vee \otimes_X M^{-1}\\
{\widetilde \Phi}_i & = (\Phi_i^\vee\otimes 1_{M_i}) \otimes 1_{M^{-1}}: {\widetilde E}
\otimes
M_i \to {\widetilde E} \\
{\widetilde \phi} & = \psi^\vee \otimes 1_{M^{-1}} : {\widetilde E}\otimes_X {M} \to
V^\vee\otimes {\mathcal O}_X
\\
{\widetilde \psi} & = \phi^\vee : V^\vee \otimes {\mathcal O}_X \to {\widetilde E} \\
\end{aligned}
\end{equation}
with $i=1,2$, determines a locally free ADHM sheaf ${\widetilde {\mathcal E}}$ of type
$(r,-e+2r(g-1),v)$ where $g$ is the
genus of $X$. ${\widetilde {\mathcal E}}$ will be called the dual of ${\mathcal E}$ in the following.
Then the following lemma is straightforward.
\begin{lemm}\label{duallemma}
Let $\delta\in \mathbb{R}$ be a stability parameter and let
${\mathcal E}$ be a locally free ADHM sheaf on $X$.
Then ${\mathcal E}$ is $\delta$-(semi)stable if and only if ${\widetilde {\mathcal E}}$ is
$(-\delta)$-(semi)stable.
\end{lemm}
\subsection{Chamber structure}\label{chamber}
This subsection summarizes the main properties of
$\delta$-stability chambers.
\begin{defi}\label{asympstab}
An ADHM sheaf ${\mathcal E}$ of type
$(r,e,v)\in\IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$ is
asymptotically (semi)stable if the following
conditions hold
\begin{itemize}
\item[$(i)$] $E$ is locally free, $\psi:V\otimes {\mathcal O}_X\to E$ is not
identically zero, and
there is no saturated proper nontrivial subobject $0\subset {\mathcal E}'\subset {\mathcal E}$
in ${\mathcal C}_{\mathcal X}$ so that $v({\mathcal E}')/r({\mathcal E}') > v/r$.
\item[$(ii)$] Any proper nontrivial subobject $0\subset {\mathcal E}'\subset {\mathcal E}$
with $v({\mathcal E}')/r({\mathcal E}') = v/r$ satisfies the slope inequality $\mu(E') \ (\leq) \ \mu(E)$.
\end{itemize}
\end{defi}
Here a subobject ${\mathcal E}'\subset {\mathcal E}$ is called saturated in the underlying
coherent sheaf $E'$ is saturated in $E$. Note that according to \cite[Lemm.
3.10]{chamberI}, any proper subobject $0\subset {\mathcal E}'\subset {\mathcal E}$ admits a canonical saturation ${\overline {{\mathcal E}'}}\subset {\mathcal E}$.
\begin{lemm}\label{asympbound}
The set of isomorphism classes of asymptotically semistable ADHM sheaves
of fixed type $(r,e,v)\in\IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$ is bounded.
\end{lemm}
{\it Proof.} The proof is based on Maruyama's boundedness theorem.
Suppose ${\mathcal E}$ is asymptotically semistable of type $(r,e,v)$, and the underlying
coherent sheaf $E$ is not semistable. Then there is a nontrivial Harder-Narasimhan
filtration
\[
0\subset E_1\subset \cdots \subset E_h = E
\]
with $h\geq 2$ so that $\mu(E_j) > \mu(E)$ and $r(E_j)<r$ for all $1\leq j \leq h-1$.
Suppose $E_j$ is $\Phi_i$-invariant, $i=1,2$, and $\mathrm{Im}(\psi)\subseteq E_j$
for some $1\leq j\leq h-1$. Then the data ${\mathcal E}_j=(E_j, \Phi_i|_{E_j\otimes_X M_i},
\phi|_{E_j\otimes_X M}, \psi)$ is subobject of
${\mathcal E}$ with
\[
v({\mathcal E}_j)/r({\mathcal E}_j) = {v\over r({\mathcal E}_j)}>{v\over r}.
\]
Since $E_j\subset E$ is saturated, it follows that ${\mathcal E}_j$ violates condition $(i)$
in definition (\ref{asympstab}). Therefore for any $1\leq j\leq h$, $E_j$ is either
not preserved by some $\Phi_i$, $i=1,2$, or it does not contain the image
of $\psi$. From this point on the proof is identical to the proof of
\cite[Prop. 2.7]{modADHM}.
\hfill $\Box$
\begin{defi}\label{genstab}
Let $\delta\in \mathbb{R}_{>0}$.
A $\delta$-semistable ADHM sheaf ${\mathcal E}$ of type
$(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$ is generic
if it is either $\delta$-stable or any proper nontrivial subobject $0\subset {\mathcal E}'\subset {\mathcal E}$
of type $(r',e',v')\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$ satisfies
\begin{equation}\label{eq:gencond}
{e'\over r'}={e\over r} \qquad {v'\over r'} = {v \over r}.
\end{equation}
The stability parameter $\delta\in \mathbb{R}_{>0}$ is called generic of type $(r,e,v)$ if
any $\delta$-semistable ADHM sheaf of type $(r,e,v)$ is generic. The stability
parameter $\delta \in \mathbb{R}_{>0}$ is called critical of type $(r,e,v)$
if there exists a nongeneric
$\delta$-semistable ADHM sheaf of type $(r,e,v)$.
\end{defi}
Lemma (\ref{boundlemma}) implies the following.
\begin{lemm}\label{asympstablemma}
For fixed $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$
there exists $\delta_\infty\in\mathbb{R}_{>0}$ so that
for all $\delta\geq \delta_\infty$
an ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$ is $\delta$-(semi)stable
if and only if it is asymptotically (semi)stable.
\end{lemm}
{\it Proof.} The proof if similar to the proof of lemma \cite[Lemm. 4.7]{chamberI}.
Some details will be provided for convenience. It is straightforward to prove that
asymptotic stability implies $\delta$-stability for sufficiently large $\delta$
using lemma (\ref{boundlemma}). The converse is slightly more involved.
First note that given any nontrivial locally free ADHM sheaf ${\mathcal E}$, any linear subspace
$V'\subset V$, determines a canonical subobject ${\mathcal E}_{V'}\subset {\mathcal E}$.
${\mathcal E}_{V'}$ is the saturation of the subobject of ${\mathcal E}$ generated by
$V'\otimes {\mathcal O}_X$ by successive applications of the ADHM morphisms $\psi, \Phi_i, \phi$.
Since ${\mathcal E}_{V'}$ is canonically determined by $V'$ and ${\mathcal E}$, lemma
(\ref{boundlemma}) implies that the set of isomorphism classes of subobjects
${\mathcal E}_{V'}$, where ${\mathcal E}$ is a $\delta$-semistable ADHM sheaf of type
$(r,e,v)$ for some $\delta>0$ is bounded. Moreover, by construction, any
subobject $0\subset {\mathcal E}'\subset {\mathcal E}$ contains the canonical subobject ${\mathcal E}_{V'}$.
Now suppose that for any $\delta>0$ there exists a $\delta$-semistable ADHM
sheaf ${\mathcal E}$ of type $(r,e,v)$ which is not asymptotically stable.
Let $0\subset {\mathcal E}'\subset {\mathcal E}$ be a saturated nontrivial proper saturated subobject
violating the asymptotic stability conditions. Note that ${\mathcal E}'$ cannot violate
condition $(ii)$ in definition (\ref{asympstab}) since ${\mathcal E}$ is $\delta$-semistable.
Therefore it must violate condition $(i)$ i.e. $v'/r'>v/r$ where
$r'=r({\mathcal E}')$. In particular $v'=v({\mathcal E}')>0$.
Then the subobject ${\mathcal E}_{V'}$ also violates condition $(i)$ since
\[
{v({\mathcal E}_{V'})\over r({\mathcal E}_{V'})} = {v'\over r({\mathcal E}_{V'})} \geq {v'\over r'}>{v/r}.
\]
Since ${\mathcal E}$ is $\delta$-semistable $\mu_\delta({\mathcal E}_{V'}) \leq
\mu_\delta({\mathcal E})$. However, as noted above, the set of isomorphism classes
of all ${\mathcal E}_{V'}$ is bounded, therefore the set of all types $(r({\mathcal E}_{V'}), d({\mathcal E}_{V'}),
v({\mathcal E}_{V'}))$ is finite. Taking $\delta$ sufficiently large, this leads to a contradiction.
\hfill $\Box$
By analogy with \cite[Lemm. 4.4]{chamberI},
\cite[Lemm. 4.6]{chamberI}, lemmas (\ref{asympstablemma}) and
(\ref{duallemma}) imply the following.
\begin{lemm}\label{chamberlemma}
Let $(r,e,v)\in\IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$ be a fixed type.
Then there is a finite set $\Delta(r,e,v) \subset \mathbb{R}$ of critical stability parameters of type
$(r,e,v)$.
Given any two stability parameters $\delta,\delta'\in \mathbb{R}$, $\delta< \delta'$
so that $[\delta, \ \delta']\cap \Delta(r,e,v)=\emptyset$, the set of $\delta$-semistable
ADHM sheaves of type $(r,e,v)$ is identical to the set of $\delta'$-semistable
ADHM sheaves of type $(r,e,v)$.
\end{lemm}
\begin{rema}\label{requalsone}
It is straightforward to check that $\Delta(1,e,v)=\{0\}$ for any $v\geq 1$.
\end{rema}
\begin{lemm}\label{specstab}
Let $(r,e,v)\in \IZ_{\geq 1} \times \IZ\times \IZ_{\geq 1}$
and let $\delta_c>0$ be a critical stability parameter of type $(r,e,v)$.
Let $\delta_\pm >0$ be stability parameters so that $\delta_-<\delta_c<\delta_+$
and $[\delta_-,\ \delta_c)\cap \Delta(r,e,v) =\emptyset$,
$(\delta_c,\ \delta_+]\cap \Delta(r,e,v) =\emptyset$.
If ${\mathcal E}$ is a $\delta_\pm$-semistable ADHM sheaf of type $(r,e,v)$, then
${\mathcal E}$ is also $\delta_c$-semistable.
\end{lemm}
\begin{defi}\label{admconfig}
Let $(r,v)\in \IZ_{\geq 1} \times \IZ_{\geq 1}$.
$(a)$ A positive admissible configuration of type $(r,v)$
is an ordered sequence of integral points
$\left(\rho_i=(r_i,v_i)\in \IZ_{\geq 1} \times \IZ_{\geq 0}\right)_{1\leq i\leq h, \, h\geq 1}$ satisfying the following conditions
\begin{itemize}
\item $\rho_1+\cdots + \rho_h = (r,v)$.
\item $(v_1+\cdots +v_i)/(r_1+\cdots+r_i) > v/r$
and $v_{i}/r_i > v_{i+1}/r_{i+1}$ for all $i=1,\ldots, h-1$.
\end{itemize}
$(b)$ A negative admissible configuration of type $(r,v)$
is an ordered sequence of integral points
$\left(\rho_i=(r_i,v_i)\in \IZ_{\geq 1} \times \IZ_{\geq 0}\right)_{1\leq i\leq h,\, h\geq 1}$ satisfying the following conditions
\begin{itemize}
\item
$\rho_1+\cdots + \rho_h = (r,v)$.
\item $(v_1+\cdots +v_i)/(r_1+\cdots+r_i) < v/r$
and $v_{i}/r_i< v_{i+1}/r_{i+1}$ for all $i=1,\ldots, h-1$.
\end{itemize}\end{defi}
\begin{rema}\label{configrema}
$(i)$
It is straightforward to prove that for fixed $(r,v)\in \IZ_{\geq 1} \times \IZ_{\geq 1}$ the set of positive, respectively negative, admissible configurations
is finite. These sets will be denoted by ${\mathcal {HN}}_\pm(r,v)$.
$(ii)$ The only positive, respectively negative admissible configuration of type $(r,v)$
with $h=1$ is $(\rho=(r,v))$.
\end{rema}
\begin{lemm}\label{HNfiltrations}
Let $\delta_c\in \mathbb{R}_{>0}$ be a critical stability parameter of
type $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$.
Then the following hold.
$(i)$ There exists $\epsilon_+ >0$,
so that $(\delta_c,\ \delta_c+\epsilon_+]\cap \Delta(r,e,v)=\emptyset$
and the following holds for any $\delta_+\in (\delta_c, \ \delta_c+\epsilon_+)$.
A locally free ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$
on $X$ is $\delta_c$-semistable if and only if
it is either $\delta_+$-semistable or there exists a unique filtration of the form
\begin{equation}\label{eq:plusfiltr}
0= {\mathcal E}_0\subset {\mathcal E}_1\subset \cdots \subset {\mathcal E}_h={\mathcal E}
\end{equation}
with $h\geq 2$ satisfying the following conditions
\begin{itemize}
\item The successive quotients ${\mathcal F}_i={\mathcal E}_i/{\mathcal E}_{i-1}$, $i=1,\ldots, h$ of the
filtration \eqref{eq:plusfiltr} are locally free ADHM sheaves with numerical types
$(r_i,e_i,v_i)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$.
$\delta_+$ is noncritical of type $(r_i,e_i,v_i)$, ${\mathcal F}_i$ is
$\delta_+$-semistable and $\mu_{\delta_c}({\mathcal F}_i) = \mu_{\delta_c}({\mathcal E})$
for all $i=1,\ldots, h$.
\item The sequence $\rho_i=(r_i,v_i)$, $i=1,\ldots, h$
is a positive admissible configuration of type $(r,e,v)$.
\end{itemize}
$(ii)$ There exists $\epsilon_- >0$,
so that $[\delta_c-\epsilon_-, \delta_c)\cap \Delta(r,e,v)=\emptyset$
and the following holds for any $\delta_-\in (\delta_c-\epsilon_-, \ \delta_c)$.
A locally free ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$
on $X$ is $\delta_c$-semistable if and only if
it is either $\delta_-$-semistable or there exists a unique filtration of the form
\begin{equation}\label{eq:minusfiltr}
0= {\mathcal E}_0\subset {\mathcal E}_1\subset \cdots \subset {\mathcal E}_h={\mathcal E}
\end{equation}
with $h\geq 2$ satisfying the following conditions
\begin{itemize}
\item The successive quotients ${\mathcal F}_i={\mathcal E}_i/{\mathcal E}_{i-1}$, $i=1,\ldots, h$ of the
filtration \eqref{eq:minusfiltr} are locally free ADHM sheaves with numerical types
$(r_i,e_i,v_i)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$.
$\delta_-$ is noncritical of type $(r_i,e_i,v_i)$, ${\mathcal F}_i$ is
$\delta_-$-semistable and $\mu_{\delta_c}({\mathcal F}_i) = \mu_{\delta_c}({\mathcal E})$
for all $i=1,\ldots, h$.
\item The sequence $\rho_i=(r_i,v_i)$, $i=1,\ldots, h$
is a negative admissible configuration of type $(r,e,v)$.
\end{itemize}
\end{lemm}
{\it Proof.}
The proof is similar to the proof of \cite[Lemm. 4.13]{chamberI}.
Details are included below for completeness.
Note that it suffices to prove statement $(i)$ since the proof of $(ii)$ is
analogous.
Let $\delta_+>\delta_c$ be an arbitrary
noncritical stability parameter
of type $(r,e,v)$ so that $(\delta_c,\delta_+]\cap \Delta(r,e,v)=\emptyset$.
Suppose ${\mathcal E}$ is a $\delta_c$-semistable ADHM
sheaf on $X$. Then ${\mathcal E}$ is either
$\delta_+$-stable or there is a Harder-Narasimhan filtration of ${\mathcal E}$ with respect to
$\delta_+$-semistability
\begin{equation}\label{eq:plusHNfiltr}
0\subset {\mathcal E}_1 \subset \cdots \subset {\mathcal E}_h = {\mathcal E}
\end{equation}
where $h\geq 2$. It is straightforward to check that ${\mathcal E}_l$, $1\leq l\leq h$
must have $r({\mathcal E}_l)\geq 1$ and the successive quotients
${\mathcal F}_l$, $0\leq l \leq h-1$ must also have $r_l\geq 1$.
Then by the general properties of Harder-Narasimhan filtrations
\begin{equation}\label{eq:HNineqB}
\mu_{\delta_+}({\mathcal E}_1)> \mu_{\delta_+}({\mathcal E}_2/{\mathcal E}_1)>\cdots >
\mu_{\delta_+}({\mathcal E}_h/{\mathcal E}_{h-1})
\end{equation}
and
\begin{equation}\label{eq:ineqAB}
\mu_{\delta_+}({\mathcal E}_l)> \mu_{\delta_+}({\mathcal E})
\end{equation}
for all $1\leq l\leq h-1$. Since ${\mathcal E}$ is $\delta_c$-semistable by assumption,
inequalities \eqref{eq:ineqAB} imply that
\begin{equation}\label{eq:HNineqX}
v({\mathcal E}_l)/r({\mathcal E}_l) > v/r
\end{equation} for all $l=1,\ldots, h$. Note that $v({\mathcal E}_l) = v_1+\cdots+ v_l$,
$r({\mathcal E}_l)=r_1+\cdots + r_l$ for any $l=1,\ldots, h$.
Moreover, using the $\delta_c$-semistability condition and
inequalities \eqref{eq:ineqAB}
we have
\begin{equation}\label{eq:ineqB}
\delta_+\left({v\over r}-{v({\mathcal E}_l)\over r(E_l)}\right)
< \mu(E_l) -\mu(E) \leq \delta_c\left({v\over r}-{v({\mathcal E}_l)\over r(E_l)}\right)
\end{equation}
for all $l=1,\ldots, h$.
Now let $\gamma >\delta_c$ be a fixed stability
parameter so that $(\delta_c,\gamma]\cap \Delta(r,e,v)=\emptyset$.
Using Grothendieck's lemma and lemma (\ref{boundlemma}), inequalities \eqref{eq:ineqB}
imply that the set of isomorphism classes of locally free ADHM sheaves ${\mathcal E}'$ on $X$ satisfying condition $(\star)$ below is bounded.
\begin{itemize}
\item[$(\star)$]
There exists a $\delta_c$-semistable
ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$ and a stability parameter $\delta_+
\in (\delta_c\ \gamma]$ so that ${\mathcal E}'\simeq {\mathcal E}_l$ for
some $l\in \{1,\ldots, h\}$, where $0\subset {\mathcal E}_1\subset\cdots\subset
{\mathcal E}_h={\mathcal E}$,
$h\geq 1$, is the Harder-Narasimhan filtration of ${\mathcal E}$
with respect to $\delta_+$-semistability.
\end{itemize}
Then it follows that the set
of numerical types $(r',e',v')$ of
locally free ADHM sheaves ${\mathcal E}'$ satisfying property $(\star)$
is finite. This implies that there exists $0< \epsilon_+<\gamma-\delta_c$
so that for any $\delta_+\in (\delta_c, \ \delta_c+\epsilon_+)$, and
any $\delta_c$-semistable ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$
inequalities \eqref{eq:ineqB} can be satisfied only if
\begin{equation}\label{eq:equal}
\mu_{\delta_c}({\mathcal E}_l) = \mu_{\delta_c}({\mathcal E})
\end{equation}
for all $l=1,\ldots, h$.
Hence also
\[
\mu_{\delta_c}({\mathcal E}_l/{\mathcal E}_{l-1}) = \mu_{\delta_c}({\mathcal E})
\]
for all $l=2,\ldots, h$. Then inequalities \eqref{eq:HNineqB}, \eqref{eq:HNineqX}
imply
that the sequence $\rho_l=(r_l,v_l)$, $l=1,\ldots, h$ is a
positive admissible configuration.
Therefore for all $\delta_+\in (\delta_c, \ \delta_c+\epsilon_+)$,
any locally free $\delta_c$-semistable ADHM sheaf ${\mathcal E}$ of type $(r,e,v)$ is either
$\delta_+$-stable or has a
Harder-Narasimhan filtration with respect to $\delta_+$-semistability
as in lemma (\ref{HNfiltrations}.$i$).
Next note that the set of numerical types
\begin{equation}\label{eq:slopeset}
{\sf S}_{\delta_c}(r,e,v) = \{(r',e',v')\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}\, |\, 0<r'\leq r ,\ 0\leq v'\leq v, \
r(e'+\delta_c v') = r'(e+\delta_c v)\}
\end{equation}
is finite. Therefore $0< \epsilon_+<\gamma-\delta_i$ above may be chosen so
that there are no critical stability parameters of type $(r',e',v')$ in the interval
$(\delta_c, \ \delta_c+\epsilon_+)$ for any
$(r',e',v')\in {\sf {S}}_{\delta_c}(r,e,v)$. In particular, $\delta_+$ is noncritical of type
$(r_i,e_i,v_i)$, $i=1,\ldots, h$ for any Harder-Narasihan filtration as above.
Conversely, suppose ${\mathcal E}$ is a locally free ADHM sheaf of type $(r,e,v)$
on $X$ which has a
filtration of the form \eqref{eq:plusfiltr} with ${\mathcal E}'$ $\delta_+$-stable
and satisfying the conditions of lemma (\ref{HNfiltrations}.$i$) for some
$\delta_+\in (\delta_c, \ \delta_c+\epsilon_+)$.
By the above choice of $\epsilon_+$, there are no critical stability parameters
of type $(r_i,e_i,v_i)$ in the interval $(\delta_c, \ \delta_c+\epsilon_+)$,
for any $i=1,\ldots,h$. Since ${\mathcal F}_i$ are $\delta_+$-semistable, lemma (\ref{specstab})
implies that ${\mathcal F}_i$ is also $\delta_c$-semistable, for any $i=1,\ldots,h$.
Hence ${\mathcal E}$ is also $\delta_c$-semistable since the ${\mathcal F}_i$ have equal $\delta_c$-slopes.
\hfill $\Box$
\subsection{Extension groups}\label{extensions}
Let ${\mathcal E}',{\mathcal E}''$ be nontrivial locally free objects in ${\mathcal C}_{\mathcal X}$ of types
$(r',e',v'), (r'',e'',v'')\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$.
Let ${\mathcal C}({\mathcal E}'',{\mathcal E}')$ be the three term complex
\begin{equation}\label{eq:hypercohA}
\begin{aligned}
0 \to \begin{array}{c} {\mathcal Hom}_{X}(E'',E') \\ \end{array}
& {\buildrel d_1\over \longrightarrow}
\begin{array}{c} {\mathcal Hom}_{X}(E''\otimes_{X}M_1 ,E') \\ \oplus \\
{\mathcal Hom}_{X}(E''\otimes_{X} M_2, E') \\ \oplus \\
{\mathcal Hom}_{X}(E''\otimes_{X} M,V'\otimes {\mathcal O}_X)\\ \oplus \\
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,E') \\ \end{array}
{\buildrel d_2\over \longrightarrow}
{\mathcal Hom}_{X}(E''\otimes_{X}M,E') \to 0 \\
\end{aligned}
\end{equation}
where
\[
\begin{aligned}
d_1(\alpha) =
(& -\alpha \circ \Phi_{1}'' +\Phi_{1}'\circ (\alpha\otimes 1_{M_1}),
-\alpha \circ \Phi_{2}''+\Phi_{2}'\circ (\alpha\otimes 1_{M_2}),\\
& \phi' \circ (\alpha
\otimes 1_M),
-\alpha\circ \psi'' )\\
\end{aligned}
\]
for any local sections $(\alpha,\alpha_\infty)$ of
the first term
and
\[
\begin{aligned}
d_2(\beta_1,\beta_2,\gamma, \delta) = &
\beta_1 \circ (\Phi''_2\otimes 1_{M_1}) -
\Phi_{2}'\circ (\beta_1\otimes 1_{M_2})
- \beta_2\circ (\Phi''_{1}\otimes 1_{M_2})\\
& + \Phi_{1}'\circ (\beta_2\otimes 1_{M_1}) +
\psi'\circ \gamma + \delta \circ \phi''\\
\end{aligned}
\]
for any local sections $(\beta_1,\beta_2,\gamma, \delta)$
of the middle term. The degrees of the three terms in
\eqref{eq:hypercohA} are $0,1,2$ respectively.
Let $C({\mathcal C}({\mathcal E}'',{\mathcal E}'))$ be the double complex obtained from ${\mathcal C}({\mathcal E}'',{\mathcal E}')$
by taking
${\check{\rm C}}$ech resolutions and let ${D}({\mathcal E}',{\mathcal E}'')$ be the diagonal complex
of $C({\mathcal C}({\mathcal E}'',{\mathcal E}'))$. Note that there is a canonical linear map
\[
\begin{aligned}
\mathrm{Hom}(V'',V') & \to D^1({\mathcal E}',{\mathcal E}'') = C^0({\mathcal C}^1({\mathcal E}'',{\mathcal E}'))\oplus
C^1({\mathcal C}^0({\mathcal E}'',{\mathcal E}'))\\
f & \to \left[\begin{array}{c}
{}^t(0,0,-(f\otimes 1_{{\mathcal O}_X})\circ \phi'', \psi'\circ (f\otimes 1_{{\mathcal O}_X}))\\
0\\ \end{array}\right]\\
\end{aligned}
\]
Given the above expressions for the differentials $d_1,d_2$ it is straightforward
to check that this map yields a morphism of complexes
\[
\begin{aligned}
\varrho: \mathrm{Hom}(V'',V')[-1] & \to {D}({\mathcal E}'', {\mathcal E}') \\
\end{aligned}
\]
Let ${\widetilde D}({\mathcal E}'',{\mathcal E}')$ denote the cone of $\varrho$.
Then the lemma below follows either by explicit $\check{\rm{C}}$ech cochain computations
as in \cite[Sect. 4]{modADHM} or using the methods of \cite{homquiv}.
\begin{lemm}\label{extlemma}
The extension groups $\mathrm{Ext}^k_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')$, $k=0,1$
are isomorphic to the cohomology groups $H^k({\widetilde D}({\mathcal E}'',{\mathcal E}'))$, $k=0,1$.
Moreover there is an exact sequence
\begin{equation}\label{eq:extseq}
\xymatrix{
0\ar[r] & \IH^0({\mathcal C}({\mathcal E}'',{\mathcal E}')) \ar[r] & \mathrm{Ext}^0_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')\ar[r]
& {Hom}(V'',V') \\
\ar[r] & \mathrm{Ext}^1_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}') \ar[r] &
\IH^1({\mathcal C}({\mathcal E}'',{\mathcal E}')) \ar[r] & 0 \\}
\end{equation}
where $\IH^k({\mathcal C}({\mathcal E}'',{\mathcal E}'))$, $k=0,1$ are hypercohomology groups of the
complex ${\mathcal C}({\mathcal E}'',{\mathcal E}')$.
\end{lemm}
\begin{coro}\label{extcoro}
Given any two locally free objects ${\mathcal E}',{\mathcal E}''$
\begin{equation}\label{eq:eulerchar}
\begin{aligned}
& \mathrm{dim}(\mathrm{Ext}^0_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')) -
\mathrm{dim}(\mathrm{Ext}^1_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')) -
\mathrm{dim}(\mathrm{Ext}^0_{{\mathcal C}_{\mathcal X}}({\mathcal E}',{\mathcal E}'')) \\ & +
\mathrm{dim}(\mathrm{Ext}^1_{{\mathcal C}_{\mathcal X}}({\mathcal E}',{\mathcal E}'')) =
v'e''-v''e'-(v'r''-v''r')(g-1) \\
\end{aligned}
\end{equation}
\end{coro}
{\it Proof.} Follows from the exact sequence \eqref{eq:extseq} and
the fact that the hypercohomology groups of the complex
${\mathcal C}({\mathcal E}'',{\mathcal E}')$ satisfy the duality relation
\[
\IH^k({\mathcal C}({\mathcal E}'',{\mathcal E}')) \simeq \IH^{3-k}({\mathcal C}({\mathcal E}',{\mathcal E}''))^\vee
\]
for $k=0,\ldots,3$.
\hfill $\Box$
\subsection{Moduli stacks}\label{moduli}
In the following let the ground field $K$ be $\mathbb{C}$.
Let ${\mathfrak {Ob}}({\mathcal X})$ denote the moduli stack of all objects of the abelian
category ${\mathcal C}_{\mathcal X}$ and let ${\mathfrak {Ob}}({\mathcal X},r,e,v)$ denote the open
and closed component of type $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$.
Standard arguments analogous to \cite[Sect. 9]{J-I}, \cite[Sect. 10]{J-I}
prove that ${\mathfrak {Ob}}({\mathcal X})$ is an algebraic
stack locally of finite type and it satisfies conditions \cite[Assumption 7.1]{J-I},
\cite[Assumption 8.1]{J-I}. Given the boundedness result
(\ref{boundlemma}), the following is also standard.
\begin{prop}\label{modstack}
For fixed type $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$ and fixed
$\delta\in \mathbb{R}_{>0}$ there is an algebraic moduli stack of finite type
${\mathfrak{M}}_\delta^{ss}({\mathcal X},r,e,v)$
of $\delta$-semistable objects of type $(r,e,v)$ of ${\mathcal C}_{\mathcal X}$.
If $\delta<\delta'$ are two stability parameters so that
$[\delta,\ \delta']\cap \Delta(r,e,v)=\emptyset$, the corresponding moduli stacks
are canonically isomorphic.
Moreover, for any $\delta\in \mathbb{R}$ there are canonical open embeddings
\begin{equation}\label{eq:opembd}
{\mathfrak{M}}_\delta^{ss}({\mathcal X},r,e,v)\hookrightarrow {\mathfrak {Ob}}({\mathcal X},r,e,v)
\hookrightarrow {\mathfrak {Ob}}({\mathcal X}).
\end{equation}
\end{prop}
\subsection{ADHM invariants}\label{invariants}
ADHM invariants will be defined applying the formalism of Joyce and
Song \cite{genDTI} to $\delta$-semistable ADHM sheaves on $X$. Given corollary
(\ref{extcoro}), the required results on Behrend constructible
functions are a straightforward generalization of the analogous
statements proven in
\cite[Sect. 7]{chamberI} for ADHM sheaves with $v=1$.
Therefore the construction of generalized Donaldson-Thomas invariants
via Behrend's constructible functions \cite{genDTI} applies to the present case.
Let ${\sf L}({\mathcal X})$ be the Lie algebra over $\IQ$ spanned by
$\{\lambda(\gamma)\,|\, \gamma\in \IZ^3\}$ with Lie bracket
\[
[\lambda(\gamma'), \lambda(\gamma'')] = (-1)^{\chi(\gamma',\gamma'')}
\chi(\gamma',\gamma'') \lambda(\gamma'+\gamma'')
\]
where
\[
\chi(\gamma',\gamma'') = v''e'-v'e''-(v''r'-v'r'')(g-1)
\]
for any $\gamma'=(r',e',v')$, $\gamma''=(r'',e'',v'')$.
Then there is a Lie algebra morphism
\begin{equation}\label{eq:Liemorphism}
\Psi: {\sf {SF}}^{\sf {ind}}_{\sf{al}}({\mathfrak {Ob}}({\mathcal X})) \to {\sf L}({\mathcal X})
\end{equation}
so that for any stack function of the form $[({\mathfrak X}, \rho)]$, whith
$\rho:{\mathfrak X} \hookrightarrow {\mathfrak {Ob}}({\mathcal X},\gamma)\hookrightarrow {\mathfrak {Ob}}({\mathcal X})$
an open embedding,
and ${\mathfrak X}$ a $\mathbb{C}^\times$-gerbe over an algebraic space
${\sf X}$,
\[
\Psi([({\mathfrak X}, \rho)]) = -\chi^B({\sf X}, \rho^*\nu) \lambda(\gamma)
\]
where $\nu$ is Behrend's constructible function of the stack ${\mathfrak {Ob}}({\mathcal X})$.
In order to define ADHM invariants note that for any $\delta\in \mathbb{R}$, the canonical
open embedding
stack ${\mathfrak{M}}_\delta^{ss}({\mathcal X},\gamma)\hookrightarrow {\mathfrak {Ob}}({\mathcal X})$ determines a stack
function ${\mathfrak d}_\delta(\gamma)\in {\underline {\sf{SF}}}({\mathfrak {Ob}}({\mathcal X}))$.
For $v=0$, the resulting stack functions are independent of stability
parameters and will be denoted by ${\mathfrak h}(\gamma)$.
According to \cite[Thm. 8.7]{J-III} the associated log stack function
\begin{equation}\label{eq:logstfctB}
{\mathfrak e}_\delta(\gamma) = \sum_{l\geq 1} {(-1)^{l-1}\over l}
\mathop{\sum_{\gamma_1+\cdots+\gamma_l=\gamma}}_{\mu_\delta(\gamma_i)=
\mu_\delta(\gamma),\ 1\leq i\leq l}
{\mathfrak d}_\delta(\gamma_1)\ast\cdots\ast {\mathfrak d}_\delta(\gamma_l)
\end{equation}
belongs to ${\sf {SF}}^{\sf {ind}}_{\sf{al}}({\mathfrak {Ob}}({\mathcal X}))$, and is supported in
${\mathfrak {Ob}}({\mathcal X},\gamma)$.
Note that for fixed $\gamma$ and $\delta$ the sum in the right hand side
is finite, therefore there are no convergence issues in the present case.
Then, for $\gamma\in \IZ_{\geq 1}\times\IZ\times
\IZ_{\geq 0}$, the $\delta$-ADHM invariant $A_\delta(\gamma)$ is defined by
\begin{equation}\label{eq:ADHMinv}
\Psi({\mathfrak e}_\delta(\gamma)) = -A_\delta(\gamma) \lambda(\gamma).
\end{equation}
Note that
${\mathfrak e}_\delta(\gamma)$ is independent of $\delta$ if $v=0$.
Then the corresponding
invariants will be denoted by $H(\gamma)$.
By analogy with \cite{genDTI}, define the invariants ${\overline A}_\delta(r,e,v)$
by the multicover formula
\begin{equation}\label{eq:multicoverA}
{\overline A}_\delta(r,e,v) = \mathop{\sum_{m\geq 1}}_{m|r,\ m|e,\ m|v}
{1\over m^2} {\overline A}_\delta(r/m, e/m,v/m).
\end{equation}
Conjecturally, ${\overline A}_\delta(r/m, e/m,v/m)$ are integral. Obviously,
for $v=0$ the alternative notation ${\overline H}(r,e)$ will be used.
\section{Wallcrossing formulas}\label{wallsection}
\subsection{Stack function identities}\label{sfidentities}
Let $\gamma=(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$ and let $\delta_c>0$ be
a critical stability parameter of type $\gamma$. Let $\delta_- <\delta_c$,
$\delta_+>\delta_c$ be stability parameters as in lemma (\ref{HNfiltrations}).
Recall that ${\mathcal {HN}}_\pm(r,v)$ denote the set of positive, respectively negative
admissible configurations of type $(r,v)$ introduced in definition (\ref{admconfig}).
For any $h\in \IZ_{\geq 2}$ let ${\mathcal {HN}}_{\pm}(\gamma,\delta_c,h)$ denote the set of ordered sequences of
triples $\left(\gamma_i =(r_i,e_i,v_i)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0} \right)_{1\leq i\leq h}$ so that $\left(\rho_i=(r_i,v_i)\right)_{1\leq i\leq h}
\in {\mathcal{HN}}_\pm(r,v)$,
\[
e_1+\cdots+e_h =e \qquad \mathrm{and}
\qquad {e_i+v_i\delta_c\over r_i} = {e+v\delta_c\over r} \qquad
\mathrm{for \ all} \ 1\leq i\leq h.
\]
More generally, given $h\in \IZ_{\geq 2}$, for any $0\leq k \leq h-1$ let
${\mathcal {HN}}_{+}(\gamma,\delta_c,h,k)$ denote the set of ordered sequences
$\left(\gamma_i =(r_i,e_i,v_i)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0} \right)_{1\leq i\leq h}$ so that
\begin{itemize}
\item $\gamma_1+\cdots +\gamma_h=\gamma$, $v_{h-k+1}=\cdots = v_h = 0$,
$v_i>0$ for $1\leq i \leq h-k$ , and
$${e_1+ v_1\delta_c \over r_1} = \cdots ={e_{h-k}+v_{h-k}\delta_c
\over r_{h-k}} = {e_{h-k+1}\over r_{h-k+1}} = \cdots = {e_h \over r_h}={e+v\delta_c\over r}$$
\item The sequence $\left(\rho_j= (r_{j}, v_{j})\right)_{1\leq j\leq h-k}$
belongs to ${\mathcal {HN}}_{+}\left(r-\sum_{i=1}^k r_i, v\right)$.
\end{itemize}
Similarly, for any $0\leq k \leq h-1$ let
${\mathcal {HN}}_{-}(\gamma,\delta_c,h,k)$ denote the set of ordered sequences
$\left(\gamma_i =(r_i,e_i,v_i)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0} \right)_{1\leq i\leq h}$ so that
\begin{itemize}
\item $\gamma_1+\cdots +\gamma_h=\gamma$, $v_1=\cdots = v_k = 0$,
$v_i>0$ for $k+1\leq i \leq h$ , and
$${e_1\over r_1} = \cdots ={e_{k}\over r_{k}} = {e_{k+1}+v_{k+1}\delta_{c}\over r_{k+1}} = \cdots = {e_h+v_h\delta_c \over r_h}={e+v\delta_c \over r}$$
\item The sequence $\left(\rho_j= (r_{k+j}, v_{k+j})\right)_{1\leq j\leq h-k}$
belongs to ${\mathcal {HN}}_{-}\left(r-\sum_{i=1}^k r_i, v\right)$.
\end{itemize}
\begin{rema}\label{configremaB}
$(i)$ Obviously, in both cases $v_i>0$ for all $1\leq i\leq h$ if $k=0$.
Moreover, \[{\mathcal{HN}}_\pm(\gamma,\delta_c,h) =
{\mathcal{HN}}_\pm(\gamma,\delta_c,h,0) \cup {\mathcal{HN}}_\pm(\gamma,\delta_c,h,1).
\]
If $k=h-1$ the condition that the sequence $(\rho_j)_{1\leq j \leq h-k}$ belong to
${\mathcal {HN}}_{\pm}\left(r-\sum_{i=1}^k r_i, v\right)$ is empty.
$(ii)$ For fixed $\gamma$ and $\delta_c>0$ it straightforward to check that the
following set is finite
\[
\bigcup_{h\geq 2} \bigcup_{0\leq k \leq h-1} {\mathcal {HN}}_{\pm}(\gamma,\delta_c,h,k),
\]
i.e. the set ${\mathcal {HN}}_{\pm}(\gamma,\delta_c,h,k)$ is nonempty
only for a finite set of pairs $(h,k)$.
\end{rema}
For any triple $\gamma'=(r',e',v')\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$
let ${\mathfrak d}_\pm(\gamma'),{\mathfrak d}_{c}(\gamma')$ be the stack functions determined by the open embeddings
${\mathfrak{M}}_{\delta_\pm}^{ss}({\mathcal X},r',e',v')\hookrightarrow {\mathfrak {Ob}}({\mathcal X})$,
respectively ${\mathfrak{M}}_{\delta_c}^{ss}({\mathcal X},r',e',v')\hookrightarrow {\mathfrak {Ob}}({\mathcal X})$.
The alternative notation ${\mathfrak h}(\gamma')$ will be used if $v'=0$.
\begin{lemm}\label{wallcrossinglemmaA}
The following relations hold in the stack function algebra
${\underline {\sf{SF}}}({\mathfrak {Ob}}({\mathcal X}))$
\begin{equation}\label{eq:wallformulaA}
\begin{aligned}
{\mathfrak d}_c(\gamma) = {\mathfrak d}_{\pm}(\gamma) + \mathop{\sum_{h\geq 2}}_{}\
\mathop{\sum_{(\gamma_i) \in {\mathcal {HN}_\pm(\gamma, \delta_c,h)}}}_{}
{\mathfrak d}_{\pm}(\gamma_1)\ast\cdots \ast{\mathfrak d}_{\pm}(\gamma_h)
\end{aligned}
\end{equation}
\begin{equation}\label{eq:wallformulaB}
\begin{aligned}
& {\mathfrak d}_-(\gamma) + \mathop{\sum_{h\geq 2}} \
\mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,\delta_c,h,0)}}_{}
{\mathfrak d}_{-}(\gamma_1)\ast\cdots \ast{\mathfrak d}_{-}(\gamma_h)
= \\
& {\mathfrak d}_c(\gamma) + \mathop{\sum_{h\geq 2}}(-1)^{h-1}
\mathop{\sum_{(\gamma_i)\in {\mathcal {HN}}(\gamma,\delta_c,h,h-1)}}_{}
{\mathfrak h}(\gamma_1) \ast \cdots \ast{\mathfrak h}(\gamma_{h-1}) \ast {\mathfrak d}_c(\gamma_h)\\
\end{aligned}
\end{equation}
\end{lemm}
{\it Proof.} Equation \eqref{eq:wallformulaA} follows directly from lemma (\ref{HNfiltrations}).
In order to prove formula \eqref{eq:wallformulaB} it will be first proven by induction
that the following
formula holds for any $l\in \IZ_{\geq 1}$.
\begin{equation}\label{eq:recformA}
\begin{aligned}
& {\mathfrak d}_-(\gamma) + \mathop{\sum_{h\geq 2}} \
\mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,\delta_c,h,0)}}_{}
{\mathfrak d}_{-}(\gamma_1)\ast\cdots \ast{\mathfrak d}_{-}(\gamma_h) = \\
& {\mathfrak d}_{c}(\gamma) + \mathop{\sum_{k=2}^l}_{}(-1)^{k-1} \
\mathop{\sum_{(\gamma_i)\in {\mathcal {HN}}_-(\gamma,\delta_c,k,k-1)}}_{}
{\mathfrak h}(\gamma_1) \ast \cdots \ast{\mathfrak h}(\gamma_{k-1}) \ast {\mathfrak d}_c(\gamma_k)\\
& + (-1)^l \mathop{\sum_{h\geq l+1}}_{} \mathop{\sum_{(\gamma_i)\in
{\mathcal{HN}}_-(\gamma, \delta_c, h, l)}}_{} {\mathfrak h}(\gamma_1)\ast
\cdots \ast {\mathfrak h}(\gamma_l) \ast {\mathfrak d}_-(\gamma_{l+1}) \ast \cdots \ast
{\mathfrak d}_-(\gamma_h)
\\
\end{aligned}
\end{equation}
First note that remark (\ref{configremaB}.$ii$) implies that
all sums in equation \eqref{eq:recformA} are finite for any $l\geq 1$.
Next, if $l=1$, equation \eqref{eq:recformA} is equivalent to \eqref{eq:wallformulaA}.
Suppose it holds for some $l\geq 1$.
Then note that equation \eqref{eq:wallformulaA} is valid for any triple $\gamma=(r,e,v)$
and any stability parameter $\delta_c$. If $\delta_c$ is not critical of type $\gamma$
as assumed above, it reduces to a trivial identity.
In particular setting $\gamma= \gamma_{l+1}$ in
equation \eqref{eq:wallformulaA} yields
\[
\begin{aligned}
{\mathfrak d}_{-}(\gamma_{l+1}) =
& {\mathfrak d}_{c}(\gamma_{l+1}) - \mathop{\sum_{m\geq 2}}_{} \
\mathop{\sum_{(\eta_i)\in
{\mathcal {HN}}_-(\gamma_{l+1}, \delta_c, m, 1)}}_{} {\mathfrak h}(\eta_1) \ast
{\mathfrak d}_-({\eta_2})\ast\cdots\ast{\mathfrak d}_-({\eta_m}) \\ & -
\mathop{\sum_{m\geq 2}}_{} \
\mathop{\sum_{(\eta_i)\in
{\mathcal {HN}}_-(\gamma_{l+1}, \delta_c, m, 0)}}_{} {\mathfrak d}_-(\eta_1) \ast
{\mathfrak d}_-({\eta_2})\ast\cdots\ast{\mathfrak d}_-({\eta_m})
\end{aligned}
\]
Using this expression, the third term in the right hand side of equation \eqref{eq:recformA}
can be rewritten as follows.
\begin{equation}\label{eq:recformB}
\begin{aligned}
& (-1)^l \mathop{\sum_{h\geq l+1}}_{} \mathop{\sum_{(\gamma_i)\in
{\mathcal{HN}}_-(\gamma, \delta_c, h, l)}}_{} {\mathfrak h}(\gamma_1)\ast
\cdots \ast {\mathfrak h}(\gamma_l) \ast {\mathfrak d}_-(\gamma_{l+1}) \ast \cdots \ast
{\mathfrak d}_-(\gamma_h) =\\
\end{aligned}
\end{equation}
\[
\begin{aligned}
& (-1)^{l}
\mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma, \delta_c, l+1, l)}}_{}
\bigg[
{\mathfrak h}(\gamma_1)\ast\cdots\ast{\mathfrak h}(\gamma_{l}) \ast{\mathfrak d}_c(\gamma_{l+1}) - \\
& \mathop{\sum_{m\geq 2}}_{}\ \mathop{\sum_{(\eta_i)\in
{\mathcal {HN}}_-(\gamma_{l+1}, \delta_c, m, 1)}}_{}
{\mathfrak h}(\gamma_1)\ast\cdots\ast{\mathfrak h}(\gamma_{l})\ast {\mathfrak h}(\eta_1)\ast
{\mathfrak d}_-({\eta_2})\ast\cdots\ast{\mathfrak d}_-({\eta_m}) \\
& -\mathop{\sum_{m\geq 2}}_{}\ \mathop{\sum_{(\eta_i)\in
{\mathcal {HN}}_-(\gamma_{l+1}, \delta_c, m, 0)}}_{}
{\mathfrak h}(\gamma_1)\ast\cdots\ast{\mathfrak h}(\gamma_{l})\ast {\mathfrak d}_-(\eta_1)\ast
{\mathfrak d}_-({\eta_2})\ast\cdots\ast{\mathfrak d}_-({\eta_m}) \bigg]\\
\end{aligned}
\]
\[
+ (-1)^l \mathop{\sum_{h\geq l+2}}_{} \mathop{\sum_{(\gamma_i)\in
{\mathcal{HN}}_-(\gamma, \delta_c, h, l)}}_{} {\mathfrak h}(\gamma_1)\ast
\cdots \ast {\mathfrak h}(\gamma_l) \ast {\mathfrak d}_-(\gamma_{l+1}) \ast \cdots \ast
{\mathfrak d}_-(\gamma_h)
\]
By construction
\[
\mathop{{\bigcup}_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma, \delta_c, l+1, l)}}_{}
{\mathcal {HN}}_-(\gamma_{l+1},\delta_c, m, j) =
{\mathcal{HN}}_-(\gamma, \delta_c, l+m, l+j)
\]
for any $m\in \IZ_{\geq 2}$, $j\in \{0,1\}$. Therefore the last two terms in the right hand side of equation \eqref{eq:recformB} cancel, and formula \eqref{eq:recformB}
reduces to
\begin{equation}\label{eq:recformC}
\begin{aligned}
& (-1)^l \mathop{\sum_{h\geq l+1}}_{} \mathop{\sum_{(\gamma_i)\in
{\mathcal{HN}}_-(\gamma, \delta_c, h, l)}}_{} {\mathfrak h}(\gamma_1)\ast
\cdots \ast {\mathfrak h}(\gamma_l) \ast {\mathfrak d}_-(\gamma_{l+1}) \ast \cdots \ast
{\mathfrak d}_-(\gamma_h) =\\
& (-1)^{l}
\mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma, \delta_c, l+1, l)}}_{}
{\mathfrak h}(\gamma_1)\ast\cdots\ast{\mathfrak h}(\gamma_{l}) \ast{\mathfrak d}_c(\gamma_{l+1}) - \\
& +(-1)^{l+1} \mathop{\sum_{h\geq l+2}}_{} \mathop{\sum_{(\gamma_i)\in
{\mathcal{HN}}_-(\gamma, \delta_c, h, l+1)}}_{} {\mathfrak h}(\gamma_1)\ast
\cdots \ast {\mathfrak h}(\gamma_{l+1}) \ast {\mathfrak d}_-(\gamma_{l+2}) \ast \cdots \ast
{\mathfrak d}_-(\gamma_h)
\end{aligned}
\end{equation}
Substituting \eqref{eq:recformC} in \eqref{eq:recformA} it follows that formula
\eqref{eq:recformA} also holds if $l$ is replaced by $(l+1)$. This concludes the inductive
proof of formula \eqref{eq:recformA}.
In order to conclude the proof of equation \eqref{eq:wallformulaB}, it suffices to observe
that for sufficiently large $l$, equation \eqref{eq:recformA} stabilizes to equation
\eqref{eq:wallformulaB} using remark (\ref{configremaB}.$ii$).
\hfill $\Box$
Now note that equations \eqref{eq:wallformulaA}, \eqref{eq:wallformulaB} yield a
recursive algorithm expressing ${\mathfrak d}_-(\gamma)$ in terms of
${\mathfrak d}_+(\gamma_i)$, $1\leq i\leq h$, $h\geq 1$. This follows observing
that in the left hand side of \eqref{eq:wallformulaB} $0<v_i<v$ for all
stack functions ${\mathfrak d}_-(\gamma_i)$ occuring in the sum
\[
\mathop{\sum_{h\geq 2}} \
\mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,\delta_c,h,0)}}_{}
{\mathfrak d}_{-}(\gamma_1)\ast\cdots \ast{\mathfrak d}_{-}(\gamma_h).
\]
Therefore, once a formula for the difference ${\mathfrak d}_-(\gamma)-{\mathfrak d}_+(\gamma)$,
has been derived for triples of the form
$\gamma=(r,e,v)$, one can recursively derive an analogous formula for triples of the form
$\gamma=(r,e,v+1)$.
For $v=1$, equations \eqref{eq:wallformulaA}, \eqref{eq:wallformulaB}
easily imply
\begin{equation}\label{eq:wallformulaD}
\begin{aligned}
{\mathfrak d}_-(\gamma) = {\mathfrak d}_+(\gamma) & +\mathop{\sum_{l\geq 2}}_{}
(-1)^{l} \mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,
\delta_c,l,l-1)}}_{} {\mathfrak h}(\gamma_1)\ast\cdots\ast[{\mathfrak d}_+(\gamma_l),{\mathfrak h}(\gamma_{l-1})]\\
\end{aligned}
\end{equation}
Employing the above recursive algorithm one can determine in principle
analogous formulas for $v\geq 2$.
Since the resulting expressions quickly become cumbersome,
explicit formulas will be given below only for $v=2$.
\begin{coro}\label{ranktwocor}
Suppose $\gamma=(r,e,2)$ with $(r,e)\in \IZ_{\geq 1}\times \IZ$.
The following relations hold in the stack function algebra
${\underline {\sf{SF}}}({\mathfrak {Ob}}({\mathcal X}))$
\begin{equation}\label{eq:wallformulaC}
\begin{aligned}
& {\mathfrak d}_-(\gamma) = {\mathfrak d}_+(\gamma) + \mathop{\sum_{l\geq 2}}_{}
(-1)^{l} \mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,
\delta_c,l,l-1)}}_{} {\mathfrak h}(\gamma_1)\ast\cdots\ast[ {\mathfrak d}_+(\gamma_{l}),
{\mathfrak h}(\gamma_{l-1}) ]\\
&
+ \mathop{ \sum_{(\gamma_1,\gamma_2)\in {\mathcal{HN}}_{+}
(\gamma, \delta_c, 2,0)}}_{} {\mathfrak d}_+(\gamma_1)\ast{\mathfrak d}_+(\gamma_2)
-\mathop{\sum_{(\gamma_1,\gamma_2)\in {\mathcal{HN}}_-(\gamma,
\delta_c,2,0)}}_{} {\mathfrak d}_-(\gamma_1) \ast {\mathfrak d}_-(\gamma_2) \\
& + \mathop{\sum_{l\geq 2}}_{}
(-1)^{l} \mathop{\sum_{(\gamma_i)\in {\mathcal{HN}}_-(\gamma,
\delta_c,l+1,l-1)}}_{}
{\mathfrak h}(\gamma_1)\ast\cdots\ast[ {\mathfrak d}_+(\gamma_{l+1})\ast {\mathfrak d}_+(\gamma_{l}),
{\mathfrak h}(\gamma_{l-1}) ]
\end{aligned}
\end{equation}
where ${\mathfrak d}_-(\gamma_1), {\mathfrak d}_-(\gamma_2)$ are given by
equation \eqref{eq:wallformulaD}.
\end{coro}
\subsection{Wallcrossing for $v=2$ invariants}\label{wallformula}
Let $\gamma=(r,e,2)$, $(r,e)\in \IZ_{\geq 1}\times \IZ$,
$\delta_c>0$ a critical stability parameter of type $\gamma$, and
$\delta_\pm$ two noncritical stability parameters
as in lemma (\ref{HNfiltrations}). The main goal of this section is
to convert the stack function relation \eqref{eq:wallformulaC} to a
wallcrossing formula for generalized Donaldson-Thomas invariants
of ADHM sheaves.
As mentioned in the introduction the alternative notation $\alpha =(r,e)$ will be used
for
pairs $(r,e)\in \IZ_{\geq 1}\times \IZ$. Using this notation, the sets
${\mathcal{HN}}_-(\alpha,v, \delta_c, h,k)$, $v\in\{1,2\}$, $k\in\{0, h-2,h-1\}$,
can be identified with sets of ordered sequences
$(\alpha_i)_{1\leq i\leq h}$ satisfying the conditions listed above theorem
(\ref{wallcrossingthmA}). For convenience, recall that ${\mathcal {HN}}_-(\alpha,v,\delta_c,l,l-1)$, $l\in \IZ_{\geq 1}$, $v\in \{1,2\}$, denotes
the set of ordered sequences
$((\alpha_i))_{1\leq i\leq l}$, $\alpha_i\in \IZ_{\geq 1}\times \IZ$,
$1\leq i\leq l$ so that
\begin{equation}\label{eq:alphadecompX}
\alpha_1+\cdots +\alpha_l=\alpha
\end{equation}
and
\begin{equation}\label{eq:slopesXA}
{e_1\over r_1} = \cdots = {e_{l-1}\over r_{l-1}}={e_l+v\delta_c\over r_l}
={e+v\delta_c\over r}
\end{equation}
Similarly, ${\mathcal {HN}}_-(\alpha,v,\delta_c,l,l-2)$,
$l\in \IZ_{\geq 2}$,
denotes the set of ordered sequences
$((\alpha_i))_{1\leq i\leq l}$, $\alpha_i\in \IZ_{\geq 1}\times \IZ$,
$1\leq i\leq l$ satisfying condition \eqref{eq:alphadecomp},
\begin{equation}\label{eq:slopesB}
{e_1\over r_1} = \cdots = {e_{l-2}\over r_{l-2}}={e_{l-1}+\delta_c\over r_{l-1}}
={e_{l}+\delta_c\over r_l} = {e+2\delta_c\over r}
\end{equation}
and $1/{r_{l-1}} < 1/{r_{l}}$.
Note that the sets ${\mathcal{HN}}_-(\alpha,2, \delta_c, h,0)$ are nonempty if and only if
$h=2$, in which case they consist of ordered pairs $(\alpha_1,\alpha_2)$ so that
$\alpha_1+\alpha_2=\alpha$, $1/r_1 < 1/r_2$, and
\[
{e_1+\delta_c\over r_1} ={e_2+\delta_c\over r_2} = {e+2\delta_c\over r}
\]
Moreover the set ${\mathcal{HN}}_-(\alpha,2, \delta_c, 1,0)$ consists of
only of the element $(\alpha)$.
It is straightforward to check that for fixed $\alpha=(r,e)$ and $\delta_c$,
the union
\begin{equation}\label{eq:unionset}
\begin{aligned}
& \qquad\qquad \bigcup_{l\geq 1}
\left[{\mathcal {HN}}_-(\alpha,2,\delta_c,l,l-1) \cup
{\mathcal {HN}}_-(\alpha,2,\delta_c,l+1,l-1)\right]\\
&\bigcup_{(\alpha_1,\alpha_2)\in
{\mathcal {HN}}_-(\alpha,2,\delta_c,2,0) }
\bigcup_{l_1\geq 1}\bigcup_{l_2\geq 1}
\left[{\mathcal {HN}}_-(\alpha_1,1,\delta_c,l_1,l_1-1) \times
{\mathcal {HN}}_-(\alpha_2,1,\delta_c,l_2,l_2-1) \right]
\\
\end{aligned}
\end{equation}
is a finite set.
Now let $0<\delta_-<\delta_c<\delta_+$ be stability parameters so that there are no
critical stability parameters of type $(\alpha,2)$ in the intervals $[\delta_-,\ \delta_c)$,
$(\delta_c,\ \delta_+]$. Since the set \eqref{eq:unionset} is finite $\delta_-, \delta_+$
can be chosen so that the same holds for all numerical types $(\alpha_i, v_i)$
in all ordered sequences in \eqref{eq:unionset}.
Then the following lemma holds.
\begin{lemm}\label{wallcrossinglemmaB}
The following relations hold in the stack function algebra ${\underline{\sf {SF}}}({\mathfrak {Ob}}({\mathcal X}))$
\begin{equation}\label{eq:wallformulaE}
\begin{aligned}
& {\mathfrak d}_-(\alpha,1)= \mathop{\sum_{l\geq 1}}_{} {(-1)^{l-1}\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,1, \delta_c, l,l-1)}}_{}
[{\mathfrak g}(\alpha_1),[\cdots[{\mathfrak g}(\alpha_{l-1}),{\mathfrak d}_+(\alpha_l,1)]\cdots]\\
\end{aligned}
\end{equation}
\begin{equation}\label{eq:wallformulaF}
\begin{aligned}
&{\mathfrak d}_-(\alpha,2) = \mathop{\sum_{l\geq 1}}_{} {(-1)^{l-1}\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l,l-1)}}_{}
[{\mathfrak g}(\alpha_1),[\cdots[{\mathfrak g}(\alpha_{l-1}),{\mathfrak d}_+(\alpha_l,2)]\cdots]\\
& +\mathop{\sum_{l\geq 1}}_{} {(-1)^{l-1}\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l+1,l-1)}}_{}
[{\mathfrak g}(\alpha_1),[\cdots[{\mathfrak g}(\alpha_{l-1}),{\mathfrak d}_+(\alpha_{l+1},1)\ast
{\mathfrak d}_+(\alpha_{l},1)]\cdots]\\
& -\mathop{\sum_{(\alpha_1,\alpha_2)\in
{\mathcal{HN}}_-(\alpha,2,\delta_c,2,0)}}_{} \mathop{\sum_{l_1\geq 1}}_{}
\mathop{\sum_{l_2\geq 1}}_{} {(-1)^{l_1-1}\over (l_1-1)!}{(-1)^{l_2-1}\over (l_2-1)!}
\mathop{\sum_{(\alpha_{1,i})\in {\mathcal{HN}}_-(\alpha_1,1, \delta_c, l_1,l_1-1)}}_{}\\
&
\mathop{\sum_{(\alpha_{2,i})\in {\mathcal{HN}}_-(\alpha_2,1, \delta_c, l_2,l_2-1)}}_{}
\big(
[{\mathfrak g}(\alpha_{1,1}),[\cdots[{\mathfrak g}(\alpha_{1,l_1-1}),{\mathfrak d}_+(\alpha_{1,l_1},1)]\cdots]\\
& \qquad \qquad\qquad \qquad \qquad \
\ast[{\mathfrak g}(\alpha_{2,1}),[\cdots[{\mathfrak g}(\alpha_{2,l_2-1}),{\mathfrak d}_+(\alpha_{2,l_2},1)]\cdots]
\big).\\
\end{aligned}
\end{equation}
\end{lemm}
{\it Proof.} Formulas \eqref{eq:wallformulaE},
\eqref{eq:wallformulaF} follow from equations \eqref{eq:wallformulaC},
\eqref{eq:wallformulaD} by repeating the computations
in the proof of \cite[Lemm. 2.6]{chamberI} in the present context.
\hfill $\Box$
{\it Proof of Theorem (\ref{wallcrossingthmA}.)}
The proof consists of two steps. First the stack function identities \eqref{eq:wallformulaE},
\eqref{eq:wallformulaF} must be converted into similar identities for the log
stack functions \eqref{eq:logstfctB}. As explained in \cite[Sect. 6.5]{J-IV},
\cite[Sect. 3.5]{genDTI},
applying the morphism
\eqref{eq:Liemorphism} to the log stack function identities \eqref{eq:wallformulaE},
\eqref{eq:wallformulaF} yields certain relations
in the universal enveloping algebra $U({\sf L}({\mathcal X}))$ of the Lie algebra ${\sf L}({\mathcal X})$.
These relations imply in turn a wallcrossing formula for generalized Donaldson-Thomas
invariants by identifying the coefficients of generators of the generators of
${\sf L}({\mathcal X})\subset U({\sf L}({\mathcal X}))$.
Given the above choice of $\delta_\pm$, for
$v=1$, equation \eqref{eq:logstfctB} reduces to ${\mathfrak e}_\pm(\gamma)=
{\mathfrak d}_\pm(\gamma)$, while for $v=2$
\begin{equation}\label{eq:logstfctA}
\begin{aligned}
{\mathfrak e}_\pm(\gamma) = {\mathfrak d}_\pm(\gamma) -{1\over 2} {\mathfrak d}_\pm(\gamma/2)\ast {\mathfrak d}_\pm(\gamma/2).
\end{aligned}
\end{equation}
The second term in the right hand side of \eqref{eq:logstfctA}
is by convention trivial unless $(r,e)$ are even.
Equations \eqref{eq:logstfctA}, \eqref{eq:wallformulaF}, \eqref{eq:wallformulaG}
yield the following identity in the universal enveloping algebra of the Lie
algebra ${\sf L}({\mathcal X})$
\begin{equation}\label{eq:univid}
\begin{aligned}
& \mathop{\sum_{\alpha}}(A_{-}(\alpha,2)-A_{+}(\alpha,2))\lambda(\alpha,2) = \\
&\mathop{\sum_{\alpha}}\mathop{\sum_{l\geq 2}}_{} {1\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l,l-1)}}_{}
\left(A_+(\alpha_l,2)\prod_{i=1}^{l-1}f_2(\alpha_i)H(\alpha_i)\right) \lambda(\alpha,2)\\
\end{aligned}
\end{equation}
\[
\begin{aligned}
&-\mathop{\sum_{\alpha}}\mathop{\sum_{l\geq 1}}_{} {1\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l+1,l-1)}}_{}
\left(A_+(\alpha_l,1)
A_+(\alpha_{l+1},1)
\prod_{i=1}^{l-1}H(\alpha_i)\right)\\
& \qquad \qquad \qquad \qquad\qquad \qquad \quad\
[\lambda(\alpha_1),[\cdots[\lambda(\alpha_{l-1}),
\lambda(\alpha_{l+1},1)\star\lambda(\alpha_{l},1)]\cdots ]\\
& +\mathop{\sum_{\alpha}}\mathop{\sum_{(\alpha_1,\alpha_2)\in
{\mathcal{HN}}_-(\alpha,2,\delta_c,2,0)}}_{} \mathop{\sum_{l_1\geq 1}}_{}
\mathop{\sum_{l_2\geq 1}}_{} {1\over (l_1-1)!}{1\over (l_2-1)!}
\mathop{\sum_{(\alpha_{1,i})\in {\mathcal{HN}}_-(\alpha_1,1, \delta_c, l_1,l_1-1)}}_{}\\
& \mathop{\sum_{(\alpha_{2,i})\in {\mathcal{HN}}_-(\alpha_2,1, \delta_c, l_2,l_2-1)}}_{}
A_+(\alpha_{1,l_1})A_+(\alpha_{2,l_2})
\prod_{i=1}^{l_1-1} f_1(\alpha_{1,i})H(\alpha_{1,i})
\prod_{i=1}^{l_2-1} f_1(\alpha_{2,i})H(\alpha_{2,i}) \\
& \qquad\qquad\qquad\qquad\qquad\quad
\lambda(\alpha_1,1)\star\lambda(\alpha_2,1)\\
&+{1\over 2}\mathop{\sum_{\alpha}} (A_{-}(\alpha/2,1)^2-A_{+}(\alpha/2,1)^2)
\lambda(\alpha/2,1)\star \lambda(\alpha/2,1) \\
&-\mathop{\sum_{\alpha}}\mathop{\sum_{l\geq 1}}_{} {1\over (l-1)!}
\mathop{\sum_{(\alpha_i)\in {\mathcal{HN}}_-(\alpha,2, \delta_c, l,l-1)}}_{}
\left(A_+(\alpha_l/2,1)^2
\prod_{i=1}^{l-1}H(\alpha_i)\right) \\
& \qquad\qquad\qquad\qquad
[\lambda(\alpha_1),[\cdots,[\lambda(\alpha_{l-1}, \lambda(\alpha_l/2,1)\star
\lambda(\alpha_l/2,1)]\cdots]\\
\end{aligned}
\]
where $\star$ denotes the associative product in the universal enveloping algebra.
By conventions the invariants of the form $A_+(\alpha/2,1)$ are trivial unless
$\alpha=2\alpha'$ for some $\alpha'=(r',e')=\IZ_{\geq 1} \times \IZ$.
Next, the
identity \cite[Eqn. 127]{J-IV} or \cite[Eqn. 45]{genDTI} yields the following relations in the universal enveloping
algebra
\[
\begin{aligned}
\lambda(\alpha_{l+1},1)\star\lambda(\alpha_{l},1) & = {1\over 2}
g(\alpha_{l+1},\alpha_{l})
\lambda(\alpha_l+\alpha_{l+1},2) + \cdots \\
\lambda(\alpha_1,1)\star\lambda(\alpha_{2},1) & = {1\over 2}g(\alpha_1,\alpha_{2})
\lambda(\alpha_1+\alpha_2,2) + \cdots \\
\lambda(\alpha/2,1)\star\lambda(\alpha/2,1) & = \cdots\\
\lambda(\alpha_l/2,1)\star\lambda(\alpha_l/2,1) & = \cdots \\
\end{aligned}
\]
where $\cdots$ stands for linear combinations of generators of $U({\sf L}({\mathcal X}))$
not in ${\sf L}({\mathcal X})$.
Since the left hand side of equation \eqref{eq:univid} must belong to the Lie algebra
${\sf L}({\mathcal X})$ according to \cite[Thm. 8.7]{J-III}, it follows that all higher order terms
must cancel. Then equation \eqref{eq:wallformulaG} follows by
straightforward computations.
\hfill $\Box$
\section{Comparison with Kontsevich-Soibelman Formula}\label{KSsect}
The goal of this section is to prove that
formula \eqref{eq:wallformulaG} is in agreement with the wallcrossing
formula of
Kontsevich and Soibelman \cite{wallcrossing}, which will be referred to as
the KS formula in the following.
As in section (\ref{wallformula}), numerical types of ADHM sheaves
will be denoted by $\gamma=(\alpha,v)$, $\alpha=
(r,e)\in \IZ_{\geq 1}\times \IZ$, $v\in \IZ_{\geq 0}$.
In order to streamline the computations, let
${\sf L}({\mathcal X})_{\leq 2}$ denote the truncation of the Lie algebra ${\sf L}({\mathcal X})$
defined by
\begin{equation}\label{eq:chargeliebracket}
\begin{aligned}
& [\lambda(\alpha_1,v_1), \lambda(\alpha_2,v_2)]_{\leq 2} =
\left\{\begin{array}{ll}
[\lambda(\alpha_1,v_1), \lambda(\alpha_2,v_2)] & \quad \mathrm{if}\ v_1+v_2\leq 2\\
0 & \quad \mathrm{otherwise}.\\
\end{array}\right. \\
\end{aligned}
\end{equation}
Furthermore, it will be more convenient to use the alternative notation
${\sf e}_\alpha=\lambda(\alpha,0)$, ${\sf f}_\alpha=\lambda(\alpha,1)$,
and ${\sf g}_{\alpha}=\lambda(\alpha,2)$.
Given a
critical stability parameter $\delta_c$ of type $(r,e,2)$, $(r,e)\in \IZ_{\geq 1}\times \IZ$,
there exist two pairs
$\alpha =(r_{\alpha}, e_{\alpha})$ and $\beta =(r_{\beta}, e_{\beta})$
with
\[
\frac{e_{\alpha} + \delta_c}{r_{\alpha}}=\frac{e_{\beta}}{r_{\beta}}= \mu_{\delta_c}(\gamma)
\]
so that any $\eta\in \IZ_{\geq 1}\times \IZ$
with $\mu_{\delta_c}(\eta) = \mu_{\delta_c}(\gamma)$ can
be uniquely written as $\eta = (q\beta,0), (\alpha+ q\beta,1),$ or
$(2\alpha+ q\beta,2)$, with $q\in \IZ_{\geq 0}$.
For any $q\in \IZ_{\geq 0}$ the following formal expressions will
be needed in the KS formula,
\begin{equation}
U_{\alpha+q\beta} = \mathrm{exp}({\sf f}_{\alpha+q\beta} + \frac{1}{4} {\sf g}_{2\alpha+2q\beta}) \ \ , \ \
U_{2\alpha+q\beta} = \mathrm{exp}({\sf g}_{2\alpha+q\beta}) \ \ , \ \
U_{q\beta} = \text{exp} (\sum_{m\geq1} \frac{{\sf e}_{mq\beta}}{m^2})\ .
\end{equation}
Moreover, let
\[
\IH = \sum_{q\geq 0} { H}(q\beta){{\sf e}_{q\beta}},
\]
where the invariants $H(\alpha)$ are defined in \eqref{eq:ADHMinv}.
Then the wallcrossing formula of Kontsevich and Soibelman
reads
\begin{equation}\label{eq:KSformA}
\begin{aligned}
& \mathrm{exp(\IH)}\prod_{{q\geq 0},\ q\downarrow} U_{2\alpha + q\beta}^{\overline{A}_{+}(2\alpha+q\beta,2)}
\prod_{{q\geq 0},\ q\downarrow} U_{\alpha+q\beta}^{A_{+}(\alpha+q\beta,1)} \\
&= \prod_{q\geq 0,\ q\uparrow} U_{\alpha+ q\beta}^{A_{-}(\alpha+q\beta,1)}
\prod_{{q\geq 0},\ q\uparrow} U_{2\alpha + q\beta}^{\overline{A}_{-}(2\alpha+q\beta,2)}
\mathrm{exp(\IH)}\end{aligned}
\end{equation}
where an up, respectively down arrow means that the factors in the corresponding
product are taken in increasing, respectively decreasing order of $q$. Note that
$\overline{A}_{\pm}(2\alpha+q\beta,2)$ are the
invariants defined in section (\ref{invariants})
by the multicover formula \eqref{eq:multicoverA}. In this case equation
\eqref{eq:multicoverA} reduces to
\[
A_{\pm}(2\alpha+q\beta,2) = \overline{A}_{\pm}(2\alpha+q\beta,2) + \frac{1}{4} A_{\pm}(\alpha+q\beta/2,1).
\]
Expanding the right hand side, equation
\eqref{eq:KSformA} yields
\begin{equation}\label{eq:KSformB}
\begin{aligned}
& \mathrm{exp}(\sum_{q\geq0} A_{-}(2\alpha+q\beta,2) {\sf g}_{2\alpha+q\beta} + \\
& \sum_{q_2>q_1\geq0}
\frac{1}{2} g(q_1\beta,q_2\beta)A_{-}(\alpha+q_1\beta,1)A_{-}(\alpha+q_2\beta,1){\sf g}_{2\alpha+(q_1+q_2)\beta} )= \\
& \mathrm{exp(\IH)}\, \mathrm{exp}(\sum_{q\geq0} A_{+}(2\alpha+q\beta,2) {\sf g}_{2\alpha+q\beta} \\
& + \sum_{q_1>q_2\geq0} \frac{1}{2} g(q_1\beta,q_2\beta)A_{+}(\alpha+q_1\beta,1)A_{+}(\alpha+q_2\beta,1) {\sf g}_{2\alpha+(q_1+q_2)\beta})\, \mathrm{exp(-\IH)},
\end{aligned}
\end{equation}
modulo terms involving ${\sf f}_{\gamma}$. These terms are omitted since they enter
$v=1$ wallcrossing formula derived in
\cite{chamberII}.
The BCH formula
\begin{equation}
\begin{aligned}
\text{exp}(A) \text{exp}(B) \text{exp} (-A) & = \text{exp}( \sum_{n=0} \frac{1}{n!} (Ad(A))^n B )\\ &
= \text{exp} ( B + [A,B] +\frac{1}{2} [A,[A,B]]+ \cdots),\\
\end{aligned}
\end{equation}
yields
\begin{equation}\label{eq:BCHg}
\begin{aligned}
& \mathrm{exp}(\IH)\, \mathrm{exp}({\sf g}_{2\alpha+q\beta})\, \mathrm{exp}(-\IH) =
\mathrm{exp}( {\sf g}_{2\alpha+q\beta} +
\sum_{q_1>0} f_2(q_1\beta)H(q_1\beta) {\sf g}_{2\alpha+(q+q_1)\beta} \\
& + \frac{1}{2 !}\sum_{q_1>0,q_2>0} f_2(q_1\beta)H(q_1\beta) f_2(q_2\beta)H(q_2 \beta) {\sf g}_{2\alpha+(q+q_1+q_2)\beta} +\cdots ) \\
& = \mathrm{exp} \Big( \sum_{l\geq0, q_i>0} \frac{1}{l !} (\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) )
{\sf g}_{2\alpha+(q+q_1+\cdots+q_l)\beta} \Big)
\end{aligned}
\end{equation}
Substituting \eqref{eq:BCHg} in \eqref{eq:KSformB}
results in
\begin{equation}\label{KSformC}
\begin{aligned}
&\mathrm{exp}\big( \sum_{q\geq0} A_{-}(2\alpha+q\beta,2){\sf g}_{2\alpha+q\beta}
+\sum_{q_2>q_1\geq0} \frac{1}{2} g(q_1\beta, q_2\beta) A_{-}(\alpha+q_1\beta,1)
A_{-}(\alpha+q_2\beta,1) {\sf g}_{2\alpha+(q_1+q_2)\beta} \Big) \\
&= \mathrm{exp}\big( \sum_{ \substack{q\geq0, l\geq 0 \\ q_i>0}} A_{+}(2\alpha+q\beta,2)
\frac{1}{l !} (\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) ) {\sf g}_{2\alpha+(q+q_1+\cdots+q_l)\beta} \\
& + \sum_{\substack{q_1' > q_2'\geq 0\\ l\geq0, q_i>0}} \frac{1}{2} g(q_1'\beta, q_2'\beta)
A_{+}(\alpha+q_1'\beta,1) A_{+}(\alpha+q_2'\beta,1) \frac{1}{l !}(\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) )
{\sf g}_{2\alpha+(q_1'+q_2'+q_1+\cdots+q_l)\beta} \Big)
\end{aligned}
\end{equation}
In order to further simplify the notation, let
\[
A_{\pm} (v\alpha+q\beta,v) \equiv A_{\pm}(q,v), \qquad {\sf g}_{2\alpha+q\beta}
\equiv {\sf g}_q .
\]
Comparing the coefficients of ${\sf g}_Q$ in \eqref{eq:KSformB},
yields
\begin{equation}\label{eq:KSformC}
\begin{aligned}
& A_{-}(Q,2) = \sum_{\substack{q'\geq0 ,\ l \geq0,\ q_i>0 \\ q'+q_1+\cdots+q_l=Q}} A_{+}(q',2)
\frac{1}{l !} (\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) ) \\
& +\frac{1}{2}\sum_{\substack{q_1' > q_2'\geq 0\\ l\geq0,\ q_i>0 \\ q_1'+q_2'+ q_1+\cdots + q_l=Q}}
g(q_1'\beta, q_2'\beta) A_{+}(q_1',1) A_{+}(q_2',1)\frac{1}{l !}(\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) ) \\
& - \frac{1}{2} \sum_{ q_2' > q_1'\geq 0,\ q_1'+q_2'=Q}\ g(q_1'\beta, q_2'\beta) A_{-}(q_1',1)A_{-}(q_2',1) \ .
\end{aligned}
\end{equation}
Using the $v=1$ wallcrossing formula \cite[Thm. 1.1]{chamberII} the last term in
\eqref{eq:KSformC} becomes
\begin{equation}\label{eq:lastterm}
\begin{aligned}
& - \frac{1}{2} \sum_{ q_2 > q_1\geq 0,\ q_1+q_2=Q}
\ g(q_1\beta, q_2\beta) A_{-}(q_1,1)A_{-}(q_2,1) \\
& = - \frac{1}{2} \sum_{ \substack{ q_2> q_1\geq 0\\ q_1+q_2=Q\\ l \geq 0,\ \tilde{l} \geq 0\\q_1' \geq 0,\ q_2' \geq 0\\ n_i>0,\ \tilde{n}_i >0 \\ q_1'+n_1+\cdots+n_l = q_1 \\
q_2'+\tilde{n}_1+\cdots+\tilde{n}_{\tilde{l}} = q_2 }} g(q_1\beta, q_2\beta) A_{+}(q_1',1)A_{+}(q_2',1)
\frac{1}{l !} (\prod_{i=1}^{l} f_1(n_i \beta) H(n_i \beta) ) \frac{1}{\tilde l !}
(\prod_{i=1}^{\tilde l} f_1(\tilde{n}_i \beta) H(\tilde{n}_i \beta) ) \ .
\end{aligned}
\end{equation}
Therefore the final wallcrossing formula for $v=2$ invariants is
\begin{equation}\label{eq:KSformD}
\begin{aligned}
& A_{-}(Q,2) = \sum_{\substack{q'\geq0,\ l \geq0,\ q_i>0 \\ q'+q_1+\cdots+q_l=Q}} A_{+}(q',2)
\frac{1}{l !} (\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) ) \\
& +\frac{1}{2}\sum_{\substack{q_1' > q_2'\geq 0\\ l\geq0,\ q_i>0 \\ q_1'+q_2'+ q_1+\cdots + q_l=Q}}
\frac{1}{2} g(q_1'\beta, q_2'\beta) A_{+}(q_1',1) A_{+}(q_2',1)\frac{1}{l !} (\prod_{i=1}^{l} f_2(q_i \beta) H(q_i \beta) ) \\
& -\frac{1}{2} \sum_{ \substack{ q_2 > q_1\geq 0\\ q_1+q_2=Q\\ l \geq 0,\ \tilde{l} \geq 0\\q_1' \geq 0,\ q_2' \geq 0\\ n_i>0, \tilde{n}_i >0 \\ q_1'+n_1+\cdots+n_l = q_1 \\
q_2'+\tilde{n}_1+\cdots+\tilde{n}_{\tilde{l}} = q_2 }} g(q_1\beta, q_2\beta) A_{+}(q_1',1)A_{+}(q_2',1)
\frac{1}{l !} (\prod_{i=1}^{l} f_1(n_i \beta) H(n_i \beta) ) \frac{1}{\tilde l !}
(\prod_{i=1}^{\tilde l} f_1(\tilde{n}_i \beta) H(\tilde{n}_i \beta) ) \ .
\end{aligned}
\end{equation}
This formula agrees with \eqref{eq:wallformulaG} since the bilinear function
$g({\quad},{\quad})$ is antisymmetric.
\section{Asymptotic invariants in the $g=0$ theory}\label{genuszero}
In this subsection $X$ will be a smooth genus 0 curve over a $\mathbb{C}$-field $K$,
and $M_1\simeq {\mathcal O}_X(d_1)$, $M_2\simeq {\mathcal O}_X(d_2)$, with
$(d_1,d_2)=(1,1)$ or $(d_1,d_2)=(0,2)$. In this case any coherent
locally free
sheaf $E$ on $X$ is isomorphic to a direct sum of line bundles.
Let $E_{\geq 0}$ denote the direct sum of all summands of non-negative
degree, and $E_{<0}$ denote the direct sum of all summands of
negative degree.
\begin{lemm}\label{zerophi}
Let ${\mathcal E}=(E,V,\Phi_1,\Phi_2,\phi,\psi)$ be a nontrivial $\delta$-semistable
ADHM sheaf of type
$(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$, for some $\delta> 0$.
Then $E_{<0}=0$
and $\phi$ is identically zero.
\end{lemm}
{\it Proof.}
Since $\delta>0$, lemma (\ref{basicprop}.$ii$) implies that $\psi$
is not identically zero.
Then obviously $E_{\geq 0}$ must be nontrivial and $\mathrm{Im}(\psi)
\subseteq E_{\geq 0}$. Since $M\simeq K_X^{-1}\simeq {\mathcal O}_X(2)$,
$E_{\geq 0}\otimes_X M\subseteq \mathrm{Ker}(\phi)$.
Moreover, since $\mathrm{deg}(M_1)\geq 0$, $\mathrm{deg}(M_2)\geq 0$,
$\Phi_i(E_{\geq 0}\otimes_X M_i) \subseteq E_{\geq 0}$.
It follows that the data
\[
{\mathcal E}_{\geq 0}=(E_{\geq 0}, V\otimes {\mathcal O}_X,\Phi_i|_{E_{\geq 0}\otimes_X M_i}, 0, \psi)
\]
is a nontrivial subobject of ${\mathcal E}$.
If $E_{<0}$ is not the zero sheaf, ${\mathcal E}_{\geq 0}$ is a proper subobject of
${\mathcal E}$. Then $\delta$-semistability condition implies
$r({\mathcal E}_{\geq 0}) < r({\mathcal E})$, hence
\begin{equation}\label{eq:posdegslope}
{d({\mathcal E}_{\geq 0})+ v({\mathcal E}_{\geq 0})\, \delta \over r({\mathcal E}_{\geq 0}) }
\leq {e+v\, \delta \over r}.
\end{equation}
However $e< d({\mathcal E}_{\geq 0})$ and $0<r({\mathcal E}_{\geq 0})<r$ under the current assumptions.
Since also $v({\mathcal E}_{\geq 0})=v$ and $\delta, d({\mathcal E}_{\geq 0})>0$, inequality
\eqref{eq:posdegslope} leads to a contradiction.
Therefore $E_{<0}=0$ and $\phi$ must be identically zero.
\hfill $\Box$
Let ${\mathcal C}^0_{\mathcal X}$ be the full abelian subcategory of ${\mathcal C}_{\mathcal X}$
consisting of ADHM sheaves ${\mathcal E}$ with $\phi=0$. For any
$\delta\in \mathbb{R}$, an object ${\mathcal E}$ of ${\mathcal C}^0_{\mathcal X}$ will be called
$\delta$-semistable
if it is $\delta$-semistable as an object of ${\mathcal C}_{\mathcal X}$.
Note that given
an object ${\mathcal E}$ of ${\mathcal C}^0_{\mathcal X}$, any subobject ${\mathcal E}'\subset {\mathcal E}$
must also belong to ${\mathcal C}^0_{\mathcal X}$. In particular all test subobjects in
definition (\ref{udeltastability}) also belong to ${\mathcal C}^0_{\mathcal X}$, and
one obtains a stability condition on the abelian category
${\mathcal C}^0_{\mathcal X}$. Then the properties
of $\delta$-stability and moduli stacks of semistable objects
in ${\mathcal C}^0_{\mathcal X}$ are analogous to those of
${\mathcal C}_{\mathcal X}$. In particular for fixed
$(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$ there are finitely many
critical stability parameters of type $(r,e,v)$ dividing the real axis into
stability chambers. The main difference between ${\mathcal C}^0_{\mathcal X}$ and
${\mathcal C}_{\mathcal X}$ is the presence of an empty chamber, as follows.
\begin{lemm}\label{emptychamber}
For any $(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 1}$ the moduli
stack of $\delta$-semistable objects of ${\mathcal C}^0_{\mathcal X}$
of type $(r,e,v)$ is empty if $\delta<0$.
\end{lemm}
{\it Proof.} Given an ADHM sheaf ${\mathcal E}=(E,V,\Phi_i,\psi)$ of
type $(r,e,v)$, it is straightforward to check that for $\delta<0$ the proper nontrivial
object $(E,0,\Phi_i,0)$ is always destabilizing if $\delta<0$.
\hfill $\Box$
\begin{lemm}\label{posdeg}
Let ${\mathcal E}$ be a $\delta$-semistable object of ${\mathcal C}^0_{\mathcal X}$ of type
$(r,e,v)\in \IZ_{\geq 1}\times \IZ\times \IZ_{\geq 0}$ for some
$\delta\geq 0$.
If $e\geq 0$, then $E_{<0}=0$ and $\phi$ is identically zero.
\end{lemm}
{\it Proof.} For $\delta>0$ and $v>0$, this obviously follows from lemma (\ref{zerophi}).
If $\delta=0$ or $v=0$ note that $E_{\geq 0}$ cannot be the zero sheaf
since $e\geq 0$. Then the proof of lemma (\ref{zerophi}) also applies to this
case as well.
\hfill $\Box$
\begin{lemm}\label{genzeroHiggs}
Let ${\mathcal E}=(E,0,\Phi_i,0,0)$ be a semistable object of ${\mathcal C}^0_{\mathcal X}$
of type $(r,e,0)$, $(r,e)\in \IZ_{\geq 1}\times \IZ$.
If $(d_1,d_2)=(1,1)$, $E$ must be isomorphic to
${\mathcal O}_X(n)^{\oplus r}$ for some $n\in \IZ$, and $\Phi_i=0$ for $i=1,2$.
If $(d_1,d_2)=(0,2)$, $E$ must be isomorphic to ${\mathcal O}_X(n)^{\oplus r}$ for some
$n\in \IZ$, and $\Phi_2=0$.
\end{lemm}
{\it Proof.} In both cases, let $E\simeq \oplus_{s=1}^r {\mathcal O}_X(n_s)$
for some $n_s\in \IZ$ so that $n_1\leq n_2\leq \cdots \leq n_r$.
Since $d_1,d_2\geq 0$, any subsheaf of the form
\[
\oplus_{s=s_0}^r {\mathcal O}_X(n_s)
\]
for some $1\leq s_0\leq r$ must be $\Phi_i$-invariant,
$i=1,2$.
Therefore the semistability condition implies
\[
{n_{s_0} +\cdots +n_r\over r-s_0+1} \leq {n_1+\ldots + n_r\over r}
\]
for any $1\leq s_0\leq r$. Then it is straightforward to check that
$n_1=\cdots = n_r =n$. The rest is obvious.
\hfill $\Box$.
\begin{coro}\label{genzeroHiggsinv}
Under the same conditions as in lemma (\ref{genzeroHiggs}),
\begin{equation}\label{eq:Higgsinv}
H(r,e) = \left\{\begin{array}{ll}
{(-1)^{d_1-1}\over r^2} & \mathrm{if}\ e=rn, \ n\in \IZ\\
& \\
0 & {\mathrm{otherwise}}. \\ \end{array}\right.
\end{equation}
\end{coro}
{\it Proof.} If $(d_1,d_2)=(1,1)$, lemma (\ref{genzeroHiggs})
implies that the moduli stack
${\mathfrak{M}}^{ss}({\mathcal X},r,e,0)$ is isomorphic to the quotient stack
$[*/GL(r)]$ if $e=rn$ for some $n\in \IZ$, and empty otherwise.
Alternatively, if $e=rn$, the moduli stack ${\mathfrak{M}}^{ss}({\mathcal X},r,e,0)$
can be identified with the moduli stack of trivially semistable
representations of dimension $r$ of a quiver consisting of only one
vertex and no arrows. Recall that the trivial semistability condition
for quiver representations is King stability with all stability
parameters associated to the vertices set to zero \cite[Ex. 7.3]{genDTI}.
If $(d_1,d_2)=(0,2)$, lemma (\ref{genzeroHiggs})
implies that the moduli stack
${\mathfrak{M}}^{ss}({\mathcal X},r,rn,0)$, $n\in \IZ$,
is isomorphic to the moduli stack
of trivially semistable representations of dimension $r$ of a quiver
consisting of one vertex and one arrow joining the unique
vertex with itself. If $e$ is not a multiple of $r$, the moduli stack ${\mathfrak{M}}^{ss}({\mathcal X},r,e,0)$
is empty.
Then
corollary (\ref{genzeroHiggsinv}) follows
by a computation very similar to \cite[Sect. 7.5.1]{genDTI}.
\hfill $\Box$
\begin{rema}\label{reqzerorem}
The same arguments as in the proof of corollary
(\ref{genzeroHiggsinv}) imply that for any $\delta>0$,
\begin{equation}\label{eq:purevinv}
A_\delta(0,0,1)=1\qquad A_\delta(0,0,2)={1\over 4}.
\end{equation}
\end{rema}
Extension groups in ${\mathcal C}^0_{\mathcal X}$ can be determined by analogy with
those of ${\mathcal C}_{\mathcal X}$. Given two locally free objects ${\mathcal E}'',{\mathcal E}'$ of
${\mathcal C}^0_{\mathcal X}$, let ${\widetilde {\mathcal C}}({\mathcal E}'',{\mathcal E}')$ be
the three term complex of locally free ${\mathcal O}_X$-modules
\begin{equation}\label{eq:zerocomplex}
\begin{aligned}
0 \to \begin{array}{c} {\mathcal Hom}_{X}(E'',E') \\ \oplus \\
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,V'\otimes {\mathcal O}_X) \\ \end{array}
& {\buildrel d_1\over \longrightarrow}
\begin{array}{c} {\mathcal Hom}_{X}(E''\otimes_{X}M_1 ,E') \\ \oplus \\
{\mathcal Hom}_{X}(E''\otimes_{X} M_2, E') \\ \oplus \\
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,E') \\ \end{array}
{\buildrel d_2\over \longrightarrow}
{\mathcal Hom}_{X}(E''\otimes_{X}M,E') \to 0 \\
\end{aligned}
\end{equation}
where
\[
\begin{aligned}
d_1(\alpha,f) =
(& -\alpha \circ \Phi_{1}'' +\Phi_{1}'\circ (\alpha\otimes 1_{M_1}),
-\alpha \circ \Phi_{2}''+\Phi_{2}'\circ (\alpha\otimes 1_{M_2}),\\
&
-\alpha\circ \psi'' +\psi' \circ f)\\
\end{aligned}
\]
for any local sections $(\alpha,f)$ of
the first term
and
\[
\begin{aligned}
d_2(\beta_1,\beta_2,\gamma) = &
\beta_1 \circ (\Phi''_2\otimes 1_{M_1}) -
\Phi_{2}'\circ (\beta_1\otimes 1_{M_2})
- \beta_2\circ (\Phi''_{1}\otimes 1_{M_2})\\
& + \Phi_{1}'\circ (\beta_2\otimes 1_{M_1}) \\
\end{aligned}
\]
for any local sections $(\beta_1,\beta_2,\gamma)$
of the middle term. The degrees of the three terms in
\eqref{eq:hypercohA} are $0,1,2$ respectively.
By analogy with lemma(\ref{extlemma}), the following holds.
\begin{lemm}\label{zeroexts}
Under the current assumptions,
$\mathrm{Ext}^k_{{\mathcal C}^0_{\mathcal X}}({\mathcal E}'',{\mathcal E}') \simeq
\IH^k({\widetilde {\mathcal C}}({\mathcal E}'',{\mathcal E}'))$ for $k=0,1$.
\end{lemm}
\begin{lemm}\label{zeroeuler}
Let ${\mathcal E}', {\mathcal E}''$ be two nontrivial locally free objects of ${\mathcal C}_{\mathcal X}^0$
of types $(r',e',v'), (r'',e'',v'')\in \IZ_{\geq 1}\times \IZ \times \IZ_{\geq 0}$.
Suppose that $E'_{<0}=0$, $E''_{<0}=0$
for both underlying locally free sheaves $E',E''$. Then
\begin{equation}\label{eq:zeroeuler}
\begin{aligned}
& \mathrm{dim}(\mathrm{Ext}^0_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')) -
\mathrm{dim}(\mathrm{Ext}^1_{{\mathcal C}_{\mathcal X}}({\mathcal E}'',{\mathcal E}')) -
\mathrm{dim}(\mathrm{Ext}^0_{{\mathcal C}_{\mathcal X}}({\mathcal E}',{\mathcal E}'')) \\ & +
\mathrm{dim}(\mathrm{Ext}^1_{{\mathcal C}_{\mathcal X}}({\mathcal E}',{\mathcal E}'')) =
v'(e''+r'')-v''(e'+r'). \\
\end{aligned}
\end{equation}
\end{lemm}
{\it Proof.}
Note that the complex \eqref{eq:zerocomplex} can be written as the cone of a morphism
of locally free complexes on $X$
\[
\varrho : {\mathcal H}[-1] \to {\mathcal V}
\]
where ${\mathcal H}$ is the complex obtained from ${\widetilde {\mathcal C}}({\mathcal E}'',{\mathcal E}')$ by
omitting all direct summands depending on $V',V''$ (as well as making some
obvious changes of signs), and ${\mathcal V}$ is
the two term complex
\[
\begin{aligned}
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,V'\otimes {\mathcal O}_X) & {\buildrel {}\over\longrightarrow}
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,E')\\
f & \longrightarrow \psi'\circ f\\
\end{aligned}
\]
with degrees $0,1$. The morphism $\varrho$ is determined by the
map
\[
\begin{aligned}
{\mathcal Hom}_{X}(E'',E') & {\buildrel {}\over\longrightarrow}
{\mathcal Hom}_{X}(V''\otimes {\mathcal O}_X,E')\\
\alpha &\longrightarrow -\alpha \circ \psi''\\
\end{aligned}
\]
Therefore there is a long exact sequence of hypercohomology groups
\begin{equation}\label{eq:zeroexseq}
\xymatrix{
0\ar[r] & \IH^0({\mathcal V}) \ar[r] & \mathrm{Ext}^0_{{\mathcal C}^0_{{\mathcal X}}}({\mathcal E}'',{\mathcal E}')\ar[r] &
\IH^0({\mathcal H}({\mathcal E}'',{\mathcal E}')) \\
\ar[r] & \IH^1({\mathcal V}) \ar[r] &\mathrm{Ext}^1_{{\mathcal C}^0_{{\mathcal X}}}({\mathcal E}'',{\mathcal E}')\ar[r] &
\IH^1({\mathcal H}({\mathcal E}'',{\mathcal E}')) \\
\ar[r] & \IH^2({\mathcal V}) \ar[r] & \cdots & \\}
\end{equation}
Since $E'_{<0}=0$ and $X$ is rational, $\IH^2({\mathcal V}) =0$. Obviously, there is
a similar exact sequence with ${\mathcal E}',{\mathcal E}''$ interchanged. Then equation
\eqref{eq:zeroeuler} easily follows observing that
\[
\IH^k({\mathcal H}({\mathcal E}'',{\mathcal E}')) \simeq \IH^{3-k}(\IH({\mathcal E}', {\mathcal E}''))^\vee
\]
for all $0\leq k \leq 3$.
\hfill $\Box$
{\it Proof of Corollary (\ref{closedform}).}
Let $\gamma=(r,e,v)\in \IZ_{\geq 1}\times \IZ_{\geq 0}\times \IZ_{\geq 0}$
be an arbitrary numerical type, and $\delta \in \mathbb{R}_{\geq 0}$.
Given any decomposition $\gamma=\gamma_1+\cdots + \gamma_l$, $l\geq 1$
so that
\[
{e_1+v_1\delta\over r_1}=\cdots = {e_l+v_l\delta\over r_l}=
{e+v\delta\over r}
\]
it is obvious that if $v_i=0$ for some $1\leq i\leq l$ then $e_i\geq 0$.
Moreover, if $\delta=0$, then $e_i\geq 0$ for all $1\leq i\leq l$.
In particular this holds for all terms in the right hand side of the defining
equation of log stack functions \eqref{eq:logstfctB}.
It also holds for all possible numerical types of Harder-Narasimhan
filtrations associated to a critical stability parameter $\delta_c\geq 0$
as in lemma (\ref{HNfiltrations}). Note that if $\delta_c=0$, the last quotient
${\mathcal F}_h$ in the Harder-Narasimhan filtration with respect to $\delta_+$-stability,
respectively the first quotient ${\mathcal F}_1$ in the Harder-Narasimhan filtration with respect
to $\delta_-$-stability is allowed to be isomorphic to the object $O_v=(0,\mathbb{C}^v,0,0,0)$,
$v\geq 1$.
In conclusion, the definition of generalized Donaldson-Thomas invariants, and
derivation of wallcrossing formulas carry over to the present set-up for semistable
objects of positive degree and stability parameters $\delta\geq 0$.
In this case the resulting invariants will be denoted by $A^0_\delta(\gamma)$,
or $A^0_\delta(\alpha, v)$ by analogy with section (\ref{wallformula}).
Lemmas (\ref{zerophi}) and (\ref{posdeg}) imply that the invariants $A^0_\delta(\alpha,2)$
satisfy the wallcrossing formula \eqref{eq:wallformulaG} at a positive critical
stability parameter $\delta_c$ of type $(\alpha,2)$.
If $\delta_c=0$, a modification of formula \eqref{eq:wallformulaG}
is required, reflecting the presence of objects isomorphic to ${\mathcal O}_v$, $v=1,2$
in the Harder-Narasimhan filtrations. Basically one has to set $\delta_c=0$
in conditions \eqref{eq:slopesA}-\eqref{eq:slopesC}, and
allow elements $(\alpha_i)_{1\leq i\leq l}$ so that $\alpha_i$, $1\leq i\leq l-1$
satisfy conditions \eqref{eq:slopesA}-\eqref{eq:slopesC}, and $\alpha_l=(0,0)$.
This will result in extra terms in the right hand side of \eqref{eq:wallformulaG}
which can be easily written down using \eqref{eq:purevinv}.
Since this is an easy exercise, explicit formulas will be omitted (see
\cite[Thm. 1.$ii$.]{chamberII} for the $v=1$ case).
Finally, note that one can also check compatibility with the Kontsevich-Soibelman
formula at $\delta_c=0$ repeating the calculations in section (\ref{KSsect}).
Then the proof of corollary (\ref{closedform}) will be based on the
KS wallcrossing formula relating $\delta$-invariants for $\delta<0$
to $\delta$-invariants with $\delta>>0$.
Let $(r,e)\in \IZ_{\geq 1}\times \IZ_{\geq 0}$
and let $\delta_+\in \mathbb{R}_{>0}\setminus \IQ$ an irrational
stability parameter so that $\delta_+$ is asymptotic
of type $(r',e')$ for all $1\leq r'\leq r$, $0\leq e'\leq e$, $1\leq v\leq 2$.
Moreover, assume that $re<\delta_+$.
Then the KS formula reads
\begin{equation}\label{eq:emptytoinftyA}
\begin{aligned}
\prod_{(r,n, v) \in \IZ_{\geq 1}\times\IZ_{\geq 0}\times \{0,1,2\} \cup \{0,0,1\} }
U_{\lambda(r,n,v)}^{{\overline {A^0}}_-(r,e,v)} =
\prod_{(r,n, v) \in \IZ_{\geq 1}\times \IZ_{\geq 0}\times \{0,1,2\} \cup \{0,0,1\} }
U_{\lambda(r,n,v)}^{{\overline {A^0}}_+(r,n,v)}
\end{aligned}
\end{equation}
where in each term the factors are ordered in increasing order of $\delta_\pm$-slopes
from left to right.
The alternative notation introduced in section (\ref{KSsect}) will be used in the
following.
Then
corollary (\ref{genzeroHiggsinv}) and equation \eqref{eq:purevinv} imply that
the left hand side of \eqref{eq:emptytoinftyA} reads
\begin{equation}\label{eq:emptytoinftyB}
\mathrm{exp}({ {\sf f}}_{00} + \frac{1}{4}{{\sf g}}_{00} )
\prod_{n=0}^{\infty}U_{{{\sf e}}_{1n}},
\end{equation}
where
\[
U_{{{\sf e}}_{rn}} = \mathrm{exp}\bigg((-1)^{d_1-1}\sum_{k=1}^{\infty} \frac{
{ {\sf e}}_{kr,kn}}{k^2}\bigg).
\]
Moreover, given the above choice of $\delta_+$,
\[
e< {\delta_+ \over r} < \cdots {e+\delta_+\over r} < {\delta_+\over r-1} < \cdots
< {e+\delta_+\over r-1} < \cdots < \delta_+ + e < {2\delta_+ \over r} < \cdots <
2\delta_+ + e.
\]
Therefore,
in the right hand side of equation \eqref{eq:emptytoinftyA},
the factors of the form $U_{{\lambda}(r',e',v)}^{{\overline A}_+(r',e',v)}$,
with $v\in \{0,1,2\}$,
and $1\leq r'\leq r$, $1\leq e'\leq e$ occur in the following
order
\begin{equation}\label{eq:emptytoinftyC}
\begin{aligned}
& \prod_{n=0}^{e}U_{{{\sf e}}_{1n}}
\prod_{n=0}^e U_{{{\sf f}}_{r,n}}^{{\overline A}_+(r,n,1)}
\prod_{n=0}^e U_{{{\sf f}}_{r-1,n}}^{{\overline A}_+(r-1,n,1)}\cdots
\prod_{n=0}^e U_{{{\sf f}}_{1,n}}^{{\overline A}_+(1,n,1)} U_{{{\sf f}}_{0,0}}^{{\overline A}_+(0,0,1)}\\
&
\prod_{n=0}^e U_{{{\sf g}}_{r,n}}^{{\overline A}_+(r,n,2)}\cdots
\prod_{n=0}^e U_{{{\sf g}}_{r-1,n}}^{{\overline A}_+(r-1,n,2)}\cdots
\prod_{n=0}^e U_{{{\sf g}}_{1,n}}^{{\overline A}_+(1,n,2)},\\
\end{aligned}
\end{equation}
where
\[
U_{{{\sf f}}_{rn}} = \mathrm{exp}({ {\sf f}}_{rn} +
\frac{1}{4} {{\sf g}}_{2r,2n}),\qquad
U_{{ {\sf g}}_{rn}} = \mathrm{exp}({ {\sf g}}_{rn}).
\]
In addition, the right hand side of \eqref{eq:emptytoinftyA} contains of course
extra factors of the form $U_{{\lambda}(r',e',v)}^{{\overline A}_+(r',e',v)}$,
with $v\in \{0,1,2\}$,
and either $ r'> r$ or $e'> e$. Some of these extra factors may in fact occur
between the factors listed in \eqref{eq:emptytoinftyC}.
However, they can be ignored for the purpose
of this computation since commutators involving such factors are again
expressed in terms of generators ${\lambda}(r',e',v)$ with
either $ r'> r$ or $e'> e$. Therefore, using the BCH formula,
\eqref{eq:emptytoinftyA}
yields
\begin{equation}\label{eq:KSconifold2}
\begin{aligned}
& (\prod_{n=0}^{e}U_{{{\sf e}}_{1n}})^{-1} \,
\mathrm{exp}({{\sf f}}_{00} + \frac{1}{4}{{\sf g}}_{00} )\,
\prod_{n=0}^{\infty}U_{{{\sf e}}_{1n}} = \\
& \mathrm{exp} \Big(\mathop{\sum_{1\leq s\leq r,\ 0\leq n \leq e}}_{}
{A}_{+}(s,n,1)\, { {\sf f}}_{sn}+
\mathop{\sum_{ 1\leq s\leq r,\ 0\leq n\leq e}}_{}{ A}_{+}(s,n,2)\,
{ {\sf g}}_{sn} +\\
& \sum_{\substack{r_1>r_2\geq 1, \ r_1+r_2\leq r,\
n_1,\ n_2\geq 0, n_1+n_2\leq e \\ \mathrm{or} \
1\leq r_1=r_2\leq r/2, \ 0\leq n_1< n_2, \ n_1+n_2\leq e \\ \mathrm{or} \ 1 \leq r_1 \leq r, \ 0 \leq n_1 \leq
e, \ r_2=n_2=0}} \frac{1}{2} (n_1-n_2+r_1-r_2) (-1)^{(n_1-n_2+r_1-r_2)} \\
& {A}_{+}
(r_1,n_1,1){ A}_{+}(r_2,n_2,1)\,
{ {\sf g}}_{r_1+r_2,n_1+n_2}+\cdots\Big)\\
\end{aligned}
\end{equation}
where $\cdots$ are terms involving generators ${\lambda}(r',e',v)$ with
either $ r'> r$ or $e'> e$.
For fixed $e\geq 1$, let $\mathcal{H}_e$ be defined by
\begin{equation}
\mathrm{exp}(\mathcal{H}_e)\equiv \prod_{n=0}^{e}U_{{\sf e}_{1,n}} =
\mathrm{exp} \bigg( (-1)^{d_1-1}\sum_{ 0\leq n \leq e ,\ k\geq 1}
\frac{{\sf e}_{k,kn}}{k^2} \bigg).
\end{equation}
Using the BCH formula, the left hand side of equation \eqref{eq:KSconifold2}
becomes
\begin{equation}\label{eq:KSconifold3}
\begin{aligned}
\mathrm{exp}\bigg({\sf f}_{00} +{1\over 4}{\sf g}_{00}
+ \sum_{j=1}^\infty {1\over j!} [\underbrace{-\mathcal{H}_e, \cdots [ -\mathcal{H}_e}_{\text{j times}},{\sf f}_{00} +{1\over 4}{\sf g}_{00} ]\cdots ]\bigg)
\end{aligned}
\end{equation}
modulo terms involving generators ${\lambda}(r',e',v)$ with
either $ r'> r$ or $e'> e$.
Next, the Lie algebra commutators
\[
\begin{aligned}
& [{\sf e}_{r_1,n_1}, {\sf f}_{r_2,n_2}] = (-1)^{n_1+r_1}(n_1+r_1)\,
{\sf f}_{r_1+r_2,n_1+n_2}\\
& [{\sf e}_{r_1,n_1}, {\sf g}_{r_2,n_2}] = 2(n_1+r_1)\,
{\sf g}_{r_1+r_2,n_1+n_2},\\
\end{aligned}
\]
yield
\[
\begin{aligned}
& [ \underbrace{-\mathcal{H}_e, \cdots [ -\mathcal{H}_e}_{\text{j times}},{\sf f}_{00}
]\cdots ] =
\mathop{\sum_{n_1,\ldots, n_j = 0}^{e}}_{}
\mathop{\sum_{k_1,\ldots, k_j \geq 1}}_{}
(-1)^{j(d_1-1)}
\prod_{i=1}^j {n_i+1\over k_i} (-1)^{(n_i+1)k_i-1} \, {\sf f}_{k_1+\cdots + k_j,
k_1n_1+\cdots + k_jn_j}\\
\end{aligned}
\]
and
\[
\begin{aligned}
& [ \underbrace{-\mathcal{H}_e, \cdots [ -\mathcal{H}_e}_{\text{j times}},{\sf g}_{00}
]\cdots ] =
\mathop{\sum_{n_1,\ldots, n_j = 0}^{e}}_{}
\mathop{\sum_{k_1,\ldots, k_j \geq 1}}_{}
(-1)^{j(d_1-1)}\prod_{i=1}^j (-2){n_i+1\over k_i} \, {\sf g}_{k_1+\cdots + k_j,
k_1n_1+\cdots + k_jn_j}\\
\end{aligned}
\]
Therefore, identifying the coefficients of the generators
${\sf f}_{rn}$ in \eqref{eq:KSconifold2}
it follows that the invariant $A_+(r',e',1)$ with $1\leq r'\leq r$ and
$0\leq e'\leq e$ equals the coefficient of the monomial $u^{r'}q^{e'+r'}$
in the expression
\[
\begin{aligned}
\sum_{j=0}^\infty {1\over j!}
\left(
\mathrm{ln}\left(\prod_{n=0}^e (1-u(-q)^{n+1})^{(-1)^{d_1-1}(n+1)}\right)\right)^j =
\prod_{n=1}^{e+1}(1-u(-q)^n)^{(-1)^{d_1-1}n}.
\end{aligned}
\]
Similarly, identifying the coefficients of the generators ${\sf g}_{rn}$ in \eqref{eq:KSconifold2}
proves that the invariant $A_+(r',e',2)$ with $1\leq r'\leq r$ and
$0\leq e'\leq e$ equals the coefficient of the monomial $u^{r'} q^{e'+r'}$
in the expression
\[
\begin{aligned}
& {1\over 4}\prod_{n=1}^{e+1}(1-uq^n)^{2(-1)^{d_1-1}n}
-
\sum_{\substack{r_1>r_2\geq 1, \ r_1+r_2\leq r,\
n_1,\ n_2\geq 0, n_1+n_2\leq e \\ \mathrm{or} \
1\leq r_1=r_2\leq r/2, \ 0\leq n_1< n_2, \ n_1+n_2\leq e
\\ \mathrm{or} \ 1 \leq r_1 \leq r, \ 0 \leq n_1 \leq
e, \ r_2=n_2=0}} \\
& \frac{1}{2} (n_1+r_1-n_2-r_2) (-1)^{(n_1+r_1-n_2-r_2)}
A_{+}(r_1,n_1,1)
A_{+}(r_2,n_2,1) q^{r_1+r_2}u^{n_1+n_2}.\\
\end{aligned}
\]
Since this holds for any $(r,e)\in \IZ_{\geq 1}\times \IZ_{\geq 0}$
(with a suitable choice of $\delta_+$), corollary (\ref{closedform})
follows.
\hfill $\Box$
|
\section{Introduction}
The analysis of spreading processes in large-scale complex networks is a
fundamental dynamical problem in network science. The relationship between
the dynamics of epidemic/information spreading and the structure of the
underlying network is crucial in many practical cases, such as the spreading
of worms in a computer network, viruses in a human population, or rumors in
a social network. Several papers approached different facets of the virus
spreading problem. A rigorous analysis of epidemic spreading in a
finite one-dimensional linear network was developed by Durrett and Liu in
\cite{DL88}. In \cite{WCWF03}, Wang et al. derived a sufficient condition to
tame an epidemic outbreak in terms of the spectral radius of the adjacency
matrix of the underlying graph. Similar results were derived by Ganesh
et al. in \cite{GMT05}, establishing a connection between the behavior of a
viral infection and the eigenvalues of the adjacency matrix of the network.
In this paper, we study the dynamics of a viral spreading in an important type of proximity
networks called Random Geometric Graphs (RGG). RGG's consist of a set of
vertices randomly distributed in a given spatial region with edges connecting pairs of nodes that
are within a given distance $r$ from each other (also called \emph{connectivity radius}). In this paper, we
derive new explicit expressions for the expected spectral
moments of the random adjacency matrix associated to an RGG. Our results allow us to derive analytical conditions under which an RGG is well-suited to tame an infection in the network.
The paper is structured as follows. In Section II, we describe random
geometric graphs and introduce several useful results concerning their
structural properties. We also present the spreading model in \cite{WCWF03}
and review an important result that relates the behavior of an initial infection
with the spectral radius of the adjacency matrix. In Section III, we study
the eigenvalue spectrum of random geometric graphs. We derive explicit
expressions for the expected spectral moments in the case of one- and
two-dimensional RGG's. In Section IV, we use these expressions to study the spectral radius of RGG's. Our results allow us to
design RGG's with the objective of taming epidemic outbreaks. Numerical
simulations in Section IV validate our results.
\section{Virus Spreading in Random Geometric Graphs}
In this section, we briefly describe random geometric graphs and introduce
several useful results concerning their structural properties (see \cite%
{Pen03} for a thorough treatment). We then describe the spreading model
introduced in \cite{WCWF03} and show how to study the behavior of an
infection in the network from the point of view of the adjacency eigenvalues.
\subsection{\label{RGG}Random Geometric Graphs}
Consider a set of $n$ nodes, $V_{n}=\{v_{1},...,v_{n}\}$, respectively
located at random positions, $\chi _{n}=\{\mathbf{x}_{1},...,\mathbf{x}_{n}\}
$, where $\mathbf{x}_{i}$ are i.i.d. random vectors uniformly distributed on
the $d$-dimensional unit torus, $\mathbb{T}^{d}$. We use the torus for
convenience, to avoid boundary effects. We then connect two nodes $%
v_{i},v_{j}\in V_{n}$ if and only if $\left\Vert \mathbf{x}_{i}-\mathbf{x}%
_{j}\right\Vert \leq r $, where $r$ is the
so-called connectivity radius. In other words, a link exists between $v_{i}$
and $v_{j}$ if and only if $v_{j}$ lies inside the sphere of radius $r\left(
n\right) $ centered at $v_{i}$. We denote this spherical region by $%
S_{i}\left( r \right) $, and the resulting random geometric
graph by $G\left( \chi _{n};r \right) $. We define a \emph{walk }of length $%
k$ from $v_{0}$ to $v_{k}$ as an ordered set of (possibly repeated) vertices
$\left( v_{0},v_{1},...,v_{k}\right) $ such that $v_{i}\sim v_{i+1},$ for $%
i=0,1,...,k-1$; if $\nu _{k}=\nu _{0}$ the walk is said to be \emph{closed}.
The \emph{degree} $d_{i}$ of a node $v_{i}$ is the number of edges
connected to it. In our case, the degrees are identical
random variables with expectation \cite{Pen03}:%
\begin{equation}
\mathbb{E}\left[ d_{i}\right] =nV^{\left( d\right) }r^{d},
\label{Expected Degree}
\end{equation}%
where $V^{\left( d\right) }$ is the volume of a $d$-dimensional unit sphere,
$V^{\left( d\right) }=\pi ^{d/2}\left/ \Gamma \left( d/2+1\right) \right. $,
and $\Gamma \left( \cdot \right) $ is the Gamma function. The \emph{%
clustering coefficient} is a measure of the
number of triangles in a given graph, where a triangle is defined by the set
of edges $\left\{ \left( i,j\right) ,\left( j,k\right) ,\left( k,i\right)
\right\} $ such that $i\sim j\sim k\sim i$. For one- and two-dimensional
RGG's we can derive an explicit expression for the expected number of triangles,
$\mathbb{E}\left[ t_{i}\right] $, touching a particular node $v_{i}$
(details are provided in Section III).
The \emph{adjacency matrix }of
an undirected graph $G,$ denoted by $A(G)=[a_{ij}]$, is defined entry-wise
by $a_{ij}=1$ if nodes $i$ and $j$ are connected, and $a_{ij}=0$ otherwise.
(Note that $a_{ii}=0$ for simple graphs.) Denote the eigenvalues of a $%
n\times n$ symmetric adjacency matrix $A(G)$ by $\lambda _{1}\leq ...\leq
\lambda _{n}$. The $k$\emph{-}th order
moment of the eigenvalue spectrum of $A(G)$ is defined as:
\begin{equation*}
m_{k}(G)=\frac{1}{N}\sum_{i=1}^{n}\lambda _{i}^{k}
\end{equation*}%
(which is also called the $k$-th \emph{order spectral moment}).
We are interested in studying asymptotic properties of the sequence $G(\chi
_{n};r\left( n\right) )$ for some sequence $\left\{ r\left( n\right) :n\in
\mathbb{N}\right\} $. In \cite{Pen03}, two particularly interesting regimes
are introduced: the \emph{thermodynamic limit} with $nr\left( n\right)
^{d}\rightarrow \alpha \in \left( 0,\infty \right) $, so that the expected
degree of a vertex tends to a constant, and the \emph{connectivity regime}
with $r\left( n\right) \rightarrow \gamma \left( \frac{\log n}{n}\right)
^{1/d}$ with a constant $\gamma $, so that the expected degree of the nodes
grows as $c\log n$. In this paper, we focus on studying the
spectral moments in the connectivity regime. In Section III, we
derive explicit expressions for the expected spectral moments of $G\left(
\chi _{n};r_{n}\right) $ for any network size $n$. We then use this
information to bound the spectral radius of the adjacency matrix of $G(\chi
_{n};r\left( n\right) )$.
\subsection{Spectral Analysis of Virus Spreading}
In this section, we briefly review an automaton model that describes the
dynamics of a viral infection in a specific network of interactions.
This model was proposed and analyzed in \cite{WCWF03}, where a connection between the growth of an initial infection in the
network and the spectral radius of the adjacency matrix was established. This model involves
several parameters. First, the infection rate $\beta $ represents the
probability of a virus at an infected node $i$ spreading to another
neighboring node $j$ during a time step. Also, we denote by $\delta $ the
probability of recovery of any infected node at each time step. For
simplicity, we consider $\beta $ and $\delta $ to be constants for all
the nodes in $G$. We also denote by $p_{i}\left[ k\right] $ the
probability that node $i$ is infected at time $k$. The evolution of the
probability of infection is modeled by means of the following system of
non-linear difference equation:%
\begin{equation}
p_{i}\left[ k+1\right] =[1-\prod_{j\in \mathcal{N}_{i}}\left( 1-\beta \,p_{j}%
\left[ k\right] \right) ]+\left( 1-\delta \right) p_{i}\left[ k\right] ,
\label{Epidemic Model}
\end{equation}%
for $i=1,...,n$, where $\mathcal{N}_{i}$ denotes the set of nodes connected
to node $i$. We are interested in studying the dynamics of the system for a low-density level of infection, i.e., $\beta \,p_{j}\left[ k\right] \ll 1$. In this regime, a sufficient condition for a small initial infection to die out
is \cite{WCWF03}:
\begin{equation}
\lambda _{\max }\left( A(G)\right) <\frac{\delta }{\beta }.
\label{Epidemic Conditions}
\end{equation}%
One can prove that (\ref{Epidemic Conditions}) is a sufficient condition for
local stability around the disease-free state. Thus, we can use condition (%
\ref{Epidemic Conditions}) to design networks with the objective of taming
initial low-density infections.
\section{Spectral Analysis of Random Geometric Graphs}
In this paper, we study the eigenvalue distribution of the random adjacency
matrix associated to $G(\chi _{n};r\left( n\right) )$\ for $n\rightarrow
\infty $. In this section, we characterize eigenvalue distribution using its sequence of spectral moments. In our derivations, we use an interesting graph-theoretical interpretation
of the spectral moments \cite{Big93}: \emph{the $k$-th spectral moment
of $G$ is proportional to the number of closed walks of length $k$ in $G$.} This result allows us to transform the algebraic problem of computing
spectral moments of the adjacency matrix into the combinatorial problem of
counting closed walks in the graph. In the following subsection, we compute the expected value of the number of
closed walks of length $k$ in $G(\chi _{n};r\left( n\right) )$.
\subsection{Spectral Moments of One-Dimensional RGG's}
As we mentioned above, we can compute the $k$-th spectral moment of a graph by counting the number of closed walks of length $k$. In the case of an RGG $G(\chi_{n};r\left( n\right) )$, this number is a random variable. In this subsection, we introduce a novel technique to compute the expected number of closed walks of length $k$. For clarity, we introduce our technique
for the first three expected spectral moments $k=1,2,3$. We then use these results to induce a general expression for higher-order moments in one-dimensional RGG's.
The first-order spectral moment is equal to the number of closed walks of
length $k=1$. Since $G(\chi _{n};r)$ is a simple graphs with no self-loops,
we have that $m_{1}\left( G(\chi _{n};r)\right) $ is a deterministic
quantity equal to $0$.
We now study the expected second moment, $\mathbb{E}\left[ m_{2}\left(
G(\chi _{n};r)\right) \right] $, by counting the number of closed walks of length two.
In simple graphs, the only possible closed walks of length two are those
that start at a given node $v_{i}$, visit a neighboring node $v_{j}\in
\mathcal{N}_{i}$, and return back to $v_{i}$. Hence, the number of closed
walks of length two starting at $v_{i}$ is equal to $d_{i}$. Thus, from (\ref%
{Expected Degree}), we have%
\begin{equation*}
\mathbb{E}\left[ m_{2}\right] =\frac{1}{n}\sum_{i=1}^{n}\mathbb{E}\left[
d_{i}\right] =nV^{\left( d\right) }r^{d}\text{,}
\end{equation*}%
where this result is valid for any dimension $d\geq 1$.
The third spectral moment is proportional to the number of closed walks of
length three in the graph. We now derive an expression for the expected
number of triangular walks starting at a given node $v_{i}$ in a one-dimensional RGG. Since all nodes
are statistically equivalent, our result is valid for any other starting
node. For simplicity in our calculations, we consider that $v_{i}$ is
located at the origin. A triangular walk starting at node $v_{i}$ exists if and
only if there exist two nodes $v_{j}$ anv $v_{k}$ such that $\left\vert x_{j}\right\vert \leq r$, $\left\vert x_{k}\right\vert \leq r$, and $\left\vert x_{k}-x_{j}\right\vert \leq r$. Also, since the random distribution of
vertices on $\mathbb{T}^{1}$ is uniform (with density $n$), the probability
of nodes $v_{j}$ and $v_{k}$ being respectively located in the differential
lengths $\left[ x_{j}+dx_{j}\right) $ and $\left[ x_{k}+dx_{k}\right) $ is
equal to $n^{2}~dx_{j}dx_{k}$. Hence, one can compute the expected number of
triangular walks starting at node $v_{i}$ as%
\begin{equation*}
\mathbb{E}\left[ t_{i}\right] =\int \int_{\left( x_{j},x_{k}\right) \in
H_{2}\left( r\left( n\right) \right) }n^{2}~dx_{j}dx_{k},
\end{equation*}%
where%
\begin{align}
H_{2}\left( r\right) & =\left\{ \left( x_{j},x_{k}\right) \in \mathbb{T}^{2}%
\text{ s.t. }\left\vert x_{j}\right\vert \leq r,\right. \label{Vol H2} \\
& \left. \text{ \ \ \ \ \ }\left\vert x_{k}-x_{j}\right\vert \leq
r,\left\vert x_{k}\right\vert \leq r\right\} . \notag
\end{align}%
Thus, $\mathbb{E}\left[ t_{i}\right] $ can be computed as $n^{2}$Vol$\left[
H_{2}\left( r\left( n\right) \right) \right] $ (where Vol$\left( H\right) $
denotes the volume contained by the polyhedron $H$.) Notice that $H_{2}\left(
r\right) $ can be defined by a set of linear inequalies; hence, $H_{2}\left(
r\right) $ is a convex polyhedron that depends on $r$. Furthermore, the set
of linear inequalities in (\ref{Vol H2}) presents a homogeneous dependency
with respect to the parameter $r$. Therefore, we can write Vol$(H_{2}\left(
r\right) )$ as $r^{2}$Vol$(H_{2}\left( 1\right) )$. Finally, one can easily
compute the volume of $H_{2}\left( 1\right) $ to be equal to $3$. Thus, the
expected third spectral moment of a one-dimensional RGG is given by%
\begin{equation*}
\mathbb{E}\left[ m_{3}\right] =\frac{1}{n}\sum_{i=1}^{n}\mathbb{E}\left[
t_{i}\right] =3n^{2}r^{2}.
\end{equation*}
In the following, we extend the above technique to compute higher-order
expected spectral moments. Denote by $W_{i}^{\left( k\right) }$ the number
of closed walks of length $k$ starting at node $v_{i}$
in $G(\chi _{n};r\left( n\right) )$. Regarding $W_{i}^{\left( k\right) }$,
we derive the following result.
\begin{theorem}
\label{Spectral Moments 1D}The expected number of closed walks
of length $k$, $W_{i}^{\left( k\right) }$, in a random geometric graph, $%
G(\chi _{n};r)$, on $\mathbb{T}^{1}$ is given by%
\begin{equation*}
\mathbb{E}\left[ W_{i}^{\left( k\right) }\right] =\left( nr\right) ^{k-1}%
\frac{1}{2\left( k-1\right) !}\sum_{j=1}^{k-2}\binom{k-1}{j-1}~E_{k-1,j},
\end{equation*}%
where $E_{k-1,j}$ are the Eulerian numbers \footnote{%
The Eulerian number $E\left( n,k\right) $ gives the number of permutations
of $\{1,2,...,n\}$ having $k$ permutation ascents \cite{GKP94}.}.
\end{theorem}
\begin{proof}
Consider a particular closed walk, $\mathbf{w}_{k}=\left(
v_{1},v_{2},v_{3},...,v_{k},v_{1}\right) $, of length $k$ starting and
ending at node $v_{1}$ (which we locate at zero for computational
convenience). A walk $\mathbf{w}_{k}$ exists if and only if
there exists a set of $k-1$ nodes, $\left\{ v_{2},v_{3},...,v_{k}\right\} ,$
such that $\left\vert x_{1}\right\vert \leq r$, $\left\vert
x_{j+1}-x_{j}\right\vert \leq r$ for $j=2,...,k-1$, and $\left\vert
x_{k}\right\vert \leq r$. Since the distribution of vertices on $\mathbb{T}%
^{1}$ is uniform (with density $n$) one can compute the expectation of $%
W_{i}^{\left( k\right) }$ as
\begin{equation*}
\mathbb{E}\left[ W_{i}^{\left( k\right) }\right] =\int_{(x_{2},...,x_{k})\in
H_{k-1}\left( r\left( n\right) \right) }n^{k-1}~dx_{2}...dx_{k},
\end{equation*}%
where%
\begin{align}
H_{k-1}\left( r\right) & =\left\{ (v_{2},v_{3},...,v_{k})\in \mathbb{T}^{k-1}%
\text{ s.t. }\left\vert v_{2}\right\vert \leq r,\right. \label{Vol Hk-1} \\
& \left. \text{ \ \ }\left\vert x_{j+1}-x_{j}\right\vert \leq r\text{ for }%
j=2,...,k-1,\right. \notag \\
& \left. \text{ \ \ \ }\left\vert x_{k}\right\vert \leq r\right\} . \notag
\end{align}%
Thus, $\mathbb{E[}W_{i}^{\left( k\right) }]$ can be computed as $n^{k-1}$Vol$%
\left[ H_{k-1}\left( r\right) \right] $, where $H_{k-1}\left( r\right) $ is
a convex polyhedron defined by a set of linear inequalities. Finally, note
that the homogeneous structure of the system of linear inequalities defining
$H_{k-1}\left( r\right) $ allows us to write Vol$(H_{k-1}\left( r\right)
)=r^{k-1}$Vol$(H_{k-1}\left( 1\right) )$. Therefore,%
\begin{equation}
\mathbb{E}\left[ W_{i}^{\left( k\right) }\right] =\left( nr\right) ^{k-1}%
\text{Vol}\left( H_{k-1}\left( 1\right) \right) . \label{Walks as Volumes}
\end{equation}%
The volume of $H_{k-1}\left( 1\right) $ is a particular number, independent
of the RGG parameters, i.e., $n$ and $r$. Furthermore, we have found an
explicit analytical expression for the volume of $H_{k}\left( 1\right) $ for
any $k\geq 1$. Although we do not provide details of our derivation, due to
space limitations, an explicit expression for the volume of $H_{k}\left(
1\right) $ is given by \cite{PJ09}:%
\begin{equation}
\text{Vol}\left( H_{k}\left( 1\right) \right) =\frac{2}{k!}\sum_{j=1}^{k-1}%
\binom{k}{j-1}~E_{k,j}, \label{Volumes as Eulers}
\end{equation}%
where $E_{d,k}$ denotes the Eulerian numbers. Substituting (%
\ref{Volumes as Eulers}) in (\ref{Walks as Volumes}) we obtain the statement
of our lemma.
\end{proof}
In \cite{Las83}, Lasserre proposed an algorithm to compute the volume of a
polyhedron defined by a set of linear inequalities. We can use this algorithm to verify the validity of (\ref{Vol Hk-1}). Applying this algorithm to the set of inequalities in (\ref{Vol Hk-1}), we compute the following
volumes for $k=1,...,10$:%
\begin{align*}
H_{1}& =2,~H_{2}=3,~H_{3}=5.333...,~H_{4}=9.58333..., \\
H_{5}& =17.6000...,~H_{6}=32.70555...,~H_{7}=61.3587..., \\
H_{8}& =115.947...,~H_{9}=220.3238...,~H_{10}=420.825...
\end{align*}%
These numerical values match perfectly with our analytical expression in
Theorem \ref{Spectral Moments 1D}.
If $nr\left( n\right) =\Omega \left( \log n\right) $ (i.e., the average
degree grows as $\log n$, or faster), one can prove that $\mathbb{E}\left[ m_{k}\right] =\left(
1+O\left( \log ^{-1}n\right) \right) ~\mathbb{E[}W_{i}^{\left( k\right) }]$.
Hence, from (\ref{Walks as Volumes}) and (\ref{Volumes as Eulers}), we have
the following closed-form expression for the asymptotic expected spectral
moments:%
\begin{equation}
\mathbb{E}\left[ m_{k}\right] \asymp \left( nr\right) ^{k-1}\frac{1}{%
2\left( k-1\right) !}\sum_{j=1}^{k-2}\binom{k-1}{j-1}~E_{k-1,j}.
\label{Expected Spectral Moments 1D}
\end{equation}
In the following table
we compare the analytical result in (\ref{Expected Spectral Moments 1D})
with numerical realizations of the empirical spectral moments. In our
simulations, we distribute $n=1000$ nodes uniformly in $\mathbb{T}^{1}$ and
choose a connectivity radius $r=0.01$ (which results in an average degree $%
\mathbb{E}[d_{i}]=20$). The second, third, and forth column in the following
table represent the analytical expectations of the spectral moments, the
empirical average of the spectral moments from 10 random realizations of the
RGG, and the corresponding empirical typical deviation, respectively.%
\begin{equation*}
\begin{tabular}{|l|lll|}
\hline
$k$ & $\mathbb{E}\left[ m_{k}\right] $ & $\text{Empirical Average}$ & $\text{%
Typical Deviation}$ \\ \hline
1 & \multicolumn{1}{|c}{0} & \multicolumn{1}{c}{1.38e-16} &
\multicolumn{1}{c|}{1.3e-15} \\
2 & \multicolumn{1}{|c}{20} & \multicolumn{1}{c}{19.9326} &
\multicolumn{1}{c|}{0.0976} \\
3 & \multicolumn{1}{|c}{300} & \multicolumn{1}{c}{297.284} &
\multicolumn{1}{c|}{4.3598} \\
4 & \multicolumn{1}{|c}{5,733} & \multicolumn{1}{c}{5,956.30} &
\multicolumn{1}{c|}{196.94} \\ \hline
\end{tabular}%
\end{equation*}%
Our numerical results present an excellent match with our analytical
predictions.
\subsection{Spectral Moments of Two-Dimensional RGG's}
In this subsection, we derive expressions for the first three
expected spectral moments of $G(\chi _{n};r\left( n\right) )$ when the nodes
are uniformly distributed in $\mathbb{T}^{2}$. The expressions for
the first and second expected spectral moments are $m_{1}=0$ and $\mathbb{E}\left[
m_{2}\right] =\pi nr^{2}$. The third spectral moment is
proportional to the number of closed walks of length three in the graph. In
the two-dimensional case, we count the number of triangular walks using a
technique that we illustrate in Fig. $\emph{1}$. In this figure, we plot two
nodes $v_{i}$ and $v_{j}$. The
parameters $\rho $ and $\phi $ in Fig. $\emph{1}$ denote the distance and
angle between these two nodes, i.e., $\rho \triangleq \left\Vert \mathbf{x}%
_{j}-\mathbf{x}_{i}\right\Vert $ and $\phi =\measuredangle \left( \mathbf{x}%
_{j}-\mathbf{x}_{i}\right) $. An edge between $v_{i}$ and $v_{j}$ exists if an
only if $v_{j}$ is located inside the circle $S_{i}\left( r\right) $. In this setting, the probability
of existence of a triangle touching both $v_{i}$ and $v_{j}$ is equal to the probability of a third node $v_{k}$ being in the
shaded area $A_{l}$ (see Fig. $\emph{1}$). This area is the result of
intersecting the circles $S_{i}\left( r\right) $ and $S_{j}\left( r\right) $%
, and the resulting probability is equal to $n~A_{l}$. The
intersecting region $A_{l}$ is a symmetric lens which area can be computed
as a function of $\rho $ and $r$ as follows:%
\begin{equation}
A_{l}\left( \rho ;r\right) =\left\{
\begin{array}{cc}
2r^{2}\cos ^{-1}\left( \frac{\rho }{2r}\right) -\frac{\rho }{2}\sqrt{%
4r^{2}-\rho ^{2}}, & \text{for }\rho \leq r, \\
0, & \text{for }\rho >r.%
\end{array}%
\right. \label{Lens Area}
\end{equation}%
Therefore, we can compute the expected number of
triangles by integrating over the set of all possible positions of $v_{j}$,
i.e., $\eta \in \left[ 0,r\right] $ and $\phi \in \lbrack 0,2\pi )$, as
follows%
\begin{equation}
\mathbb{E}\left[ t_{i}\right] =\int_{\rho =0}^{r}\int_{\phi =0}^{2\pi
}n^{2}A_{l}\left( \rho ;r\right) ~\rho ~d\rho ~d\phi .
\label{Triangle Integral}
\end{equation}%
After substituting (\ref{Lens Area}) in (\ref{Triangle Integral}), we can
explicitly solve the resulting integral to be%
\begin{equation}
\mathbb{E}\left[ t_{i}\right] =\left( \pi -\frac{3\sqrt{3}}{4}\right) \pi
\left( nr^{2}\right) ^{2}\approx 5.78\left( nr^{2}\right) ^{2}.
\label{Triangles 2D}
\end{equation}%
Consequently, we have the following expression for the third
expected spectral moment $\mathbb{E}\left[ m_{3}\right] =\frac{1}{n}%
\sum_{i=1}^{n}\mathbb{E}\left[ t_{i}\right] =\mathbb{E}\left[ t_{i}\right] $.
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{Fig1.eps}
\caption{This figure illustrates the technique proposed in Section III.B to count the number of triangular walks in a two-dimensional RGG.}
\end{figure}
In the following, we extend the technique introduced above to compute closed walks of arbitrary length. Denote by $W_{i}^{\left(
k\right) }$ the number of closed walks of length $k$ starting at node $v_{1}$
in $G(\chi _{n};r\left( n\right) )$. The idea behind our technique is
illustrated in Fig. $\emph{2}$, where we represent a particular closed
walk of length $6$. We denote this walk by\ $\mathbf{w}%
_{k}=\left( v_{1},v_{2},...,v_{k-1},v_{k},v_{1}\right) $. We define the
following set of relative distances and angles between every pair of
connected vertices: $\rho_{i}\triangleq \left\Vert \mathbf{x}_{i+1}-\mathbf{x}%
_{i}\right\Vert $ and $\phi _{i}=\measuredangle \left( \mathbf{x}_{i+1}-%
\mathbf{x}_{i}\right) $ for $i=1,...,k-2$. We also define the following parameter%
\begin{equation}
\rho =\left\vert \sum_{j=1}^{k-2}\rho_{j}e^{\mathbf{i}\alpha _{j}}\right\vert ,
\label{Rho parameter}
\end{equation}%
($\mathbf{i}=\sqrt{-1}$) which is the resulting distance between nodes $v_{k-1}
$ and $v_{1}$ given a particular set of distances and angles $\left\{ \left(
r_{i},\phi _{i}\right) \right\} _{i=1,...,k-2}$ (see Fig. $\emph{2}$). In
this setting, the conditional probability of existence of a walk $\mathbf{w}%
_{k}=\left( v_{1},v_{2},...,v_{k-1},v_{k},v_{1}\right) $ given the set of
relative positions, $\left\{ \left( r_{i},\phi _{i}\right) \right\}
_{i=1,...,k-2}$, is equal to the probability of $v_{k}$ being in the shaded
area $A_{l}$ in Fig. $\emph{2}$. We have an expression for this area in (\ref%
{Lens Area}), where $\rho $ is defined in (\ref{Rho parameter}).
Finally, we can compute the expectation of $W_{i}^{\left( k\right) }$ by
performing an integration over the set of all possible positions (i.e., $%
\rho _{j}\in \left[ 0,r\right] $ and $\phi _{j}\in \lbrack 0,2\pi )$ for $%
j=2,...,k-1$), as follows%
\begin{equation*}
\mathbb{E}\left[ W_{i}^{\left( k\right) }\right] =n^{k-1}\int_{\left(
\mathbf{\eta ,\varphi }\right) \in C_{k-2}}A_{l}\left( \rho ;r\right)
~\prod_{j=2}^{k-1}\eta _{j}~d\mathbf{\eta ~}d\mathbf{\varphi ,}
\end{equation*}%
where $\mathbf{\eta =}\left( \rho _{2},...,\rho _{k-1}\right) $, $\mathbf{%
\varphi }=\left( \phi _{2},...,\phi _{k-1}\right) $, and $%
C_{k-2}=\{\left( \mathbf{\eta ,\varphi }\right) :\mathbf{\eta \in }\left[ 0,r%
\right] ^{k-2}$ and $\mathbf{\varphi }\in \lbrack 0,2\pi )^{k-2}\}$.
Although a closed-form for the above expression can only be computed for $%
k \leq 3$, we can always find a good approximation via numerical integration.
For example, the integration for $k=4$ gives us $\mathbb{E[}W_{i}^{\left(
4\right) }]\approx 14.2511\left( nr^{2}\right) ^{3}$.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{Fig2.eps}
\caption{This figure illustrates the technique proposed in Section III.B to count the number of closed walks of length $k$ in a two-dimensional RGG.}
\end{figure}
In the following table, we compare our analytical results with numerical
realizations of the empirical spectral moments of a two-dimensional RGG. In
our simulations, we distribute $n=1000$ nodes uniformly on $\mathbb{T}^{2}$
and choose a connectivity radius $r=\sqrt{50/\pi n}\approx 0.1784$ (which
results in an average degree $\mathbb{E}[d_{i}]=50$). The second, third, and
forth columns in the following table represent the analytical expectation of
the spectral moments, the empirical average from 10 random realizations, and
the corresponding empirical typical deviation, respectively.%
\begin{equation*}
\begin{tabular}{|l|lll|}
\hline
$k$ & $\mathbb{E}\left[ m_{k}\right] $ & $\text{Empirical Average}$ & $\text{%
Typical Deviation}$ \\ \hline
1 & \multicolumn{1}{|c}{0} & \multicolumn{1}{c}{-9.2e-16} &
\multicolumn{1}{c|}{1.1e-15} \\
2 & \multicolumn{1}{|c}{50} & \multicolumn{1}{c}{50.0820} &
\multicolumn{1}{c|}{0.3908} \\
3 & \multicolumn{1}{|c}{1,464.1} & \multicolumn{1}{c}{1,475.8} &
\multicolumn{1}{c|}{37.3777} \\
4 & \multicolumn{1}{|c}{59,452} & \multicolumn{1}{c}{60,127} &
\multicolumn{1}{c|}{2,955.3} \\ \hline
\end{tabular}%
\end{equation*}%
Our numerical results present an excellent match with our analytical
predictions.
In the following section, we use the results introduced in this section to
study the spreading of an infection in a random geometric network.
\section{Spectral Analysis of Virus Spreading}
In this section, we use the expressions for the expected spectral moments to
design random geometric networks to tame an initial viral infection in the
network. In our design problem, we consider that the size of the network $n$
and the parameters in (\ref{Epidemic Model}), i.e., $\beta $ and $\delta $,
are given. Hence, our design problem is reduced to studying the range of
values of $r$ for which the RGG is well-suited to tame an initial viral
infection.
A sufficient condition for local stability around the disease-free state was
given in (\ref{Epidemic Conditions}). Thus, we have to find the range of
values of $r$ for which the associated spectral radius $\lambda _{\max }$ is
smaller than the ratio $\delta /\beta $. In the following subsection, we
show how to derive an analytical upper bound for the spectral radius based
on the expected spectral moments.
\subsection{Analytical Upper Bound for the Spectral Radius}
In order to upper-bound the spectral radius, we use Wigner's high-order
moment method \cite{Wig58}. This method provides a probabilistic upper bound
based on the asymptotic behavior of the $k$-th expected spectral moments for
large $k$. We present the details for a one-dimensional RGG, although the
same technique can be applied to RGG's in higher dimensions. For a
one-dimensional RGG in the connectivity regime, we derived an explicit
expression for the expected spectral moments in (\ref{Expected Spectral
Moments 1D}). A logarithmic plot of Vol$\left( H_{k}\left( 1\right) \right) $ for $k=1,2,...,9$ unveils that Vol$\left( H_{k}\left( 1\right) \right) \rightarrow \beta_{1}c_{1}^{k}$ for large-order moments (a line in logarithmic
scale), where, from a numerical fitting, we find that $\beta _{1}=0.35$ and $%
c_{1}=1.9192$. Therefore, from (\ref{Expected Spectral Moments 1D}) we have%
\begin{equation*}
\mathbb{E}\left[ m_{k}\right] \asymp \beta _{1}\left( c_{1}nr\right)
^{k},
\end{equation*}%
for large $k$.
For even-order expected spectral moments (i.e., $k=2s$ for $s\in \mathbb{N}$%
), the following holds%
\begin{equation*}
\mathbb{E}\left[ m_{2s}\right] =\frac{1}{n}\sum_{i=1}^{n}\mathbb{E}[\lambda
_{i}^{2s}]\geq \frac{1}{n}\mathbb{E}[\lambda _{\max }^{2s}].
\end{equation*}%
Define $f\left( n\right) =n^{1-\delta }\log n$; thus, for any $\varepsilon
,\delta >0$ (and $c_{1}=1.9192$), we can apply Markov%
\'{}%
s inequality as follows%
\begin{eqnarray*}
\mathbb{P}\left( \lambda _{\max }^{2s}\geq (c_{1}nr+\varepsilon rf\left(
n\right) )^{2s}\right) &\leq &\frac{\mathbb{E}[\lambda _{\max }^{2s}]}{%
(c_{1}nr+\varepsilon rf\left( n\right) )^{2s}} \\
&\leq &\frac{n~\mathbb{E}\left[ m_{2s}\right] }{(c_{1}nr+\varepsilon
rf\left( n\right) )^{2s}},
\end{eqnarray*}%
For large $s$, one can prove that \cite{Pre07}%
\begin{equation*}
\mathbb{P}\left( \lambda _{\max }\geq c_{1}nr+\varepsilon rf\left( n\right)
\right) \leq n\beta _{1}\exp \left( -\frac{\varepsilon }{c_{1}}sn^{-\delta }\log
n\right) .
\end{equation*}%
Assuming that $s$ grows as $\beta _{2}n^{\delta }$, for $\beta _{2},\delta >0
$, we have%
\begin{equation*}
\mathbb{P}\left( \lambda _{\max }\geq c_{1}nr+\varepsilon rf\left( n\right)
\right) \leq n\beta _{1}\exp \left( -\frac{\beta _{2}\varepsilon }{c_{1}}%
\log n\right) =o\left( 1\right) ,
\end{equation*}%
for all sufficiently large $\varepsilon $. Thus,%
\begin{equation}
\lim_{n\rightarrow \infty }\mathbb{P}\left( \lambda _{\max
}<c_{1}nr+\varepsilon rn^{1-\delta }\log n\right) =1. \label{Upper Bound}
\end{equation}%
In other words, $\lambda _{\max }$ is upper-bounded by $cnr+\varepsilon
rn^{1-\delta }\log n$ with probability $1$ for $n\rightarrow \infty $. In
practice, for a large (but finite) $n$, we can use $1.9192~nr$ as an upper
bound of $\lambda _{\max }$. In Fig. 4, we plot the empirical
spectral radius of an RGG with $n=1000$ and $r\left( n\right) =\bar{d}/2n$,
with expected degrees $\bar{d}=[$10:1:100$]$ (circles in the figure). We also plot the values of our
analytical upper bound, $1.9192~nr$, in solid line.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Fig4.eps}
\caption{Comparison between the empirical spectral radius of an RGG (circles in the plot) and the values of our analytical upper bound (solid line) for $n=1000$ and $r\left( n\right) =\bar{d}/2n$, with expected degrees $\bar{d}=[$10:1:100$]$.}
\end{figure}
The technique introduced in this subsection is also valid for RGG's in
higher-dimensions. In general, one can prove that for a $d$-dimensional RGG that the expected spectral moment grows as $\mathbb{E}\left[ m_{k}\right]
\rightarrow \beta _{d}\left( c_{d}nr^{d}\right) ^{k}$. Applying Wigner's
high-order moment method to this sequence, one can derive a probabilistic
upper bound similar to (\ref{Upper Bound}). In particular, we have that $\lambda _{\max }<c_{d}nr^{d}$ for large $n
$ with high probability. In the following subsection, we use our results to
design the connectivity radius of an RGG in order to tame an initial viral
infection.
\subsection{Spectral Radius Design}
Once the spectral radius is upper-bounded, our design problem becomes
trivial. Since (\ref{Epidemic Conditions}) represents a sufficient condition
for local stability around the disease-free state, we have the following
condition to tame an initial viral infection for a $d$-dimensional RGG:%
\begin{equation*}
\lambda _{\max }\left( G(\chi _{n};r\right) )<c_{d}nr^{d}<\frac{\delta }{%
\beta },
\end{equation*}%
which implies the following design condition for the connectivity radius:%
\begin{equation}
r<\left( \frac{\delta }{\beta c_{d}n}\right) ^{1/d}, \label{Radius design}
\end{equation}%
where $c_{d}$ is a positive constant that depends on the dimension of $%
\mathbb{T}^{d}$. For example, in the one-dimensional case, we have $%
c_{1}=1.9192$; hence, (\ref{Radius design}) becomes $r<\delta /\left(
1.9192~\beta n\right) $. We now validate this result with several numerical
simulations of a viral infection in a one-dimensional RGG.
Consider an RGG with $n=1000$ nodes and a connectivity radius of $r=0.005$
(which implies an average degree of $10$). The resulting spectral radius in
this RGG is $\lambda _{\max }=17.2629$. In our numerical simulations, we
choose the initial probability of infection to be $p_{i}\left[ 0\right] \sim
0.01$\textsf{Unif}$[0,1]$; hence, approximately $1\%$ of the nodes in the
network are initially infected. In our first experiment, we choose a rate of
infection $\beta =0.020$, and a recovery rate $\delta =0.018$. Since the
sufficient condition for viral control in (\ref{Radius design}) is not
satisfied, we cannot guarantee an initial infection to be tamed. In
Fig. $\emph{5}$ we show an image of the evolution of the probability of
infection for this case. This figure is a color map for the simultaneos
evolution of $p_{i}\left[ n\right] $ for $i=1,...,1000$. Each horizontal
line represents the value of $p_{i}\left[ n\right] $ for a particular $i$. In
this color map, blue represents a zero value, green and yellow tones
represent intermediate values, and red represents values close to one. On the other hand,
if we increase the recovery rate to $\delta =0.35$ keeping the rest of
parameters fixed, we have that $\delta /\beta =17.50>\lambda _{\max }$ and
we satisfy condition (\ref{Epidemic Conditions}). Hence, the probability of
infection of every node is guaranteed to converge towards zero. In Fig. $%
\emph{6}$, we observe the color map for the evolution of the probability of
infection in this case, where we clearly observe how $p_{i}\left[ n\right]
\rightarrow 0$ for all $i$. Hence, this latter RGG is well-suited to tame
initial viral infections.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Fig5.eps}
\caption{Color map representing the evolution of the probabilities of infection $p_{i}\left[ n\right] $ for $i=1,...,1000$ in an RGG with $n=1000$ nodes, connectivity radius $r=0.005$, rate of infection $\beta =0.020$, and recovery rate $\delta =0.018$. Each horizontal line represents the value of $p_{i}\left[ n\right] $ for a particular $i$. In this color map, blue represents a zero value, green and yellow tones represent intermediate values, and red represents values close to one. In this case, we observe an epidemic outbreak.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Fig6.eps}
\caption{Color map representing the evolution of the probabilities of infection $p_{i}\left[ n\right]$ when we increase the recovery rate to $\delta =0.35$ (the rest of parameters are the same as we used for Fig. 5). We observe how the probability of infection of every node converges towards zero in this case. }
\end{figure}
\section{Conclusions}
In this paper, we have studied the spreading of a viral infection in a random geometric graph from a spectral point of view. We have focused our attention on studying the eigenvalue distribution of the adjacency matrix. We have derived, for the first time, explicit expressions for the spectral moments of the adjacency matrix as a function of the density of nodes and the connectivity radius. We have then applied our results to the problem of viral spreading in a network with a low-density infection. Using our expressions, we have derived upper bounds for the spectral radius of the adjacency matrix. Finally, we have applied this upper bound to design random geometric graphs that are well-suited to tame an initial low-density infection. Our numerical results match our predictions with high accuracy.
|
\section{Introduction}
\label{sect_intro}
The vast
energy released in accretion by a supermassive black hole means that
active galactic nuclei (AGN) could have a fundamental role in
the formation and evolution of galaxies (see \citealt{cattaneo09}). This
depends however on the nature of the physical mechanisms that
couple the AGN to its host galaxy. One such possible AGN feedback
mechanism is that of a massive outflow, launched somewhere close to
the accreting black hole. It has already been shown that such flows
might be able to account for observed correlations between properties
of nuclear black holes and their host galaxies (see e.g. \citealt{king03b,king05}).
Direct evidence for AGN outflows and quantification of their
properties, however, requires interpretation of observational
data. Perhaps the best evidence for energetically important outflows
around AGN comes from the detection of X-ray absorption lines identified
with significantly blueshifted line transitions of highly ionized
material (see \citealt{turner09} for a recent review of
X-ray observations of AGN). The most promising origin for mass-loss in
the immediate vicinity of an accreting supermassive black hole is an
accretion disk wind. Hydrodynamical simulations have demonstrated
that such flows are plausible (e.g. \citealt{proga04}). However, these
simulations also suggest that disk winds are likely to be
very complex and thus to interpret their spectroscopic signatures
quantitatively requires detailed synthetic spectra for comparison.
In previous work (\citealt{sim05b}, \citealt{sim08}, hereafter
Paper~I), we developed
a numerical code for computing synthetic X-ray spectra for outflow models.
We showed that simply-parameterized disk wind models could readily
account for strong
blueshifted absorption features associated with Fe~{\sc xxv} and
Fe~{\sc xxvi} and we explored some of the effects of the outflow
density, geometry and ionization state on these absorption line properties.
Our models also confirmed that a very highly ionized outflow could
affect the X-ray spectrum in more subtle ways. In particular, the blueshifted
Fe absorption lines have associated emission features that are broad
and can develop extended red-skewed wings owing to the effects of
electron scattering in the flow (see also
\citealt{laming04,laurent07}). We concluded that, although complex,
outflow models have the promise to simultaneously
account for several of the observed X-ray spectroscopic features of
AGN. We illustrated this via a direct comparison to observations
of the well-known narrow line Seyfert 1 galaxy Mrk~766.
However, that work was limited to the treatment of only the atomic
physics necessary for the most highly ionized atomic species
(specifically, He- and H-like ions). Although AGN provide copious
ionizing radiation,
the region responsible for the primary
X-ray emission is assumed to be small and centrally concentrated. Therefore
more distant portions of an outflow -- or those shielded from the
X-ray source by other material -- need not be so significantly
ionized. Thus a more realistic study requires that a wider range of
ionization conditions can be considered. In particular, photoelectric
absorption and/or fluorescent
emission by lower ionization states of Fe may affect the X-ray
spectrum, both in the Fe~K region and at lower X-ray energies.
For lower typical ionization conditions, the relative importance of
the lighter elements also grows significantly and they can lead to strong
photoelectric absorption for energies $\sim 1$~keV and associated
discreet line features in the soft X-ray spectrum as have been
reported in observations
(e.g. \citealt{pounds03,pounds05,pounds07,pounds09}).
Here, we extend our Monte Carlo radiative transfer code
described in Paper~I to treat the physics of L- and M-shell ions of
astrophysically abundant elements. We also substantially improve the
treatment of the ionization state and introduce a simple means of
estimating the kinetic temperature in the outflow, thereby eliminating
this as a free parameter of the models. In addition, we modify the
means by which the orientation-dependent
spectra are extracted from our simulations in order to
suppress the level of Monte Carlo noise.
These improvements to the code
and atomic data used are described in Section~\ref{sect:method}. In
Section~\ref{sect:iontest} we demonstrate that our improved treatment
of ionization leads to quantitatively good agreement with well-known 1D
codes and in Section~\ref{sect_example} we discuss results from a
calculation made with the improved code. Finally, in
Section~\ref{sect_pg12}, we discuss the comparison of spectra computed
with our models to observations of the quasar PG1211+143 before
drawing conclusions in Section~\ref{sect_conc}.
\section{Method}
\label{sect:method}
\subsection{Code overview}
The code operates by performing a sequence of Monte Carlo radiative
transfer simulations employing an indivisible packet scheme (see
e.g. \citealt{lucy02,lucy03}). In each simulation, we follow the
propagation of Monte Carlo
quanta (``$r$-packets'') representing bundles of X-ray photons through
simply-parameterized models for accretion disk winds. During the
$r$-packet propagation,
the effects of Compton scattering, photoelectric absorption,
bound-bound line interactions (for which the Sobolev approximation is adopted)
and free-free absorption are simulated in detail. Interactions
with matter (specifically bound-bound and bound-free transitions) are
treated using the Macro Atom
formalism introduced by \citet{lucy02}. This
approach allows for a full treatment of line scattering, recombination
and fluorescence as required to produce realistic spectra for
comparison with observations. The code has also been extended to use
the
method of \citet{lucy02,lucy03} for treating both the
radiative heating and cooling of matter.
In this approach, when the MC quanta
undergo physical events in which the energy they represent is used to heat the
plasma (e.g. by free-free absorption of photons), they are instantaneously
converted to packets of thermal kinetic energy (so-called $k$-packets).
Assuming thermal
equilibrium, these $k$-packets are then eliminated by randomly sampling the
available cooling processes (see Section~\ref{section:tb}). Depending on the cooling process
selected, the $k$-packet may be converted to an $r$-packet which is
then free to propagate. The $r$-packets created in this way
simulate the cooling radiation emitted
by the wind.
During each Monte Carlo simulation, the physical
wind properties (e.g. temperature, ionization conditions etc.) are
held fixed. At the end of each such simulation, the histories of the Monte
Carlo quanta are used to make improved estimates of the wind
properties assuming radiative, thermal and ionization equilibrium and
then the Monte Carlo experiment is repeated. In this way,
the wind properties are iterated to consistency with the radiation
field. After the iteration cycle, a final Monte Carlo simulation is
performed from which viewing-angle dependent spectra are extracted
(see Section~\ref{section:extract}).
\subsection{Ionization}
To provide a more complete description of the ionization state of the
outflow, the modified nebular approximation adopted in Paper~I has
been replaced with a detailed computation of ionization and
recombination rates that are then used to solve for the steady-state
ionization fractions.
The following sections describe the processes
that are included.
\subsubsection{Photoionization}
\label{sect:Gamma}
Bound-free absorption involving the ejection of an outer shell
electron is included for ground states and low-energy metastable
levels of all ions.
During the Monte Carlo simulation, the packet
trajectories are used to record estimators for the photoionization
rate coefficient ($\gamma$) for each bound-free process in each of the
computational grid cells of wind properties. The estimator for the
rate coefficient of bound-free absorption from level $j$ in a cell $p$ is
given by
\begin{equation}
\gamma_{j,k} = \frac{1}{V_{p} \; \Delta t}\sum_{\nu > \nu_0} \frac{a_{j}(\nu)}{h
\nu} \epsilon \, \mbox{d} s
\end{equation}
where $V_{p}$ is the wind cell volume, $a_{j}$ is the photoabsorption
cross-section for level $j$ (see Section~\ref{sect_atomicdata}) and the
summation runs over all Monte Carlo quanta trajectories inside cell
$p$ when the quanta have co-moving frame frequency $\nu$ that is greater than the
threshold $\nu_0$. $\epsilon$ is the co-moving frame energy of the
quantum and $\mbox{d} s$ is the trajectory length.
$\Delta t$ is the time interval represented
by the Monte Carlo experiment. The Doppler factor correction term for
transforming the photon path length from the observer frame to the
fluid frame (see e.g. \citealt{lucy05}) is
neglected for computational expediency.
At the end of a Monte Carlo simulation, the $\gamma$
estimators for levels of each ion are combined to make a
photoionization rate estimator for each ion ($I$) in each cell ($p$) defined by
\begin{equation}
\Gamma_{I,k} = \sum_{j \in I} n_{j,k} \gamma_{j,k} / \sum_{j \in I} n_{j,k}
\end{equation}
where the summations run over all level $j$ of the ion
for which bound-free absorption has been included in the Monte Carlo
simulation.
The level populations, $n_{j,k}$ are discussed in Section~\ref{sect:excitation}.
For L- and M-shell ions, photoabsorption involving inner
shell electrons is also included. In the treatment of these
processes, details of the outer electron
configuration are neglected and all levels of the ion are assigned the
same cross section.
Following inner shell photoabsorption, the newly made
ion is generally in a highly-excited state that may either decay
radiatively or via ejection of one or more Auger electrons.
In the case of K-shell photoionization of L-shell ions, at most one Auger electron is expected to
be produced -- thus the ionization state may ultimately increase by either one or two.
To incorporate this in the ionization
balance, Monte Carlo estimators are used to obtain a rate
coefficient for K-shell photoabsorption for each ion
\begin{equation}
\Gamma^K_{I,k} = \frac{1}{V_{p} \; \Delta t}\sum_{\nu > \nu_K} \frac{a^{K}_{I}(\nu)}{h
\nu} \epsilon \, \mbox{d} s
\end{equation}
where the summation now runs over all packets having frequency about
the K-shell edge threshold, $\nu_K$ and $a^{K}$ is the K-shell
photoabsorption cross-section. To obtain separate rates for single and
double ionization, the mean probability ($p_{I}[1]$) of K-shell absorption being
followed by ejection of a single Auger electron is required. This is computed using
\begin{equation}
p_{I}[1] = < \frac{\sum_i A^{a}_{ji}}{\sum_i A_{ji}\beta_{ji} + \sum_i
A^{a}_{ji}} >_{j}
\end{equation}
where $< ... >_{j}$ is a weighted mean over all
states $j$ of the ionization state $I+1$ that are accessible by
K-shell photoabsorption of ion $I$, $\sum_i A^a_{ji}$ is the total autoionization
rate out of level $j$ and $\sum_i A_{ji} \beta_{ji}$ is the sum of
radiative decay rates incorporating the Sobolev escape probabilities
$\beta_{ji}$ over all downward transitions from level $j$. In
practice, only states reached by K-shell photoabsorption from the
ground configuration of ion $I$ are included in the average and these
are weighted with their statistical weights.
The estimated rate at
which K-shell photoabsorption in an L-shell ion increases the ionization state by one is
then given by
\begin{equation}
\Gamma^K_{I,k} (1 - p_{I}[1])
\end{equation}
and, for double ionization,
\begin{equation}
\Gamma^K_{I,k} p_{I}[1] \; \; \; .
\end{equation}
Data sources for $A_{ji}$ and $A^{a}_{ji}$ are discussed in Section~\ref{sect_atomicdata}.
K- and L-shell photoionization of M-shell ions, are treated in a
similar manner, recording a $\Gamma^{K/L}$ estimator for each
such process in each grid cell. Owing to the complexity of the
subsequent decays of the vacancy state (which lead to the ejection of
up to three Auger electrons for K-shell ionization of the M-shell ions
we consider),
the Auger ejection probabilities $p_{I}[1]$, $p_{I}[2]$ and
$p_{I}[3]$, are not computed from the atomic data set used by the code but taken from
the yields tabulated by \cite{kaastra93}. This approach obviates
the need to follow the decay chains of the vacancy states in
detail but comes at the expense of assuming that all the relevant
radiative transitions are optically thin such that the M-shell ion populations
are unaffected by photon trapping.
\subsubsection{Collisional ionization and recombination processes}
In addition to the photoionization and Auger processes described
above, the ionization balance includes collisional ionization, both
direct ($C_{DI}$) and due to collisional excitation followed by
autoionization ($C_{EA}$), and both radiative ($\alpha^{r}$) and di-electronic
($\alpha^{d}$) recombination (see Section~\ref{sect_atomicdata} for
data sources).
\subsubsection{Ionization equilibrium}
At the end of each Monte Carlo simulation, the computed
photoionization rate coefficients ($\Gamma$; see
Section~\ref{sect:Gamma}) are used to solve for an improved set of ion populations
for each element, assuming ionization equilibrium and including all the
processes described above. Since the recombination and collisional
ionization terms depend on the adopted kinetic temperature ($T_{e}$), the
ionization balance is iterated to consistency with the calculation of
$T_{e}$ (see below) for fixed
Monte Carlo estimators.
\subsection{Thermal balance}
\label{section:tb}
In Paper~I, $T_{e}$ was assumed to be uniform and
treated as a model parameter. In order to address more general and
realistic wind conditions, here we relax this assumption and obtain
an estimate of $T_{e}$ as a function of position via a simplified
treatment of the heating and cooling rates in the wind and the
assumption of thermal equilibrium.
\subsubsection{Heating rates}
The treatment includes heating by
Compton scattering, free-free absorption and bound-free absorption of
the X-ray radiation field described by the Monte Carlo
simulations. Estimators for each of these heating processes in each
grid cell can be readily constructed following \citet{lucy03}. For
Compton scattering, the heating rate in a grid cell ($p$) is given by
\begin{equation}
H^C_p = \frac{n_{e}}{V_{p} \; \Delta t} \sum \bar{f}(\nu) \, \sigma(\nu) \,
\epsilon \, \mbox{d} s
\end{equation}
where the summation runs over all packet trajectories within the grid
cell $p$, $\sigma(\nu)$
is the Compton cross-section for the frequency ($\nu$) of the packet
and $\bar{f}(\nu)$ is the mean energy lost per Compton scattering
event. The free-free heating rate is
\begin{equation}
H^{ff}_p = \frac{1}{V_{p} \; \Delta t} \sum \kappa_{ff}(\nu) \,
\epsilon \, \mbox{d} s \; \; .
\end{equation}
Again, the summation runs over all packet trajectories in the cell.
The
free-free absorption coefficient is given by
\begin{equation}
\kappa_{ff}(\nu) = 3.69 \times 10^8 \nu^{-3} T_{e}^{-1/2} n_{e} (1 -
e^{-h\nu/kT_e}) \sum_{I} Z_{I}^2 N_{I} \; \; \mbox{cm$^{-1}$}
\label{eqn:ffop}
\end{equation}
where the summation runs over all ions ($I$) and $Z_{I}$ is the ion
charge.
Heating rates for photoabsorption are obtained in a similar fashion to
the ionization rate estimators in \ref{sect:Gamma}. For bound-free
absorption involving outer shells, level-by-level estimators are
recorded and used to obtain a total heating rate for each ion $I$,
\begin{equation}
H^{bf}_{I,k} = \sum_{j \in I} n_{j,k} h^{bf}_{j,k}
\end{equation}
where
\begin{equation}
h^{bf}_{j,k} = \frac{1}{V_{p} \; \Delta t}\sum_{\nu > \nu_0}
{a_{j}(\nu)} \left({1 - \frac{\nu_{0}}{\nu}}\right) \epsilon \, \mbox{d} s \; \;.
\end{equation}
Similarly, the heating rate due to K-shell photoabsorption by ion $I$ is given by
\begin{equation}
H^{K}_{I,k} = \frac{N_{I}}{V_{p} \; \Delta t}\sum_{\nu > \nu_K}
{a^{K}_{I}(\nu)} \left({1 - \frac{\nu_{K}}{\nu}}\right) \epsilon \, \mbox{d} s \; \;.
\end{equation}
All other possible heating sources are neglected, including any contributions
from outside the spectral region of the simulated radiation field or
by any non-radiative processes.
\subsubsection{Cooling rates}
\label{sect:cooling}
Cooling rates are required for both the treatment of $k$-packets (see
Section~\ref{sect:kpkt}) and to obtain an estimate of the local kinetic temperature.
Cooling due to radiative recombination, electron
collisional excitation, free-free emission, Compton cooling by
low-energy photons and the adiabatic expansion of the outflow are
included in this calculation.
The cooling rates due to spontaneous radiative recombination
$C^{bf}_{I,k}$ and electron collisional excitation $C^{cl}_{I,k}$ are obtained
following exactly Lucy (2003; his equations 31 and 33).
The free-free cooling rate is
\begin{equation}
C^{ff}_{p} = 1.426 \times 10^{-27} {T_{e}^{1/2}} n_{e} \sum_{I} Z_{I}^2
N_{I} \; \mbox{ergs~cm$^{-3}$~s$^{-1}$}
\end{equation}
where we follow \citet{lucy03} in setting the mean Gaunt factor to
one. The adiabatic
rate cooling rate is approximated by
\begin{equation}
C^{a}_{p} = 1.5 n_{g} k T_{e} {\mbox{\boldmath$\nabla\cdot v$}}
\end{equation}
where $n_{g}$ is the total particle density and {\boldmath$\nabla\cdot
v$} is the divergence of the velocity, evaluated at the midpoint of the cell.
The cooling rate for a population of non-relativistic electrons due to Compton up-scattering of low-energy photons ($h
\nu << k T_{e}$) can be estimated from the local energy density of such
photons $U_{\gamma}$ via
\begin{equation}
C^C_{p} = 4 \sigma_{T} U_{\gamma} \frac{k T_{e}}{m_{e} c}
\end{equation}
where $\sigma_T$ is the Thomson cross-section (see \cite{frankbook}, p. 180).
The cooling rate due to Compton up-scattering of photons is expected to
be dominated by interactions with the low energy photons
originating from the accretion disk since they should significantly
outnumber the X-ray photons for which our detailed radiative transfer
calculations are performed. An accurate treatment of Compton cooling
would therefore require that the entire bolometric light output of
the system be considered. However, since our primary concern is only
to obtain a reasonable estimate of the kinetic temperature, we
avoid this complication and derive an estimate for the local value of
$U_{\gamma}$ from the total bolometric luminosity of the source
($L_{\mbox{\scriptsize bol}}$, which is treated as an input parameter
for the model) and an assumption of spherical
geometric dilution, i.e.
\begin{equation}
U_{\gamma} = \frac{L_{\mbox{\scriptsize bol}}}{4 \pi r^2 c}
\label{eqn:compcool}
\end{equation}
where $r$ is the distance from the centre of the grid cell to the
coordinate origin. This expression is only a crude estimate for $U_{\gamma}$
since it neglects both the geometry of the emission
regions for low-energy photons (i.e. it assumes a centrally
concentrated point source) and the opacity of the
outflow to low-energy photons. Nevertheless, it provides a convenient
estimate of the Compton cooling rate
that introduces no additional computational demands
and should be reasonably accurate, at least for the
inner (hottest) regions of the wind.
\subsubsection{The kinetic temperature determination}
\label{sect:get_tb}
To obtain the kinetic temperature ($T_{e}$) for each grid cell, the
radiative heating rates computed from the Monte Carlo simulation are
compared with cooling rates to estimate the temperature at which
thermal equilibrium,
\begin{equation}
H^C_p + H^{ff}_p + \sum_{I} (H^{bf}_{I,k} + H^{K}_{I,k})
= C^{ff}_{p} + C^{a}_{p} + C^{C}_p + \sum_{I} (C^{bf}_{I,k} + C^{cl}_{I,k})
\label{eqn:tb}
\end{equation}
is established.
In practice, the thermal balance equation (equation~\ref{eqn:tb}) is solved
iteratively with the ionization balance equations to obtain
self-consistent estimates of both the
ionization fractions and $T_{e}$ for each grid cell.
\subsubsection{Excitation}
\label{sect:excitation}
In principle, the excitation state should be obtained from the
statistical equilibrium equations. However, to do so would require
computation and storage of
estimators for all bound-bound transitions in every cell which is
prohibitive for large computational grids. Therefore, we adopt a very
simplified treatment of excitation, namely we use the Boltzmann
distribution at the local kinetic temperature
\begin{equation}
\frac{n_{i}}{n_{\mbox{\scriptsize g.s.}}} = \frac{g_{i}}{g_{\mbox{\scriptsize g.s.}}} \exp(-\epsilon_{i}/kT_{e})
\end{equation}
where $n_{i}$ is the population of a state with statistical weight
$g_{i}$ and energy $\epsilon_{i}$, relative to the grounds state (g.s.).
For the wind models we will describe in Sections~\ref{sect_example} and \ref{sect_pg12},
our simplified treatment of excitation is not expected to
significantly affect the Fe K$\alpha$ region of the spectrum since the line
features which form there are mostly associated with low excitation
states of high-ionization state material. The soft regions ($\simlt
1$~keV) are more likely to be affected owing to the large numbers of
transitions between excited states of L-shell Fe and Ni ions which
contribute to the opacity at these photon energies.
\subsection{Monte Carlo simulations}
The radiation transport simulations are performed using the
scheme described in Sections 3.1 to 3.3 of Paper~I, modified as
described below.
\subsubsection{Initialization of packets}
As in Paper~I, an initial power-law spectrum of packet energies is adopted but
the input spectrum now extends from 0.1~keV up to 511~keV to allow
comparison with
observational constraints from instruments with significant effective
area at relatively high photon energies.
To account for the effects of a small but non-zero angular size of the
X-ray emission region, the packets are no longer initialized exactly
at the coordinate origin but in a spherical region around the
origin. The radial extent of the emission region, $r_{er}$, is a
parameter in the model and is generally chosen to be several
gravitational radii, as appropriate for an X-ray
emission region of size comparable to the innermost
radii of an accretion disk.
\subsubsection{Propagation of packets}
Compton scattering, bound-free absorption and line absorption are
treated exactly as described in Paper~I, except that bound-free
absorption by a few low-lying metastable levels is included for most
ions.
Free-free opacity from all ions is included using
equation~\ref{eqn:ffop}. Following free-free absorption, packets are
always converted to thermal energy ($k$-packets; see \citealt{lucy03}).
K-shell photoionization is included as an opacity source for
all L-shell and M-shell ions. Details of the outer electron
configuration are neglected such that all states of an ion contribute
to the same opacity term. When K-shell photoabsorption occurs in the
simulations, the packet either activates a macro atom or is converted
to a $k$-packet -- this is exactly analogous to the
treatment of outer-shell bound-free processes which are discussed in
detail by \cite{lucy03}. Here, the probability of $k$-packet
conversion is simply given by
\begin{equation}
p_{k} = 1 - \nu_K/\nu
\end{equation}
where $\nu$ is the absorption frequency and $\nu_K$ is the
relevant K-shell absorption edge frequency.
If the outcome of K-shell photoionization is activation of a macro
atom, the state activated must have a K-shell vacancy in its
configuration and will
generally be highly excited \footnote{Conversely,
following activation by outer-shell bound-free absorption, macro atoms
are always assumed to be in the ground
state of the newly formed ion.}.
Which of the K-shell vacancy
states in the newly formed ion is activated is chosen randomly by
sampling the statistical weights of the viable states in the
atomic data set.
\subsubsection{Macro Atom processes}
The macro atom treatment used in Paper~I has been extended to
incorporate autoionization, bound-free transitions from excited states and K-shell
photoionization.
Autoionization allows activated macro atom states to de-excite by
conversion to a $k$-packet and to make internal transition to higher
ionization states. Following the argumentation of \cite{lucy02}, the
deactivation probability (macro atom $\rightarrow$ $k$-packet)
\begin{equation}
p^D = N_{i} A_{a} (\epsilon_i - \epsilon_f)
\end{equation}
where $A_{a}$ is the rate coefficient ([s$^{-1}$]) for
autoionization from initial state $i$ to final state $f$,
$\epsilon_i$ and $\epsilon_f$ are the energies (excitation
plus ionization) of the states, and $N_{i}$ is the
normalization factor. Similarly, the
probability of making an internal macro atom jump $i \rightarrow f$ is
\begin{equation}
p^J_{i \rightarrow f} = N_{i} A_a \epsilon_f \; \; .
\end{equation}
This formulation requires that the autoionization process
connects two well-defined states in the atomic data set. In our
calculations, this is done for L-shell ions using level-to-level
autoionization rates (see Section~\ref{sect_atomicdata} for data
sources).
However, for the M-shell ions
we do not consider the target level in detail but assume
that the energy flow in the autoionization process is dominated by the
energy carried by the Auger
electron -- i.e. we assume that $\epsilon_f << \epsilon_i$ such that
$p^J/p^D \approx 0$. This has the advantage that we do not require
atomic level-to-level autoionization rates but only total Auger-widths
for vacancy states, as are available from the literature (see Section~{\ref{sect_atomicdata}).
For L-shell ions di-electronic recombination, is also
included in the macro atom scheme.
Di-electronic rate coefficients, $\alpha_d$, are obtained
from $A_a$ in the usual manner and used to formulate an
internal transition probability between the states
\begin{equation}
p^{J}_{i \rightarrow f} = N_{f} \alpha_d \epsilon_f n_{e}
\end{equation}
where $n_{e}$ is the local electron number density. Note that, as for
e.g. bound-free absorption, there is no macro-atom deactivation
probability associated with $\alpha_d$.
The $\gamma$ estimators described in Section~\ref{sect:Gamma} are used
to compute bound-free macro-atom internal transition probabilities
following \cite{lucy03}. In all such cases, it is assumed that the
relevant upper state is the ground state of the target ion. Similarly,
the $\gamma^{K}$ estimators are used to include internal transitions
associated with inner shell ionization. These are applied to all
energy levels of the absorbing ion lying below the first ionization
potential and connect to K-vacancy states of the next higher
ionization state. The choice of which K-vacancy state to activate
following an internal transition is made
by randomly sampling their statistical weights.
L-shell ionization of M-shell ions is not included in the macro atom treatment.
\subsection{Accretion disk}
\label{sect:disk}
We assume that the black hole is surrounded by an optically thick
accretion disk that lies in the $xy$-plane and extends from an inner
radius $r_{d}$ outwards throughout the simulation domain. $r_{d}$ is
treated as a model parameter but set equal to 6~$r_{g}$ (where $r_{g} = G M_{bh} / c^2$) for all the
simulations performed here.
During the Monte Carlo simulations, we also assume that all photon
packets that strike the disk are lost to the X-ray regime and they
are removed from the simulation.
For our particular choice of source
geometry, the disk subtends a modest solid angle as seen
by the X-ray source ($\sim 2.8$~sr),
meaning that a fraction of the primary X-ray photons strike the inner
regions of the disk
directly, without interaction in the wind.\footnote{This fraction, however, is
not physically meaningful in our model since it is merely
determined by our simple choice for the geometry of the X-ray
emitting region.}
The total flux striking the disk, however, is greater than this since
the wind effectively causes irradiation
of the outer parts of the disk
via scattering and emission in the wind.
Physically, these photons will interact
with the disk and may lead to an additional component of disk
reflection in the X-ray spectrum. We neglect any such component here
but it will be investigated in a future study.
\subsection{Thermal emission and $k$-packets}
\label{sect:kpkt}
In paper~I, it was assumed that energy packets converted to
$k$-packets would be lost to the hard X-ray band. However, since
cooling rates are now computed in order to impose thermal equilibrium (Section~\ref{sect:cooling}),
these are now used to re-emit the packet, thus simulating the thermal emission of the outflow.
When conversion to a $k$-packet occurs,
the cooling terms described in Section~\ref{sect:cooling} are first
randomly sampled to choose a cooling process. If $C^{a}$ is chosen,
the energy is assumed to be lost to the radiation field.
The collisional cooling processes ($C^{cl}$) lead to activation of a
macro atom state, the particular state being chosen by sampling the
terms contributing to $C^{cl}$ (see \citealt{lucy03}). The full macro
atom machinery then determines the outcome of this event.
Choosing $C^{fb}$ or
$C^{ff}$ leads to conversion to photon packets. Both
processes are assumed to emit isotropically in the co-moving frame and
the emission frequency is selected using equations 26 and 41 of
\cite{lucy03}. Note that this emission rule for bound-free processes
is not exact here since it assumes a simplified cross-section shape
but it is significantly less computationally demanding that sampling
the real emissivity and is adequate for current purposes.
\subsection{Extraction of spectra}
\label{section:extract}
Although spectra may be obtained by directly binning the emergent
packets by frequency and direction of travel, as done in Paper~I, this
has the disadvantage that very large numbers of quanta must be
simulated to suppress Monte Carlo noise, particularly when the spectrum
is expected to depend strongly on the observer line-of-sight. As
discussed by e.g. \cite{lucy99}, that approach is far
from optimal and does not make full use of the packet trajectories
during the Monte Carlo simulation.
Ideally, volume-based Monte Carlo estimators should be used to obtain
emissivities throughout the volume of the simulation
and used to obtain spectra via a
formal solution of the radiative transfer equation
(see discussion e.g. \citealt{lucy99}). However, this approach is very
demanding of memory resources: for our atomic data set (with $\> 10^5$ atomic processes) and
computational
grids (100$\times$100 wind cells), it ideally requires the storage of $\sim
10^9$ floating-point estimators which is prohibitive. Therefore we
extract spectra using a approach similar to that implemented in the
disk-wind simulations of \cite{long02} which, although less efficient,
does not make significant demands on memory consumption.
Viewing-angle dependent spectra are only extracted during the final
Monte Carlo simulation, once the ionization and thermal properties of
the outflow have been iterated to consistency with the radiation
field. The set of lines-of-sight for which spectra are required is
pre-specified and then every {\it physical} event (meaning emission or
scattering of a photon-packet) that occurs in the
Monte Carlo simulations is used to compute a contribution to each of
the spectra in the following manner.
When a physical event takes place in the Monte Carlo simulation, the
propagation of the packet is temporarily suspended and the probability
per unit solid angle
of the the associated photon-packet having been scattered into each of
the requested lines-of-sight is computed, $\mbox{d} P(${\boldmath
$n$}$)/\mbox{d} \Omega$ where the vector {\boldmath $n$} identifies the
line-of-sight. Depending on the physical process responsible for the
packet emission/scattering, $\mbox{d} P(${\boldmath
$n$}$)/\mbox{d} \Omega$ may depend on previous properties of the packet
(e.g. its incoming direction in the case of Compton scattering) or on
the local outflow properties (e.g. the Sobolev escape probability in
the case of line emission). We then integrate the total optical depth,
$\tau(${\boldmath $n$}$)$,
from the position of the physical event to the edge of the wind in the
direction {\boldmath $n$}. $\tau(${\boldmath $n$}$)$ is computed
exactly as in the Monte Carlo simulation,
incorporating contributions from all physical processes included in
the code and allowing for Doppler shifts to bring the photons into
resonance with spectral lines. With $\mbox{d} P(${\boldmath
$n$}$)/\mbox{d} \Omega$ and $\tau(${\boldmath $n$}$)$ computed, we then add
an energy contribution from the event to the spectrum associated with
direction {\boldmath $n$} of
\begin{equation}
\epsilon_{rf} \frac{\mbox{d} P(\mbox{\boldmath $n$})}{\mbox{d} \Omega} e^{-\tau(\mbox{\boldmath $n$})}
\end{equation}
where $\epsilon_{rf}$ is the rest-frame energy of the packet and the
contribution is added to the frequency bin associated with the
rest-frame frequency of the packet. This calculation is repeated for
each of the lines-of-sight for which spectra have been requested and
then the Monte Carlo simulation proceeds until the next physical event
occurs. In this way, we obtain high signal-to-noise spectra without
resorting to coarse angular binning of the emergent spectra or requiring
the storage of prohibitively large numbers of Monte Carlo estimators.
\subsection{Rescaling of model parameters}
\label{sect_rescale}
Although our numerical simulations are performed adopting a
particular value for the mass of the central black hole
($M_{\mbox{\scriptsize bh}}$), the spectra obtained are applicable for
other $M_{\mbox{\scriptsize bh}}$-values following
a rescaling of the dimensional wind
parameters. Specifically, the ionization state, optical depths and
velocity-law in the wind are preserved under a global rescaling where
all the model luminosity parameters ($L_{\mbox{\scriptsize bol}}$,
$L_{X}$) scale with the Eddington
luminosity ($L_{\mbox{\scriptsize Edd}}$),
all lengths ($r_{\mbox{\scriptsize min}}$, $r_{\mbox{\scriptsize max}}$,
$d$, $R_{v}$, $r_{d}$ and $r_{\mbox{\scriptsize em}}$) are scaled to
the gravitational radius ($r_{g}$) and the mass-loss rate ($\dot{M}$) is scaled
to the Eddington accretion rate ($\dot{M}_{\mbox{\scriptsize Edd}}$).
\subsection{Atomic data}
\label{sect_atomicdata}
Table~\ref{tab_atoms} lists the elements and ionization stages for
which atomic models are
included in the radiative transfer calculations. Compared to Paper~I,
the L-shell ions of the astrophysically abundant intermediate-mass and
iron group elements have been added (Mg, Si, S, Ar, Ca, Fe and
Ni). The highest M-shell ions (down to the Cl-like ion) have also been added for
Fe and Ni.
The solar element abundances of \citet{asplund05} are adopted.
\begin{table}
\caption{Elements and ions that are included in the radiative transfer
calculations.}
\label{tab_atoms}
\begin{tabular}{ccccc}
Element & Ions &~~~~~~~~ & Element & Ions\\ \hline
C & {\sc iv -- vii} & &S & {\sc viii -- xvii}\\
N & {\sc v -- viii} & &Ar &{\sc x -- xix}\\
O & {\sc vi -- ix} & &Ca &{\sc xii -- xxi}\\
Ne & {\sc viii -- xi} & &Fe & {\sc x -- xxvii}\\
Mg & {\sc iv -- xiii} & &Ni & {\sc xii -- xxix}\\
Si & {\sc vi -- xv} & & \\ \hline
\end{tabular}
\end{table}
In all cases, inner-shell photoionization cross-sections were taken
from the fits by \cite{verner95}. We note that, although convenient,
those data do not account for smoothing of the K edges by broad
resonances below threshold (see
\citealt{palmeri02,kallman04}). Depending on the ionization state, this
effect could be critical for quantitative analysis of the region around the
Fe K edge but it is not important for the He- and H-like ions
which dominant in most of the wind models we describe in
Sections~\ref{sect_example} and \ref{sect_pg12}.
For outer-shell photoionization from
ground configurations (and also low-lying metastable states),
cross-sections were taken from \cite{verner96} and a simple hydrogenic
approximation was used for cross-sections from more highly excited states.
Atomic energy levels, bound-bound $A$-values and electron collision
strengths were extracted from the CHIANTI atomic database, version 5
\citep{dere97, landi06} for all the included L- and M-shell ions of C to
Fe.
For the M-shell ions of Fe, the data were limited to those of the
configurations included in the study by \cite{mendoza04}. In all
cases, the $T_{e}$ dependence of the electron collision strength was
neglected and the low-temperature limit adopted.
Autoionization rates for K-vacancy states in L-shell ions of Fe
were taken from \cite{palmeri03} and Auger widths for the K-vacancy states of the M-shell Fe
ions from \cite{mendoza04}.
The total
autoionization rates for the K-vacancy states of Mg, Si, S, Ar and Ca were taken from
\cite{palmeri08} and the branching ratios of the autoionization
process into the specific final states of the target ion were assumed
to be the same for the iso-electronic Fe ion.
For the included L- and M-shell ions of Ni ({\sc xii -- xix}), energy
levels, Einstein A-values and total Auger widths were taken from
\cite{palmeri08b}. Again, the L-shell autoionization branching ratios were
assumed to be the same as in the iso-electronic Fe ion.
Radiative recombination rates are taken from \cite{gu03}, where
available, or else from \cite{verner96b} (H-, He- and Li-like ions of
C, N, O and Ne and Na-like ions of Fe and Ni),
\cite{arnaudray} (Fe~{\sc xvi}) or \cite{shull82,shull82a}, otherwise.
Di-electronic rates were obtained from \cite{gu03b}, supplemented by
those from \cite{arnaudray} (M-shell Fe ions), \cite{shull82} (M-shell
Ni ions) and \cite{arnaudroth} (He- and Li-like C, N, O and Ne).
Collisional ionization rate coefficients ($C_{DI}$ and $C_{EA}$) are
obtained from \cite{arnaudray} for Fe and \cite{arnaudroth} for all
other elements.
\subsection{Wind models}
The modification described above mean that the code is now capable of
simulating all the necessary atomic physics to compute X-ray
spectra for wind models and several such calculations will be
presented in the Sections below.
For the moment, we continue to work with the
simply-parameterized, smooth outflow models exactly as described in
Paper~I. This approach allows us to investigate outflow signatures and
their sensitivity to the major wind parameters in a relatively
simple manner. In the
future, however, we plan to extend our studies
to models that go beyond the smooth,
steady-state flow prescriptions we adopt here. Numerical simulations
clearly suggest that outflows are likely to have complex structure
that will affect the observed spectra (see
e.g. \citealt{proga04,schurch09}) and time variation in the outflow
properties may well have a role in explaining the observed time
dependence of some spectral features.
\section{Test of ionization balance}
\label{sect:iontest}
The most significant improvements in the physics of the current code
compared to that described in Paper~I are the extension to lower
ionization states and the substantially more sophisticated treatment
of the ionization balance. To test these improvements, we have made a
simple comparison with the ionization balance in the well-known
photoionization
radiative transfer code {\sc cloudy} (version 07.02, last described
by \citealt{ferland98}). For the test, we used a wind model with very
low column density such that it would be optically thin and every cell
would be illuminated with the pure power-law continuum from the X-ray
point source. A primary photon power-law index of $\Gamma = 1.2$ was
used for the test and the kinetic temperature was forced to be
$10^6$~K everywhere. Ionization fractions were computed using
the full machinery of our code and the results for Fe ions are plotted in
Figure~\ref{fig:ioncompare} versus an ionization parameter defined as
\begin{equation}
\xi = \frac{L_{0.1 - 50}}{r^2 n_H}
\end{equation}
where $L_{0.1 - 50}$ is the source luminosity between 0.1 and
50~keV. The Fe ionization fractions obtained with {\sc cloudy} for a
low column-density spherical shell are over-plotted for comparison.
\begin{figure}
\epsfig{file=fig1.eps, width=8cm}
\caption{
Ionization fractions as a function of ionization parameter
for Fe {\sc xvi} -- {\sc xxvii} computed for
optically thin conditions with {\sc cloudy} (solid lines) and our
Monte Carlo code (diamond symbols). In both calculations, a power-law
continuum with photon index $\Gamma = 1.2$ and a kinetic temperature
of $T_{e} = 10^6$~K were adopted.
}
\label{fig:ioncompare}
\end{figure}
The agreement between the ionization calculations in the two codes is
very good, better than 10 per cent for most of the K- and L-shell
ions of iron. This precision is comparable to the accuracy of the
source atomic data. Our calculations become less reliable for
ionization parameters much below $\xi \sim 100$~ergs~cm~s$^{-1}$ since our
current atomic data set does not contain all the necessary ions (in
our test calculation, Fe~{\sc ix} -- which we neglect --
has a significant role in the {\sc cloudy} ionization balance below $\xi \sim 50$~ergs~cm~s$^{-1}$).
\section{Example calculations}
\label{sect_example}
\subsection{Model parameters}
Here we present details of two outflow models computed with the
improved code. The first, Model~A has the same outflow parameters as
the example model described in Paper~I. However, the physical
properties of the flow differ slightly from Paper~I since the kinetic
temperature is no longer a parameter but is determined as described in
Section~{\ref{sect:get_tb}. Also, since the new code version
allows for a finite spatial extent of the primary X-ray
emission region and occultation of the
scattered/reprocessed light (see Section~\ref{sect:disk}), the relative
amplitudes of the direct and scattered/reprocessed components of the
spectrum are altered (see Section~\ref{sect:spec}).
Our second model, Model~B, has parameters chosen to illustrate the
effects of lower ionization state material in the wind as can now be
treated with the improved code version. We adopted a black-hole
mass of $10^7$~M$_{\odot}$
and an outflow that is wider than those considered
in Paper I, extending across a factor of three range of launching
radius on the disk (50 -- 150 $r_{g}$). Since the ionizing source is
centrally concentrated this means that the ionization gradient across
the wind (which was already present in Paper I) now extends to even
lower ionization states around the outer edge. We have also chosen a
more gradually accelerated outflow with lower terminal
velocity. Specifically, we adopt a larger value of the
velocity-law acceleration length parameter $R_{v}$ (which specifies how far downstream the
outflow reaches one half of the terminal speed; see equation 1 of
Paper~I) and a smaller value of $f_{v}$
(which relates the terminal speed to the escape speed at the base of a
streamline; see equation 2 of Paper~I).
The X-ray source parameters and the wind mass-loss rate for
Model~B were
motivated by previous studies of the bright quasar PG1211+143 with
which more detailed comparisons are made in Section~\ref{sect_pg12}.
The primary X-ray source luminosity and the total bolometric
luminosity were chosen to give
ratios to the Eddington luminosity ($L_{\mbox{\scriptsize Edd}}$) similar to those
inferred for PG1211+143: adopting a typical 2 -- 10~keV luminosity
of $10^{44}$~ergs~s$^{-1}$ and bolometric luminosity of
$4 \times 10^{45}$~ergs~s$^{-1}$
\citep{pounds03} yields $L_{X}/L_{\mbox{\scriptsize Edd}} \sim
0.02$ and $L_{\mbox{\scriptsize bol}}/L_{\mbox{\scriptsize Edd}} \sim
0.8$ for a black hole mass $\sim 4 \times 10^7$~M$_{\odot}$
\citep{kaspi00} and accretion efficiency of $\sim 0.06$.
The power-law index ($\Gamma \sim 1.8$)
and mass-loss rate ($\dot{M} \sim \dot{M}_{\mbox{\scriptsize Edd}}$)
were also chosen to be roughly appropriate for PG1211+143, as
motivated by the discussion of \citet{pounds03}.
The complete set of model parameters is given in
Table~\ref{tab_param}; definitions of the geometry and velocity-law
parameters are given in Paper~I.
For both models, we placed the innermost edge for the accretion disk
at the last stable
orbit for a Schwarzschild black hole ($6 r_{g}$) and assumed a
comparable size for the region of X-ray emission (again $6
r_{g}$).
\begin{table*}
\caption{Inputs parameters for the example model.}
\label{tab_param}
\begin{tabular}{lll}\\ \hline
Parameter & Model A & Model B \\ \hline \hline
mass of central object, $M_{bh}$ & $4.3 \times 10^6$ M$_{\odot}$ &$10^7$ M$_{\odot}$ \\
bolometric source luminosity, $L_{\mbox{\scriptsize bol}}$ & $2.5 \times 10^{44}$
ergs~s$^{-1}$ ($\sim 0.6
L_{\mbox{\scriptsize Edd}}$)& $10^{45}$ ergs~s$^{-1}$ ($\sim 0.8
L_{\mbox{\scriptsize Edd}}$)\\
source luminosity (2 -- 10 keV), $L_{X}$ & $10^{43}$
ergs~s$^{-1}$ ($\sim 0.02
L_{\mbox{\scriptsize Edd}}$) & $2.5 \times 10^{43}$
ergs~s$^{-1}$ ($\sim 0.02
L_{\mbox{\scriptsize Edd}}$)\\
source power-law photon index, $\Gamma$ & $2.38$ & $1.8$ \\
range of source photon energies in simulation & 0.1 -- 511~keV & 0.1 -- 511~keV \\
size of emission region, $r_{er}$ & $6 r_g = 3.8 \times 10^{12}$ cm & $6 r_g = 8.9 \times 10^{12}$ cm \\
inner radius of disk,$r_{d}$ & $6 r_g = 3.8 \times 10^{12}$ cm & $6 r_g = 8.9 \times 10^{12}$ cm \\
inner launch radius, $r_{\mbox{\scriptsize min}}$ & 100 r$_g = 6.4
\times 10^{13}$ cm & 50 r$_g = 7.4 \times 10^{13}$ cm \\
outer launch radius, $r_{\mbox{\scriptsize max}}$ & $1.5
r_{\mbox{\scriptsize min}}$ & $3 r_{\mbox{\scriptsize min}}$ \\
distance to wind focus, $d$ & $r_{\mbox{\scriptsize min}}$ & $r_{\mbox{\scriptsize min}}$ \\
terminal velocity parameter, $f_{v}$ & 1.0 & 0.5 \\
velocity scale length, $R_{v}$& $r_{\mbox{\scriptsize min}}$ & $3 r_{\mbox{\scriptsize min}}$\\
velocity exponent, $\beta$& 1.0 & 1.0 \\
launch velocity, $v_{0}$ & 0.0 & 0.0 \\
wind mass-loss rate, $\dot{M}$ & 0.1 M$_{\odot}$
yr$^{-1}$ ($\sim 0.6 \dot{M}_{\mbox{\scriptsize Edd}}$)$^{a}$
& 0.38 M$_{\odot}$
yr$^{-1}$ ($\sim \dot{M}_{\mbox{\scriptsize Edd}}$)$^{a}$\\
mass-loss exponent, $k$ & -1.0 & -1.0 \\
outer radius of simulation grid & {$5 \times 10^{16}$ cm}& {$5 \times 10^{16}$ cm}\\
3D Cartesian RT grid cells & {$180 \times 180
\times 180$} & {$180 \times 180 \times 180$}\\
2D wind grid zones & {$100 \times 100$} &{$100 \times 100$} \\
\hline
\end{tabular} \\
$^{a}$ The Eddington accretion rate, $\dot{M}_{\mbox{\scriptsize
Edd}}$ was computed assuming a radiative efficiency of six percent.
\end{table*}
\subsection{Thermal and ionization structure}
Figure~\ref{fig:thermalstate} shows the distribution of kinetic
temperature and Fe ionization states in the example models.
For Model~A,
the spatial variation of the ionization state is qualitatively similar
to that discussed in Paper~I: the inner edge of the wind
is almost fully ionized and gradients of decreasing ionization occur
both along the outflow and across the base of the flow. Model~B is similar but,
owing to the
larger radial extent of the flow launching region in this model, the
gradient across the flow is particularly well-developed such that a
substantial region on the outside of the wind is dominated by L-shell
Fe ions.
As expected, the regions of the wind that are most strongly ionized
are also most strongly heated by the X-ray source leading to
significant kinetic temperature variations across the flow (right
panels of Figure~\ref{fig:thermalstate}).
\begin{figure*}
\epsfig{file=fig2a.ps, width=5.8cm,angle=90}
\hspace{-1.1cm}
\epsfig{file=fig2b.ps, width=5.8cm,angle=90}\\
\epsfig{file=fig2c.ps, width=5.8cm,angle=90}
\hspace{-1.1cm}
\epsfig{file=fig2d.ps, width=5.8cm,angle=90}
\caption{
Distribution of mean Fe ionization state (left) and
kinetic temperature (right) for
Models~A (top) and B (bottom).
The dashed lines indicate the five
lines-of-sight for which spectra are shown in Figures~3 and 4.
}
\label{fig:thermalstate}
\end{figure*}
\subsection{Computed spectra}
\label{sect:spec}
For both the example models, we computed emergent spectra for twenty
lines-of-sight uniformly sampling the orientation ($0 < \mu <
1$, where $\theta = \cos^{-1} \mu$ is the
angle between the line-of-sight and the
rotation axis). Figures~\ref{fig:examplespec_old} and
\ref{fig:examplespec} show the computed spectra for five of these
lines-of-sight for Models~A and B, respectively.
The orientations of the chosen lines-of-sight are indicated by the dashed
lines in Figure~\ref{fig:thermalstate}.
As in Paper~I, the propagation
histories of the Monte Carlo quanta are used to divide the emergent
spectrum into a direct component (representing photons that reached
the observer without any interactions) and a scattered/reprocessed
component derived from the quanta that underwent at least one
interaction in the wind. These contribution are separately plotted in
the left panels of Figures~\ref{fig:examplespec_old} and \ref{fig:examplespec}.
The right panels show the 2 -- 10 keV region of the spectra in
greater detail.
\begin{figure*}
\epsfig{file=fig3.eps, width=12cm}
\caption{
Spectra for five observer lines-of-sight computed from Model~A
(from top to bottom, $\mu = 0.875$, 0.675, 0.575, 0.475 and 0.275
where $\theta = \cos^{-1} \mu$ is the angle of the line-of-sight relative to the polar
axis). The left panels show the spectrum from 0.1 -- 200 keV while the
right panels show the 2 -- 10 keV region in detail for the same lines-of-sight.
Note that the abscissa is plotted logarithmically in the left panels but linearly in
the right panels.
In the left panels, the total spectrum is shown in black, the spectrum
of direct photons in red and the scattered/reprocessed spectrum in
blue.
In the right panels, the total spectrum computed with the new code
version is plotted in black while the spectra from the example model in
Paper~I are shown in green, for comparison.
The vertical lines indicate the mean rest energies
of the Fe~{\sc xxv}/{\sc xxvi} K$\alpha$ transitions ($\sim$6.7/6.97~keV).
All the spectra are normalised to the incident X-ray spectrum which
has a photon power-law index, $\Gamma$=2.38. Monte Carlo noise from
the simulations is
present in all the spectra and is responsible for the small-scale
fluctuations.
}
\label{fig:examplespec_old}
\end{figure*}
\begin{figure*}
\epsfig{file=fig4.eps, width=12cm}
\caption{
As Figure 3 but showing spectra from Model~B. No comparison spectra
are shown in the right panels.
All the spectra are normalised to the incident X-ray spectrum which
has a photon power-law index, $\Gamma$=1.8.
}
\label{fig:examplespec}
\end{figure*}
The first line of sight shown for both models (top panels of
Figures~\ref{fig:examplespec_old} and \ref{fig:examplespec})
corresponds to an observer
at sufficiently low inclination that no portion of the wind obscures
the primary X-ray source.
Thus the direct component of radiation is
unaffected by the wind (red line in top left panels). However, since
the outflow scatters light from other lines-of-sight and produces its
own thermal emission, the direct light is supplemented by a
significant component of scattered/reprocessed radiation. As one would
expect, this scattered/reprocessed spectrum is qualitatively similar
to that obtained from standard disk reflection models
(e.g. \citealt{ross05}).
In particular, a moderately strong Fe~K$\alpha$ emission
line appears in both models.
This is predominantly formed by Fe~{\sc xxv} and {\sc
xxvi} since most of
the wind reflection seen from this line-of-sight occurs in the
highly-ionized inner surface of the outflow. The emission line profile
is Doppler broadened and, owing to the substantial outflow velocities,
develops an electron scattering wing \citep{auer72,laurent07} that
skews the profile to the red (for both models,
it extends down to $\sim 5$~keV for this line-of-sight).
At softer energies, the spectrum of scattered Monte Carlo quanta also
contains a variety of other emission features. These are considerably
stronger in Model~B owing to the greater range of ionization
conditions in the wind. In this model, weak but distinct
K$\alpha$ emission lines of O, Si and S
($\sim 0.65, 2.0$ and 2.6~keV, respectively) are present together with
forests of blended lines from the L-shell ions of Fe.
The emission profiles of these features are broadened by the same
mechanisms which affect the Fe K line leading to rather complex
spectra.
The scattered/reprocessed spectrum also adds a
distinct ``Compton hump'' to the total spectrum, causing it to have a
broad peak around 20 -- 30 keV; at even harder energies the scattered/reprocessed
spectrum bends down (owing to the energy dependence of the Compton
cross-section) such that the flat, primary X-ray spectrum becomes
increasingly dominant above about 50~keV.
For this inclination angle, the spectrum computed for Model~A is very
similar to that obtained for the same outflow parameters using the
code version in Paper~I (see upper right panel of
Figure~\ref{fig:examplespec_old}). The main difference is a modest
offset in normalization that arises from the reduced amplitude of the
scattered component of radiation owing to occultation by the accretion
disk which is included in the new calculations (see Section~\ref{sect:disk}).
The second line of sight considered ($\mu = 0.675$)
passes through the upper-most
layers of the outflow models which are hottest and most
highly-ionized. The remaining three lines-of-sight ($\mu = 0.575$, 0.475
and 0.275) pass through
increasingly denser and cooler parts of the wind such that the spectra
become more complex.
As in Paper~I, lines of sight through the wind show
weak, narrow, blueshifted absorption features associated with highly ionized
material, most importantly the K$\alpha$ lines of Fe~{\sc xxv} and
{\sc xxvi}. In Model~A, the spectra for $\mu = 0.675$ (and lower)
are characteristically similar to
that of the example model in Paper~I (see comparison in
Figure~\ref{fig:examplespec_old}) but there are some differences in
detail. In particular, with the improved code version
the amplitude of the component of scattered/reprocessed
light is reduced by disk occultation, the mean ionization state is
slightly lower and the absorption features are sharper. This last
effect arises because of the angular binning of MC quanta used to
extract the spectra in Paper~I: since the line-shifts depend on
inclination, spectra obtained via angular binning will tend to have
slightly smeared line profiles.
In Model B,
the
$\mu = 0.675$
line of sight
has an
integrated hydrogen column density of $N_{H} \sim 2 \times 10^{23}$~cm$^{-2}$, only a
factor of a few smaller than that
suggested for the most highly ionized material in the spectrum of
PG1211+143 ($\sim 5 \times 10^{23}$~cm$^{-2}$, \citealt{pounds03}).
However, the material
along this line of sight is hot and almost fully ionized
(see Figure~\ref{fig:thermalstate}) such that Compton scattering is the
only important opacity source. This reduces the flux of the direct
component of the radiation and causes it to curve upwards at high
energies, a consequence of the photon-energy dependence of the Compton
cross section. Although the scattered/reprocessed component of the
light still shows a significant Compton hump, this upwards curvature of
the direct component partially cancels it out leaving only a modest
bump in the total spectrum. The only significant absorption
feature imprinted by the wind for this line of sight is from the trace
population of Fe~{\sc xxvi} and only the K$\alpha$ line is
strong enough to be plausibly detectable with current instruments. The
emission features in the scattered/reprocessed spectrum are
qualitatively very similar to those for the lower inclination angle
described above except that the features are slightly broader thanks
to the greater line-of-sight components of both the outflow and
rotational velocity fields for this orientation. In addition, the
Fe~K$\alpha$ line shows two distinct peaks, one close to the rest
K$\alpha$ energy of
Fe~{\sc xxvi} and the other to that of Fe~{\sc xxv} and the high L-shell ions.
The third to fifth lines-of-sight for which spectra are
shown for Model B (Figure~\ref{fig:examplespec})
have sufficiently high column densities
($N_{H} \sim 10^{24}$, $\sim 4 \times 10^{24}$ and
$\sim 10^{25}$~cm$^{-2}$, respectively) that the direct
component of radiation is almost completely blocked except at rather
high energies ($\simgt 30$~keV). For the highest inclination angle
considered ($\mu = 0.275$), the reprocessed component of light
dominates even well above 100~keV such that the total spectrum of the
wind shows a
downturn at these energies, a property that can be constrained by
high-energy observations (see discussion by \citealt{fabian95}).
However, we stress that for the lower inclination angles in this model,
and indeed all the angles considered for Model~A,
the direct component rises sufficiently rapidly with energy
that there is still significant flux emerging above 200~keV.
For the $\mu = 0.575$, 0.475 and 0.275 lines-of-sight, the Model~B spectrum is progressively
more strongly affected by absorption in the wind. The Fe~K$\alpha$
absorption line becomes deeper and gradually switches from being
dominated by Fe~{\sc xxvi} to Fe~{\sc xxv} at higher inclination
angles. The associated K$\alpha$ emission line remains strong and its
apparent red wing becomes increasingly extended for high inclination angles.
As expected, the Fe~K$\alpha$ absorption is also accompanied by blueshifted
absorption by Fe~K$\beta$ ($\sim 9.0$~keV) and Ni~K$\alpha$ (Ni~{\sc
xxvii} at $\sim 8.5$~keV and {\sc xxviii} at $\sim 8.8$~keV). There is
also significant absorption around the blueshifted Fe~K edge which,
although somewhat smeared out by the Doppler shifts, causes the
distinct drop in flux at around 10~keV.
At softer energies, blueshifted absorption features associated with
the abundant light elements become increasingly prominent as the
inclination angle is increased.
For an inclination of $\mu = 0.475$ (fourth row in
Figure~{\ref{fig:thermalstate}), weak narrow K$\alpha$ absorption
lines of
S~{\sc xvi} ($\sim 2.8$~keV), Si~{\sc xiv} ($\sim 2.2$~keV), Mg~{\sc
xii} ($\sim 1.6$~keV), Ne~{\sc x} ($\sim 1.1$~keV) and O~{\sc viii} ($\sim 0.71$~keV)
are all present but their equivalent widths are small, generally
less than a few eV. Weak Fe~{\sc xxiv}
lines also appear at $\sim 1.2$ and 1.3~keV due to blueshifted
transitions between its ground 2p$^5$ configuration
and states of the excited 2p$^{4}$~($^{3}$P)3s configuration.
At the higher inclination angle of $\mu = 0.275$ (fifth row in
Figure~\ref{fig:examplespec}), the absorption spectrum is even
more complex. The K$\alpha$ transitions mentioned above are all still
present and generally somewhat stronger. Very weak K$\alpha$ lines due
to Ar and Ca also manifest ($\sim 3.6$ and 4.4~keV, respectively), as
do the K$\beta$ lines of the H-like ions of Mg, Si and S. Although
O~{\sc viii} K$\beta$ absorption is also significant in the
simulation, this feature is blended with comparably strong lines of
Fe~{\sc xviii} and {\sc xix}. Deep absorption occurs between 1.0 and
1.3~keV owing to the forest of lines arising in transitions from the low-lying $n=2$ configurations of Fe and Ni L-shell ions.
Furthermore, there is also significant bound-free absorption by the
ground states of the K-shell ions of O, Mg, Si and S and of the
L-shell ions of Fe that causes a drop in the flux above $\sim 1.5$~keV.
\section{Preliminary comparison with PG1211+143}
\label{sect_pg12}
Thanks to the extension of the capabilities of the code to deal with
lower ionization conditions, we can now use the models to compare
much more realistic theoretical spectra with observations to
establishing how readily outflow models can explain observed
spectroscopic features and, if real, what the physical properties of
these flows might be. For a fully detailed comparison a large grid of
theoretical models covering the physically interesting parameter space
must be computed. This will be an important next step in our studies
but we begin here by making a preliminary comparison with the observed
spectrum of the bright quasar PG1211+143.
\subsection{Observational data}
Amongst the first direct evidence for highly ionized massive outflows
from AGN was a report by \cite{pounds03,pounds05} based on the
analysis of
{\it XMM-Newton}
X-ray spectra of PG1211+143. They reported the detection of several
strong absorption features which they associated with blueshifted
K$\alpha$ transitions of C, N, O, Ne, Mg, S and Fe. To quantify the absorption
features in the (1 -- 10~keV) {European Photon Imaging Camera} (EPIC) pn spectrum they fit an {\sc xstar} \citep{kallman01}
absorption model and found that a high column density ($N_{H}
\sim 5 \times 10^{23}$~cm$^{-2}$) and a high ionization parameter
($\log (\xi / \mbox{ergs~cm~s$^{-1}$}) \sim 3.4$) were required to explain both the claimed Fe and S~K$\alpha$
absorption lines as originating in a rapid ($v \sim 0.08$c)
outflow. Based on the proposed identification of the lower ionization
Mg line in the EPIC pn spectrum and lines attributed to the lighter
elements in their contemporaneous {Reflection Grating
Spectrometer} (RGS) spectra, they also postulated
that additional absorbing components must be present in the outflow
with significantly lower column density and ionization parameter
(although comparable outflow speeds).
Broad
emission was also identified in the EPIC pn spectrum, extending from
$\sim 3$ to 7~keV.
Although this emission feature can be
fit as
an extreme relativistic Fe~K$\alpha$ emission line
originating from reflection by the
inner most regions of an ionized accretion disk \citep{pounds03},
it has more recently been suggested that it
may be P~Cygni emission physically associated with
the blue-shifted K$\alpha$ absorption (e.g. \citealt{pounds09}) as
would be expected from a wind with a large covering fraction.
Subsequent re-observations of PG1211+143 with {\it XMM-Newton} were
made in 2004 \citep{pounds07} and 2007 \citep{pounds09}. These showed
that, although the spectral properties of PG1211+143 are significantly
time-variable, the fast outflow of highly ionized gas is persistent
and confirmed that it is likely to be an energetically important
component of the system.
Given the complexity of the available data for PG1211+143 and the wide
range of spectroscopic features identified in the 2001 {\it XMM} data first reported
by \cite{pounds03}, this object provides an excellent point of
comparison with our synthetic spectra for outflows. In particular,
while previous efforts to quantify the spectra have generally adopted
distinct fit components for the claimed absorption lines and the
apparently broad emission redward of $\sim$7~keV, our wind models allow us
to explore the relationship between these features with a set of
physical models.
The X-ray spectrum of PG1211+143 shows a strong excess for energies
$\simlt 1$~keV relative to an extrapolation of a power-law fit to the
harder energy spectrum (see e.g. \citealt{pounds03}).
The origin of soft excesses in
AGN spectra remains unclear: although it has been suggested that they
may be a consequence of absorption by relatively low ionization
material and data showing strong soft excesses
(including those of PG1211+143; see \citealt{pounds07,pounds09})
have been successfully fit assuming an absorption-dominated origin,
the geometry and origin of the absorbing material is not
well-constrained.
In view of this
uncertainty, we restrict our fitting to the 2 -- 10~keV spectrum of
PG1211+143 to avoid
biasing our fit by a choice of particular model for the origin of the
soft excess. However, we will qualitatively discuss the lower-energy absorption
features predicted by the model compared to those
reported in the PG1211+143 spectra.
In the following, we will make quantitative comparisons with the
data set compiled by \cite{pounds09} by stacking {\it XMM} EPIC pn
data from 2001, 2004 and 2007. We have rebinned the data to the
energy-dependent half-width at half-maximum (HWHM) of the EPIC pn
instrument.
\subsection{Choice of model parameters}
We began by making a comparison of the stacked 2 -- 10~keV data of
PG1211+143 with our example spectra (described in
Section~\ref{sect_example}). As discussed above, Model~B already
indicated that many of the discreet features identified in the
spectrum of PG1211+143 can qualitatively be reproduced in our theoretical
spectra. To produce a strong, significantly blueshifted Fe
{\sc xxv}/{\sc xxvi} absorption feature, the example models suggest
that a line-of-sight passing through some portion of the inner, highly
ionized edge of the flow is required ($\mu \sim 0.4$ --
0.6 for the example model geometric parameters). Although Model~B
already produces strong K$\alpha$ absorption, the blueshift is
slightly too small compared to that suggested for PG1211+143
\citep{pounds03,pounds05,pounds06}. Therefore, we first increased the outflow
velocity parameter ($f_{v}$) from $1/2$ to $2/3$, a value intermediate
between those used for the two example models.
Upon quantitative comparison, the Model~B spectra also differ in detail
from the observations of PG1211+143 in that the lower ionization
features are somewhat too weak for the appropriate inclination angles
(e.g. the equivalent width of
the S~{\sc xvi} Ly$\alpha$ line is only $\simlt 5$eV in the model
while \cite{pounds05} and \cite{pounds06} report equivalent widths of
$\sim 40$ and $\sim 24$eV, respectively).
Thus, models with lower typical
ionization state are desirable.
In principle, a full grid of models exploring all the outflow
parameters should be explored to find the best match to the observed
spectrum. Here, however, we restrict ourselves to varying parameters
that are expected to directly affect the mean ionization state.
We therefore
created a small grid of six models that explores the effects of
higher wind mass-loss rates ($\dot{M}$) and longer velocity-law scale
lengths ($R_{v}$) while keeping all other parameters the same as for Model~B.
Relatively modest numbers of Monte Carlo quanta were
used for these simulations in order to reduce the computation demands
(only $\sim 10^6$ packets were used for the calculation of the spectra).
The quantitative effect of $\dot{M}$ on the models was discussed in
Paper~I and that discussion remains applicable here. In particular,
increasing $\dot{M}$ leads to higher densities and therefore lower
ionization states for a fixed X-ray source luminosity. In general, it
also leads to stronger absorption features thanks to the increased
column density for a fixed line of sight. Our grid of models explored
$\dot{M}$-values of 0.38, 0.76 and 1.52~M$_{\odot}$~yr$^{-1}$
corresponding to $\sim 1$, 2 and 4 $\dot{M}_{\mbox{\scriptsize Edd}}$ for our
adopted value of $M_{\mbox{\scriptsize bh}} \sim 10^7 M_{\odot}$.
The second parameter varied, $R_{v}$, affects the spectrum for two
distinct reasons. First, a larger value of $R_{v}$ means that the flow
accelerates more gradually such that the material in the accelerating
region is more dense for fixed values of $\dot{M}$ and
$f_{v}$. Secondly, the larger value of $R_{v}$ reduces the outflow
velocity-gradient such that the Sobolev optical depths of the line
transitions become larger in the acceleration region. For our grid we
considered $R_{v}$-values of 3 and 5~$r_{\mbox{\scriptsize min}}$.
As when making our comparison with the average spectrum of Mrk~766 in
Paper~I, we constructed a multiplicative
table (``mtable'') model for use with {\sc xspec} (version 11, \citealt{arnaud96})
from the ratio of our computed spectra to the primary power-law
adopted in the radiative transfer simulations. We then attempted to fit the
PG1211+143 data using a model comprising of a power-law (with
free normalization [$N$] and power-law index [$\Gamma$]) combined with this mtable
(which contains a total of 120 spectra from the grid of 3 $\times$ 2
$\times$ 20 in $\dot{M}$, $R_{v}$ and $\mu$). Given the strong but complex
dependence of the spectral features to the line-of-sight ($\mu$) we
did not allow this parameter to be fit directly but stepped through
the $\mu$-values of the grid manually, freezing this parameter and
then fitting the remaining four quantities ($N$, $\Gamma$, $\dot{M}$
and $R_{v}$).
We did not allow for any component of systematic
error in any of our fits as this is not expected to have a significant effect
in the fit to the 2 -- 10~keV region (see the discussion in \citealt{miller09}).
Although other physical processes are expected to have a
role in shaping the X-ray spectra of AGN (in particular, some degree
of disk reflection and some warm absorber component seem all but
inevitable), we did not include any additional component in this fit
since our objective is to establish whether our disk wind models alone could be
the dominant component that accounts for the major spectra features in the
2 -- 10~keV band.
For this model, the best fit was obtained for $\mu = 0.525$ (frozen),
$\Gamma = 1.9$, $\dot{M} = 0.68$~M$_{\odot}$~yr$^{-1}$ and $R_{v} =
3.2 \;
r_{\mbox{\scriptsize min}}$ and yielded $\chi^2/$d.o.f of
163/132. Although an imperfect fit, this suggests that the majority of
the spectroscopic features are well accounted for by the model (a pure
power-law fit to the same data yields $\chi^2/$d.o.f of 389/134).
To make a detailed comparison, we computed spectra for eighteen more models
adopting parameters close to those favoured from the {\sc xspec}
fits and extending the range in $R_{v}$
($\Gamma = [1.7,1.9]$, $R_{v} = [3,5,8]r_{\mbox{\scriptsize min}}$, $\dot{M} = [0.53,0.65,0.76]$~M$_{\odot}$~yr$^{-1}$). For these simulations a larger number of
MC quanta ($\sim 7 \times 10^6$) were used to compute the spectra. The
spectra obtained from each of these simulations are qualitatively
similar although the amplitude of the spectra features varies
slightly.
We fit spectra from these eighteen models to the 2 -- 10~keV spectrum in
a similar manner to that described above -- the best fitting spectrum
obtained from this model grid ($\chi^2/$d.o.f of 160/132) was for
parameters close to that of the model with
$\Gamma = 1.9$,
$R_{v} = 8r_{\mbox{\scriptsize min}}$ and $\dot{M} = 0.53$~M$_{\odot}$~yr$^{-1}$ (for
$\mu = 0.575$); therefore we use the spectra from this model as the
basis for our comparison (see below).
\subsection{Discussion}
Figure~\ref{fig:pg1211} compares the Fe K$\alpha$ spectral region for
the model with $R_{v} = 8r_{\mbox{\scriptsize min}}$, $\dot{M} =
0.53$~M$_{\odot}$~yr$^{-1}$
and $\mu = 0.575$
with the unfolded data of PG1211+143
\citep{pounds09}.
The comparison is shown for the photon energy scale observed at the
Earth, accounting for the source redshift of $z = 0.0809$ \citep{marziani96}.
The agreement between model and data is generally very good and shows
that the
disk wind paradigm readily produces the important features that
characterize these data -- blueshifted Fe K$\alpha$ absorption and broad
red shifted K$\alpha$
emission with a red-skewed wing -- with strengths comparable to those
observed. In agreement with the data, the model also predicts weak
S~{\sc xvi} absorption around 2.7~keV.
There are, however, some
clear discrepancies.
For example,
additional narrow components of emission may
be required in the emission line profile (see discussion by \citealt{pounds09})
and the model
seems to predict too little absorption at around 8~keV.
However, none of these details suggest
that the outflow paradigm is inappropriate for these data and may
simply indicate physically distinct contributions to the spectrum.
E.g. any very narrow emission
components required by the data
are unlikely to be associated with the high velocity wind
but can be more readily attributed to reflection by low-velocity
material.
\begin{figure}
\epsfig{file=fig5.eps, width=8cm}
\caption{
Comparison of the Pounds et al. (2009)
stacked data for PG1211+143 (black crosses) and the wind model (black histogram)
described in the text.
The wind parameters are all identical to Model~B (Table 2)
except for $\Gamma = 1.9$, $R_{v} = 8
r_{\mbox{\scriptsize min}}$ and $\dot{M} =$ 0.53~M$_{\odot}$~yr$^{-1}$. The
model spectrum shown is for a viewing inclination of $\mu = 0.575$.
The red and blue histograms show the contributions to the model
spectrum from direct and scattered/reprocessed radiation, respectively.
The photon energy scale of the model spectrum has been adjusted to
account for the redshift of PG1211+143 ($z = 0.0809$).
}
\label{fig:pg1211}
\end{figure}
Since the source is clearly complex, fitting the 2 -- 10~keV spectrum
to a wind model alone yields parameters that
should only by regarded as indicative of those appropriate for PG1211+143.
To make a quantitative comparison of our models with the lower
energy spectra for PG1211+143 would requires significantly
more complex fits of the data to be performed to address the strong
soft excess and account for any physically distinct,
lower-ionization contribution to absorption. However, at minimum, the
highly-ionized wind should be able to imprint discreet blueshifted absorption and
broad emission
features in the softer spectrum consistent with those that have been
claimed from previous analyses of the PG1211+143 RGS spectrum
(e.g. \citealt{pounds03,pounds07}). \footnote{We note, however, that
the exact combination of wind model and
observer line-of-sight which fit the
stacked EPIC pn data of \cite{pounds09} (which combines observations from
2001, 2004 and 2007) is not expected to be
perfectly applicable for
comparison with the RGS absorption features discussed by
\cite{pounds03} since it is known that the
spectroscopic signatures of outflow are time-variable.}
To investigate this, we show in
Figure~\ref{fig:pg1211_wide} the 10 -- 35\AA~spectra (c.f. figures~6
-- 9 of \citealt{pounds03}) for the same
wind model used in Figure~\ref{fig:pg1211}.
The spectrum is shown for the line-of-sight ($\mu = 0.575$)
that best fit the Fe~K region of the stacked data of \cite{pounds09}
and also for two higher inclination angles ($\mu = 0.475$ and
0.375).
The wavelength of blueshifted absorption lines for which
identifications were claimed by \cite{pounds03} are indicated in the
figure. Several of these -- in
particular those corresponding to relatively high ionization states
(e.g. Ne~{\sc x}, O~{\sc viii}, N~{\sc vii}, C~{\sc vi}) --
are clearly present in the model with strength and blueshift depending
sensitively on the observer's inclination.
Fe~{\sc xxiv} imprints a clear additional absorption feature for $\mu =
0.475$ -- although they did not discuss this in detail, we note that
the fit in figure~9 of \cite{pounds03} does include an absorption
line around 11~\AA.
The model also predicts
significantly broadened O~{\sc viii} Lyman $\alpha$ emission,
qualitatively similar to that discussed by \cite{pounds07,pounds09} in their
more recent analyses of the PG1211+143 RGS data.
Thus our
outflow models are able to simultaneously account for
observable features in both the Fe~K band and the softer-energy parts
of the spectrum and there are good prospects that more detailed
fitting of the complete set of spectral data will allow more robust
constraints to be placed on the outflow model parameters.
The less ionized species (Ne~{\sc ix} and O~{\sc vii}) are present in
only very small quantities in the model and do not imprint lines in the
spectra shown.
The presence of these lines in the data suggests that a wider
range of ionization states than in the model used here
is required to explain the complete 2001 RGS spectrum. This may be found
by exploring a greater range of our model parameter-space but may also
require us to go beyond our assumption of a smooth, steady-state flow. A
more realistic model incorporating flow inhomogeneities will
inevitably have a wider range of densities, and therefore
ionization state. Such effects will be investigated in future studies.
\begin{figure}
\epsfig{file=fig6.eps, width=8cm}
\caption{
The RGS region of the spectra of the same wind model shown in
Figure~\ref{fig:pg1211}.
The black histogram shows the model spectrum for $\mu = 0.575$ (same
inclination as used for the fit to the stacked EPIC pn data shown in
Figure~\ref{fig:pg1211}). The green and orange histograms show the spectra
for $\mu = 0.475$ and $0.375$, respectively. Each spectrum is
shown as a ratio to the input pure power-law spectrum which had
$\Gamma = 1.9$ and renormalised to 1.0 around 30\AA.
The identifications with red tick marks
indicate the observed energies of lines identified in the observations of PG1211+143 by
Pounds et al. (2003) that are distinct in the model spectra. The black
tick marks indicate lines identified in the observations that do
not have significant strength in any of the model spectra.
Fe~{\sc xxiv} (blue tick mark) is responsible for the feature are 11~\AA~ in the $\mu =
0.475$ spectrum which was not discussed by Pounds et al. (2003).
}
\label{fig:pg1211_wide}
\end{figure}
The spectral model used here is too simplistic to address the strong soft
excess observed in PG1211+143.
However, we note that the wind model does clearly lead to an excess of
emission below $\sim 1$~keV suggesting that reprocessing by the fast
flow may account for a modest fraction of the excess soft-energy emission
(c.f. \citealt{king03}). Much of this contribution is due to heavily
blended forests of emission lines (including those of L-shell Fe) but
there are also moderately strong discreet features, such as
the broad O~{\sc viii} Ly$\alpha$ emission
line mentioned above.
We defer more detailed studies including both a wider
exploration of the model parameter space and the interplay of
wind-formed spectral features with other components of the system to
later work but the simple comparisons presented here already support
the notion that the Fe~K absorption {\it and} emission features could be
predominantly formed in a fast outflow as in the picture presented by
\cite{pounds09}. Scaled to the black hole mass of PG1211+143 ($4
\times 10^7$~M$_{\odot}$, \citealt{kaspi00}; see
Section~\ref{sect_rescale}), the model with which we have
compared has a mass-loss
rate of $\dot{M} = 2.1$~M$_{\odot}$~yr$^{-1}$ and a total covering
fraction\footnote{Although our wind has a total covering
fraction of $b = 0.7$, this value cannot be directly compared
to that obtained from the much simpler model of \cite{pounds09}
-- there is significant diversity in the line-of-sight properties
of our model and some of the lines-of-sight through the wind have
only very low optical depth.} of $b = 0.7$. These are comparable to the properties inferred
by \cite{pounds09} and support their conclusion that the flow is very
likely to be energetically important.
\section{Conclusions and future work}
\label{sect_conc}
Outflows from the accretion disks around supermassive black holes are
a promising explanation for the blueshifted absorption line features
that have been identified in the X-ray spectra of AGN.
They are also theoretically expected in high Eddington ratio sources for
which line-driven winds
have been modelled by e.g. \citep{proga04}.
AGN outflows may be massive enough to have significant implications for our
understanding of accretion by supermassive black holes
and likely have implications for the interpretation of a wide range of both absorption and emission features
in the high energy spectra of AGN \citep{turner09}.
However, to
properly interpret the observed spectroscopic features,
modelling based on realistic theoretical spectra for outflow models is
required in order to quantify the flow properties.
Here, we have made a significant step towards this goal by
extended the code described in Paper~I to incorporate the
physics of L- and M-shell ions and to
compute both the ionization and thermal state of the outflow in
detail. This allows us to explore a wider range of
plausible outflow conditions including, in particular, less
highly-ionized and more optically thick flows.
An example calculation for one of our simply-parameterized wind
models illustrates that, for mass-loss rates comparable to the
Eddington accretion rate, a wide range of physical conditions are
likely to be present in a fast outflow such that a very diverse range
of spectral signatures are possible. As in Paper~I, we find that considerable
complexity can be introduced to the Fe~K$\alpha$ region where both
narrow absorption and broad, red-skewed emission lines are predicted to
form. However, such an outflow can also significantly affect the
spectrum at softer energies -- narrow, highly blueshifted
absorption lines of lighter
elements and lower ionization states of Fe are to be expected for many
line-of-sight through an outflow.
Emission features associated with these lines also form
which can be
broadened and skewed in a manner similar to the Fe~K$\alpha$ line,
potentially allowing the wind model to account for broad Fe~L features
as have been reported in the spectrum of 1H0707-495
(see e.g. \citealt{fabian09,zoghbi09}).
These features in the soft band are
generally weaker that those in the Fe~K$\alpha$ region and their
interpretation is more likely to be complicated owing to the wide range
of physical components that may affect the softer bands (e.g. the
soft excess and/or warm absorbers).
As proof-of-concept, we have presented a simple comparison of
our outflow model spectra with the well-known quasar PG1211+143. We
found that the most important features that have been identified in
its hard X-ray spectra, namely a strong absorption line at
$\sim 7$~keV and a broad emission feature peaking around $6$~keV, can be
simultaneously well-matched by our theoretical spectra. Although we
have not fully explored the possible parameter-space of outflow models
nor the interplay with other physically motivated
phenomena (such as disk reflection), our results support the
interpretation of \cite{pounds09} that the absorption and emission
components are likely physically related and form in a fast outflow
that has a substantial covering fraction and a
mass-loss rate comparable to the Eddington accretion rate.
Our next step will be to attempt detailed model fits to one or more
AGN using our radiative transfer code. This will allow us to more
fully test the outflow paradigm and quantify the range of
outflow properties which are suggested by observations, a critical step in
establishing the role of highly-ionized outflows in the AGN phenomenon.
\section*{Acknowledgments}
SAS thanks Caroline D'Angelo for many useful discussions and
helpful suggestions.
TJT acknowledges NASA Grant NNX09AO92G.
We thank the anonymous referee for several constructive comments.
\bibliographystyle{mn2e}
|
\section{Introduction}
The purpose of this paper is to study the convergence and complexity
of adaptive finite element computations for a class of elliptic
partial differential equations of second order and to apply our
general approach to three problems: a nonsymmetric problem, a
nonlinear problem, and an eigenvalue problem with an unbounded
coefficient. One technical tool for motivating this work is the
relationship between the general problem and a linear symmetric
elliptic problem, which is derived from some perturbation arguments
(see Theorem \ref{thm-general-boundary} and Lemma
\ref{lemma-bound-general} ).
Since Babu{\v s}ka and Vogelius \cite{babuska-vogelius-84} gave an
analysis of an adaptive finite element method (AFEM) for linear
symmetric elliptic problems in $1D$, there are a number of work on
the convergence and complexity of adaptive finite element methods in
the literature. For instance, D{\" o}rfler \cite{dorfler-96}
presented the first multidimensional convergence result, which has
been improved and generalized in
\cite{binev-dehmen-devore-04,cascon-kreuzer-nochetto-siebert-07,Dai-Xu-Zhou-08,
mekchay-nochetto-05,morin-nochetto-siebert-00,morin-nochetto-siebert-02,morin-siebrt-veeser-08,
stevenson-06a}.
For a nonsymmetric problem, in particular, Mekchay and Nochetto
\cite{mekchay-nochetto-05} imposed a quasi-orthogonality property
instead of the Pythagoras equality to prove the convergence of AFEM
while Morin, Siebrt, and Veeser \cite{morin-siebrt-veeser-08}
showed the convergence of error and estimator simultaneously with
the strict error reduction and derived the convergence of the
estimator by exploiting the (discrete) local lower but not the upper
bound. To our best knowledge, however, there has been no any work on
the complexity of AFEM for nonsymmetric elliptic problems in the
literature. In this paper, we can get the convergence and optimal
complexity of nonsymmetric problems from our general approach. For a
nonlinear problem,
Chen, Holst and Xu \cite{chen-holst-xu} proved the convergence of
an adaptive finite element algorithm for Poisson-Boltzmann equation
while we are able to obtain the convergence and optimal complexity
of AFEM for a class of nonlinear problems now. For a smooth
coefficient eigenvalue problem, Dai, Xu, and Zhou
\cite{Dai-Xu-Zhou-08} gave the convergence and optimal complexity
of AFEM for symmetric elliptic eigenvalue problems with piecewise
smooth coefficients (see, also convergence analysis of a special
case \cite{Garau-Morin-Zuppa-08,Giani-Graham-09}). In this paper we
will derive similar results for unbounded coefficient eigenvalue
problems from our general conclusions, too. We mention that a
similar perturbation approach was used in \cite{Dai-Xu-Zhou-08}.
This paper is organized as follows. In section 2 we review some
existing results on the convergence and complexity analysis of AFEM
for the typical problem. In section 3 we generalize results to a
general model problem by using a perturbation argument. In section 4
and section 5, we provide three typical applications for
illustration, including theory and numerics.
\section{Adaptive FEM for a typical problem}\label{sc:linear
boundary} In this section, we review some existing results on the
convergence and complexity analysis of AFEM for a boundary value
problem in the literature.
Let $\Omega\subset \mathbb{R}^d(d\ge 2)$ be a bounded polytopic
domain. We shall use the standard notation for Sobolev spaces
$W^{s,p}(\Omega)$ and their associated norms and seminorms, see,
e.g., \cite{adams-75,ciarlet-lions-91}. For $p=2$, we denote
$H^s(\Omega)=W^{s,2}(\Omega)$ and $H^1_0(\Omega)=\{v\in H^1(\Omega):
v\mid_{\partial\Omega}=0\}$, where $v\mid_{\partial\Omega}=0$ is understood in
the sense of trace, $\|\cdot\|_{s,\Omega}= \|\cdot\|_{s,2,\Omega}$.
Throughout this
paper, we shall use $C$ to denote a generic positive constant which
may stand for different values at its different occurrences.
We will also use $A\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} B$ to mean that $A\le C B$ for some constant
$C$ that is independent of mesh parameters. All constants involved
are independent of mesh sizes.
\subsection{A boundary value problem}
Consider a homogeneous boundary value
problem:
\begin{equation}\label{problem}
\left\{\begin{array}{rl}
Lu :=-\nabla \cdot (\mathbf{A}\nabla u)&=f \,\,\, \mbox{in} \quad \Omega,\\
u &= 0\,\,\,\mbox{on}~~\partial\Omega,
\end{array}\right.
\end{equation}
where $\mathbf{A}:\Omega\rightarrow \mathbb{R}^{d\times d}$ is piecewise
Lipschitz over initial triangulation $\mathcal {T}_{0}$, for $x\in
\Omega$ matrix $\mathbf{A(x)}$ is symmetric and positive definite
with smallest eigenvalue uniformly bounded away from 0, and $f\in
L^{2}(\Omega)$.
\begin{remark}\label{remark-boundary}
The choice of homogeneous boundary condition is made for ease of
presentation, since similar results are valid for other boundary
conditions \cite{cascon-kreuzer-nochetto-siebert-07}.
\end{remark}
The weak form of (\ref{problem}) reads as follows: Find $u\in
H^1_0(\Omega)$ such that
\begin{eqnarray} \label{variation}
a(u, v)= (f, v) \qquad\forall v\in H^1_0(\Omega),
\end{eqnarray}
where $a(\cdot,\cdot)=(\mathbf{A}\nabla \cdot,\nabla \cdot)$. It is
seen that $a(\cdot,\cdot)$ is bounded and coercive on
$H^1_0(\Omega)$, i.e., for any $w, v\in H^1(\Omega)$ there exist
constants $0<c_a\le C_a<\infty$ such that
\begin{eqnarray*}
|a(w, v)| \leq C_a \|w\|_{1, \Omega} \|v\|_{1, \Omega} \quad
\textnormal{and} \quad c_a\|v\|^2_{1,\Omega} \leq a(v,v) ~~\forall
v\in H^1_0(\Omega).
\end{eqnarray*}
The energy norm $\|\cdot\|_{a,\Omega}$ , which is equivalent to
$\|\cdot\|_{1,\Omega}$ , is defined by
$\|w\|_{a,\Omega}=\sqrt{a(w,w)}$ . It is known that
(\ref{variation}) is well-posed, that is, there exists a unique
solution for any $f\in H^{-1}(\Omega)$.
Let $\{\mathcal{T}_h\}$ be a shape regular family of nested
conforming meshes over $\Omega$: there exists a constant
$\gamma^{\ast}$ such that
\begin{eqnarray*}
\frac{h_\tau}{\rho_\tau} \leq \gamma^{\ast} ~~~ \forall \tau \in
\bigcup_h\mathcal{T}_h,
\end{eqnarray*}
where, for each $\tau\in \mathcal{T}_h$, $h_\tau$ is the diameter of
$\tau$, $\rho_\tau$ is the diameter of the biggest ball contained
in $\tau$, and
$h=\max\{h_\tau: \tau\in \mathcal{T}_h\}$. Let $\mathcal{E}_h$ denote
the set of interior sides (edges or faces) of $\mathcal{T}_h$. Let
$S_0^h(\Omega) \subset H_0^1(\Omega)$ be a family of nested finite
element spaces consisting of continuous piecewise polynomials
over $\mathcal{T}_h$ of fixed degree $n \geq 1$, which vanish on
$\partial\Omega$.
Define the Galerkin-projection $P_h:H^1_0(\Omega) \rightarrow
S_0^h(\Omega)$ by
\begin{eqnarray}\label{projection}
a(u-P_h u, v) =0 \quad\forall v\in S_0^h(\Omega).
\end{eqnarray}
For any $u\in H_0^1(\Omega)$, there apparently hold:
\begin{equation*}\label{stable1}
\|P_hu\|_{a,\Omega}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \|u\|_{a,\Omega}
\quad \textnormal{and} \quad
\lim_{h\rightarrow 0} \|u-P_hu\|_{a,\Omega}=0.
\end{equation*}
Now we introduce the following quantity:
\begin{eqnarray*}
\rho_{_{\Omega}}(h)=\sup_{ f\in L^2(\Omega),\|f\|_{0,\Omega}=1}
\inf_{v\in S_0^h(\Omega)}\|L^{-1}f-v\|_{a,\Omega},
\end{eqnarray*}
then $\rho_{_{\Omega}}(h)\to 0$ as $h\to 0$ (see, e.g.,
\cite{babuska-osborn-89,xu-zhou-00}).
A standard finite element scheme for (\ref{variation}) is: Find
$u_h \in S_0^h(\Omega)$ satisfying
\begin{eqnarray}\label{dis-fem}
a(u_h, v) =(f, v) \qquad \forall v \in S_0^h(\Omega).
\end{eqnarray}
By definition (\ref{projection}), we know that $u_h=P_hu$.
By a contradiction
argument, we have (c.f., e.g., \cite{zhou-07})
\begin{lemma}\label{lemma1}
As operators over $H^1_0(\Omega)$, there holds
\begin{eqnarray*}
\label{operator-property}
\lim_{h\rightarrow 0}\|\mathcal {K}(I-P_h)\|=0
\end{eqnarray*}
if $\mathcal {K}$ is a compact operator over $H^1_0(\Omega)$.
\end{lemma}
\subsection{Adaptive algorithm}
Given an initial triangulation $\mathcal{T}_0$, we shall generate a
sequence of nested conforming triangulations $\mathcal{T}_k$ using
the following loop:
\begin{eqnarray*}\label{loop}
{\bf SOLVE}\rightarrow {\bf ESTIMATE}\rightarrow {\bf
MARK}\rightarrow {\bf REFINE}.
\end{eqnarray*}
More precisely to get $\mathcal{T}_{k+1}$ from
$\mathcal{T}_k$ we first solve the discrete equation to get $u_k$ on
$\mathcal{T}_k$. The error is estimated using $u_k$ and used to mark
a set of elements that are to be refined. Elements are refined in
such a way that the triangulation is still shape regular and
conforming.
We assume that the solutions of finite-dimensional problems can be
solved to any accuracy efficiently.\footnote{By the similar
perturbation argument, indeed, it will be seen that some
approximations to the finite-dimensional problem will be
sufficient.} Examples of such optimal solvers are multigrid method
or multigrid-based preconditioned conjugate gradient method.
Now we review the residual type a posteriori error estimators for
finite element solutions of (\ref{problem}). Let $\mathbb{T}$ denote
the class of all conforming refinements by bisection of
$\mathcal{T}_0$. For $\mathcal{T}_h \in \mathbb{T}$ and any $v\in
S^h_0(\Omega)$ we define the element residual
$\tilde{\mathcal{R}}_{\tau}(v)$ and the jump residual
$\tilde{J}_e(v)$ by
\begin{eqnarray*}\label{residual}
\tilde{\mathcal{R}}_{\tau}(v) &:=& f-L v = f +\nabla \cdot (\mathbf{A}\nabla v)\qquad \mbox{in}~ \tau\in
\mathcal{T}_h,\\
\tilde{J}_e(v) &:=& -\mathbf{A} \nabla v^{+}\cdot \nu^{+} - \mathbf{A} \nabla v^{-}\cdot
\nu^{-} := [[\mathbf{A}\nabla v]]_e \cdot \nu_e ~~~~ \mbox{on}~ e\in \mathcal{E}_h,
\end{eqnarray*}
where $e$ is the common side of elements $\tau^+$ and $\tau^-$ with
unit outward normals $\nu^+$ and $\nu^-$, respectively, and
$\nu_e=\nu^-$. Let $\omega_e$ be the union of elements which share
the side $e$ and $\omega_\tau$ be the union of elements sharing a
side with $\tau$.
For $\tau\in \mathcal{T}_h$, we define the local error indicator
$\tilde{\eta}_h(v, \tau)$ by
\begin{eqnarray*}\label{error-indicator}
\tilde{\eta}^2_h(v, \tau) := h_\tau^2\|\tilde{\mathcal{R}}_{\tau}(v)\|_{0,\tau}^2
+ \sum_{e\in \mathcal{E}_h,e\subset\partial \tau
} h_e \|\tilde{J}_e(v)\|_{0,e}^2
\end{eqnarray*}
and the oscillation $\widetilde{osc}_h(v,\tau)$ by
\begin{eqnarray*}\label{local-oscillation}
\widetilde{osc}^2_h(v,\tau) :=
h_\tau^2\|\tilde{\mathcal{R}}_{\tau}(v)-\overline{\tilde{\mathcal{R}}_{\tau}(v)}\|_{0,\tau}^2
+ \sum_{e\in \mathcal{E}_h,e\subset\partial \tau
} h_e \|\tilde{J}_e(v)-\overline{\tilde{J}_e(v)}\|_{0,e}^2,
\end{eqnarray*}
where $\overline{w}$ is the $L^2$-projection of $w\in L^2(\Omega)$
to polynomials of some degree on $\tau$ or $e$.
Given a subset $\omega \subset \Omega$, we define the error
estimator $\tilde{\eta}_h(v, \omega)$ and the oscillation $\widetilde{osc}_h(v,\omega)$ by
\begin{eqnarray*}
\tilde{\eta}^2_h(v, \omega) := \sum_{\tau\in \mathcal{T}_h, \tau
\subset \omega}
\tilde{\eta}^2_h(v, \tau) \quad \textnormal{and} \quad
\widetilde{osc}^2_h(v, \omega) := \sum_{\tau\in \mathcal{T}_h, \tau
\subset \omega}
\widetilde{osc}^2_h(v, \tau).
\end{eqnarray*}
For $\tau\in \mathcal{T}_h$, we also need
notation
\begin{eqnarray*}
\eta^2_h(\mathbf{A},\tau)
:=h_\tau^2(\|\mbox{div}\mathbf{A}\|^2_{0,\infty,\tau}+h^{-2}_{\tau}\|\mathbf{A}\|^2_{0,\infty,\omega_{\tau}})
\end{eqnarray*}
and
\begin{eqnarray*}
osc^2_h(\mathbf{A},\tau)
:=h_\tau^2(\|\mbox{div}\mathbf{A}-\overline{\mbox{div}\mathbf{A}}\|^2_{0,\infty,\tau}+
h^{-2}_{\tau}\|\mathbf{A}-\bar{\mathbf{A}}\|^2_{0,\infty,\omega_{\tau}}),
\end{eqnarray*}
where $\overline{v}$ is the best $L^{\infty}$-approximation in the
space of discontinuous polynomials of some degree.
Given a subset $\omega \subset \Omega$ we finally set
\begin{eqnarray*}
\eta_h(\mathbf{A},\omega):=\max_{\tau\in \mathcal{T}_h, \tau \subset
\omega}\eta_h(\mathbf{A},\tau) \quad \textnormal{and} \quad
osc_h(\mathbf{A},\omega):=\max_{\tau\in \mathcal{T}_h, \tau \subset
\omega}osc_h(\mathbf{A},\tau).
\end{eqnarray*}
We now recall the well-known upper and lower bounds for the energy
error in terms of the residual-type estimator (see, e.g.,
\cite{mekchay-nochetto-05,morin-nochetto-siebert-02,verfurth-96}).
\begin{theorem}
\textnormal{(Global a posterior upper and lower bounds)}. Let $u \in
H^1_0(\Omega)$ be the solution of (\ref{variation}) and $u_h \in
S^h_0(\Omega)$ be the solution of (\ref{dis-fem}). Then there exist
constants $\tilde{C}_1$, $\tilde{C}_2$ and $\tilde{C}_3>0$ depending
only on the shape regularity $\gamma^{\ast}$, $C_a $ and $c_a$ such
that
\begin{eqnarray}\label{boundary-upper}
\|u - u_h \|^2_{a,\Omega} \leq \tilde{C}_1 \tilde{\eta}^2_h(u_h, \mathcal{T}_h)
\end{eqnarray}
and
\begin{eqnarray}\label{boundary-lower}
~~~~~~ \tilde{C}_2 \tilde{\eta}^2_h (u_h, \mathcal{T}_h)\le
\|u-u_h\|_{a,\Omega}^2+
\tilde{C}_3\widetilde{osc}^2_h(u_h,\mathcal{T}_h).
\end{eqnarray}
\end{theorem}
We replace the subscript $h$ by an iteration counter called $k$ and
call the adaptive algorithm without oscillation marking as {\bf
Algorithm $D_0$}, which is defined as follows:
Choose a parameter $0 < \theta <1:$
\begin{enumerate}
\item Pick any initial mesh $\mathcal{T}_0$, and let $k=0$.
\item
Solve the system on $\mathcal{T}_0$ for the discrete solution $u_0$.
\item Compute the local indicators ${\tilde \eta}_k$.
\item
Construct $\mathcal{M}_k \subset
\mathcal{T}_k$ by {\bf Marking Strategy $E_0$ } and parameter
$\theta$.
\item Refine $\mathcal{T}_k$ to get a new conforming mesh $\mathcal{T}_{k+1}$ by Procedure {\bf REFINE}.
\item Solve the system on $\mathcal{T}_{k+1}$ for the discrete
solution $u_{k+1}$.
\item Let $k=k+1$ and go to Step 3.
\end{enumerate}
The marking strategy, which we call {\bf Marking Strategy $E_0$}, is
crucial for our adaptive methods. Now it can be stated by:
Given a parameter $0<\theta < 1$ :
\begin{enumerate}
\item Construct a minimal subset $\mathcal{M}_k$ of
$\mathcal{T}_k$ by selecting some elements
in $\mathcal{T}_k$ such that
\begin{eqnarray*}
\tilde{\eta}_k(u_k, \mathcal{M}_k) \geq \theta
\tilde{\eta}_k(u_k,\mathcal{T}_k).
\end{eqnarray*}
\item Mark all the elements in $\mathcal{M}_k$.
\end{enumerate}
Due to \cite{cascon-kreuzer-nochetto-siebert-07}, the procedure {\bf
REFINE} here is not required to satisfy the Interior Node Property
of \cite{mekchay-nochetto-05,morin-nochetto-siebert-02}.
Given a fixed number $b\geq 1$, for any $\mathcal{T}_k \in
\mathbb{T}$ and a subset $\mathcal{M}_k\subset \mathcal{T}_k$ of
marked elements,
\begin{eqnarray*}
\mathcal{T}_{k+1}={\bf REFINE}(\mathcal{T}_k,\mathcal{M}_k)
\end{eqnarray*}
outputs a conforming triangulation $\mathcal{T}_{k+1}\in
\mathbb{T}$, where at least all elements of $\mathcal{M}_k$ are
bisected $b$ times. We define $R_{\mathcal{T}_k\rightarrow
\mathcal{T}_{k+1}}:=\mathcal{T}_k\backslash(\mathcal{T}_k\cap
\mathcal{T}_{k+1})$ as the set of refined elements, thus
$\mathcal{M}_k\subset R_{\mathcal{T}_k\rightarrow
\mathcal{T}_{k+1}}$.
\begin{lemma}
\textnormal{(Complexity of Refine)}. Assume that $\mathcal{T}_0$
verifies condition (b) of section 4 in \cite{Stevenson-08}. For
$k\geq 0$ let $\{\mathcal{T}_k\}_{k\geq 0}$ be any sequence of
refinements of $\mathcal{T}_0$ where $\mathcal{T}_{k+1}$ is
generated from $\mathcal{T}_{k}$ by $\mathcal{T}_{k+1}={\bf
REFINE}(\mathcal{T}_{k},\mathcal{M}_k)$ with a subset
$\mathcal{M}_k\subset \mathcal{T}_k$. Then
\begin{eqnarray}\label{complexity}
\#\mathcal{T}_k-\#\mathcal{T}_0 \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \sum_{j=0}^{k-1}\#\mathcal{M}_j
\quad \forall k\geq 1
\end{eqnarray}
is valid, where the hidden constant depends on $\mathcal{T}_0$ and
b.
\end{lemma}
The convergence of {\bf Algorithm $D_0$} is shown in
\cite{cascon-kreuzer-nochetto-siebert-07}.
\begin{theorem}\label{convergence-boundary}
Let $\{u_k\}_{k\in \mathbb{N}_0}$ be a sequence of finite element
solutions corresponding to a sequence of nested finite element
spaces $\{S^k_0(\Omega)\}_{k\in \mathbb{N}_0}$ produced by {\bf Algorithm $D_0$}.
Then there exist constants $\tilde{\gamma}>0$ and $\tilde{\xi}\in
(0,1)$ depending only on the shape regularity of meshes, the data
and the marking parameter $\theta$, such that for any two
consecutive iterates we have
\begin{eqnarray*}\label{convergence-neq}
&&\|u-u_{k+1}\|^2_{a,\Omega} + \tilde{\gamma}
\tilde{\eta}^2_{k+1}(u_{k+1},\mathcal{T}_{k+1}) \nonumber \\
&\leq & \tilde{\xi}^2\big( \|u-u_k\|^2_{a,\Omega} +
\tilde{\gamma} \tilde{\eta}^2_k(u_k, \mathcal{T}_k)\big).
\end{eqnarray*}
Indeed, constant ${\tilde\gamma}$ has the following form
\begin{eqnarray}\label{gamma-boundary}
\tilde{\gamma} := \frac{1}{(1 + \delta^{-1})\Lambda_1
\eta^2_0(\mathbf{A},\mathcal{T}_0)},
\end{eqnarray}
where
$\eta^2_0(\mathbf{A},\mathcal{T}_0):=\eta^2_{\mathcal{T}_0}(\mathbf{A},\mathcal{T}_0)$,
$\Lambda_1 :=(d+1)C_0^2/c_a$ with $C_0$ some positive constant and constant $\delta \in (0,1)$.
\end{theorem}
Following \cite{cascon-kreuzer-nochetto-siebert-07,Dai-Xu-Zhou-08},
we have a link between nonlinear approximation theory and the {\bf
AFEM} through the marking strategy as follows.
\begin{lemma}
\label{complexity-general-optimal-marking}\textnormal{(Optimal
Marking).} Let $u_k \in S^k_0(\Omega)$ and $u_{k+1} \in
S^{k+1}_0(\Omega)$ be finite element solutions of (\ref{variation})
over a conforming mesh $\mathcal{T}_k$ and its refinement
$\mathcal{T}_{k+1}$ with marked element $\mathcal{M}_k$. Suppose
that they satisfy the decrease property
\begin{eqnarray*}
&&\|u - u_{k+1} \|_{a,\Omega}^2 + \tilde{\gamma_{\ast}}
\widetilde{osc}^2_{k+1}(u_{k+1},\mathcal{T}_{k+1}) \nonumber\\
& \leq& \tilde{\beta_{\ast}}^2 \big(\|u-u_k\|_{a,\Omega}^2
+ \tilde{\gamma_{\ast}}\widetilde{osc}^2_k(u_k,\mathcal{T}_k)\big)
\end{eqnarray*}
with constants $\tilde{\gamma_{\ast}}>0$ and
$\tilde{\beta_{\ast}}\in (0,\sqrt{\frac{1}{2}})$. Then the set
$\mathcal{R}:=R_{\mathcal{T}_k\rightarrow \mathcal{T}_{k+1}}$
satisfies the following inequality
\begin{eqnarray*}\label{complexity-general-optimal-marking-neq1}
\tilde{\eta}_k(u_k,\mathcal {R})\geq
\hat{\theta}\tilde{\eta}_k(u_k,\mathcal{T}_k)
\end{eqnarray*}
with $\hat{\theta}^2 =
\frac{\tilde{C}_2(1-2\tilde{\beta_{\ast}}^2)}{\tilde{C}_0 (
\tilde{C}_1 + (1 + 2 C \tilde{C}_1) \tilde{\gamma_{\ast}})}$ ,
where $C=\Lambda_1 osc^2_0(\mathbf{A},\mathcal{T}_0)$
\textnormal{and} $\tilde{C}_0 = \max(1,
\frac{\tilde{C}_3}{\tilde{\gamma_{\ast}}})$.
\end{lemma}
\section{A general framework}
Let $u\in H^1_0(\Omega)$ satisfy
\begin{eqnarray}\label{Gvariation}
a(u, v)+(Vu,v)= (\ell u, v) \qquad\forall v\in H^1_0(\Omega),
\end{eqnarray}
where $\ell:H^1_0(\Omega)\rightarrow L^{2}(\Omega)$ is a bounded
operator and $V:H^1_0(\Omega)\rightarrow L^{2}(\Omega)$ is a linear
bounded operator.
Let $K: L^2(\Omega)\rightarrow H^1_0(\Omega)$ be the operator
defined by
\begin{eqnarray*}\label{def-operator-k}
a(Kw, v) = (w, v)~~~~\forall w, v \in L^2(\Omega).
\end{eqnarray*}
Then $K$ is a compact operator and (\ref{Gvariation}) becomes as
\begin{eqnarray*}\label{operator-K-1}
u+KVu=K\ell u.
\end{eqnarray*}
Let $u_h\in S^h_0(\Omega)$ be a solution of disctetization
\begin{eqnarray}\label{Gdis-fem}
a(u_h, v)+(Vu_h,v) =(\ell_{h}u_h, v) ~~~~ \forall v \in
S^h_0(\Omega),
\end{eqnarray}
where $\ell_{h}:S^h_0(\Omega)\rightarrow L^{2}(\Omega)$ is some
bounded operator. Note that we may view $\ell_{h}$ as a perturbation
to $\ell$, for which we assume that there exists $\kappa_1(h)\in
(0,1)$ such that
\begin{eqnarray}\label{assume}
\|K(\ell u- \ell_h u_h)\|_{a,\Omega}=\mathcal
{O}(\kappa_1(h))\|u-u_h\|_{a,\Omega},
\end{eqnarray}
where $\kappa_1(h)\rightarrow 0$ as $h\rightarrow 0$.
Note that (\ref{Gdis-fem}) can be written as
\begin{eqnarray*}\label{operator-K-2}
u_h+P_h KVu_h=P_hK\ell_h u_h,
\end{eqnarray*}
where $P_h$ is defined by (\ref{projection}). We have for
$w^h=K\ell_{h}u_h-KVu_h$ that
\begin{eqnarray}\label{u-w}
u_h=P_hw^h.
\end{eqnarray}
\begin{theorem}\label{thm-general-boundary}
There exists $\kappa(h)\in (0,1)$ such that $\kappa(h)\rightarrow 0$ as $h\rightarrow
0$ and
\begin{eqnarray}\label{general-boundary-neq}
\|u-u_h\|_{a,\Omega}= \|w^h - P_h w^h\|_{a,\Omega} +\mathcal
{O}(\kappa(h))\|u-u_h\|_{a,\Omega}.
\end{eqnarray}
\end{theorem}
\begin{proof}
By definition, we have
\begin{eqnarray*}
u-w^h=K\ell u - K Vu - (K \ell_h u_h - KVu_h) = K(\ell u-\ell_h u_h) +
KV(u_h-u).
\end{eqnarray*}
Let $\kappa_2(h)=\|KV(I-P_h)\|.$ Since $KV :
H^1_0(\Omega)\rightarrow H^1_0(\Omega)$ is compact, we get from
Lemma \ref{lemma1} that
$\kappa_2(h)\rightarrow 0$ as $h\rightarrow 0$. Note that
\begin{eqnarray*}
KV(u_h-u)=KV(I-P_h)(u_h-u),
\end{eqnarray*}
we obtain
\begin{eqnarray}\label{KV-property}
\|KV(u_h-u)\|_{a,\Omega}=\mathcal
{O}(\kappa_2(h))\|u-u_h\|_{a,\Omega}.
\end{eqnarray}
Set $\kappa(h)=\kappa_1(h)+\kappa_2(h)$, we
have that $\kappa(h)\rightarrow 0$ as $h\rightarrow 0$ and
\begin{eqnarray}\label{thm1-neq1}
\|u-w^h\|_{a,\Omega}\leq \tilde{C}\kappa(h)\|u-u_h\|_{a,\Omega}.
\end{eqnarray}
Since (\ref{u-w}) implies
\begin{eqnarray*}
u - u_h = w^h-P_h w^h +u-w^h,
\end{eqnarray*}
we get (\ref{general-boundary-neq}) from (\ref{thm1-neq1}). This
completes the proof.
\end{proof}
Theorem \ref{thm-general-boundary} sets up a relationship between
the error estimates of finite element approximations of the general
problem and the associated typical finite element boundary value
solutions, from which various a posteriori error estimators for the
general problem can be easily obtained since the a posteriori error
estimators for the typical boundary value problem have been
well-constructed. In fact, Theorem \ref{thm-general-boundary}
implies that up to the high order term, the error of the general
problem is equivalent to that of the typical problem with $\ell_h
u_h-Vu_h$ as a source
term. However, the high order term can not be estimated
easily in the analysis of convergence and optimal complexity of AFEM
for the general problem, for instance, for a nonsymmetric problem, a
nonlinear problem and an unbounded coefficient eigenvalue problem.
\subsection{Adaptive algorithm}
Following the element residual $\tilde{\mathcal{R}}_{\tau}(u_h)$
and the jump residual $\tilde{J}_e(u_h)$ for (\ref{dis-fem}), we
define the element residual $\mathcal{R}_{\tau}(u_h)$ and the jump
residual $J_e(u_h)$ for (\ref{Gdis-fem}) as follows:
\begin{eqnarray*}\label{Gresidual}
\mathcal{R}_{\tau}(u_h) &:=& \ell_{h}u_h-Vu_h-L u_h =\ell_{h}u_h-Vu_h +\nabla \cdot (\mathbf{A}\nabla u_h)~~~~ \mbox{in}~ \tau\in
\mathcal{T}_h,\\
J_e(u_h) &:=& -\mathbf{A} \nabla u_h^{+}\cdot \nu^{+} - \mathbf{A} \nabla u_h^{-}\cdot
\nu^{-} := [[\mathbf{A}\nabla u_h]]_e \cdot \nu_e ~~~~ \mbox{on}~ e\in
\mathcal{E}_h.
\end{eqnarray*}
For $\tau\in \mathcal{T}_h$, we define the local error indicator
$\eta_h(u_h, \tau)$ by
\begin{eqnarray*}\label{Gerror-indicator}
\eta^2_h(u_h, \tau) := h_\tau^2\|\mathcal{R}_{\tau}(u_h)\|_{0,\tau}^2
+ \sum_{e\in \mathcal{E}_h,e\subset\partial \tau
} h_e \|J_e(u_h)\|_{0,e}^2
\end{eqnarray*}
and the oscillation $osc_h(u_h,\tau)$ by
\begin{eqnarray*}\label{Glocal-oscillation}
osc^2_h(u_h,\tau) :=
h_\tau^2\|\mathcal{R}_{\tau}(u_h)-\overline{\mathcal{R}_{\tau}(u_h)}\|_{0,\tau}^2
+ \sum_{e\in \mathcal{E}_h,e\subset\partial \tau
} h_e \|J_e(u_h)-\overline{J_e(u_h)}\|_{0,e}^2,
\end{eqnarray*}
where $e$ , $\nu^+$ and $\nu^-$ are defined as those in section
\ref{sc:linear boundary}.
Given a subset $\omega \subset \Omega$, we define the error
estimator $\eta_h(u_h, \omega)$ by
\begin{eqnarray}\label{Gerror-estimator}
\eta^2_h(u_h, \omega) := \sum_{\tau\in \mathcal{T}_h, \tau \subset \omega}
\eta^2_h(u_h, \tau)
\end{eqnarray}
and the oscillation $osc_h(u_h,\omega)$ by
\begin{eqnarray}\label{Goscilliation}
osc^2_h(u_h, \omega) := \sum_{\tau\in \mathcal{T}_h, \tau \subset \omega}
osc^2_h(u_h, \tau).
\end{eqnarray}
Let $h_0\in (0,1)$ be the mesh size of the initial mesh
$\mathcal{T}_0$ and define $$\tilde{\kappa}(h_0):=\sup_{h\in
(0,h_0]}\kappa(h).$$ Obviously, $\tilde{\kappa}(h_0) \ll 1$ if
$h_0\ll 1$.
To analyze the convergence and complexity of finite element
approximations, we need to establish some relationship between the
two level approximations. We use $\mathcal{T}_H$ to denote a coarse
mesh and $\mathcal{T}_h$ to denote a refined mesh of
$\mathcal{T}_H$. Recall that $w^h=K(\ell_h u_h-Vu_h)$ and
$w^H=K(\ell_H u_H-Vu_H)$.
\begin{lemma}\label{lemma-bound-general}
Let $h, H \in (0, h_0]$, then
\begin{eqnarray}\label{lemma-bound-general-conc-1}
\|u - u_h \|_{a, \Omega}= \|w^H - P_h w^H \|_{a, \Omega}
+ \mathcal{O} (\tilde{\kappa}(h_0))\left( \|u - u_h
\|_{a, \Omega}+ \|u - u_H\|_{a, \Omega}\right),~
\end{eqnarray}
\begin{eqnarray}\label{lemma-bound-general-conc-3}
\eta_{h}(u_h, \mathcal{T}_h) = \tilde{\eta}_{h}(P_h w^H,
\mathcal{T}_h) + \mathcal{O}
(\tilde{\kappa}(h_0))\left( \|u - u_h\|_{a, \Omega} + \|u -
u_H\|_{a, \Omega}\right),
\end{eqnarray}
and
\begin{eqnarray}\label{lemma-bound-general-conc-2}
osc_h(u_h,\mathcal{T}_h)=\widetilde{osc}_h(P_h w^H, \mathcal{T}_h) +
\mathcal{O} (\tilde{\kappa}(h_0))\left( \|u - u_h\|_{a, \Omega} +
\|u - u_H\|_{a, \Omega}\right).
\end{eqnarray}
\end{lemma}
\begin{proof}
First, we prove (\ref{lemma-bound-general-conc-1}). It follows that
\begin{eqnarray*} \|P_h (w^h - w^H)+ u - w^H
\|_{a,\Omega} &\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \| w^h - w^H\|_{a,\Omega} + \|u -
w^H\|_{a,\Omega}\\ &\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \|u - w^H\|_{a,\Omega} + \|u -
w^h\|_{a,\Omega},
\end{eqnarray*}
which together with (\ref{thm1-neq1}) implies
\begin{eqnarray*}
\|P_h (w^h - w^H) + w^H - u\|_{a,\Omega}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \kappa(H) \|u -
u_H\|_{a,\Omega} + \kappa(h)\|u - u_h\|_{a,\Omega}.
\end{eqnarray*}
Namely,
\begin{eqnarray}\label{lem-bound-general1}
\|P_h (w^h - w^H) + w^H - u\|_{a,\Omega}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \tilde{\kappa}(h_0)(\|u
- u_H\|_{a,\Omega} + \|u - u_h\|_{a,\Omega}).
\end{eqnarray}
Observing that identity (\ref{u-w}) leads to
\begin{eqnarray*}\label{lem-bound-general2}
u-u_h=w^H - P_h w^H+P_h(w^H -w^h)+u-w^H,
\end{eqnarray*}
we then obtain (\ref{lemma-bound-general-conc-1}) from
(\ref{lem-bound-general1}).
Next, we turn to prove (\ref{lemma-bound-general-conc-2}). Due to
$Lw^h=\ell_hu_h-Vu_h$ and $Lw^H=\ell_Hu_H-Vu_H$, we know that $w^h
- w^H$ is the solution of typical boundary value problem with
$\ell_hu_h-\ell_Hu_H+Vu_H-Vu_h$ as a source term. Since
\begin{eqnarray*}
\tilde{\mathcal{R}}_{\tau}(P_h(w^h-w^H))=\ell_hu_h-\ell_Hu_H+Vu_H-Vu_h-L(P_h(w^h-w^H)),
\end{eqnarray*}
we have
\begin{eqnarray}\label{temp4}
& &\widetilde{osc}^2_h(P_h(w^h-w^H),\mathcal{T}_h) =\sum_{\tau\in
\mathcal{T}_h}\widetilde{osc}^2_h(E,\tau)\nonumber\\
&=&\sum_{\tau\in \mathcal{T}_h}\big(h^2_{\tau}\|\tilde{\mathcal{R}}_{\tau}(E)-\overline{\tilde{\mathcal{R}}_{\tau}(E)}\|^2_{0,\tau}
+\sum_{e\in \mathcal{E}_h,e\subset\partial \tau
} h_e
\|\tilde{J}_e(E)-\overline{\tilde{J}_e(E)}\|^2_{0,e}\big)\nonumber\\
&\leq&\sum_{\tau\in
\mathcal{T}_h}h^2_{\tau}\|\tilde{\mathcal{R}}_{\tau}(E)+LE-\overline{(\tilde{\mathcal{R}}_{\tau}(E)+LE)}\|^2_{0,\tau}\nonumber\\
&& +\sum_{\tau\in\mathcal{T}_h}\big(h^2_{\tau}\|LE-\overline{LE}\|^2_{0,\tau}+
\sum_{e\in
\mathcal{E}_h,e\subset\partial \tau } h_e
\|\tilde{J}_e(E)-\overline{\tilde{J}_e(E)}\|^2_{0,e}\big),
\end{eqnarray}
where $E=P_h(w^h-w^H)$.
Following the proof of Proposition 3.3 in
\cite{cascon-kreuzer-nochetto-siebert-07}, we see that
$$\sum_{\tau\in\mathcal{T}_h}\big(h^2_{\tau}\|LE-\overline{LE}\|^2_{0,\tau}+
\sum_{e\in
\mathcal{E}_h,e\subset\partial \tau } h_e
\|\tilde{J}_e(E)-\overline{\tilde{J}_e(E)}\|^2_{0,e}\big)$$ can be
bounded by
\begin{eqnarray*}
\sum_{\tau\in\mathcal{T}_h}C_0^2
osc^2_h(\mathbf{A},\tau)\|P_h(w^h-w^H)\|^2_{1,\omega_{\tau}}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}
osc^2_h(\mathbf{A},\mathcal{T}_h)\|P_h(w^h-w^H)\|^2_{a,\Omega}.
\end{eqnarray*}Hence using the fact
$osc_h(\mathbf{A},\mathcal{T}_h)\leq
osc_0(\mathbf{A},\mathcal{T}_0)$, we obtain
\begin{eqnarray}\label{temp_osc}
&&\sum_{\tau\in\mathcal{T}_h}\big(h^2_{\tau}\|LE-\overline{LE}\|^2_{0,\tau}+
\sum_{e\in
\mathcal{E}_h,e\subset\partial \tau } h_e
\|\tilde{J}_e(E)-\overline{\tilde{J}_e(E)}\|^2_{0,e}\big)\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&
osc^2_0(\mathbf{A},\mathcal{T}_0)\|P_h(w^h-w^H)\|^2_{a,\Omega}.
\end{eqnarray}
Using the inverse inequality, the bounded property of $V$ and
(\ref{assume}), we get
\begin{eqnarray}\label{rightterm}
& &\big(\sum_{\tau\in
\mathcal{T}_h}h^2_{\tau}\|\tilde{\mathcal{R}}_{\tau}(E)+LE-
\overline{(\tilde{\mathcal{R}}_{\tau}(E)+LE)}\|^2_{0,\tau}\big)^{1/2}\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&\big(\sum_{\tau\in
\mathcal{T}_h}\|h_{\tau}(\ell_hu_h-\ell_Hu_H+Vu_H-Vu_h)\|^2_{0,\tau}\big)^{1/2}\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \|K(\ell_hu_h-\ell_Hu_H)\|_{a,\Omega}+h\|u_H-u_h\|_{a,\Omega}\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&\|K(\ell_hu_h-\ell u)\|_{a,\Omega}+\|K(\ell_Hu_H-\ell
u)\|_{a,\Omega}\nonumber\\
& & +h\|u-u_H\|_{a,\Omega}+h\|u-u_h\|_{a,\Omega}\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&\tilde{\kappa}(h_0)\left( \|u - u_h\|_{a, \Omega} + \|u -
u_H\|_{a, \Omega}\right).
\end{eqnarray}
Note that
\begin{eqnarray*}
& &\|P_h (w^h - w^H)\|_{a,\Omega} \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \| w^h - w^H\|_{a,\Omega} \nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \|u-w^h\|_{a,\Omega}+\|u-w^H\|_{a,\Omega},
\end{eqnarray*}
which together with (\ref{thm1-neq1}) implies
\begin{eqnarray}\label{temp1}
\|P_h (w^h - w^H)\|_{a,\Omega} &\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&\tilde{\kappa}(h_0)\left( \|u -
u_h\|_{a, \Omega} + \|u - u_H\|_{a, \Omega}\right).
\end{eqnarray}
Combing (\ref{temp4}), (\ref{temp_osc}), (\ref{rightterm}) and
(\ref{temp1}), we conclude that
\begin{eqnarray}\label{temp5}
\widetilde{osc}_h(P_h(w^h-w^H),\mathcal{T}_h)&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&
\tilde{\kappa}(h_0)\left(
\|u - u_h\|_{a, \Omega} + \|u - u_H\|_{a, \Omega}\right).
\end{eqnarray}
Due to $u_h=P_h w^H+P_h(w^h-w^H)$,
we obtain from the definition of oscillation that
\begin{eqnarray}\label{temp6}
\widetilde{osc}_h(P_h w^h,\mathcal{T}_h)\leq\widetilde{osc}_h(P_h
w^H,\mathcal{T}_h) + \widetilde{osc}_h(P_h (w^h-w^H),\mathcal{T}_h).
\end{eqnarray}
Hence from
$\widetilde{osc}_h(u_h,\mathcal{T}_h)=osc_h(u_h,\mathcal{T}_h)$,
(\ref{temp5}) and (\ref{temp6}), we arrive at
(\ref{lemma-bound-general-conc-2}).
Finally, we prove (\ref{lemma-bound-general-conc-3}). By
(\ref{boundary-lower}) and (\ref{temp5}), we have
\begin{eqnarray}\label{lem-bound-general5}
& &\tilde{\eta}_h(P_h( w^h - w^H), \mathcal{T}_h)\nonumber\\ &\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&
\| (w^h - w^H) -P_h(w^h - w^H)\|_{a, \Omega}+
\widetilde{osc}_h(P_h(w^h - w^H), \mathcal{T}_h)
\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} & \|u - w^h\|_{a, \Omega} + \|u - w^H\|_{a, \Omega} +
\tilde{\kappa}(h_0)\left( \|u - u_h\|_{a, \Omega} + \|u - u_H\|_{a,
\Omega}\right)\nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}&\tilde{\kappa}(h_0)\left( \|u - u_h\|_{a, \Omega} + \|u -
u_H\|_{a, \Omega}\right).
\end{eqnarray}
From (\ref{lem-bound-general5}) and the fact that
\begin{eqnarray*}
\tilde{\eta}_h(P_h w^h, \mathcal{T}_h) = \tilde{\eta}_h(P_h w^H +
P_h( w^h - w^H), \mathcal{T}_h),
\end{eqnarray*}
we obtain
\begin{eqnarray*}
\tilde{\eta}_h(P_h w^h, \mathcal{T}_h) = \tilde{\eta}_h(P_h w^H,
\mathcal{T}_h)
+ \mathcal{O} (\tilde{\kappa}(h_0))\left( \|u - u_h\|_{a, \Omega} + \|u -
u_H\|_{a, \Omega}\right),
\end{eqnarray*}
which is nothing but (\ref{lemma-bound-general-conc-3}) since
$\tilde{\eta}_h(P_h w^h, \mathcal{T}_h) =
\eta_h(u_h,\mathcal{T}_h)$.
\end{proof}
\begin{theorem}
Let $h_0 \ll 1$ and $h \in (0, h_0].$ There exist constants $C_1,
C_2$ and $C_3$, which only depend on the shape regularity constant
$\gamma^{\ast}$, $C_a$ and $c_a$ such that
\begin{eqnarray}\label{upper-bound}
\|u-u_h\|^2_{a,\Omega} \leq C_1 \eta^2_h(u_h, \mathcal{T}_h)
\end{eqnarray}
and
\begin{eqnarray}\label{lower-bound}
~~~~~~C_2 \eta^2_h(u_h, \mathcal{T}_h) \le
\|u-u_h\|_{a,\Omega}^2+ C_3osc_h^2(u_h, \mathcal {T}_h).
\end{eqnarray}
\end{theorem}
\begin{proof} Recall that $L w^h = \ell_{h}u_h-Vu_h$.
From (\ref{boundary-upper}) and
(\ref{boundary-lower}) we have
\begin{eqnarray}\label{auxiliary-boundary-problem-upper}
\|w^h - P_h w^h \|^2_{a,\Omega} \leq \tilde{C}_1 \tilde{\eta}^2_h (P_h w^h, \mathcal{T}_h)
\end{eqnarray}
and
\begin{eqnarray}\label{auxiliary-boundary-problem-lower}
\tilde{C}_2 \tilde{\eta}^2_h (P_h w^h, \mathcal{T}_h) &\le&
\|w^h-P_h
w^h\|_{a,\Omega}^2+\tilde{C}_3\widetilde{osc}^2_h(P_h w^h,
\mathcal{T}_h) .
\end{eqnarray}
Thus we obtain (\ref{upper-bound}) and (\ref{lower-bound}) from
(\ref{u-w}), (\ref{general-boundary-neq}),
(\ref{auxiliary-boundary-problem-upper}) and
(\ref{auxiliary-boundary-problem-lower}). In particular, we may
choose $C_1$, $C_2$ and $C_3$ satisfying
\begin{eqnarray}\label{coef-eigen-bound}
C_1 = \tilde{C}_1 (1+ \tilde{C} \tilde{\kappa}(h_0))^2, ~~C_2 = \tilde{C}_2 (1 - \tilde{C}
\tilde{\kappa}(h_0))^2, ~~C_3 = \tilde{C}_3 (1 - \tilde{C}\tilde{\kappa}(h_0))^2.
\end{eqnarray}
\end{proof}
\begin{remark}
The requirement $h_0\ll 1$ is somehow reasonable for finite element
approximations of (\ref{Gvariation}). We can refer to
\cite{mekchay-nochetto-05} for the initial mesh size requirement in
adaptive finite element computations for nonsymmetirc boundary value
problems.
\end{remark}\vskip 0.2cm
Now we address step {\bf MARK} of solving (\ref{Gdis-fem}) in
detail, which we call {\bf Marking Strategy $E$}. Similar to {\bf
Marking Strategy $E_0$} for (\ref{dis-fem}), we define {\bf Marking
Strategy $E$} for (\ref{Gdis-fem}) to enforce error reduction as
follows:
Given a parameter $0<\theta < 1$:
\begin{enumerate}
\item Construct a minimal subset $\mathcal{M}_k$ of $\mathcal{T}_k$ by selecting some elements in
$\mathcal{T}_k$ such that
\begin{eqnarray*}
\eta_k(u_k, \mathcal{M}_k) \geq \theta
\eta_k(u_k, \mathcal{T}_k).
\end{eqnarray*}
\item Mark all the elements in $\mathcal{M}_k$.
\end{enumerate}
The adaptive algorithm of solving (\ref{Gdis-fem}), which we call
{\bf Algorithm $D$}, is nothing but {\bf Algorithm $D_0$} when {\bf
Marking Strategy $E_0$} is replaced by {\bf Marking Strategy $E$}.
\subsection{Convergence}\label{section-convergence}
We now prove that {\bf Algorithm $D$} of
(\ref{Gdis-fem}) is a contraction with respect to the sum of the
energy error plus the scaled error estimator.
\begin{theorem}\label{error-reduction}
Let $\theta \in (0,1)$ and $\{u_k\}_{k\in\mathbb{N}_0}$ be a
sequence of finite element
solutions corresponding to a sequence of nested finite element
spaces $\{S^k_0(\Omega)\}_{k\in \mathbb{N}_0}$ produced by {\bf Algorithm
$D$}.
Then there exist constants
$\gamma>0$ and $\xi \in (0,1)$
depending only on the shape regularity constant $\gamma^{\ast}$, $ C_a $, $c_a$ and the marking parameter
$\theta$ such that
\begin{eqnarray}\label{error-reduction-neq1}
& & \|u-u_{k+1}\|_{a,\Omega}^2+\gamma \eta^2_{k+1}(u_{k+1}, \mathcal{T}_{k+1})\nonumber\\
&\leq& \xi^2 \big(\|u-u_k\|_{a,\Omega}^2
+\gamma \eta^2_k(u_k, \mathcal{T}_k)\big).
\end{eqnarray}
Here,
\begin{eqnarray}\label{gamma}
\gamma := \frac{\tilde{\gamma}}{1 - C_4
\delta_1^{-1}\tilde{\kappa}^2(h_0)}
\end{eqnarray}
with $C_4$ a positive constant, provided $h_0\ll 1$.
\end{theorem}
\begin{proof}
For convenience, we use $u_h$, $u_H$ to denote $u_{k+1}$ and
$u_{k}$, respectively. Thus we only need to prove that for $u_h$ and
$u_H$, there holds,
\begin{eqnarray*}\label{error-reduction-neq-2}
~ \|u-u_h\|_{a,\Omega}^2 + \gamma \eta^2_{h}(u_{h},
\mathcal{T}_h)\leq \xi^2 \big(\|u-u_H\|_{a,\Omega}^2 +\gamma
\eta^2_{H}(u_H, \mathcal{T}_H)\big).
\end{eqnarray*}
We conclude from Theorem \ref{convergence-boundary}, $w^h=K(\ell_h
u_h-Vu_h)$ and $ w^H=K(\ell_H u_H-Vu_H)$ that there exist constants
$\tilde{\gamma}>0$ and $\tilde{\xi}\in (0,1)$ satisfying
\begin{eqnarray*}
& & \|w^H-P_h w^H\|_{a,\Omega}^2 +\tilde{\gamma}\tilde{\eta}^2_h(P_h
w^H, \mathcal{T}_h) \nonumber\\
&\leq& \tilde{\xi}^2 \big( \|w^H- P_H w^H \|_{a,\Omega}^2+ \tilde{\gamma}
\tilde{\eta}^2_H(P_H w^H,\mathcal{T}_H)\big).
\end{eqnarray*}
Hence use the fact that $u_H=P_Hw^H$, we obtain
\begin{eqnarray}\label{error-reduction-neq-4}
&& \|w^H-P_h
w^H\|_{a,\Omega}^2 +\tilde{\gamma}\tilde{\eta}^2_h(P_h w^H,
\mathcal{T}_h) \nonumber\\&\leq& \tilde{\xi}^2 \big( \|w^H- u_H
\|_{a,\Omega}^2+ \tilde{\gamma} \eta^2_H(u_H, \mathcal{T}_H)\big).
\end{eqnarray}
By (\ref{lemma-bound-general-conc-1}) and
(\ref{lemma-bound-general-conc-3}), there exists a constant
$\hat{C}>0$ such that
\begin{eqnarray*}
& &\|u - u_h\|_{a, \Omega}^2 + \tilde{\gamma}\eta^2_h(u_h,
\mathcal{T}_h)\nonumber\\
&\leq& (1+\delta_1) \|w^H - P_h w^H\|_{a, \Omega}^2+ (1 + \delta_1) \tilde{\gamma} \tilde{\eta}^2_h(P_h w^H,
\mathcal{T}_h) \nonumber\\
&&+ \hat{C} (1+\delta_1^{-1})\tilde{\kappa}^2(h_0) (\|u - u_h\|_{a, \Omega}^2 + \|u -
u_H\|_{a, \Omega}^2)\nonumber\\
&&+\hat{C} (1+\delta_1^{-1}) \tilde{\kappa}^2(h_0)
\tilde{\gamma}(\|u-u_h\|_{a, \Omega}^2+\|u - u_H\|_{a, \Omega}^2),
\end{eqnarray*}
where the Young's inequality is used and $\delta_1\in (0,1)$
satisfies
\begin{eqnarray}\label{error-reduction-neq-delta}
(1 + \delta_1) \tilde{\xi}^2 <1.
\end{eqnarray}
It thus follows from (\ref{error-reduction-neq-4}),
(\ref{thm1-neq1}), and identity
$\tilde{\eta}_H(P_H w^H, \mathcal{T}_H) = \eta_H(u_H,
\mathcal{T}_H)$ that there exists a positive constant $C^{\ast}$
depending on $\hat{C}$ and $\tilde{\gamma}$ such that
\begin{eqnarray*}
& & \|u - u_h\|_{a, \Omega}^2 + \tilde{\gamma}\eta^2_h(u_h,
\mathcal{T}_h)\nonumber\\
&\le& (1+ \delta_1) \tilde{\xi}^2 \big(\|w^H- u_H \|_{a, \Omega}^2+
\tilde{\gamma} \eta^2_H(u_H, \mathcal{T}_H)\big)\nonumber\\
& & + C^{\ast} \delta_1^{-1} \tilde{\kappa}^2(h_0) (\|u-u_h\|_{a,
\Omega}^2+\|u - u_H\|_{a, \Omega}^2)\nonumber\\
&\leq& (1+ \delta_1) \tilde{\xi}^2 \left(\big( 1 +
\tilde{C}\tilde{\kappa}(h_0)\big)^2 \|u -
u_H\|_{a, \Omega}^2 + \tilde{\gamma} \eta^2_H(u_H, \mathcal{T}_H)\right) \nonumber\\
&& + C^{\ast} \delta_1^{-1}
\tilde{\kappa}^2(h_0)\left(\|u-u_h\|_{a, \Omega}^2+\|u - u_H\|_{a,
\Omega}^2\right).
\end{eqnarray*}
Hence, if $h_0\ll 1$, then there exists a positive constant $C_4$
depending on $C^{\ast}$ and $\tilde{C}$ such that
\begin{eqnarray*}
&&\|u - u_h\|_{a, \Omega}^2 + \tilde{\gamma}\eta^2_h(u_h,
\mathcal{T}_h)\nonumber\\ &\leq& (1+
\delta_1) \tilde{\xi}^2\left(\|u - u_H\|^2_{a, \Omega} +
\tilde{\gamma}
\eta^2_H(u_H, \mathcal{T}_H)\right)\nonumber\\
&& + C_4 \tilde{\kappa}(h_0) \|u - u_H\|^2_{a, \Omega} + C_4
\delta_1^{-1} \tilde{\kappa}^2(h_0) \|u - u_h\|_{a, \Omega}^2.
\end{eqnarray*}
Consequently,
\begin{eqnarray*}
&&\big(1 - C_4 \delta_1^{-1} \tilde{\kappa}^2(h_0)\big) \|u -
u_h\|_{a,
\Omega}^2 + \tilde{\gamma}\eta^2_h(u_h, \mathcal{T}_h)\nonumber\\
& \leq &\big((1+ \delta_1) \tilde{\xi}^2 + C_4
\tilde{\kappa}(h_0)\big)\|u - u_H\|_{a, \Omega}^2 + (1+ \delta_1) \tilde{\xi}^2 \tilde{\gamma}
\eta^2_H(u_H, \mathcal{T}_H),
\end{eqnarray*}
that is
\begin{eqnarray*}
& &\|u -u_h\|_{a, \Omega}^2 + \frac{\tilde{\gamma}}{1 - C_4
\delta_1^{-1} \tilde{\kappa}^2(h_0)}\eta^2_h(u_h, \mathcal{T}_h)\nonumber\\
~~~~ &\leq& \frac{ (1+ \delta_1) \tilde{\xi}^2 + C_4
\tilde{\kappa}(h_0) }{1 - C_4 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}\|u -u_H\|_{a, \Omega}^2
+\frac{(1+ \delta_1) \tilde{\xi}^2 \tilde{\gamma}}{1 - C_4 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}\eta^2_H(u_H, \mathcal{T}_H).
\end{eqnarray*}
Since $h_0\ll 1$ implies ${\tilde r}(h_0)\ll 1$, we have that the
constant $\xi$ defined by
\begin{eqnarray*}
\xi := \left(\frac{ (1+ \delta_1) \tilde{\xi}^2 + C_4
\tilde{\kappa}(h_0)}{1 - C_4 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}\right)^{1/2}
\end{eqnarray*}
satisfying $\xi \in (0,1)$ if $h_0\ll 1$. Therefore,
\begin{eqnarray*}
&&\|u -u_h\|_a^2 + \frac{\tilde{\gamma}}{1 - C_4
\delta_1^{-1}\tilde{\kappa}^2(h_0)}\eta^2_h(u_h, \mathcal{T}_h)\nonumber\\
& \leq& \xi^2\left(\|u -u_H\|_{a, \Omega}^2 +
\frac{(1+ \delta_1) \tilde{\xi}^2 \tilde{\gamma}}{ (1+ \delta_1) \tilde{\xi}^2 +
C_4
\tilde{\kappa}(h_0) } \eta^2_H(u_H, \mathcal{T}_H)\right).
\end{eqnarray*}
Finally, we arrive at (\ref{error-reduction-neq1}) by using the fact
that
$$ \frac{(1+
\delta_1) \tilde{\xi}^2 \tilde{\gamma}}{ (1+ \delta_1) \tilde{\xi}^2
+ C_4
\tilde{\kappa}(h_0) }<\gamma.$$
This
completes the proof.
\end{proof}
\subsection{Complexity}\label{optimal-complexity}
We shall study the complexity in a class of functions defined by
\begin{eqnarray*}
\mathcal{A}_{\gamma}^s:=\{v \in H^1_0(\Omega): |v|_{s,\gamma} <
\infty \},
\end{eqnarray*}
where $\gamma>0$ is some constant,
\begin{eqnarray*}
|v|_{s,\gamma} = \sup_{\varepsilon
>0}\varepsilon \inf_{\{\mathcal{T}_k\subset \mathcal{T}_0:
\inf (\|v-v_k\|_{a,\Omega}^2 + (\gamma +1) osc^2_k(v_k,
\mathcal{T}_k))^{1/2} \leq \varepsilon\}} \big(\#\mathcal{T}_k - \#
\mathcal{T}_0\big)^s
\end{eqnarray*}
and $\mathcal{T}_k\subset \mathcal{T}_0$ means $\mathcal{T}_k$ is a
refinement of $\mathcal{T}_0$. It is seen from the definition that,
for all $\gamma>0$, $\mathcal{A}_{\gamma}^s = \mathcal{A}_{1}^s$.
For simplicity, here and hereafter, we use $\mathcal{A}^s$ to stand
for $\mathcal{A}_{1}^s$, and use $|v|_{s}$ to denote
$|v|_{s, \gamma}$. So $\mathcal{A}^s$ is the class of
functions that can be approximated within a given tolerance
$\varepsilon$ by continuous piecewise polynomial functions over a
partition $\mathcal{T}_k$ with number of degrees of freedom
$\#\mathcal{T}_k-\# \mathcal{T}_0 \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \varepsilon^{-1/s}
|v|_{s}^{1/s}$.
In order to give the proof of the complexity of {\bf Algorithm D}
for solving (\ref{Gdis-fem}), we need some preparations. Recall that
associated with $u_k$, the solution of (\ref{Gdis-fem}) in each mesh
$\mathcal{T}_k$, $w^k=K(\ell_k u_k-Vu_k)$ satisfies
\begin{eqnarray}\label{auxiliary-boundary-eq}
a(w^k, v) = (\ell_ku_k-Vu_k, v) ~~~~~ \forall v\in
H^1_0(\Omega).
\end{eqnarray}
\iffalse
By Theorem \ref{thm-error-estimator}, we may choose
$\tilde{C}_1$ and $C_1$, $\tilde{C}_1$ and $C_1$ that have the
following relationship,
\begin{eqnarray}\label{complexity-coef-eigen-bound}
C_1 =\tilde{C}_1 (1+ \tilde{C} {\tilde r}(h_0)), ~~~ C_2 = \tilde{C}_2(1 - \tilde{C}
{\tilde r}(h_0)).
\end{eqnarray}
\fi
Using the similar procedure as in the proof of Theorem
\ref{error-reduction}, we have
\begin{lemma}\label{complexity-general-boundary}
Let $u_k$ and $u_{k+1}$ be discrete
solutions of (\ref{Gdis-fem}) over a
conforming mesh $\mathcal{T}_k$ and its refinement
$\mathcal{T}_{k+1}$ with marked
set $\mathcal{M}_k$.
Suppose that they
satisfy the following property
\begin{eqnarray*}
&&\|u - u_{k+1} \|_{a,\Omega}^2 + \gamma_{\ast} osc^2_{k+1}(u_{k+1},\mathcal{T}_{k+1}) \nonumber\\
& \leq& \beta_{\ast}^2 \big(\|u-u_k\|_{a,\Omega}^2 + \gamma_{\ast} osc^2_k(u_k,\mathcal{T}_k)\big),
\end{eqnarray*}
where $ \gamma_{\ast}$ and $\beta_{\ast}$ are some positive constants.
Then for problem (\ref{auxiliary-boundary-eq}), we have
\begin{eqnarray*}\label{lemma-complexity-general-bound-conc2}
&& \|w^k - P_{k+1} w^k \|_{a,\Omega}^2 + \tilde{\gamma_{\ast}}
\widetilde{osc}_{k+1}^2(P_{k+1} w^k,\mathcal{T}_{k+1})\nonumber\\
&\leq& \tilde{\beta_{\ast}}^2 \big( \|w^k -P_k w^k\|^2_{a,\Omega}
+ \tilde{\gamma_{\ast}} \widetilde{osc}^2_k(P_k
w^k,\mathcal{T}_k)\big)
\end{eqnarray*}
with
\begin{eqnarray}\label{complexity-general-boundary-beta-gamma}
\tilde{\beta_{\ast}} := \left(\frac{ (1+ \delta_1) \beta_{\ast}^2 +
C_5
\tilde{\kappa}(h_0) }{1 - C_5 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}\right)^{1/2}, \quad \tilde{\gamma_{\ast}} :=
\frac{\gamma_{\ast}}{1 - C_5 \delta_1^{-1}
\tilde{\kappa}^2(h_0)},
\end{eqnarray}
where $C_5 $ is some positive constant and $\delta_1 \in (0, 1)$ is
some constant as in the proof of Theorem \ref{error-reduction}.
\end{lemma}
\begin{corollary}\label{optimal-marking}
Let $u_k$ and $u_{k+1}$ be as those in Lemma
\ref{complexity-general-boundary} . Suppose that they satisfy the
decrease property
\begin{eqnarray*}
&&\|u - u_{k+1} \|_{a,\Omega}^2 + \gamma_{\ast} osc^2_{k+1}(u_{k+1},\mathcal{T}_{k+1}) \nonumber\\
& \leq& \beta_{\ast}^2 \big(\|u-u_k\|_{a,\Omega}^2 + \gamma_{\ast} osc^2_k(u_k,\mathcal{T}_k)\big)
\end{eqnarray*}
with constants $\gamma_{\ast}>0$ and $\beta_{\ast}\in
(0,\sqrt{\frac{1}{2}})$. Then the set
$\mathcal{R}:=R_{\mathcal{T}_k\rightarrow \mathcal{T}_{k+1}}$
satisfies the following inequality
\begin{eqnarray*}\label{optimal-marking-neq1}
\eta_k(u_k,\mathcal{R})\geq \hat{\theta}\eta_k(u_k,\mathcal{T}_k)
\end{eqnarray*}
with $\hat{\theta}^2 =
\frac{\tilde{C}_2(1-2\tilde{\beta_{\ast}}^2)}{\tilde{C}_0 (
\tilde{C}_1 + (1 + 2 C \tilde{C}_1) \tilde{\gamma_{\ast}})}$ and
$\tilde{C}_0 = \max(1, \frac{\tilde{C}_3}{\tilde{\gamma_{\ast}}})$,
where $\tilde{\beta_{\ast}}$ and $\tilde{\gamma_{\ast}}$ are defined
in (\ref{complexity-general-boundary-beta-gamma}) with $\delta_1$
being chosen such that $\tilde{\beta_{\ast}}^2\in (0,\frac{1}{2})$.
\end{corollary}
\begin{proof}
It is a direct consequence of combining $u_k=P_k w^k$ with Lemma
\ref{complexity-general-optimal-marking} and Lemma
\ref{complexity-general-boundary}.
\end{proof}
The key to relate the best mesh with AFEM triangulations is the fact
that procedure {\bf MARK} selects the marked set $\mathcal{M}_k$
with minimal cardinality.
\begin{lemma}
\textnormal{(Cardinality of $\mathcal{M}_k$).} Let $u\in
\mathcal{A}^s$, $\mathcal{T}_k$ be a conforming partition obtained
from $\mathcal{T}_0$, and $\theta$ satisfies $\theta\in (0,\frac{C_2
\gamma}{ C_3(C_1 + (1 + 2C C_1)\gamma )})$. Then the following
estimate is valid:
\begin{eqnarray}\label{DOF}
\# \mathcal{M}_k \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}
\left(\|u-u_k\|_{a,\Omega}^2 + \gamma osc^2_k(u_k, \mathcal{T}_k)\right)^{-1/2s} |u|_{s}^{1/s},
\end{eqnarray}
where the hidden constant depends on the discrepancy between
$\theta$ and $\frac{C_2 \gamma}{C_3( C_1 + (1 + 2CC_1)\gamma) }$
with $C$ defined in Lemma \ref{complexity-general-optimal-marking}.
\end{lemma}
\begin{proof}
Let $\alpha, \alpha_1 \in (0,1)$ satisfy $\alpha_1\in (0,\alpha) $
and
$$\theta < \frac{C_2 \gamma}{ C_3(C_1 + (1 + 2CC_1)\gamma)
}(1-\alpha^2).$$
Choose $\delta_1\in (0,1)$ to satisfy
(\ref{error-reduction-neq-delta}) and
\begin{eqnarray}\label{thm-complexity-delta-cond-1}
(1+\delta_1)^2 \alpha_1^2 \leq \alpha^2,
\end{eqnarray}
which implies
\begin{eqnarray}\label{thm-complexity-delta-cond-3}
(1+\delta_1) \alpha_1^2 <1.
\end{eqnarray}
Set
$$\varepsilon = \frac{1}{\sqrt{2}} \alpha_1 \big(\|u-u_k\|_{a,\Omega}^2 +
\gamma osc^2_k(u_k,
\mathcal{T}_k)\big)^{1/2}$$ and let $\mathcal{T}_{\varepsilon}$ be a refinement of
$\mathcal{T}_0$ with minimal degrees of freedom satisfying
\begin{eqnarray}\label{complexity-optimal-neq0}
\|u - u_{\varepsilon}\|_{a,\Omega}^2 + (\gamma + 1) osc^2_{\varepsilon}(u_{\varepsilon}, \mathcal{T}_{\varepsilon}) \leq
\varepsilon^2.
\end{eqnarray}
It follows from the definition of $\mathcal{A}^s$ that
\begin{eqnarray*}\label{upper-bound-dof-neq1}
~~~~\#\mathcal{T}_{\varepsilon} - \# \mathcal{T}_0 \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} {\varepsilon}^{-1/s} |u|_{s}^{1/s}.
\end{eqnarray*}
Let $\mathcal{T}_{\ast}=\mathcal{T}_{\varepsilon}\oplus \mathcal{T}_k$ be the smallest common refinement of
$\mathcal{T}_k$ and $\mathcal{T}_{\varepsilon}$.
Note that $w^{\varepsilon}=K(\ell_{\varepsilon}
u_{\varepsilon}-Vu_{\varepsilon})$ satisfies
\begin{eqnarray*}\label{complexity-boundary-problem-2}
L w^{\varepsilon} = \ell_{\varepsilon}
u_{\varepsilon}-Vu_{\varepsilon},
\end{eqnarray*}
we get from the definition of oscillation and Young's inequality
that
\begin{eqnarray*}
\widetilde{osc}^2_{\ast}(P_{\ast}w^{\varepsilon},\tau)
&\leq&
2\widetilde{osc}^2_{\ast}(P_{\varepsilon}w^{\varepsilon},\tau)+2C_0^2osc_{\ast}^2(\mathbf{A},\tau)\|P_{\varepsilon}w^{\varepsilon}-
P_{\ast}w^{\varepsilon}\|^2_{1,\omega_{\tau}} ~~\forall \tau\in\mathcal
{T}_{\ast},
\end{eqnarray*}
which together with the monotonicity property
$osc_{\ast}(\mathbf{A},\mathcal{T}_{\ast})\leq
osc_0(\mathbf{A},\mathcal{T}_0)$ yields
\begin{eqnarray*}
\widetilde{osc}^2_{\ast}(P_{\ast}w^{\varepsilon},\mathcal{T}_{\ast})
&\leq&
2\widetilde{osc}^2_{\ast}(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\ast})+2C\|P_{\varepsilon}w^{\varepsilon}-
P_{\ast}w^{\varepsilon}\|^2_{a,\Omega},
\end{eqnarray*}
where $C=\Lambda_1 osc^2_0(\mathbf{A},\mathcal{T}_0)$. Due to the
orthogonality
\begin{eqnarray*}\label{ortho-relation}
\|w^{\varepsilon}-P_{\ast}w^{\varepsilon}\|^2_{a, \Omega} =
\|w^{\varepsilon} - P_{\varepsilon}w^{\varepsilon}\|^2_{a, \Omega}
- \|P_{\ast}w^{\varepsilon} -
P_{\varepsilon}w^{\varepsilon}\|_{a, \Omega}^2,
\end{eqnarray*}
we arrive at
\begin{eqnarray*}
&&\| w^{\varepsilon} - P_{\ast}
w^{\varepsilon}\|_{a,\Omega}^2 + \frac{1}{2 C}\widetilde{osc}^2_{\ast}(P_{\ast}w^{\varepsilon},\mathcal{T}_{\ast})\nonumber\\
&\leq & \|w^{\varepsilon} - P_{\varepsilon}
w^{\varepsilon}\|_{a,\Omega}^2+ \frac{1}{C}
osc^2_{\varepsilon}(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\varepsilon}).
\end{eqnarray*}
Since (\ref{gamma-boundary}) implies $\tilde{\gamma} \leq \frac{1}{2
C}$, we obtain that
\begin{eqnarray*}
&&\| w^{\varepsilon} - P_{\ast}
w^{\varepsilon}\|_{a,\Omega}^2 + \tilde{\gamma}\widetilde{osc}^2_{\ast}(P_{\ast}w^{\varepsilon},\mathcal{T}_{\ast})\nonumber\\
&\leq & \|w^{\varepsilon} - P_{\varepsilon} w^{\varepsilon}\|_{a,\Omega}^2
+ \frac{1}{ C}
osc^2_{\varepsilon}(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\varepsilon})\nonumber\\
&\leq & \|w^{\varepsilon} - P_{\varepsilon}
w^{\varepsilon}\|_{a,\Omega}^2+ (\tilde{\gamma} + \sigma)
osc^2_{\varepsilon}(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\varepsilon})
\end{eqnarray*}
with $\sigma = \frac{1}{C} - \tilde{\gamma} \in (0, 1) $.
Applying the similar argument in the proof of Theorem
\ref{error-reduction} when (\ref{lemma-bound-general-conc-3}) is
replaced by (\ref{lemma-bound-general-conc-2}), we then get
\begin{eqnarray}\label{complexity-optimal-neq2}
&&\|u - u_{\ast}\|_{a,\Omega}^2 + \gamma osc^2_{\ast}(u_{\ast},
\mathcal{T}_{\ast})\nonumber\\
& \leq& \alpha_0^2\left(
\|u - u_{\varepsilon}\|_{a,\Omega}^2 + (\gamma + \sigma)osc^2_{\varepsilon}
(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\varepsilon})\right) \nonumber\\
&\leq & \alpha_0^2\left(\|u - u_{\varepsilon}\|_{a,\Omega}^2
+ (\gamma + 1)
osc^2_{\varepsilon}
(P_{\varepsilon}w^{\varepsilon},\mathcal{T}_{\varepsilon})\right),
\end{eqnarray}
where
\begin{eqnarray*}
\alpha_0^2 := \frac{ (1+ \delta_1) + C_4
\tilde{\kappa}(h_0) }{1 - C_4 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}
\end{eqnarray*}
and $C_4$ is the constant appearing in the proof of Theorem
\ref{error-reduction}. Thus, by (\ref{complexity-optimal-neq0}) and
(\ref{complexity-optimal-neq2}), it follows
\begin{eqnarray*} \|u - u_{\ast}\|_{a,\Omega}^2 + \gamma osc^2_{\ast}(u_{\ast},
\mathcal{T}_{\ast}) \leq \check{\alpha}^2 \big(\|u-u_k\|^2_{a,\Omega} + \gamma osc^2_k(u_k,
\mathcal{T}_k)\big)
\end{eqnarray*}
with $\check{\alpha} = \frac{1}{\sqrt{2}} \alpha_0 \alpha_1$. In
view of (\ref{thm-complexity-delta-cond-3}), we have
$\check{\alpha}^2\in (0,\frac{1}{2})$ when $h_0\ll 1$. Let $\mathcal
{R}:=R_{\mathcal{T}_k\rightarrow \mathcal{T}_{\ast}}$, by Corollary
\ref{optimal-marking}, we have that
$\mathcal{T}_{\ast}$ satisfies
\begin{eqnarray*}
\eta_k(u_k, \mathcal {R}) \geq \check{\theta}
\eta_k(u_k,\mathcal{T}_k),
\end{eqnarray*}
where $\check{\theta}^2 =
\frac{\tilde{C}_2(1-2\hat{\alpha}^2)}{\tilde{C}_0 ( \tilde{C}_1 + (1
+ 2 C \tilde{C}_1)\hat{\gamma})}, \quad \hat{\gamma}=
\frac{\gamma}{1 - C_5 \delta_1^{-1} \tilde{\kappa}^2(h_0)}$,
$\tilde{C}_0 = \max(1, \frac{\tilde{C}_3}{\hat{\gamma}})$, and
\begin{eqnarray*}
\hat{\alpha}^2= \frac{ (1+ \delta_1)\check{\alpha}^2 + C_5
\tilde{\kappa}(h_0) }{1 - C_5 \delta_1^{-1}
\tilde{\kappa}^2(h_0)}.
\end{eqnarray*}
It follows from the
definition of $\gamma$ (see (\ref{gamma})) and $\tilde{\gamma}$ (see (\ref{gamma-boundary}))
that $\hat{\gamma}<1$ and hence ${\tilde C}_0 =
\frac{\tilde{C}_3}{\hat{\gamma}}.$
Since $h_0 \ll 1$,
we obtain that $\hat{\gamma}>\gamma$ and $\hat{\alpha}\in
(0,\frac{1}{\sqrt{2}}\alpha)$ from (\ref{thm-complexity-delta-cond-1}).
It is easy to see from (\ref{coef-eigen-bound}) and
$\hat{\gamma}>\gamma$ that
\begin{eqnarray*}
&&\check{\theta}^2=\frac{\tilde{C}_2(1-2\hat{\alpha}^2)}{\frac{\tilde{C}_3}{\hat{\gamma}}
(\tilde{C}_1 + (1 + 2 C\tilde{C}_1)\hat{\gamma} )}
\geq \frac{\tilde{C}_2}{\tilde{C}_3(
\frac{\tilde{C}_1}{\hat{\gamma}} + 1 + 2 C\tilde{C}_1
)}(1-\alpha^2)\nonumber\\
&=&
\frac{\frac{C_2}{(1-\tilde{C}\tilde{\kappa}(h_0))^2}}{\frac{C_3}{(1-\tilde{C}\tilde{\kappa}(h_0))^2}
(\frac{C_1}{\hat{\gamma}((1+\tilde{C}\tilde{\kappa}(h_0))^2)}+1
+2C\frac{C_1}{(1+\tilde{C}\tilde{\kappa}(h_0))^2})}(1-\alpha^2)
\nonumber\\&\geq& \frac{C_2 }{C_3 ( \frac{C_1}{\gamma} + (1 + 2 C
C_1) )}(1-\alpha^2) = \frac{C_2 \gamma }{C_3 ( C_1 + (1 + 2 C
C_1)\gamma )}(1-\alpha^2)
> \theta
\end{eqnarray*}
when $h_0 \ll 1$. Thus
\begin{eqnarray*}
\#\mathcal{M}_k &\leq&
\#\mathcal{R}
\leq \#\mathcal{T}_{\ast} - \#\mathcal{T}_k
\leq \#\mathcal{T}_{\varepsilon}- \#\mathcal{T}_0 \nonumber\\
&\leq&(\frac{1}{\sqrt{2}}\alpha_1)^{-1/s}
\left(\|u - u_k\|_{a,\Omega}^2 + \gamma osc^2_k(u_k,
\mathcal{T}_k)\right) ^{-1/2s} |u|_{s}^{1/s},
\end{eqnarray*}
which is the desired estimate (\ref{DOF}) with an explicit
dependence on the discrepancy between $\theta$ and $\frac{C_2
\gamma}{C_3( C_1 + (1 + 2CC_1)\gamma )}$ via $\alpha_1$. This
completes the proof.
\end{proof}
As a consequence, we obtain the optimal complexity as follows.
\begin{theorem}\label{thm-optimal-complexity}
Let $u \in \mathcal{A}^s$ and $\{u_k\}_{k\in \mathbb{N}_0}$ be a
sequence of finite element
solutions corresponding to a sequence of nested finite element spaces $\{{S^k_0(\Omega)}\}_{k\in \mathbb{N}_0}$
produced by {\bf Algorithm $D$}.
Then
\begin{eqnarray*}
\|u-u_k\|_{a,\Omega}^2 + \gamma osc^2_k(u_k, \mathcal{T}_k) \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}
(\#\mathcal{T}_k
-\#\mathcal{T}_0)^{-2s}|u|_s^2,
\end{eqnarray*}
where the hidden constant depends on the exact solution u and
the discrepancy between
$\theta$ and $\frac{C_2 \gamma}{C_3( C_1 + (1 + 2 C C_1)\gamma)
}$.
\end{theorem}
\begin{proof}
It follows from (\ref{complexity}) and (\ref{DOF}) that
\begin{eqnarray*}
& &\#\mathcal{T}_k - \#\mathcal{T}_0 \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \sum_{j=0}^{k-1}
\#\mathcal{M}_j \nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \sum_{j=0}^{k-1}\left(\|u - u_j\|_{a,\Omega}^2 + \gamma osc^2_j(u_j,
\mathcal{T}_j)\right)^{-1/2s}|u|_{s}^{1/s}.
\end{eqnarray*}
Note that (\ref{lower-bound}) implies
\begin{eqnarray*}
\|u - u_j\|^2_{a, \Omega} + \gamma \eta^2_j(u_j, \mathcal{T}_j) \leq
\check{C} \big( \|u - u_j\|^2_{a, \Omega} + \gamma osc^2_j(u_j,
\mathcal{T}_j)\big),
\end{eqnarray*}
where $ \check{C} = \max(1 + \frac{\gamma}{C_2}, \frac{C_3}{C_2}).$
It then turns out
\begin{eqnarray*}
\#\mathcal{T}_k - \#\mathcal{T}_0 &\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& \sum_{j=0}^{k-1}
\left(\|u - u_j\|_{a,\Omega}^2 + \gamma \eta^2_j(u_j,
\mathcal{T}_j)\right)^{-1/2s}|u|_{s}^{1/s}.
\end{eqnarray*}
Due to (\ref{error-reduction-neq1}), we obtain for $0\leq j < k$
that
\begin{eqnarray*}
\|u-u_k\|_{a,\Omega}^2 + \gamma \eta^2_k(u_k, \mathcal{T}_k)\leq
\xi^{2(k-j)} \left( \|u-u_j\|_{a,\Omega}^2 + \gamma \eta^2_j(u_j,
\mathcal{T}_j)\right).
\end{eqnarray*}
Consequently,
\begin{eqnarray*}
\#\mathcal{T}_k-\#\mathcal{T}_0
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& |u|_{s}^{1/s} \left(\|u-u_k\|_{a,\Omega}^2 + \gamma \eta^2_k(u_k, \mathcal{T}_k)\right)^{-1/2s}
\sum_{j=0}^{k-1}\xi^{\frac{k-j}{s}} \nonumber\\
&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}& |u|_{s}^{1/s} \left(\|u-u_k\|_{a,\Omega}^2 + \gamma \eta^2_k(u_k, \mathcal{T}_k)\right)^{-1/2s},
\end{eqnarray*}
the last inequality holds because of the fact $\xi<1$.
Since $osc_k(u_k, \mathcal{T}_k) \leq \eta_k(u_k,\mathcal{T}_k)$,
we
arrive at
\begin{eqnarray*}
\#\mathcal{T}_k-\#\mathcal{T}_0
\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \left(\|u-u_k\|_{a,\Omega}^2 +
\gamma osc_k^2(u_k, \mathcal{T}_k)\right)^{-1/2s}|u|_{s}^{1/s}.
\end{eqnarray*}
This completes the proof.
\end{proof}
\section{Applications}
In this section, we provide three typical examples to show that our
general theory is quite useful.
\subsection{A nonsymmetric problem}\label{nonsymmetric}
The first example is a nonsymmetric elliptic partial differential
equation of second order. We consider the following problem: Find $u
\in H^1_0(\Omega)$ such that
\begin{eqnarray}\label{application2}
\left\{\begin{array}{rl}
-\nabla \cdot (\mathbf{A}\nabla u)+ {\bf b} \cdot \nabla u + cu&=f \,\,\, \mbox{in} \quad \Omega,\\
u &= 0\,\,\,\mbox{on}~~\partial\Omega,
\end{array}\right.
\end{eqnarray}
where $\Omega \subset \mathbb{R}^d (d\ge 2)$ is a bounded ploytopic
domain, $\mathbf{A}:\Omega\rightarrow \mathbb{R}^{d\times d}$ is
piecewise Lipschitz over initial triangulation $\mathcal {T}_{0}$,
for $x\in \Omega$ matrix $\mathbf{A(x)}$ is symmetric and positive
definite with smallest eigenvalue uniformly bounded away from 0,
${\bf b} \in [L^{\infty}(\Omega)]^d $ is divergence free , $c \in
L^{\infty}(\Omega),$ and $f \in L^2(\Omega)$ .
A finite element discretization of (\ref{application2}) reads: Find
$u_h \in S^h_0(\Omega)$ satisfying
\begin{eqnarray}\label{variation-2}
(\mathbf{A}\nabla u_h,\nabla v)+({\bf b}\cdot \nabla
u_h,v)+(cu_h,v)= (f,v)~~~~~\forall v \in S^h_0(\Omega).
\end{eqnarray}
It is seen that (\ref{variation-2}) is a special case of
(\ref{Gdis-fem}), in which $Vu:={\bf b} \cdot \nabla u + cu$ and
$\ell u=\ell_h u_h=f$. Consequently, $\kappa_1(h)=0$,
$w^h=K(f-Vu_h)$ and
\begin{eqnarray*}
u-w^h = KV(u_h-u) = KV(I-P_h)(u_h-u).
\end{eqnarray*}
Obviously, $V:H^1_0(\Omega)\rightarrow L^2(\Omega)$ is a
linear bounded operator and $KV$ is a compact operator over
$H^1_0(\Omega)$. We have the conclusion of Theorem
\ref{thm-general-boundary}.
In this application, the element residual and jump residual become
\begin{eqnarray*}
\mathcal{R}_{\tau}(u_h) &:=& f -{\bf b} \cdot \nabla u_h - cu_h+\nabla \cdot (\mathbf{A}\nabla u_h)~~~~ \mbox{in}~ \tau\in
\mathcal{T}_h,\\
J_e(u_h) &:=&[[\mathbf{A}\nabla u_h]]_e \cdot \nu_e ~~~~~~~~~~~~~~~~~~~ \mbox{on}~ e\in
\mathcal{E}_h
\end{eqnarray*}
while the corresponding error estimator $\eta_h(u_h, \mathcal{T}_h)$
and the oscillation $osc_h(u_h,\mathcal{T}_h)$ are defined by
(\ref{Gerror-estimator}) and (\ref{Goscilliation}), respectively.
Thus Theorem \ref{error-reduction} and Theorem
\ref{thm-optimal-complexity} ensure the convergence and optimal
complexity of AFEM for nonsymmetric problem (\ref{application2}).
\subsection{A nonlinear problem}\label{nonlinear}
In this subsection, we derive the convergence and optimal complexity
of AFEM for a nonlinear problem from our general theory.
Consider the following nonlinear problem: Find $u \in H^1_0(\Omega)$
such that
\begin{eqnarray}\label{application3}
\left\{\begin{array}{rl}
\mathcal {L}u:=-\Delta u+f(x ,u)&=0 \,\,\, \mbox{in} \quad \Omega,\\
u &= 0\,\,\,\mbox{on}~~\partial\Omega,
\end{array}\right.
\end{eqnarray}
where $f(x,y)$ is a smooth function on
$\mathbb{R}^3\times\mathbb{R}^1$.
For convenience, we shall drop the dependence of variable $x$ in
$f(x,u)$ in the following exposition. We assume that $u\in
H^1_0(\Omega)\cap H^{1+s}(\Omega)$ for some $s\in (0,1]$.
For any $w\in
H^1_0(\Omega)\cap H^{1+s}(\Omega)$, the linearized operator
$\mathcal {L}'_w$ at $w$ (namely, the Fr{\' e}chet derivative of
$\mathcal {L}$ at $w$) is then given by
\begin{eqnarray*} \mathcal {L}'_w=-\Delta + f'(w).
\end{eqnarray*}
We assume that $\mathcal {L}'_w:H^1_0(\Omega)\rightarrow
H^{-1}(\Omega)$ is an isomorphism. As a result, $u\in
H^1_0(\Omega)\cap H^{1+s}(\Omega)$ must be an isolated solution of
(\ref{application3}). The associated finite element scheme for
(\ref{application3}) reads: Find $u_h\in S^h_0(\Omega)$ satisfying
\begin{eqnarray}\label{variation-3}
(\nabla u_h,\nabla v)+(f(u_h),v)=0~~~~\forall v\in S^h_0(\Omega).
\end{eqnarray}
Let $a(\cdot ,\cdot )=(\nabla \cdot,\nabla \cdot)$, $K=(-\Delta)^{-1}: L^2(\Omega)\rightarrow H^1_0(\Omega)$,
$V=0$ and $\ell_h w = -f(w)$ for any $w\in S^h_0(\Omega) $, then (\ref{variation-3}) becomes
(\ref{Gdis-fem}).
As usual, to analyze the finite element approximation of nonlinear
problem (\ref{variation-3}), we require mesh $\mathcal {T}_h$ to
satisfy that there exists $\varsigma\ge 1$ such that (c.f.
\cite{xu-zhou-01})
\begin{eqnarray*}
h^{\varsigma}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} h(x) ~~ x\in\Omega,
\end{eqnarray*}
where $h(x)$ is the diameter $h_{\tau}$ of the element $\tau$
containing $x$. We consider the case of that $S^h_0(\Omega)$ is the
conforming piecewise linear finite element space associated with
$\mathcal {T}_h$. We assume that $\varsigma<2s$. Thus we can choose
$p\in (3, 6\varsigma/(3\varsigma-2s)]$ and obtain from Theorem 3.1
and Theorem 3.2 of \cite{xu-zhou-01} that
\begin{lemma}\label{fem-nonlinear-lemma}
If $h\ll 1$, then
\begin{eqnarray*}
\|u-u_h\|_{1,\Omega}+h^s\|u_h\|_{0,\infty,\Omega}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} h^s
\end{eqnarray*}
and
\begin{eqnarray*}
\|u-u_h\|_{0,\Omega}\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} r(h)\|u-u_h\|_{1,\Omega},
\end{eqnarray*}
where $r(h)\rightarrow 0$ as $h\rightarrow 0$.
\end{lemma}
Now we shall show that Theorem 3.1 is applicable for
(\ref{application3}). Since $K$ is monotone and $f(x,y)$ is smooth,
we have from Lemma \ref{fem-nonlinear-lemma} that
\begin{eqnarray*}
& &\|K(f(u)-f(u_h))\|_{a,\Omega} \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} \|K(u-u_h)\|_{a,\Omega}
\nonumber\\&\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}} & \|u-u_h\|_{0,\Omega} \mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}
r(h)\|u-u_h\|_{a,\Omega}.
\end{eqnarray*}
Therefore we have (\ref{general-boundary-neq}) when we choose
$\kappa_1(h)=r(h)$ and $\kappa_2(h)=0$.
In this application, the element residual and jump residual become:
\begin{eqnarray*}
\mathcal{R}_{\tau}(u_h) &:=& -f(u_h)+\Delta u_h \qquad \mbox{in}~ \tau\in
\mathcal{T}_h,\\
J_e(u_h) &:=& -\nabla u_h^{+}\cdot \nu^{+} - \nabla u_h^{-}\cdot
\nu^{-} := [[\nabla u_h]]_e \cdot \nu_e ~~~~ \mbox{on}~ e\in
\mathcal{E}_h
\end{eqnarray*}
and the corresponding error estimator $\eta_h(u_h, \mathcal{T}_h)$
and the oscillation $osc_h(u_h,\mathcal{T}_h)$ are defined by
(\ref{Gerror-estimator}) and (\ref{Goscilliation}), respectively.
Then Theorem \ref{error-reduction} and Theorem
\ref{thm-optimal-complexity} ensure the convergence and optimal
complexity of AFEM for nonlinear problem (\ref{application3}).
\subsection{An unbounded coefficient problem}\label{nonsmooth} Finally,
we investigate a nonlinear eigenvalue problem, of which a
coefficient is unbounded. It is known that electronic structure
computations require solving the following Kohn-Sham equations
\cite{Beck-00,gong-shen-zhang-zhou-08,Kohn-Sham-65}
\begin{eqnarray}\label{KS}
\left(-\frac{1}{2}\Delta-\sum_{j=1}^{N_{atom}}\frac{Z_j}{|x-r_j|}+\int_{\mathbb{R}^3}\frac{\rho(y)}{|x-y|}d
y+V_{xc}(\rho)\right)u_i=\lambda_i u_i ~~in \quad \mathbb{R}^3,
\end{eqnarray}
where $N_{atom}$ is the total number of atoms in the system, $Z_j$
is the valance charge of this ion (nucleus plus core electrons),
$r_j$ is the position of the $j$-th atom $(j=1,\cdots ,N_{atom})$,
$$
\rho=\sum_{i=1}^{N_{occ}}c_i|u_i|^2
$$
with $u_i$ the $i$-th smallest eigenfunction, $c_i$ the number of
electrons on the i-th orbit, and $N_{occ}$ the total number of the
occupied orbits.
The central computation in solving the Kohn-Sham
equation is the repeated solution of the following eigenvalue
problem: Find $(\lambda,u)\in \mathbb{R}\times H^1_0(\Omega)$ such
that
\begin{eqnarray}\label{application1}
\left\{\begin{array}{rl}
-\frac{1}{2}\Delta u+ Vu &= \lambda u \quad \mbox{in}~~ \Omega,\\
\|u\|_{0,\Omega} &= 1,
\end{array}\right.
\end{eqnarray}
where $\Omega$ is a bounded domain in $\mathbb{R}^3$, $V=V_{ne}+V_0$
is the so-called effective potential. Here, $V_0 \in
L^{\infty}(\Omega)$ and
\begin{eqnarray*}
V_{ne}(x)=-\sum_{j=1}^{N_{atom}}\frac{Z_j}{|x-r_j|}.
\end{eqnarray*}
A finite element discretization of
(\ref{application1}) reads: Find $(\lambda_h,u_h)\in
\mathbb{R}\times S^h_0(\Omega)$ such that
\begin{eqnarray}\label{variation-1}
\frac{1}{2}(\nabla u_h,\nabla
v)+(Vu_h,v)=\lambda_h(u_h,v)~~~~~\forall v \in S^h_0(\Omega).
\end{eqnarray}
Let $\ell_h : S^h_0(\Omega)\rightarrow L^2(\Omega)$ be defined by
\begin{eqnarray*}
\ell_h v=\lambda_h v ~~~~\forall v \in S^h_0(\Omega),
\end{eqnarray*}
then (\ref{variation-1}) is a special case of (\ref{Gdis-fem})
when $a(\cdot,\cdot)=\frac{1}{2}(\nabla \cdot, \nabla \cdot)$ and $K=\frac{1}{2}(-\Delta)^{-1}:
L^2(\Omega)\rightarrow H^1_0(\Omega)$.
Using the uncertainty principle lemma (see, e.g.,
\cite{Reed-Simon-75})
\begin{eqnarray*}
\int_{\mathbb{R}^3} \frac{w^2(x)}{|x|^2}\leq
4\int_{\mathbb{R}^3}|\nabla w|^2~~~~\forall w \in
C^{\infty}_{0}(\mathbb{R}^3)
\end{eqnarray*}
and the fact that $C^{\infty}_0(\Omega)$ is dense in
$H^1_0(\Omega)$, we obtain
\begin{eqnarray*}
\int_{\Omega} \frac{w^2(x)}{|x|^2}\leq 4\int_{\Omega}|\nabla
w|^2~~~~\forall w \in H^1_0(\Omega).
\end{eqnarray*}
Then for any $w \in H^1_0(\Omega)$, we have
\begin{eqnarray*}
\|V_{ne} w+V_0w\|_{0,\Omega}&\leq& C\|w\|_{1,\Omega},
\end{eqnarray*}
namely, $V$ is a bounded operator over $H^1_0(\Omega)$. Thus $KV$ is
a compact operator over $H^1_0(\Omega)$.
We consider the case of that $(\lambda,u)\in \mathbb{R}\times
H^1_0(\Omega)$ is some simple eigenpair of (\ref{application1}) with
$\|u\|_{0,\Omega}=1$. Note that for $\ell v:=\lambda v ~~\forall
v\in H^1_0(\Omega)$, there holds
\begin{eqnarray*}
K(\ell u-\ell_h u_h) = \lambda K(u-u_h)+(\lambda-\lambda_h)Ku_h.
\end{eqnarray*}
So if $(\lambda_h,u_h)\in \mathbb{R}\times S^h_0(\Omega)$ is the
associated finite element eigenpair of (\ref{variation-1}) with
$\|u_h\|_{0,\Omega}=1$ that satisfy $$
\|u-u_h\|_{0,\Omega}+|\lambda-\lambda_h|\mathrel{\raise2pt\hbox{${\mathop<\limits_{\raise1pt\hbox{\mbox{$\sim$}}}}$}}
\kappa_1(h)\|u-u_h\|_{a,\Omega}, $$ we then have (c.f.
\cite{Dai-Xu-Zhou-08})
\begin{eqnarray*}
\| K(\ell u-\ell_h u_h)\|_{a,\Omega}
= O(\kappa_1(h))\|u - u_h\|_{a,\Omega},
\end{eqnarray*}
where $\kappa_1(h):= \rho_{_{\Omega}}(h)+\|u-u_h\|_{a,\Omega}$
satisfying $\kappa_1(h)\rightarrow 0$ as $h\rightarrow 0$.
In this application, the element residual and jump residual become:
\begin{eqnarray*}
\mathcal{R}_{\tau}(u_h) &:=&\lambda_hu_h-Vu_h +\frac{1}{2}\Delta u_h~~~~ \mbox{in}~ \tau\in
\mathcal{T}_h,\\
J_e(u_h) &:=& [[\frac{1}{2}\nabla u_h]]_e \cdot \nu_e ~~~~~~~~~~~~~~ \mbox{on}~ e\in
\mathcal{E}_h
\end{eqnarray*}
and the corresponding error estimator $\eta_h(u_h, \mathcal{T}_h)$
and the oscillation $osc_h(u_h,\mathcal{T}_h)$ are defined by
(\ref{Gerror-estimator}) and (\ref{Goscilliation}), respectively.
Then Theorem \ref{error-reduction} and Theorem
\ref{thm-optimal-complexity} ensure the convergence and optimal
complexity of AFEM for unbounded coefficient problem
(\ref{application1}) (c.f. \cite{Dai-Xu-Zhou-08}).
\section{Numerical examples}
In this section we will report some numerical results to illustrate
our theory. Our numerical results were carried out on LSSC-II in the
State Key Laboratory of Scientific and Engineering Computing,
Chinese Academy of Sciences, and our codes were based on the toolbox
PHG of the State Key Laboratory of Scientific and Engineering
Computing, Chinese Academy of Sciences.
{\bf Example 1}. We consider (\ref{application2}) when the
homogenous Dirichlet boundary condition is replaced by $u = g$ on
$\partial \Omega$ and $\Omega = (0,1)^3$ with the isotropic
diffusion coefficient $\mathbf{A} = \epsilon I$, $\epsilon =
10^{-2}$, convection velocity $\mathbf{b} = (2,3,4)$, and c = 0
(c.f. \cite{Knobloch-Tobiska-03} for a 2D case and Remark
\ref{remark-boundary}). The exact solution is given by
\begin{eqnarray*}
u=\left(x^3-\exp\big(\frac{2(x-1)}{\epsilon}\big)\right)
\left(y^2-\exp\big(\frac{3(y-1)}{\epsilon}\big)\right)
\left(z-\exp\big(\frac{4(z-1)}{\epsilon}\big)\right).
\end{eqnarray*}
For small $\epsilon >0$ the solution has the typical layer behavior
in the neighbourhood of $x=1$, $y=1$, $z=1$, respectively.
The Dirichlet boundary condition $g(x,y,z)$ on $\partial
\Omega$ is given by
\begin{equation*}\label{boundary}
g(x,y,z)=
\left\{\begin{array}{rcl} \displaystyle 0 ~~~~~~~~ x=1~~ or~~ y=1 ~~or~~ z=1, \\
u(x,y,z) ~~~~ x=0~~ or~~ y=0~~ or~~ z=0.
\end{array}\right.
\end{equation*}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/general_P1_mesh.ps}}\end{picture}
\begin{center} Z=0.0 \end{center}\caption{The cross-section of an
adaptive mesh of {\bf Example 1} using linear finite
elements}\label{general_P1_mesh}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/general_P2_mesh.ps}}\end{picture}
\begin{center} Z=0.0 \end{center}\caption{The cross-section of an
adaptive mesh of {\bf Example 1} using quadratic finite
elements}\label{general_P2_mesh}
\end{minipage}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0) \resizebox*{5cm}{5cm}{\includegraphics{eps-figures/general_P1.eps}}\end{picture}
\caption{The convergence curves
of {\bf Example 1}
using linear finite elements}\label{general_P1}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm} {\includegraphics{eps-figures/general_P2.eps}}\end{picture}
\caption{The convergence curves
of {\bf Example 1}
using quadratic finite elements }\label{general_P2}
\end{minipage}
\end{center}
\end{figure}
Some adaptively refined meshes
are displayed in Fig. \ref{general_P1_mesh} and Fig.
\ref{general_P2_mesh}. Our numerical results are presented in Fig.
\ref{general_P1} and Fig. \ref{general_P2}. It is shown from Fig.
\ref{general_P2} that $\|u - u_h\|_1$ is proportional to the a
posteriori error estimators, which indicates the efficiency of the a
posteriori error estimators given in section \ref{nonsymmetric}.
Besides, it is also seen from Fig. \ref{general_P1} and Fig.
\ref{general_P2} that, by using linear finite elements and
quadratic finite elements, the convergence curves of errors are
approximately parallel to the line with slope $-1/3$ and the line
with slope $-2/3$, respectively. These mean that the approximation
error of the exact solution has optimal convergence rate, which
coincides with our theory in section \ref{section-convergence}.
{\bf Example 2}. Consider the following nonlinear problem:
\begin{eqnarray*}\label{example3}
\left\{\begin{array}{rl}
-\Delta u+ u^3&=f \,\,\, \mbox{in} \quad \Omega,\\
u &= 0\,\,\,\mbox{on}~~\partial\Omega,
\end{array}\right.
\end{eqnarray*}
where $\Omega = (0,1)^3$. The exact solution is given by $u=\sin(\pi
x_1)\sin(\pi x_2)\sin(\pi x_3)/(x_1^2+x_2^2+x_3^2)^{1/2}$.
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0) \resizebox*{5.0cm}{5.0cm}{\includegraphics{eps-figures/nonlinear_P1_mesh.ps}}\end{picture}
\caption{The cross-section of an adaptive mesh of {\bf Example 2} using linear finite
elements}\label{nonlinear_P1_mesh}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5.0cm}{5.0cm} {\includegraphics{eps-figures/nonlinear_P2_mesh.ps}}\end{picture}
\caption{The cross-section of an adaptive mesh of {\bf Example 2}
using quadratic finite elements}\label{nonlinear_P2_mesh}
\end{minipage}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0) \resizebox*{5.0cm}{5.0cm}{\includegraphics{eps-figures/nonlinear_P1.eps}}\end{picture}
\caption{The convergence curves
of {\bf Example 2}
using linear finite elements }\label{nonlinear_P1}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5.0cm}{5.0cm} {\includegraphics{eps-figures/nonlinear_P2.eps}}\end{picture}
\caption{The convergence curves
of {\bf Example 2}
using quadratic finite elements }\label{nonlinear_P2}
\end{minipage}
\end{center}
\end{figure}
Fig. \ref{nonlinear_P1_mesh} and Fig. \ref{nonlinear_P2_mesh} are
two adaptively refined meshes, which show that the error indicator
is good. It is shown from Fig. \ref{nonlinear_P1} and Fig.
\ref{nonlinear_P2} that $\|u - u_h\|_1$ is proportional to the a
posteriori error estimators, which implies the a posteriori error
estimators given in section \ref{nonlinear} are efficient. Besides,
similar conclusions to that of Example 1 can be obtained from Fig.
\ref{nonlinear_P1} and Fig. \ref{nonlinear_P2}, too.
{\bf Example 3}. Consider the Kohn-Sham equation for helium atoms:
\begin{eqnarray*}\label{atom-He}
\left(-\frac {1}{2}\Delta -\frac {2}{|x|} +
\int\frac{\rho(y)}{|x-y|}dy+ V_{xc}\right) u = \lambda u
~~\mbox{in}~ \mathbb{R}^3,
\end{eqnarray*}
and $\int_{\mathbb{R}^3}{|u|^2}=1$, here $\rho=2|u|^2.$ In our
computation of the ground state energy, we solve the following
nonlinear eigenvalue problem: Find $(\lambda,u)\in \mathbb{R}\times
H^1_0(\Omega)$ such that $\int_{\Omega}{|u|^2}dx=1$ and
\begin{equation}\label{example1}
\left\{\begin{array}{rcl} \displaystyle \left(-\frac {1}{2}\Delta -
\frac {2}{|x|}+ \int\frac{\rho(y)}{|x-y|}dy+ V_{xc}\right) u &=& \lambda u ~~\mbox{in} ~\Omega, \\
u &=& 0 ~~ \mbox{on}~\partial\Omega,
\end{array}\right.
\end{equation}
where $\Omega=(-10.0, 10.0)^3$, and $V_{xc}(\rho) =
-\frac{3}{2}\alpha(\frac{3}{\pi}\rho)^{\frac{1}{3}}$ with
$\alpha=0.77298$. Since (\ref{example1}) is a nonlinear eigenvalue
problem, we need to linearize and solve them iteratively, which is
called the self-consistent approach
\cite{Beck-00,gong-shen-zhang-zhou-08,Kohn-Sham-65,perdew-zunger-81}.
In our computation, a Broyden-type quasi-Newton method
\cite{srivastava} were used.
In 1989, White \cite{White-89} computed helium atoms over uniform
cubic grids and obtained ground state energy -2.8522 a.u. by using
500,000 finite element bases. While the ground state energy of
helium atoms in Software package fhi98PP \cite{Fuchs-Scheffler-99}
is -2.8346 a.u., which we take as a reference.
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/energy_P1.eps}}\end{picture}
\caption{The ground state energy using linear finite elements
}\label{energy_P1}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/energy_P2.eps}}\end{picture}
\caption{The ground state energy using quadratic finite elements}
\label{energy_P2}
\end{minipage}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/He_P1_mesh.ps}}\end{picture}
\begin{center} Z=0.0 \end{center}\caption{The cross-section of an
adaptive mesh of {\bf Example 3} using linear finite
elements}\label{he_P1_mesh}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/He_P2_mesh.ps}}\end{picture}
\begin{center} Z=0.0 \end{center}\caption{The cross-section of an
adaptive mesh of {\bf Example 3} using quadratic finite
elements}\label{he_P2_mesh}
\end{minipage}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{1cm}
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/He_P1.eps}}\end{picture} \caption{The
convergence curve
of {\bf Example 3} using linear finite elements }\label{he_P1}
\end{minipage}
\hfill
\begin{minipage}[t]{5.0cm}
\begin{picture}(5.0,5.0)\resizebox*{5cm}{5cm}
{\includegraphics{eps-figures/He_P2.eps}}\end{picture} \caption{The
convergence curve
of {\bf Example 3} using quadratic finite elements} \label{he_P2}
\end{minipage}
\end{center}
\end{figure}
Our results are displayed in Fig. \ref{energy_P1}, Fig.
\ref{energy_P2}, Fig. \ref{he_P1_mesh}, Fig. \ref{he_P2_mesh}, Fig.
\ref{he_P1}, and Fig. \ref{he_P2}. It is seen from Fig.
\ref{energy_P2} that the ground state energy in our computation is
close to the reference with less 100,000 degrees of freedom when the
quadratic finite element discretization is used. Some cross-sections
of the adaptively refined meshes are displayed in Fig.
\ref{he_P1_mesh} and Fig.\ref{he_P2_mesh}. Since we do not have the
exact solution, we list the convergence curves of the a posteriori
error estimators in Fig. \ref{he_P1} and Fig. \ref{he_P2} only. It
is shown from these figures that the a posteriori error estimators
given in section \ref{nonsmooth} are efficient.\vskip 0.2cm
{\sc Acknowledgements.} The authors would like to thank Mr. Huajie
Chen, Dr. Xiaoying Dai, and Prof. Lihua Shen for their stimulating
discussions and fruitful cooperations that have motivated this work.
|
\section{Introduction}
In recent years, it has been increasingly recognized that many of the clues to the problem of
galaxy formation are preserved in galaxy outskirts \citep[e.g.][and references therein]{johnston08}.
The current consensus is that large spirals begin as small fluctuations in the early Universe and
grow by in situ star formation and hierarchical merging \citep[e.g.][]{wr78}. Subsequently, once
a spiral is the dominant component in such mergers, it continues accreting and disrupting
sub-halos falling into its potential well. With the accumulation of accretion and disruption events,
massive spirals build up stellar and dark matter halos \citep{abadi06}. The complete disruption
of the sub-halos may take several orbits, distributing thus the tidally stripped stars over a broad
range of distances from the main galaxy. As galaxies are predicted to assemble their outskirts
from the disruption of a large number of satellites, these regions are expected to possess
significant density and chemical substructures \citep{font06}.
Observationally however, the nature and the origin of those regions remains elusive.
Much of what we know about their properties is based on observations of the Local Group
massive spirals \citep[e.g.][and references therein]{fb02}. Recent surveys find evidence that
the Galaxy stellar halo is divisible into two components, with a moderately flat inner regions
showing a modest prograde rotation, whereas the outer regions are less chemically evolved
and exhibit a nearly spherical distribution with a retrograde rotation \citep{carollo08}.
Wide-field imaging data indicates that the stellar halo of the Galaxy is highly structured
\citep{ibata03,yanny03,belokurov07}, suggesting that a large fraction of the Halo has
been accreted from satellites \citep{bell08}. The outer regions of Andromeda have been recently
observed to contain even more substructure and streams than observed around the Milky Way
\citep{ibata07}, suggesting that its accretion history may have been more active than the
suspected quiet one of the Milky Way \citep{mouhcine05a,hammer07}. It is completely unknown
however if these properties are generic features of the outskirts of spirals, or are reflecting
peculiar assembly histories.
We have seen recently a dramatic progress in large-scale mapping of the Milky Way and
Andromeda, with a number of large surveys measuring photometric, kinematic, and chemical
properties of individual stars over a wide galactic volume. Although those surveys will
significantly advance our understanding of the assembly of these galaxies, it cannot be
assumed that a sample of two galaxies will provide the definitive solution to the nature and
origin of the stellar content in the outskirts of spirals. The fundamental next step to fully exploit
the Local Group surveys and thereby to establish a comprehensive picture of the assembly
histories of spirals is to determine whether the Local Group massive spirals are suitably typical
by studying giant spirals beyond the Local Group.
A number of surveys of the low surface brightness outskirts of spirals beyond the Local Group
have been conducted recently. These surveys were however either sampling limited galactic
volumes \citep{mouhcine05b,blandhawthorn05,mouhcine07,dejong07}, or too shallow to detect
the old stellar tracers \citep{davidge06,davidge07}, thus severely hampering their impact.
Measurements of galaxy outskirts have been also attempted using integrated light
\citep{morrison94}, and have succeeded on detecting low surface brightness tidal streams in
the outskirts of a few nearby disk galaxies \citep{martinez-delgado08,martinez-delgado09}.
Those measurements are however affected by large uncertainties
\citep{dejong09,martinez-delgado08}, and are able to detect stellar structures down to a surface
brightness limit of ${\rm \mu_{I} \sim 28\, mag\, arcseec^{-2}}$ at the faintest \citep{zheng99},
many magnitudes brighter than the bulk of substructure detected in the outskirts of the Local
Group spirals \citep{belokurov07,ibata07}.
Characterizing comprehensively the outskirts of spirals beyond the Local Group requires resolving
panoramically the stellar content of those regions, giving access to the extremely low surface
brightness structures. The required observations are however extremely challenging due to the
photometric depth one has to reach, i.e., $I\sim 26.-28.5$, to resolve old giant stars in galaxies
beyond the Local Group, restricting this approach (using the present-day instrumentation) to the
small number of spirals closer than $\sim 10-12\,$Mpc.
Because the theoretical predictions are inherently statistical in nature, and due to the stochastic
nature of halo formation, constraining the assembly history of galaxy outskirts must rely in large
part on measuring the demographics of their stellar populations. To this end, we have initiated
a survey to resolve panoramically the stellar content of galaxy outskirts for a sample of nearby,
i.e., up to $\sim10$ Mpc, highly inclined spirals, distributed over a broad range of masses,
i.e., circular velocities ranging from ${\rm \sim 80\,km\,s^{-1}}$ up to ${\rm \sim 230\,km\,s^{-1}}$,
and morphologies, i.e., ranging from Sa to Sd.
As part of this effort, we have targeted the edge-on galaxy NGC~891, often considered as
the nearest prime analog of the Galaxy. Our group recently used deep optical imaging data,
obtained with the Advanced Camera for Surveys on board the Hubble Space Telescope
(HST/ACS), to investigate the properties of the resolved extra-planar stellar populations over
the south-east quadrant of NGC~891, extending up to $\sim10$\,kpc from the galactic plane
\citep{mouhcine07,ibata09,rejkuba09,harris09}.
Succinctly summarized, those studies indicate that the thick disk of NGC~891 shares comparable
structural and chemical properties to its Galactic counterpart. The stellar populations beyond
the thick disk appear to possess significant small-scale variations in the stellar metallicity.
Interestingly, those regions are found to be dominated by stars significantly more chemically
enriched than those populating the regions of comparable heights from the plane of the Galaxy.
In the present contribution, we report the first results of our survey. Detailed analysis of the
properties of the stellar content of different galactic components, the search for substructures,
the determination of metallicity distribution functions and their spatial variation, and the
properties of the globular cluster system will be reported in forthcoming papers.
The layout of this paper is as follow; in \S~\ref{data} we present briefly the data set.
In \S~\ref{analysis} we report the discovery of a new morphological structures around NGC~891,
and discuss the implications of our finding for the formation and evolution of spiral galaxies.
Throughout the paper we use an intrinsic distance modulus for NGC~891 of
$(m-M)_{\circ} = 29.94$, with the tip of the red giant branch located at $I\,=\,25.84 \pm 0.04$ mag
\citep{mouhcine07,tg05}. NGC~891 is located at relatively low Galactic latitude
($\ell=140.38^\circ$, $b=-17.42^\circ$), and therefore it suffers from significant (though not large)
extinction from foreground dust: $E(B-V)= 0.065$ \citep{schlegel98}.
\section{Data}
\label{data}
The Subaru Prime Focus Camera (Suprime-Cam) on the 8.2-m Subaru Telescope is a mosaic
camera of ten-chip $2048 \times 4096$ charge-coupled devices, which
covers a ${\rm 34\,arcmin \times 27\,arcmin}$ field of view with a pixel scale of
0.20\,arcsec \citep{miyazaki02}. The instrument was used to image the outer regions of NGC~891.
In the following, we will describe briefly the observations and the data reduction, the full details
will be reported elsewhere.
The observations were taken in the Johnson visual V-band and Gunn $i$-band. We have obtained
a total of 10 hours of good quality data in the V-band and 11.27 hours in the $i$-band, with seeing
better than 0.6 arcsec, allowing us to resolve approximately the brightest two magnitudes of the
red giant branch (RGB hereafter) of metal-poor stellar populations at the distance of NGC~891.
By covering a vast ${\rm \sim 90\,kpc \times 90\,kpc}$ region around NGC~891, the survey allows
us to distinguish local density enhancements from large-scale structure of the halo of the galaxy
and/or the foreground distribution of stars.
The images were pre-processed using the the Cambridge Astronomical Survey Unit photometric
pipeline. The final stacks were created by summing all pixels, weighted by the estimated seeing
on each frame. The DAOPHOT/ALLSTAR software suite \citep{stetson87} was used to detect all
sources down to $3\sigma$ above the sky. Isolated bright stars were identified over the frame to
serve as point spread function (PSF) templates. The PSF was modelled as a Moffat function;
experiments showed that allowing the PSF to vary spatially over the field did not improve
significantly the fit to bright stars, so we adopted a spatially constant PSF. Finally, the instrumental
magnitudes were shifted to agree with the calibrated photometry presented in \citet{rejkuba09}.
The analysis of the photometric errors indicates that the average $i$-band error varies from
$\sigma_i \approx 0.05$ at $i \approx 25.8-26.0$, i.e., the bright end of the RGB, to
$\sigma_i \approx 0.15$ at $i \approx 26.8-27.0$, i.e., approximately a magnitude below
the tip of the RGB, while the average color error varies from $\sigma_{(V-i)} \approx 0.12$
to $\sigma_{(V-i)} \approx 0.18$ within the same magnitude range. To estimate the photometric
completeness of the survey, we used the artificial star simulations performed for our analysis of
the HST/ACS deep images of NGC~891, and described in full detail in
\citet{rejkuba09}. By comparing directly the number of stars detected in both HST/ACS and
Supreme-Cam images as a function of magnitude and spatial density of RGB stars, it is
possible to compute the photometric completeness of our survey. For stars located above
${\rm \sim 6\,kpc}$ from the galactic plane, i.e., regions where the crowding is not at all
important (see \citet{rejkuba09} for more details), the data are well above $\sim 80$ per cent
complete for the upper one magnitude of the giant branch, i.e., $i_{\circ} \la 27$.
Objects were classified as artefacts, galaxies or stars according to their morphological structure
on the images. The point source catalog consists of NGC~891 RGB stars, NGC~891 asymptotic
giant branch (AGB hereafter) stars, Galactic foreground stars, and unresolved background
galaxies. In order to isolate stars to probe the spatial distribution of the stellar populations in
the outskirts of NGC~891, we applied a series of magnitude and color cuts. RGB stars are
defined as those with foreground extinction-corrected $i$-band magnitude fainter than 25.8 and
brighter than 27, and with (V-$i$)$_{\circ}$ color redder than 1.1, while AGB stars are selected
as those with foreground extinction-corrected $i$-band magnitude ranging between 25.0 and
25.8 and foreground extinction-corrected (V-$i$)$_{\circ}$ colors redder than 1.2.
\section{The panoramic landscape around NGC~891}
\label{analysis}
Fig.~1 shows the surface density map of RGB stars across the surveyed area.
Numerous enhancements of the surface density of RGB stars are visible with the most striking
being the large-scale complex of arcing loops and streams wrapped around the galaxy.
About a half-dozen arc-like features are visible in the density map of old stars, extending up
to $\sim 40\,$kpc west and $\sim 30\,$kpc east of the galactic stellar disk. On the left side,
corresponding to north-eastern part of the galaxy, it is possible to trace the full extension of
a giant loop turning around, falling toward the plane of the galaxy, and extending further to
the south. The spatial density of AGB stars does not exhibit however any obvious
enhancements associated with the large-scale complex of streams of old stars.
Note that there is no evidence for a significant clumping in the distribution in of extended
sources across our field, Galaxy clusters and other large-scale structure are unlikely to
dominate the counts at the angular scales of the observed features \citep{couch93}.
It is tempting to suggest that the large-scale network of old star streams is connected, originating
from a single accretion event, however it is impossible to argue this firmly based on the stellar
density map alone. Nevertheless, the arc-like features exhibit similar shapes, angular lengths,
radii of curvature, separation from the parent galaxy, and levels of stellar density enhancements.
In addition, the stellar loops appear to be distributed and to cross each other in a pattern
strikingly similar to the classical rosette-shaped loops predicted for tidally disrupting dwarf
galaxies in $N$-body simulations \citep[e.g.][]{ibata98,law05}. These properties provide
circumstantial evidence that the arclike features around the galaxy may all come from a single
accretion event (detailed modeling of the stellar stream will be presented elsewhere).
The discovery of the NGC~891 giant stream in the first deep, panoramic survey of the Milky
Way's nearest analog, together with the previous discoveries of tidal streams in the Milky Way,
Andromeda, NGC~5907, and NGC~4013, suggests that halo substructure in the form of tidal
streams may be a generic property of massive spirals, and that the formation of galaxies
continues at a moderate rate up to the present day.
Fig.~1 shows that the spatial density of old stars varies along the streams, however nowhere
among these features is there a region dense enough that it could be identified as the
remaining core of the disrupted progenitor dwarf satellite. The stellar streams are most likely
the fossil remnants of a totally cannibalized system, although the core of the disrupted satellite
could possibly be hidden behind the disk of the galaxy. Deep radio observations of NGC~891
have reported the presence of a large gaseous filament and a number of counter-rotating
clouds in its halo, extending to comparable distances from the disk as the stellar stream
\citep{oosterloo07}. The structure and the kinematics of these gaseous structures appear to
favor a scenario in which they are the results of a flyby interaction with the gas-rich satellite
UGC~1807 \citep{oosterloo07,mapelli08}. Fig.~2 shows the color-magnitude diagram of
stars along the streams, with fiducials of RGB sequences of three Galactic globular clusters
from \citet{dacosta90} superimposed. The morphology of the the color-magnitude diagram
indicates clearly that the stream progenitor contained a broad range of stellar populations,
with a mean metallicity comparable to that of 47 Tuc, i.e., ${\rm [Fe/H]\sim-0.7}$, and ages
older than a few Gyr with hardly any young stellar components, in sharp contrast with the
typical stellar population of gas-rich dwarf galaxies \citep{skillman}. This indicates that it is
highly unlikely that the large-scale complex of loops and streams in the halo of NGC~891
could be associated with the gas-rich dwarf companion.
The second striking feature visible in the surface density map of RGB stars is the extended
super-thick envelope surrounding the galaxy. The defining high surface brightness components
of a spiral, i.e., the bulge and the disk, are embedded in a vast flattened cocoon-like structure
extending vertically up to $\sim 15\,$kpc, and radially up to $\sim 40\,$kpc, with a structure
quite unlike any classical notions of thick disk or halo. This structure is strikingly flatter than
the inner component of the Galactic halo that dominates out to $\sim10$\,kpc above the
disk \citep{preston91,carollo08}. An additional highly flattened component of the Galactic halo
has been recently detected \citep{morrison09}. This halo flat component is however suspected
to be confined within very close distances from the Galactic plane \citep{preston91,morrison09}
A stellar structure resembling the Galaxy thick disk has been resolved in NGC~891, with a
vertical scale height of $h_Z\sim 1.45$ kpc \citep{ibata09}. Stars in this morphological structure
show no metallicity gradients both vertically and radially, and cover a wide range in
metallicity, similar to what is observed for the Galaxy \citep{gilmore95,ivezic08}. However, at
vertical distances larger than $\sim 5$ kpc from the plane of the galaxy, stars belong to a more
extended structure \citep{ibata09}, and possess a negative metallicity gradient, albeit a mild
one \citep{rejkuba09}. The envelope surrounding NGC~891 is therefore highly unlikely an
extension of the thick disk. Thus it appears that the flattened super-thick envelope surrounding
the galaxy is a previously unknown component of NGC~891.
Our previous investigations of the properties of the stellar content of the south-east quadrant,
sampling stars of the super-thick envelope, find strong evidence for large amounts of
chemically-distinct sub-structures \citep{ibata09}. This suggests that those regions are
populated by numerous accretion remnants that are spread over a large volume and are
still far from being fully phase mixed. Since it is natural to expect that the stellar populations
in this quadrant are representative of those populating the super-thick structure, this indicates
that the stellar structure surrounding the galaxy was assembled from the tidal disruption of
many accreted satellites.
The super-thick structure presented here is not just a peculiarity of NGC 891: a preliminary
analysis by our group of deep panoramic observations of the nearby edge-on galaxy
NGC~2683 shows clearly the presence of an almost identical structure surrounding the high
surface brightness components of that galaxy. Detailed kinematical measurements have
revealed that a number of stellar substructures present at large radii of Andromeda co-rotate
with the inner stellar disk, and are distributed in a gigantic flattened ``extended disk'' structure,
extending radially from the high surface brightness inner disk out to $\sim 40\,$kpc from the
galaxy center \citep{ibata05}, i.e., strikingly similar to the radial extent of the super-thick stellar
structure surrounding NGC 891. Given that we do not observe Andromeda edge-on, it is
possible that the ``extended disk'' in that galaxy has a similar vertical extent to the new
morphological structure identified in NGC~891 and NGC~2683.
The debris resulting from the disruption of a dwarf galaxy on a prograde orbit that is coplanar
with the host galaxy disk is expected to relax into an extended rotating disk \citep{penarrubia06}.
The detection of the Monoceros stream in the Galaxy \citep{yanny03,ibata03}, the presence of
an extended-disk around Andromeda \citep{ibata05}, and the stellar stream around NGC~4013
\citep{martinez-delgado09}, all suspected to be the remnants of the disruption of satellites
moving in low inclined orbits, suggest that the accretion of satellites on low-inclination and low
eccentricity orbits may be a relatively common process in the formation of spirals \citep{pohlen04}.
Numerical simulations in the framework of the hierarchical cold dark formation scenario predict
that large spirals experience several mergers of dwarfs with masses ranging from 1\% to 10\%
of the host mass \citep{gao04}, with many on low-eccentricity orbits \citep{ghigna98}.
The present work suggests that a super-thick stellar envelope formed by numerous accretions
may be a common feature of large spirals.
\section{Acknowledgements}
This work is based on data obtained at the Subaru Telescope, operated by the National
Astronomical Observatory of Japan.
|
Subsets and Splits